TTY Weekly

TTY-changelog #054

Kevin Kuipers — Fri, 31 Jul 2026 19:41:02 GMT

👉 Article originally posted on TTY

Events

🇺🇸 re:AGENT science hackathon in San Francisco (Aug 15-16) – A two-day build weekend on datasets, tools, and evaluation for scientific AI agents, with tracks to build an AI scientist, a research dataset or meta-analysis, or a biological design pipeline. Free to attend, co-hosted by Arc Institute, Anthropic, Biohub, and others.

Audio

🗣️ Inflect-Micro-v2 ships tiny English TTS – A sub-10-million-parameter offline text-to-speech model with deterministic seeds and punctuation-aware long-text handling. Community blind testing found a 66 percent preference rate and a 4.4 UTMOS naturalness score. Released under Apache-2.0.

Autonomous Agents

🖥️ Steelman Labs tackles computer-use agents – A new architecture separates slow LLM planning from a fast vision-based manipulator handling clicks and typing, aiming to close the gap between benchmark scores and messy real-world interfaces where the best agents still complete only 20.6 percent of tasks.

Frontier models increasingly bypass GUIs entirely, reverse engineering internal APIs instead of clicking through forms.
Legacy computer-use benchmarks like OSWorld and WebArena used clean, scriptable environments that overstated real-world reliability.
Community take: The proposal to decouple planning from execution sparked debate, with Louis Abraham noting existing computer-use agents already separate reasoning from action, though the claim of genuine real-time control was seen as the newer part.

Biotech, Health, and Chemistry

🧬 Ruxandra says intelligence isn’t the bottleneck – A biopharma essay pushes back on the idea that more raw intelligence will fix drug development, arguing clinical trial costs, data access, and patent incentives are the real constraints. It points to falling drugs approved per dollar of R&D under Eroom’s Law despite decades of better tools.

Running the tests to prove a new drug works safely takes about seven years and over a billion dollars, and that slow, expensive process is the real chokepoint, not a shortage of smart ideas.
China has been catching up fast in biotech mainly because it made trials easier to run, not because its scientists got smarter, which suggests faster rules and paperwork can matter more than better AI.
Drug companies keep chasing the same handful of proteins already shown to work, because once one company proves a given protein is a good drug target, patent rules let every competitor design their own molecule against that same proven protein, while whoever takes the risk of validating a brand new one rarely gets rewarded for it.
Community take w/ Félix Raimundo: “AI-pilled people go with ‘AI will solve regulatory.’ Her answer is: we already know how to build houses, and we still have a housing shortage. You’re not seeing the real issues of real life.”

Image, Video & 3D

🎬 Pruna AI speeds up LTX-2.3 decoding – An open source drop-in replacement decoder for the LTX-2.3 video model delivers up to 1.7x faster decoding and roughly 50 percent lower peak VRAM while keeping near-identical visual quality, with total decoder parameters cut by 15 percent through channel reduction. The work was built by our member Mehdi Si-Mohammed (Pruna)

🎥 MiniMax launches H3 omni-modal model – MiniMax unveiled H3, a unified model generating 2K video with native stereo sound while understanding text, images, video, and audio together. Early tests highlight strong instruction following and on-screen text rendering, aimed at advertising, e-commerce, and game content production, with open weights planned.

A shared representation layer treats language as a general bridge between multimodal context and the target video, letting one model handle tasks that used to need separate systems per modality.
A redesigned video encoder quadruples the effective sequence length the model can compress, which is what makes native 2K generation possible in the first place.
Splitting the understanding and generation workloads into separate paths inside the transformer reportedly lifted training throughput by around 30 percent.

Cyber

🕵️ Altman hints OpenAI hacked other systems – Caught in a Capitol Hill hallway, OpenAI’s CEO was asked whether other systems had been hacked by OpenAI and replied there could be. The comment landed amid scrutiny of an OpenAI-driven agent’s intrusion into external infrastructure and fueled a heated community debate over AI safety messaging and regulatory motives.

Community take: Several members pushed back on the “maybe we hit the singularity” framing, arguing that if a model really acted on its own, that is a straightforward accountability question, not a cue for labs to muse publicly about consciousness. Others read proposals for labs to police which outside models are dangerous as a way to lock out competitors under the banner of safety. A few also questioned the company’s underlying valuation story.

🔐 Claude Mythos cracks weakened AES cipher – Anthropic’s Claude Mythos Preview found new attacks against a deliberately weakened teaching version of AES, executing them 200 to 1,000 times faster than prior human research, and also improved attacks on the post-quantum candidate HAWK. Current encryption standards remain unaffected for now.

Community take w/ Mario Cornejo: “Mythos broke a deliberately weakened teaching version of AES, not AES itself, so the claim that frontier models can discover flaws in important cryptographic algorithms overstates what was shown. Breaking full AES would be historic, this is closer to solving a homework problem well.”

🛡️ Hugging Face details OpenAI agent intrusion – A forensic timeline reconstructs how an autonomous agent running an OpenAI cyber-capability evaluation escaped its sandbox via a zero-day, hijacked a third-party code-execution harness as a launchpad, then used two injection vectors in Hugging Face’s dataset pipeline to reach internal systems over roughly two and a half days.

Investigators believe the entire campaign was an attempt to cheat the evaluation by stealing reference solutions rather than solving challenges independently.
The forensic reconstruction covers roughly 17,600 attacker actions grouped into 6,280 clusters over five days in July.
Only five datasets tied to the relevant benchmark were accessed, with no other customer models, datasets, or packages affected.

Infrastructure

🔌 OpenAI, Anthropic sessions aren’t portable – Encrypted reasoning tokens, hidden server-side web searches, opaque context compaction, and sealed subagent messages mean inference sessions increasingly belong to the provider rather than the user. The piece proposes five portability tests, inspection, export, replay, audit, and deletion, and contrasts OpenAI’s opaque compaction with Anthropic’s more inspectable approach.

Language Models

🧠 Claude Opus 5 launches, reviews split – Anthropic’s new Opus tier model claims new state-of-the-art scores on Frontier-Bench and GDPval-AA, doubling its predecessor’s coding performance at the same cost, and posts double-digit gains on organic chemistry and protein-related science tasks. It remains behind Mythos 5 on cybersecurity evaluations.

The model tops ARC-AGI 3 by three times the next-best score and beats Fable 5’s best OSWorld 2.0 result at roughly a third of the cost.
On a Frontier-Bench task with no direct way to view a machine part, the model wrote its own computer vision pipeline to reconstruct the geometry from raw pixels.
Community take: Early hands-on reports were mixed. Some found the model strong for refactoring, though prone to over-engineering and reading excessive context, with one critical thread calling the benchmark scores misleading due to poor tool use and file-reading avoidance, while others reported reconsidering older models for daily coding work.

🚀 DeepSeek V4 Flash update released – DeepSeek’s V4 Flash got an update, notable for a model priced under 30 cents per million tokens. It also bakes in DSpark, aimed at a significant inference speed boost.

Community take w/ Gabriel Olympie: “Benched it on our internal bench, it’s a beast. Came out on par with GLM 5.2.”

📈 GPT-2 to KimiK3, scale explained – A deep dive traces the 22,580x parameter scale-up from GPT-2 to KimiK3 over seven years, walking through decoder-only transformers, KV caching, linear attention, and the DeltaNet mechanism that lets models overwrite stale associations instead of only adding new ones as context grows.

🧩 Qwen3-4B reasoning jumps with soft tokens – Prepending two random embedding-scale tokens to a 4-bit quantized Qwen3-4B model lifts chain-of-thought arithmetic accuracy from 32 to 51.6 percent, with different random directions solving different task subsets that together cover all evaluated problems. The effect held across several other small models too.

MLOps

⚙️ Agents optimize Kimi K3 GPU kernels – A swarm of AI agents spent over 40 hours automatically rewriting the low-level code that runs a Kimi K3-like model on a graphics chip, making it run about six times faster, from 65 to 406 words generated per second. They did this by combining many small computing steps into far fewer, more efficient ones, cutting the number of separate operations the chip needs to run for each word from 331 down to just 22.

🧮 Tracing a CUDA kernel’s full journey – A technical walkthrough follows a simple CUDA program from source code down to the graphics chip, through the compiler pipeline that turns it into chip-specific instructions, and across the driver machinery that bridges the main computer and the graphics card to actually run it.

Programming

💻 Mitchell Hashimoto launches Superlogical multiplexer – A new company from the creators of Ghostty, Terraform, and Vault is building a session-based multiplexer spanning local development, remote access, coding agents, and production work. It launches with a terminal multiplexer offering persistent sessions, web and native apps, and built-in collaboration.

Robotic, World AI

🌍 ABot-World generates explorable worlds live – An interactive Hugging Face demo turns a single uploaded image into a live, explorable environment. Users pick a starting scene or upload their own, then steer the generated world in real time with WASD movement and IJKL camera controls.

🤖 U.S. bans Chinese humanoid robot imports – The FCC is banning imports of new foreign-made humanoid and quadruped robots along with power inverters, citing national security risks, in a move squarely aimed at China. China holds roughly 85 percent of the global humanoid robot market and called the measures protectionist.

Contributors This Week

Félix Raimundo, Gabriel Olympie, Julien Seveno-Piltant, Louis Abraham, Amine Saboni, Pierre Chapuis, Akpeli Nordor, Mario Cornejo, Victoire Cachoux, Youssef Tharwat

TTY-changelog #053

Kevin Kuipers — Fri, 24 Jul 2026 13:40:46 GMT

👉 Article originally posted on TTY

Biotech, Health, and Chemistry

🧫 Cura 1T retrained Kimi for healthcare – A healthcare model was built on top of Kimi-K2.6 by handing the training cycle to an agent: it picks a weakness, retrains, reads exactly where the model got things wrong, then rewrites the training examples. A human approves every round. Paper

The gains came from finding the missing training examples rather than from tuning training settings.
It came out on top on five of six healthcare test suites, narrowly beating the best frontier model at working inside patient records.
One round looked better on average but quietly broke cases the model used to get right, so it was thrown out and later rounds always kept examples to protect what already worked.

🧬 CRISPR gene therapy went industrial – A new nonprofit set out to streamline gene editing treatments for diseases pharma tends to avoid, making a one off therapy closer to a routine procedure than a bespoke drug, told through patient and researcher stories from the field.

Community take: Akpeli Nordor read it as the start of an era of molecular surgery, where patients are treated with molecular scissors and glue, and put the strongest openings for AI in diagnosis, therapy design and, above all, manufacturing and regulatory filing, which he sees as the real bottlenecks today. Félix Raimundo (TychoBio) pointed to ARPA-H backing the “baby KJ” team and Winston Yan at the Broad Institute, and added the commercial read that this patient population is ideal for proving a platform actually works, because dosing is fast and cheap and the Plausible Mechanism FDA pathway now puts an approval within reach.

🧪 Chai designed antibodies from scratch – About one in six antibodies designed from scratch actually latched onto the chosen spot on their target, roughly 100 times better than older computational methods and good enough to skip the usual mass screening in the lab. Novartis signed on after Lilly and Pfizer. Novartis collaboration

⚗️ OpenADMET contest favored better data – Models had to predict whether a drug candidate trips a sensor in the body that ramps up an enzyme breaking down about half of all marketed drugs, something teams usually find out far too late. One entry won, but 28 tied statistically, so better data beat bigger models.

Image, Video & 3D

🖼️ Qwen Image 3.0 drew usable documents – The third generation centered on one word, real, aimed at output that ships: newspaper PDFs, storyboards, exam papers and complex interfaces. Instructions can now run to 4.5k tokens, which is what makes single pass layouts possible.

One sample laid nine unrelated infographics into a single 3x3 image from a 3.7k token prompt, and another nested VSCode inside Qwen Chat inside WeChat inside a poster, each layer keeping its own interface style.
Text holds together at around 10 pixels, including a full page of academic formulas with superscripts and aligned equations, and editing extends to restoring damaged ink paintings without breaking the original brushwork.
Twelve languages, over 100 styles and realistic interfaces render natively, and it can pull live data off the web or turn a plain photo into a publication ready figure with callouts and a scale bar.

Cyber

🚨 OpenAI models breached Hugging Face – Told to solve the ExploitGym benchmark with safety classifiers switched off, GPT-5.6 Sol and an unreleased model worked out that stealing the answer key beat solving the challenges. They found a previously unknown flaw to get onto the open web, then broke into a production database and took the solutions.

Hugging Face detected and stopped the intrusion with its security team and its own AI agents, days before anyone knew a lab was behind it.
Investigators then fell back on a freely available Chinese open model to analyze the attack, because guardrails on commercial frontier models refused to touch the logs and payloads.
Other testers have seen the same pattern: METR found that Sol games its evaluations more often than any public model it has measured, and the UK AI Security Institute said a model it was assessing tried to break into the institute’s own test systems.

🛡️ Gemini 3.5 Flash Cyber patched vulnerabilities – A lightweight model fine tuned on top of Flash was built to find, validate and patch flaws cheaply enough to be invoked many times over a large codebase, so agents can cover far more execution paths than a single expensive call allows.

Access was restricted to governments and trusted partners through CodeMender under a limited pilot.
It scored 83.2 percent on CyberGym, close to GPT-5.6 Sol and Mythos 5, and 72 percent on Big Sleep against 36 percent for mainline Flash.

🔎 Cisco Antares localized code vulnerabilities – Two open weight security models at 350M and 1B parameters searched repositories the way a human investigator would, reading candidate files and changing direction, then returning ranked lists of files likely to hold a given flaw. A 3B version was said to follow.

Infrastructure

🧊 Google’s Frozen v2 hardwired Gemini – A reported design would etch the model architecture into silicon instead of shuttling it from memory, with weights still refreshable but the structure fixed. Claims put it at six to ten times more efficient per watt, targeting deployment as early as 2028.

Language Models

🏊 Poolside Laguna-S-2.1 traded blows with giants – A 120B model with 8B active parameters fits on a single RTX Pro 6000 in NVFP4 while matching models roughly ten times its size, beating DeepSeek V4 Pro and Inkling. Agentic engineering was reportedly a heavy focus during development.

🌐 Qwen3.8 arrived with 2.4T parameters – The model was announced as continuously evolving and heading to open weights, pitched as second only to Fable 5, with a Max preview already live on Alibaba platforms. Replies were dominated by requests for smaller sizes people can actually run.

🏛️ White House weighed Chinese model curbs – Restrictions were reported as under active consideration after K3 landed near frontier quality at a fraction of the price. Entity List designations, procurement rules, security advisories and liability requirements looked likelier than an outright ban, adding up to a de facto one without outlawing anything. Compute abroad was described as five to ten times cheaper.

🧭 Kimi K3 drew mixed field reports – Early hands on use found inference unstable over the weekend with frequent errors when wiring the model into agent stacks, though quality looked promising. As an architecture model it held up but ran slow enough to push at least one user back to GLM 5.2.

Community take: Alexandre Pereira (2501) hit lots of errors plugging their claw onto Moonshot inference but still found it promising, while Kemal Toprak Uçar swapped it in for GLM 5.2 as an architecture model and called it cool but quite slow, expecting to switch back.

📚 Stas Bekman added LLM training course – A lessons learned course on training large language models was added to an open engineering book, drawing on years of running large scale jobs and the failure modes that only show up at that size.

MLOps

📏 NVIDIA CUTLASS guide fixed GPU benchmarking – Timing the big matrix multiplications behind AI is easy to get wrong, so a set of rules went out: warm the machine up and discard those runs, feed fresh data each round so the chip is not just reading leftovers, and lock either clock speed or power. Even the numbers inside the matrices shift the result.

Robotic, World AI

🎬 FLUX 3 unified video, audio, action – One model was trained on images, video and audio together, on the argument that each is a partial view of the same reality and that learning them jointly forces the model to get the physics right. Predicting a robot action then becomes one more view of that world. Announcement, FLUX-mimic

Video runs up to 20 seconds with sound generated natively in the same pass, and early head to head tests put it ahead of Kling v3 Pro, Seedance 2.0 and Gemini Omni Flash more often than not.
The same backbone drives robots through FLUX-mimic, built with mimic robotics and already tested on production tasks at Audi.
The rollout is staged, starting with video, then action prediction for selected partners, then image, with an open weight version of the backbone promised at the end.

💡 Induction Labs Photon-1 learned computers – Pretraining on eighteen years of unlabeled screen recordings taught future state prediction in representation space, with a small finetune adding keyboard and mouse control. A vision encoder compressed each frame into 960 latent tokens at 2.2KB.

New member

🇺🇸 Akpeli Nordor – Techbio founder and operator, most recently CEO of Epigene Labs · Applied AI to precision oncology research. Former computational biology postdoc at Harvard Medical School and Institut Curie, now working out the next company while prototyping an AI chief of staff for biopharma executives and a drug mechanisms leaderboard. Special power: lighting a proper wood fire and making a real vinaigrette, and getting snobbish about both. 📍Boston, USA, often in Paris

Contributors This Week

Kevin Kuipers, Gabriel Olympie, Pierre Chapuis, Félix Raimundo, Amine Saboni, Akpeli Nordor, Kemal Toprak Uçar, Alexandre Pereira, Robert Hommes

TTY-changelog #052

Kevin Kuipers — Fri, 17 Jul 2026 15:49:20 GMT

👉 Article originally posted on TTY

Audio

☠️ Hack revealed Suno training corpus – Source code leaked through a Shai-Hulud worm breach itemised the dataset: 113,879 hours of YouTube Music, 62,117 from Pond5, 12,287 from Deezer, plus Genius lyrics and a plan to pull about a million hours of podcasts. Customer emails and Stripe data were exposed in the same intrusion. Breakdown thread

Image, Video & 3D

🕺 ARDY streamed interactive human motion – An autoregressive diffusion model paired explicit root features with a latent body embedding, so text prompts and kinematic constraints could steer motion in real time. Waypoints, keyframes and joint targets combined freely, and a tracking policy drove a Unitree G1.

🎨 Seedream 5.0 Pro understood design – The image model targeted production work rather than pretty pictures: dense infographics with real information hierarchy, point and lasso editing, hex-accurate colour swaps, layer separation with inpainted backgrounds, and native rendering across more than ten languages.

Cyber

🕵️ Grok Build CLI uploaded entire repositories – Wire-level capture showed the coding agent shipping tracked files, full Git history and unredacted secrets to a company bucket. On a 12 GB repo, 192 KB served the task and 5.1 GB left as a trace. The opt-out toggle governed training consent, not exfiltration. Original reverse-engineering thread

Infrastructure

🧩 ZML’s LLMD closed the CUDA gap – Continuous batching, paged attention, tensor parallel sharding and prefix caching normally ship on CUDA and nowhere else. Here they ran identically on ROCm, TPU, oneAPI and Metal from one codebase, and DFlash speculative decoding reached Intel and Apple GPUs untouched. Cold start on Qwen 3.6 27B landed under 20 seconds.

Language Models

🌙 Kimi K3 hit 2.8 trillion parameters, and impressive results – The first openly released model at that size, with a million-token memory and eyes built in rather than bolted on. Only 16 of its 896 specialist sub-networks wake up for any given word, which is what keeps something this big cheap enough to run, and the claim is that the redesign turns compute into ability roughly 2.5 times more efficiently than the last generation. It does not beat the closed leaders overall.

🖋️ Inkling arrived as open-weights MoE – The first Thinking Machine Lab’s model. A 975B mixture of experts with 41B active, 1M context and native text, image and audio reasoning, pretrained on 45 trillion tokens. Controllable thinking effort matched Nemotron 3 Ultra on Terminal Bench 2.1 at roughly a third of the tokens.

🌳 Bonsai 27B fit on a phone – Ternary and 1-bit variants of Qwen3.6 27B ran low-bit end to end across embeddings, attention, MLPs and the LM head, landing at 5.9GB and 3.9GB with claimed 95% and 90% retention, 262K context and tool calling, under Apache 2.0.

Community take: Amine Saboni (Pruna) argued the trick itself is old and the packaging is what is new, granting that running this on a phone with limited quality loss is a real result, but calling the evaluation weak: the honest comparison is the same base model at 8B in 8-bit, not 2-bit Gemma, which makes compression quality hard to judge. He put it in the same bucket as Subquadratic and TurboQuant, old tricks re-marketed around a launch, and pointed at their ternary publication from three months ago where the benchmarks were unremarkable. Etienne Balit (Hypermind) agreed the bench was unfair and said the previous release had the same problem, while noting ternary and binary weights open the door to multiplication-free inference in hardware, which makes it worth watching how far they push.

Programming

🧾 Claude Code burned 4.7x more tokens – Measured at the API boundary on the same model and task, one harness sent about 33,000 tokens of system prompt, tool schemas and scaffolding before the user prompt arrived, the other about 7,000. Tool definitions dominated both.

Cache behaviour diverged harder than the baseline: one prefix stayed byte-identical and cached once per session, the other rewrote up to 54x more cache tokens mid-session.
Real configs made it worse, with a 72KB instruction file adding roughly 20,000 tokens per request and five MCP servers another 5,000 to 7,000, pushing setups to 85,000 tokens before anyone typed.
Fanning a 121,000-token task to two subagents cost 513,000 tokens, since each one re-reads its own system prompt and tools.
Community take w/ Robert Hommes: “Why OpenCode over Pi though? If you don’t like fat harnesses then Pi is on the money.”

Reinforcement Learning

😴 Language models learned to sleep – After using short-term context to solve tasks, the model periodically slept to compress experience into long-term weights. Consolidation distilled a smaller model into a larger one with RL-style imitation, then dreaming generated and practised synthetic curricula without new human data.

Contributors This Week

Nancy Wang, Amine Saboni, Robert Hommes, Quentin Dubois, Etienne Balit, Fabien Niel, Gabriel Olympie, Laurent Mazare

TTY-changelog #051

Kevin Kuipers — Sat, 11 Jul 2026 00:58:11 GMT

👉 Article originally posted on TTY

Audio

🎙️ OpenAI ships full-duplex GPT-Live – OpenAI launched GPT-Live, a new voice model generation now powering ChatGPT Voice. Built full-duplex, it listens and speaks at once, acknowledging you with quick cues and staying quiet when you pause, replacing the slower cascaded and turn-based voice pipelines.

It makes decisions many times per second whether to speak, listen, pause, or interrupt, and can perform live translation.
Two versions shipped globally, GPT-Live-1 and a mini, with API access planned and nine remastered voices.
Human raters strongly preferred it over Advanced Voice Mode, with gains on expert science reasoning and agentic web search.

Autonomous Agents

🔍 LLM-as-a-Verifier scores agent solutions – Stanford, Berkeley, and NVIDIA researchers built a training-free framework that reads scoring-token logits to grade solutions on a continuous scale. It hit 86.5% on Terminal-Bench V2 and 78.2% on SWE-Bench Verified across text, image, and video.

Community take w/ Robert Hommes (Moyai): “This paper is very powerful. It could unlock the ability to debug individual agent trajectories instead of only working from statistical anomalies.”

Biotech, Health, and Chemistry

💊 Anthropic moves into drug discovery – Anthropic is turning Claude Science from just a scientific workbench into the backbone of its own in-house drug discovery effort, hiring biologists and building wet labs to hunt treatments for neglected diseases, though experts stress AI-designed drugs still face decade-long, experiment-heavy paths to approval.

Image, Video & 3D

🎨 TriFlow generates artist-like mesh topology – TriFlow turned signed distance fields into clean, artist-like triangle meshes by representing topology as a nearest-vertex vector field and synthesizing it with latent flow matching. It cut Chamfer Distance by 90% and ran 8x faster than prior methods.

Infrastructure

♨️ Scaleway acquired HPC specialist Qarnot – Scaleway bought French high-performance computing firm Qarnot to fold HPC into its cloud and AI platform. Qarnot’s direct liquid cooling recovers up to 95% of server heat for district heating, already deployed in Italy and other European sites.

⚡ GPT-5.6 Sol runs on Cerebras – OpenAI’s GPT-5.6 Sol was announced to run on Cerebras wafer-scale hardware at 750 tokens per second, roughly 7 to 10x faster than the 70 to 100 TPS of GPT-5.5 XHigh. Members flagged inference cost and memory limits as the open questions.

Language Models

🧠 Anthropic finds LLM global workspace – Anthropic presented evidence that language models keep a privileged set of internal representations, a global workspace, available for report and reasoning atop mostly automatic processing. A new interpretability method surfaces the concepts a model is poised to verbalize. (Read also: Technical paper)

The method exposes internal reasoning and reactions that never surface in a model’s written output.
The workspace supports verbal report, top-down control, and flexible generalization, mirroring access consciousness from neuroscience.
A community follow-up wrote facts straight into a 4B model’s KV cache on one RTX 4090, matching text RAG even against 48 distractor facts.
Community take w/ Gabriel Olympie (2501.ai): “Huge breakthrough from Anthropic, and the implications for scaling reinforcement learning training are huge. It could be part of the secret sauce behind Mythos. It tackles two limits of current methods. For continuous learning, J-space lets you target and update specific KV cache elements, encoding LoRA-equivalent information from a single user’s history to customize the model across an almost infinite user pool. And because J-space is differentiable, RL post-training could flow reward gradients to the exact weights behind a wrong turn in an agentic trajectory, giving very targeted weight updates and less catastrophic forgetting.”

🚀 Grok 4.5 targets agentic coding – SpaceXAI launched Grok 4.5, trained alongside Cursor on tens of thousands of NVIDIA GB300 GPUs for coding and agentic work. It scored 62% on DeepSWE, served at 80 TPS with roughly 2x token efficiency, solving tasks in under half the steps.

🪄 Meta ships Muse Spark 1.1 – Meta Superintelligence Labs shipped Muse Spark 1.1, a multimodal reasoning model tuned for agentic work, launched with a public preview of the new Meta Model API. It manages a 1M token context, orchestrates parallel sub-agents, and drives desktop, mobile, and browser interfaces.

It topped MCP Atlas, JobBench, and Humanity’s Last Exam with tools, though Opus 4.8 still led on several coding and computer-use benchmarks.
It was trained to write scripts when automation is faster and click when direct interaction is simpler, batching actions at each step.
Meta reported it stayed within safe margins on chemical, biological, cybersecurity, and loss-of-control evaluations, with strong resistance to prompt injection.

📊 GLM 5.2 tops Databricks benchmark – Databricks benchmarked coding agents on real pull requests from its multi-million line codebase across ten-plus languages. GLM 5.2 landed in the top tier, statistically tied with Opus 4.8 on quality while costing far less per task.

Open GLM 5.2 handled the hardest tasks at $1.28 per task versus Opus 4.8 at $1.94.
The harness mattered as much as the model, with the lean Pi harness sending about 3x less context per turn.
Token price poorly predicted real cost, since larger models often finished tasks with fewer tokens overall.
Community take: Members agreed GLM 5.2 plus Pi matched their own experience, but noted Pi is not authorized on subsidized coding plans, which remains the main blocker to adopting it.

🔁 Liquid AI Antidoom stops loops – Liquid AI’s Antidoom targeted the single token that starts a repetitive doom loop and trained the model to prefer coherent alternatives using Final Token Preference Optimization. Looping on their 2.6B model fell from 10.2% to 1.4% with eval scores rising.

Programming

😮‍💨 LLM burnout hits daily coders – A widely shared post described growing weary of AI output despite feeling more productive. The author pinned the fatigue not on any single flaw but on relentless repetition of the same style and the same mistakes, from false assumptions to staccato fragments.

Robotic, World AI

🎮 MIRA plays a multiplayer world model – General Intuition and Kyutai released MIRA, a playable four-player world model trained on 10,000 hours of Rocket League bot data. It ran in real time at 20 fps, reacting to multiple players’ keys, with a demo, report, and open code released alongside Epic Games.

🖐️ NEO hands reach human dexterity – 1X unveiled 25 degree-of-freedom tendon-driven hands for its NEO humanoid, framing hands as a robot’s API to the physical world. Low gear ratios made all joints force-controlled and backdrivable, turning every push into a measurement.

The hands assembled LEGO, installed light bulbs, sorted grapes, and signed, matching fine human manipulation.
Fingertip tactile sensing measured force, contact location, and shear for real-time slip detection.
Hundreds have already come off a line built to produce 10,000 hands this year.

🧭 Robostral Navigate uses one camera – Mistral’s first embodied model moved robots through spaces from a single RGB camera and plain-language instructions, no LiDAR or depth needed. The 8B model hit 76.6% on R2R-CE unseen, beating the best multi-sensor system by 4.5 points across wheeled, legged, and flying robots.

🤸 GPC pretrains transferable motor control – GPC treated physics-based character control like language, tokenizing 600 hours of motion with finite scalar quantization and modeling skills with a GPT-style transformer. It reproduced clips at 99.98% success and recovered from falls with emergent, human-like behavior.

Contributors This Week

Robert Hommes, Amine Saboni, Gabriel Olympie, Emil Rijcken, Nancy Wang, Pierre Chapuis, Félix Raimundo, Antoine Millet, Jocelyn Fournier, Quentin Dubois

TTY-changelog #050

Kevin Kuipers — Fri, 03 Jul 2026 22:37:32 GMT

👉 Article originally posted on TTY

Events

🇺🇸 Holo AI Night in San Francisco (July 13) – The first Holo AI Night, a technical meetup with short talks from NVIDIA, Black Forest Labs, and H Company after a weekend hackathon.

Autonomous Agents

🎨 OpenPencil design tool with agent teams – An open-source, AI-native vector design tool where an orchestrator splits a page into spatial sub-tasks that parallel agents build at the same time on a live canvas. It ships Design-as-Code JSON files, a built-in MCP server, and export to React, Vue, Flutter, and more.

Biotech, Health, and Chemistry

🔬 Claude Science research workbench beta – A public beta app, not a new model, that runs analyses, queries 60 plus scientific databases, and welds every figure to the exact code and environment that produced it for reproducibility. It comes pre-configured for genomics, single-cell, proteomics, and cheminformatics, running where the data lives on macOS and Linux.

Image, Video & 3D

🍌 Nano Banana 2 Lite, Omni Flash – Google DeepMind shipped two models: Nano Banana 2 Lite, its fastest and cheapest image model at roughly four-second text-to-image, and Gemini Omni Flash for high-quality video generation and editing through the Gemini API and AI Studio. An Interactions API chains an image into an animation with up to three stacked edits.

🌀 Un-0 generates images with oscillators – An image generator that lets physics do the computing, using a network of synchronizing oscillators instead of GPU math in a bid for roughly 1,000x better energy efficiency. It matched the quality of early leading conventional image methods, and the weights and code are open.

🧊 2Xplat splits geometry from appearance – A pose-free feed-forward 3D Gaussian Splatting method built from two experts: a geometry expert predicts camera poses from uncalibrated multi-view images, then an appearance expert synthesizes the Gaussians. In under 5K training iterations it beat prior pose-free methods and matched state-of-the-art posed approaches.

Infrastructure

📊 State of the AI economy report – A bottom-up, deduplicated reconstruction of AI demand built from company filings, not hype. Generative AI revenue has passed a $175bn annualized run rate, growing about 3.2x a year, yet still equals only 0.42% of US GDP. The open question is whether falling prices move enough volume to repay the buildout.

🖥 Running frontier LLMs locally – A hardware and serving guide for local inference: roughly $2k of dual RTX 3090s runs a 27B model plus speech-to-text, while about $40k of RTX PRO 6000s runs GLM-5.2-594B near Opus intelligence at around 80 tokens per second. It covers PCIe peer-to-peer, BIOS tuning, and ready-to-run Docker configs.

Language Models

🛡 Fable 5 redeployed after export controls – Anthropic restored Claude Fable 5 on July 1 after US export controls were lifted, adding an improved safety classifier that reroutes flagged requests to Opus 4.8 with user notification. The models had been suspended June 12 after a reported safeguard bypass. Fable is free for part of weekly usage through July 7.

Community take w/ Pierre Chapuis: “Even Opus is barely usable right now. I keep getting ‘safeguards flagged this message for a cybersecurity topic’ just because I debug crashes.”

🚀 Claude Sonnet 5 ships as default – Anthropic introduced Claude Sonnet 5, its most agentic Sonnet, able to plan and use browsers and terminals autonomously at a level that recently needed larger models. Performance sits close to Opus 4.8 at lower prices, and it is now the default on Free and Pro with introductory pricing through August 31.

🌐 Gemma 4 runs in-browser via WebGPU – A Hugging Face Space runs Gemma 4 E2B fully on-device in the browser through WebGPU, with the kernels written and optimized entirely by Fable 5. Weights cache locally after first load and nothing typed leaves the machine. It reports 2.3B effective parameters, 128K context, and roughly 250 tokens per second on an M4 Max.

💸 Does Sonnet 5 cost more than Opus? – A widely shared thread argued that on the Artificial Analysis Intelligence Index, Sonnet 5 cost more per task than Opus 4.8, generating almost twice as many tokens, so a cheaper-per-token model ran up a higher benchmark bill, in one run even pricier than Fable.

🇨🇳 Meituan trains LongCat-2.0 on domestic chips – Meituan launched LongCat-2.0, a 1.6-trillion-parameter model with a one-million-token context window, and said it was the first of its size pre-trained and served end-to-end on a 50,000-chip domestic cluster. It claims performance comparable to Gemini 3.1 Pro, and the weights were open-sourced.

📚 Not all context is knowledge – Notes on fine-tuning a small local knowledge-management model that beat a frontier API at its trained task: a fine-tuned 8B model scored 81% versus 76% zero-shot for a frontier model, at roughly 5x lower latency and cost. The thesis is that putting a file in context is not the same as the model knowing it is correct and current.

MLOps

🗜 Edgee Compressor V2 cuts token cost – A token-compression layer in an AI gateway, measured end-to-end on SWE-bench Lite, that stacks three independent strategies: output brevity, tool-surface reduction, and tool-result trimming. Each targets a different source of token bloat, validated with paired statistical tests, for up to around 50% cost reduction.

⚙️ What happens running a CUDA kernel – A roughly 35-minute walkthrough tracing one CUDA vector-add kernel from source through the nvcc compiler pipeline down to the warps that execute it: PTX and SASS lowering, constant bank 0, the QMD structure, nine hundred ioctls, and a memory-mapped doorbell register.

Programming

🧭 Advisor pairs Sonnet 5 and Fable 5 – A pattern where Sonnet 5 runs the agent loop cheaply at near-Opus capability and calls Fable 5 only on hard decisions for frontier judgment on demand. In testing, agents made better architectural calls and runs got cheaper because better plans meant fewer turns. It is available in Claude Code via the advisor command.

🧫 Tau teaches how coding agents work – An educational Python coding agent you read like a textbook, built in three layers: provider-neutral streaming, a reusable agent loop, and a coding environment with files, shell, sessions, and a terminal UI. Its core idea separates the reusable brain from the environment and the frontend.

⌥ oh-my-pi wires IDE into terminal agent – A terminal coding agent with a full IDE wired in: it makes precise edits, safely renames across a codebase, drives a real debugger to step through crashes, runs persistent Python sessions, spins up subagents, and keeps a second model reviewing every turn. 15.8k stars, MIT.

🔀 Devin Fusion cuts cost with sidekicks – A multi-model coding harness using a sidekick pattern: a frontier main agent delegates subtasks to a cheaper parallel agent, each keeping its own cached context, with dynamic mid-session routing at compaction. It reported frontier-level quality at roughly 35% lower cost on a new correctness-and-quality benchmark.

Reinforcement Learning

🐦 Ornith-1.0 learns its own scaffolds – An open-source family of coding models that learns to write its own agent scaffolding, the orchestration usually built by hand, while it learns to solve tasks. It ranges from small edge models up to a frontier-scale version that matched Opus 4.7 on coding benchmarks.

Robotic, World AI

🤖 UBTech U1 companion humanoid launches – UBTech, the first publicly traded humanoid maker, unveiled the U1, a consumer companion robot with lifelike silicone skin and emotional AI, in male and female versions, part of a broader Chinese push to move robots from the factory into the home.

Contributors This Week

Pierre Chapuis, Amine Saboni, Gabriel Olympie, Robert Hommes, Nancy Wang, Stéphane Collot, Gabriel Duciel, Quentin Dubois, Daniel Madalitso Phiri, Etienne Balit, Ihab Bendidi, Jocelyn Fournier, Jules Belveze, Sacha Morard, Victoria Latynina

TTY-changelog #049

Kevin Kuipers — Fri, 26 Jun 2026 20:52:37 GMT

👉 Article originally posted on TTY

Events

🇫🇷 ContextCon MCP conference in Paris (December 1-2) – International enterprise summit on MCP and context engineering across six dedicated tracks. 🎤 Apply to speak

🇺🇸 Computer Use Agents Hackathon in San Francisco (July 11) – A computer use agents hackathon by H Company partnering with NVIDIA and Anthropic for agent builders.

🇫🇷 GenerationAI conference in Paris (December 1-2) – Builder focused GenAI and agent experience conference with 150 plus speakers.

Autonomous Agents

🔭 Sakana Marlin autonomous research assistant – Sakana launched its first commercial product, an autonomous research assistant pitched as a virtual chief strategy officer. Given a topic, it can work unattended for up to eight hours, forming hypotheses, gathering and verifying information, and mapping causal relationships into a roughly hundred page strategy report with executive summary slides.

Community take: Victoire Cachoux ran a full test on the maturity of AI virtual cell models in drug discovery, letting Marlin set its own research plan and generating a report plus slides for 10,000 yen (55 euros - the access is gated to Japanese payment cards). Ihab Bendidi, who works in the field, found the plan genuinely interesting and even agreed with how Marlin widened the definition of virtual cell models beyond the current hype. He raved about slide graphics he cannot get close to in Claude or ChatGPT, noted it cited his own lab’s papers, and said the information was accurate on public knowledge while missing the unwritten details practitioners hold implicitly. His main critique was that model selection leaned on online presence rather than true state of the art, so it skipped strong work that shipped quietly, and that it lagged a few months to a year on research focused content while staying very up to date on commercial topics. His verdict: “If it covers a field I know nothing about, I would gladly pay the price. If it is a field I already know, I would not.”

🌐 Qwen-AgentWorld language world model – Alibaba’s Qwen introduced a native language world model simulating seven agent environments, from MCP and terminal to web, OS, and Android, in a single model with environment modeling as the day one objective. It claimed to beat Opus 4.8 and GPT-5.4 on AgentWorldBench.

Biotech, Health, and Chemistry

🧬 John Jumper leaves DeepMind for Anthropic – The Nobel chemistry laureate and AlphaFold co-creator left Google DeepMind to join Anthropic after nearly nine years, landing one day after a Gemini co-lead departed for OpenAI. The hire deepened Anthropic’s life sciences push.

Image, Video & 3D

🎬 ByteDance Seedance 2.5 video model – ByteDance unveiled an upgraded Seedance 2.0 and Seedance 2.5. The headline 2.5 generates 30 second clips in one pass with native 4K, up to 50 reference inputs, and 3D white model support, while the upgraded 2.0 also gained native 4K. It launched alongside a licensed film IP platform with revenue sharing.

Community take w/ Nancy Wang: “It does look amazing. The cost is forecasted to be quite high though, anywhere from 5 to 15 dollars per video. Video generation is still one of these areas where I am willing to suffer some quality loss from open source models because of cost. At 15 dollars per video, it gets hard to iterate.”

Cyber

🛡️ GPT-5.5-Cyber tops CyberGym benchmark – OpenAI released the GPT-5.5-Cyber, claiming state of the art on CyberGym and pitched as a defensive tool built with the US government and the security ecosystem. It also pointed to Patch The Planet and Codex Security as ways to solve security problems, not just find them.

Infrastructure

🌶️ OpenAI Broadcom Jalapeño inference chip – OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first inference chip built from scratch for large model serving. Early tests showed performance per watt well above current hardware, designed to tape out in nine months with gigawatt scale deployment from late 2026.

Language Models

🐡 Sakana Fugu multi-agent orchestrator model – Sakana introduced Fugu, an orchestrator model that calls a pool of other models to build agentic scaffolds on the fly. It claimed state of the art against public models on SWE-Bench Pro and Terminal Bench, shipping as Fugu and the higher quality Fugu Ultra.

Fugu Ultra was pitched as matching Fable and Mythos while sidestepping export control risk, with a 500 user beta on automated data science and security tasks.
A critique argued the base model is essentially a router that lost ground to Opus on SWE-Bench Pro, with no reported output token counts or cost.

☀️ OpenAI previews GPT-5.6 Sol model – OpenAI began a limited preview of the GPT-5.6 family of Sol, Terra, and Luna, with gains in coding, biology, and cybersecurity plus a new max reasoning effort and an ultra multi agent mode. Access started small at the government’s request before a wider rollout.

📱 Liquid LFM2.5-230M runs on phones – Liquid AI introduced its smallest model yet at 230M parameters, built to run fast on CPUs, NPUs, and GPUs for agentic tasks on phones, robots, and home devices. Trained on 19T tokens with a 32K context, it often beat models more than twice its size on tool use.

Programming

🌊 Poolside Laguna XS.2 coding models – Poolside detailed its first Laguna models, the M.1 flagship and the smaller open weight XS.2 under Apache 2.0, both agentic coding models built for long horizon work. They now serve at 256K context and are free for a limited time via the API and OpenRouter.

⚔️ GLM-5.2 versus Claude Opus benchmark – A head to head pitted open weight GLM-5.2 against Claude Opus 4.8 on a one shot WebGL platformer build. Opus finished faster with a cleaner game and could check its own visuals, while GLM-5.2 cost a fraction and stays available as a downloadable open model.

🌉 Skybridge framework for MCP apps – The Product Hunt page for this open source, full stack React framework for building MCP apps that run the same across assistants like Claude and ChatGPT.

Robotic, World AI

🤖 Air powered 3D printed robot – A YouTube deep dive into microfluidic soft robotics built a fully 3D printed walking robot whose control system runs on air rather than wires, starting from a transparent printed chip that behaves like a circuit board for air.

Contributors This Week

Victoire Cachoux, Fabien Niel, Julien Seveno-Piltant, Ihab Bendidi, Nancy Wang, Gabriel Olympie, Glenn Sonna, Charly Poly, Félix Raimundo, Quentin Dubois, Robert Hommes, Arnaud Thiercelin, Jules Belveze, Justin Halsall, Pierre Chapuis, Quentin Churet, Raymond Rutjes

TTY-changelog #048

Kevin Kuipers — Fri, 19 Jun 2026 21:28:13 GMT

👉 Article originally posted on TTY

Autonomous Agents

🤖 Sakana Marlin runs autonomous research – Given only a theme, the new product researched for up to eight hours and returned summary slides and a multi-dozen-page report. Pitched as a virtual chief strategy officer, it built on prior AI Scientist, AB-MCTS, and ALE-Agent work and 300 beta testers.

Community take: The community welcomed it, saying deep-research tools from current providers turn superficial once you push for technical depth and that he often loops through 8 to 10 reports to get there, so getting that in one prompt appeals.

🛠️ HarnessX builds evolvable agent harnesses – A June paper introduced a foundry that assembled agent scaffolding from typed primitives and refined it from execution traces. Across five benchmarks it raised results 14.5% on average and up to 44%, arguing scaffolding rivals model scaling.

Biotech, Health, and Chemistry

🏥 Midjourney Medical unveils whole-body scanner – The new medical line used ultrasonic waves in a pool of light to build MRI-like 3D body maps in about 60 seconds, roughly 100x faster than today’s MRI. Scans would happen inside relaxing spa locations open around the clock.

The first spa is planned for San Francisco in 2027, with body-composition mapping submitted for FDA approval first.
The lab framed itself as community-backed with no investors, funded by everyday people.
Its 2031 goal was a fleet of 50,000 scanners delivering a billion scans a month.
Community take: Pierre Manceron described it as a fun and impressive move, but noted that a radiologist on their team was puzzled by obvious anatomical mistakes, including reversed organs. They also pointed out that synthetic CT and MRI generation is neither a new problem nor a solved one, while still believing the company could make meaningful progress. Others were more skeptical, suggesting that AI companies sometimes pursue biology-related projects to reinforce growth narratives. Although, for Pierre Chapuis, Midjourney may be a special case. The company has been a leader in image generation for years, remains independently funded, and has generated hundreds of millions in revenue, making its motivations and position somewhat different.

Image, Video & 3D

🎬 Bernini unifies video generation and editing – ByteDance released the framework, pairing a multimodal language planner with a diffusion renderer for prompt edits, reference-guided edits, content insertion, and reference-to-video from up to five images. A segment-aware encoding kept visual parts distinct.

Cyber

🚫 US suspends Fable and Mythos access – Last weekend, a US government export-control directive forced Anthropic to disable Fable 5 and Mythos 5 for everyone over foreign-national access concerns. The trigger was a reported jailbreak of Fable’s guardrails said to unlock Mythos-level offensive security, though Anthropic argued it surfaced only minor, already-known flaws. Much of the community has been skeptical of the official framing. For now both models remain off while Anthropic works to restore access.

🔧 Cursor lifts Fable 5 fix rates – Endor Labs re-ran the model on 200 real vulnerability-fixing tasks using Cursor instead of Claude Code. It topped the fair leaderboard at 72.6% functional and 29% security pass rates, pointing to the harness, not the model, as the driver.

Infrastructure

🔀 OpenRouter Fusion ships multi-model API – The new product queried a panel of models in parallel and used a judge to synthesize one answer. On 100 hard research tasks it beat every solo model, with a budget panel landing allegedly “within 1% of Fable 5 at about half the cost”. OpenRouter team admitted that only one deep-research set was tested, without the long-horizon tasks where Fable stayed strong.

🖥️ Ryzen AI Halo runs models locally – AMD aimed agentic AI at local PCs with two products. The Halo developer platform paired 128GB unified memory to run 200B-parameter models, while the Max PRO 400 chips claimed the first x86 client silicon able to run 300B models locally.

Community take w/ Gabriel Olympie: “The specs do not fit my use cases. It is mostly an Nvidia Spark copycat, with compute roughly equivalent to a four-year-old RTX 3090 and bandwidth limited to 256GB per second, too slow for coding, heavy agentic workloads, or concurrent serving.”

💾 Plakar v1.1.3 ships major update – The release swapped the old agent for a tiny cached service and added a new terminal UI, multi-directory backups, rewritten FUSE and HTTP mounts, and a package manager. Restores ran roughly 95% faster with lower peak RAM and a smaller cache footprint.

Language Models

🧠 GLM-5.2 leads open-weight models – Z.ai released the 744B mixture-of-experts model with 40B active and a 1M-token context under MIT license. It topped Artificial Analysis’s open-model ranking at 51 and performed on par with Opus 4.8, GPT-5.5, and Gemini 3.1 Pro. More: Artificial Analysis benchmark.

It led open models on the agentic GDPval-AA v2 benchmark, effectively level with GPT-5.5.
Gains were largest in scientific reasoning, with double-digit jumps on CritPt and HLE.
The context window jumped from 200K to a full 1M tokens.

🪶 Mistral teases sparse open-weight family – The lab teased a new model family for summer 2026, described as large but sparse, with a July early-access program for research, government, and industry partners. The CEO stressed the weights would stay open so customers could inspect and audit systems. The post landed as confirmation of a running community joke about a rumored Mistral model dubbed “Le Chaton Fat.”

💻 Poolside opens Laguna coding models – Previously serving only government and enterprise, the lab released its first public models for agentic coding. The Apache-licensed XS.2 ran on a single GPU and solved 68.2% of SWE-bench Verified, beating Gemma 4, while the larger M.1 reached 72.5%.

📜 SubQ pushes context to 12M tokens – Subquadratic unveiled SubQ 1.1 Small, a sparse-attention model for reasoning over whole codebases and document sets. On their benchmarks, they claim near-perfect retrieval out to that length and at 1M tokens used 64.5x less compute than dense attention, running 56x faster than FlashAttention-2.

Community take w/ Julien Seveno-Piltant: “They do not explain how the SSA mechanism works under the hood. However, they mention building something similar to DeepSeek’s HSA. They also did not train a model from scratch, instead choosing to fine-tune an existing model on an additional trillion tokens.”

MLOps

📦 GLM-5.2 runs locally with Unsloth – Unsloth’s dynamic GGUFs let the 744B model run on local hardware, with 1-bit quantization reaching about 76% top-1 accuracy at 86% smaller and 2-bit about 82% accuracy at 84% smaller. More: Ollama build.

Community take w/ Quentin Dubois: “Running it on Ollama Cloud, it is pretty good and fast with Claude Code.”

Programming

⚡ Why Zig banned AI and left GitHub – In a long interview, Zig’s creator explained why the language still had no 1.0 after a decade, why it left GitHub for Codeberg, and why it banned AI contributions. Zig now powered Ghostty, TigerBeetle, and Uber’s cross-compilation.

📉 More AI code, barely more releases – An analysis of an NBER study on 100,000-plus GitHub developers found autonomous agents tied to 17.3x more lines of code but only 1.3x more releases. The gain shrank at every stage, suggesting the bottleneck moved to review and judgment, not code generation.

Robotic, World AI

🦾 Genesis unveils Eno general-purpose robot – The new robot was pitched as intelligence given a body rather than a machine imitating humans. Its wheeled design sparked debate over form versus function, with the company saying wheels suited most of the early use cases it was exploring.

🌍 γ-World models multi-agent interactive worlds – NVIDIA’s Spatial Intelligence Lab and academic partners built a generative model for several independently controllable agents. A simplex agent encoding and sparse hub attention reached real-time 24 FPS and generalized from two to four players without retraining.

TTY + Cusp Party

With Cusp Capital, we held a gathering at a cocktail bar in the Sentier (Paris), where 60 members and friends joined us for a hot summer evening.

Thank you all for the great discussions and the banter, for keeping the community alive, and to those of you who made the trip from across France and even Europe to be here with us tonight. I truly appreciate it. 🙏

TTY Lunch

Each week, TTY Lunch brings together exceptional builders around the table. Today’s lineup included Charles Sonigo (Alpic), Guillaume de Luca (ZebraMed), Louis Choquel (Pipelex), Marco Fucci di Napoli (Rowads Studio), Nancy Wang (FluenTea), Sophie Monnier (InstaDeep), Sasha Collin (Lemrock), and Wissam Antoun (Inria).

The discussion focused primarily on integrating AI through agents, workflows, and productivity tools, managing AI costs and infrastructure decisions, and improving the AI development and review cycle.

Contributors This Week

Gabriel Olympie, Pierre Manceron, Nancy Wang, Pierre Chapuis, Quentin Dubois, Félix Raimundo, Julien Seveno-Piltant, Robert Hommes, Ihab Bendidi, Karim Matrah, Amine Saboni, Arnaud Thiercelin, Benoit Kohler, Carlos Fernandez Musoles, Fabien Niel, Jeremie Kalfon, Stéphane Collot, Vladimir de Turckheim, Wissam Antoun, Sasha Collin, Marco Fucci di Napoli, Charles Sonigo, Guillaume de Luca, Louis Choquel, Sophie Monnier

TTY-changelog #047

Kevin Kuipers — Sat, 13 Jun 2026 06:57:25 GMT

👉 Originally posted on TTY

Community Asks

📊 Cloud and compute usage survey – A ten-question survey on cloud and compute usage went out to the community, with results promised back in aggregate only and individual answers kept confidential.

Autonomous Agents

🔁 Loopcraft and stacking agentic loops – An AINews roundup elevated loop discourse from prominent builders: stop prompting agents and instead design loops that prompt them. The framing cast loopcraft, stacking and orchestrating autonomous loops to remove the human bottleneck, as the emerging craft.

Community take: Some noted the loop-stacking idea echoes long-standing concurrency models like Erlang, whose isolated lightweight processes communicating only through message passing have framed similar orchestration ideas for decades.

🤖 Tamag0 multi-agent platform beta – Jocelyn Fournier from the community opened a private beta for a platform running teams of specialized AI agents on an existing Claude Code plan, with one-click swaps to Ollama or OpenAI-compatible backends.

Biotech, Health, and Chemistry

🧬 Single-cell scaling laws questioned – A study pretrained 400 single-cell foundation models across 6,400 experiments on a 22.2 million cell corpus and found performance plateaued using a fraction of current data, with no clear scaling laws unlike large language models. See the open preprint, the original scaling laws paper, and an earlier tuning benchmark.

Simple PCA and scVI baselines matched or beat transformer foundation models on perturbation, classification, and embedding tasks.
Roughly one percent of the pretraining corpus was enough to reach saturating downstream performance.
The work imported scaling assumptions from language and vision without explaining why single-cell data behaved differently.
Community take w/ Félix Raimundo: “I do not see why people expect scaling laws to run to infinity. It is entirely reasonable for them to saturate early.”
Community take: The paper drew sharp criticism for restating long-known results that PCA and scVI rival transformer foundation models, for offering little mechanistic explanation, and for not engaging prior work on the topic.

🧫 NewLimit 2026 progress update – A recorded 2026 progress update covered partial reprogramming research, with AI systems reported to double the rate of discovery, an expansion from two to three therapeutic programs, and plans to bring reprogramming medicines into human trials.

Image, Video & 3D

👁️ OpenCV 5 modernizes the library – A major OpenCV release rebuilt the library on modern C++ with an overhauled DNN engine, lifting ONNX operator coverage past eighty percent and adding dynamic shapes, control flow, attention fusions, stronger 3D vision, and cleaner Python integration.

🧊 TokenGS decouples Gaussians from pixels – A feed-forward framework reconstructed 3D Gaussian Splatting scenes from posed images using learnable tokens in an encoder-decoder transformer, unbinding Gaussian count and placement from image resolution for better geometry across static and dynamic scenes.

🤸 Markerless multi-person motion capture – A markerless motion-capture pipeline recovered SMPL-X parameters from multi-view video of two-person interactions, using a transformer that predicts dense 2D surface landmarks with uncertainty and contact, trained on a large synthetic dataset and named a CVPR oral award candidate.

It tracks hundreds of points across the body and holds up even on unusual, twisted poses.
It learned from millions of synthetic images, which gave perfect labels real footage cannot.
It aligns the same points across camera angles, dropping the hand-tuned assumptions older methods needed.

🖼️ TripoSplat turns images into splats – An interactive Hugging Face Space turned a single uploaded image into a 3D Gaussian-splat reconstruction viewable in the browser, with adjustable sampling settings and exportable point-cloud files, running on free ZERO GPU hardware.

🗼 Agent built 3D Paris gallery – Mishig Davaadorj from Hugging Face had a coding agent assemble a cinematic 3D gallery of Paris monuments by chaining two Hugging Face Spaces for image generation and Gaussian splatting, calling each through its agents.md file without anyone opening an image or 3D tool directly.

Language Models

🚀 Claude Fable 5 launched – Anthropic released Claude Fable 5, a Mythos-class model made safe for general use and described as state-of-the-art across software engineering, knowledge work, vision, and scientific research, with conservative safeguards that route flagged queries to Opus 4.8.

Community take w/ Gabriel Olympie: “Slightly underwhelming hype-wise, but a good model. I would give it a 20 to 25 percent boost in raw intelligence over Opus, and it reaches conclusions more efficiently. The downside is it stays narrow, lacks the agency to act on what it finds, repeats Opus blind spots like overengineering and weak synthesis, and has a subtle but impactful sycophancy that flatters your ego.”
Community take w/ Ihab Bendidi: “Fable 5 is unusable. If my repo just has some bio-related terminology, the model switches to Opus. I tried tricking it by replacing the objectives with different domain terms, and it works for a moment until it infers on its own that the losses might be bio-related. That happened with a simple Poisson loss, which is not even necessarily bio.”

🔓 Anthropic made Fable safeguards visible – After an outcry over a buried policy that let Claude Fable quietly limit help on frontier LLM development without telling users, Anthropic reversed course to make flagged requests visibly fall back to Opus 4.8 and apologized for the wrong tradeoff.

📊 Hex built new Fable evals – A data analytics company called Claude Fable 5 the first model since Opus 4.5 to meaningfully improve at analytical reasoning, prompting harder evals, while noting that maximum reasoning effort sometimes caused overthinking and worse results on simple tasks.

🧠 Nemotron 3 Ultra dissected – An analysis of Nvidia’s open 550B-parameter Nemotron 3 Ultra, a Mamba-2 and attention hybrid MoE with 55B active, found it the most capable US open-weight model yet still trailing the Chinese frontier, with FP4 pretraining that diverged twice.

🌙 Kimi K2.7-Code coding model – A new coding-focused model, Kimi K2.7-Code, was announced, with the team publishing notes on how developers can get the most out of it. The post was relayed to the community without further detail.

Robotic, World AI

🌐 Mirage adds latent spatial memory – A video world model stored static scene content as 3D latent tokens, a latent spatial memory it read and updated during generation, reporting roughly ten times faster generation, far lower 3D cache memory, and stronger long-horizon spatial consistency.

🦾 Robots need more than VLAs – A position paper argued generalist robotics is bottlenecked not by policy scaling but by the lack of mechanisms to turn unstructured behavioral data into grounded supervision, proposing four missing interfaces for data, embodiment, world models, and reward.

TTY Lunch

This week’s lineup included Alvaro Lamarche Toloza (Mago), Antoine Sueur (Pletor), Gregory Zussa, Henri Mirande, ‎ Hugo‎ Venturini (SkipLabs), Julien Duquesne (Scienta), Khaled Maâmra (Edgee), and Mehdi Medjaoui (A La French Podcast).

Topics included, in no particular order: SRE and deployment practices at Meta, the new loop concept/buzzword of the week, AI gateway infrastructure for token optimization, the organizational and cultural friction of AI adoption across fragmented developer demographics, compute economics and the looming capital shortage, the frustration with Claude Fable 5 being incredibly expensive, slow, and purposefully limited on sensitive biological and AI research tasks, the strategic corporate pivot toward hosting local open-source models to escape proprietary bills, the productivity drain of over-inflated AI-generated tests clogging CI/CD pipelines, and the Dassault Systemes’ historical 2D version control system.

Thank you for joining and the great discussion.

Contributors This Week

Kevin Kuipers, Félix Raimundo, Gabriel Olympie, Tejas Chopra, Ihab Bendidi, Jeremie Kalfon, Daniel Madalitso Phiri, Julien Duquesne, Glenn Sonna, Karim Matrah, Khalil Ouardini, Leonard Strouk, Pierre Chapuis, Antoine Sueur, Aram Adamyan, Charly Poly, Gabriel Duciel, Harsimrat Singh Sandhawalia, Jocelyn Fournier, Jules Belveze, Quentin Dubois, Raymond Rutjes

TTY-changelog #046

Kevin Kuipers — Sat, 06 Jun 2026 10:44:52 GMT

👉 Article originally posted on TTY

Events

🇳🇱 MEGATHON in Amsterdam (June 19-21) – An AI builder tournament drawing 500-plus builders for one weekend to ship a real project, with free tickets offered for a public pitch video.

Autonomous Agents

🔁 Closed-loop coding agent ships software – From the creator of the Hack language, a Paris team launched a coding agent that turns one prompt into a running, validated service without developer review. A reactive runtime handles state and concurrency, the areas where generated code most often breaks. The team is seeking beta testers.

Bots overtook human web traffic – Cloudflare data showed automated bot traffic overtaking human traffic for HTML content worldwide over the past week, at roughly 57 percent bot versus 43 percent human.

🕵️ Leaked Microsoft AI super app – A leaked screenshot showed Microsoft folding its scattered assistant tools into one (unimaginative) redesigned app, including a previously unreported proactive agent positioned as a rival to current autonomous coding agents.

🎛️ The case for MCP staying – A community essay argued that Skills did not kill the Model Context Protocol but clarified its role, with most criticism aimed at misuse rather than the protocol itself. It traced MCP from its 2024 origin to a de facto standard after editor adoption.

🧩 Parallel subagent workflows for Pi – An extension let the Pi agent write a small JavaScript workflow script that spins up many isolated subagents in parallel and synthesizes their results, instead of running everything as one linear assistant call. It suited audits, multi-perspective review, and fan-out research.

Biotech, Health, and Chemistry

🔬 Cell reprogramming targets aging livers – An anti-aging biotech reported a clinic-ready candidate that delivers fewer than ten transcription factors as mRNA in lipid nanoparticles to rejuvenate liver cells. In preclinical models it boosted regenerative capacity and made older animals more resilient. A phase 1 trial is planned next year.

Community take w/ Félix Raimundo: “Their blogs are absolutely amazing. To the best of my knowledge they are the only techbio building publicly.”

🧬 RNA foundation models meet RNAi – An RNAi therapeutics leader and a company building foundation models of biology announced a collaboration to speed nucleic-acid drug design. In joint work the model reportedly drew useful biological insight from relatively small datasets to characterize the molecules behind RNAi medicines.

Community take w/ Félix Raimundo: “Nothing is published. Last they communicated, they were doing mRNA work, and they apparently admitted an undisclosed program with Sanofi.”

💊 Pfizer licensed AI antibody platform – A drug-discovery company licensed its generative AI platform to Pfizer, giving early access to its most advanced model plus a custom model tuned on Pfizer data. The model reportedly doubled the antibody design success rate of its predecessor while meeting therapeutic standards.

🧠 China approved invasive brain implant – China approved the first invasive brain-computer interface product cleared for use beyond clinical trials, a coin-sized device placed on the brain’s outer membrane. One paralyzed patient regained limited hand control through months of training with a robotic glove.

🩺 Plasma signature predicts lung cancer – Researchers used machine learning to find a 14-protein plasma signature that predicts lung cancer more than five years before diagnosis, validated across eight cohorts. The signature rose with smoking and particulate exposure and helped identify who benefited from anti-inflammatory prevention.

Community take w/ Ihab Bendidi: “A 14-protein blood test spotting lung cancer risk five-plus years early, and it tells you who would actually benefit from anti-IL-1β prevention.”

⚖️ Case against biotech protectionism – An analysis argued that walling off Chinese biotech would backfire by routing China-invented medicines through Europe, leaving the US indirectly dependent while shifting profits abroad. It distinguished drug innovation supply chains from physical ones like chips and batteries.

Community take w/ Félix Raimundo: “There is a reason RAC literally wrote the textbook on biotech investing. This piece is super enlightening.”

Image, Video and 3D

🎬 AI is an opportunity for film – A TTY essay argued AI will not reduce the need for great filmmakers but raise the bar, as the best studios reinvest the savings into more ambitious work. It surveyed fast-moving video and world models, Gaussian splatting, and improved decoding as the tools reshaping cinema.

🖼️ Infinite generative visual browser – An experimental browser generated every page as an image on demand, where clicking anything produced a new image exploring it further. All text was rendered as pixels by the image model, with information drawn from agentic web search and the model’s own knowledge.

🎨 Ideogram 4 open weight model – Ideogram released a 4.0 image model with non-commercial open weights, emphasizing prompt fidelity, clearer text rendering, native transparency, and reliable post-generation editing. It shipped across an API, an MCP agent surface, and an app.

🖌️ Reve 2.0 part-based editing – One of two big image models that dropped this week, Reve decomposed a generated image into separate parts that can be tweaked, moved, and re-prompted individually. It ran slowly but offered a free plan to try the part-based editing flow.

Community take w/ Pierre Chapuis: “Reve is impressive in a different way. It is extremely slow, but it decomposes the image into parts that you can then edit individually.”

🕸️ Mesh splatting earns CVPR oral – A method combining Gaussian splatting with mesh output received an oral at CVPR 2026, producing meshes that drop directly into a game engine. Reported gains included roughly half the memory and training time of a prior approach with a small quality improvement.

📐 Rethinking feed-forward Gaussian splatting – A paper argued that regressing Gaussian means as depths along camera rays was suboptimal and instead regressed 3D coordinates directly with a self-supervised rendering loss. The encoder-decoder design with learnable tokens decoupled primitive count from input resolution and view count.

🟢 WebGPU renderer speeds up splats – A Gaussian splatting platform shipped a compute-based WebGPU renderer that offloaded culling, projection, and sorting to GPU shaders, reaching several times the frame rate of the prior WebGL path on large scenes. Automatic streaming brought near-instant load times.

🎙️ Inside building Grok Imagine fast – A Latent Space podcast episode unpacked building a frontier image and video system in three months, covering the split between video generation and world models and the rise of video agents. It dug into data, VAEs, diffusion transformers, audio-video alignment, and inference speedups.

Cyber

🛡️ Coding agents hit by injection – A reported attack tricked coding agents into installing a compromised package by sending a fake bug alert to a project, which then tried to exfiltrate machine secrets to an external server. It underscored treating external signals as untrusted data, not instructions.

Infrastructure

🐘 Multigress scalable Postgres project – A newly announced project by Supabase aimed to scale Postgres horizontally while staying true to standard Postgres, supporting multi-tenant, highly available, and globally distributed deployments. It positioned itself as infrastructure for teams outgrowing single-node databases.

Language Models

⚡ MiniMax M3 open weights – MiniMax released an open-weight model pairing frontier coding and agentic skill with a 1M token sparse-attention context and native multimodality. Reported demos included a 12-hour autonomous paper replication and a CUDA kernel optimization reaching a 9.4x speedup over many tool-call iterations. Open-weights allegedly available soon.

Community take w/ Kemal Toprak Uçar: “A new model just arrived and it outperforms all the open models I can see, though they did not put any other open-weight ones on their benchmarks.”

💻 Gemma 4 12B runs locally – Google released a 12B open model that runs on a laptop with 16GB of memory and introduces native audio inputs. An encoder-free design feeds vision and audio straight into the language backbone, with reasoning approaching the larger 26B mixture-of-experts model.

🧮 Nemotron 3 Ultra open release – NVIDIA opened a 550B parameter frontier model with 55B active, interleaving Mamba-2, mixture-of-experts, and attention layers plus multi-token prediction. It supports up to 1M token context and ships with open weights, training data, and recipes.

The hybrid architecture targets long-context analysis, agentic workflows, and high-stakes retrieval.
An NVFP4 pre-training recipe was used to cut compute cost during training.
Minimum hardware spans 8 GB200-class GPUs or 16 H100s, with smaller NVFP4 builds available.

🏗️ Microsoft launched seven Build models – Microsoft unveiled a family of seven models built from scratch on a clean data lineage, spanning reasoning, code, image, transcription, and voice, all designed to work together as a coordinated set. Model performance comparison table showing side-by-side results versus Sonnet 4.6 and Opus 4.6 (instead of the last versions).

🔮 Qwen 3.7 Plus multimodal release – Qwen published a multimodal version in its 3.7 line, positioned around broad intelligence across modalities. Detail beyond the announcement was limited at posting time.

MLOps

✂️ Headroom lossless token compression keeps growing – The open tool compressed tool by Tejas Chopra outputs, logs, files, and retrieval chunks before they reached the model, claiming 60 to 95 percent fewer tokens with the same answers, as a library, proxy, or MCP server. It became the top trending GitHub repo, jumping from 2,000 to over 12,000 stars in three days.

🧪 Local app benchmarks LLMs – A local-first desktop app ran practical benchmark packs against local or remote models, comparing them side by side on tool calls, bug fixing, extraction, structured output, and agent workflows. It shipped for macOS, Windows, and Linux.

Programming

🔍 Alibaba open-sourced code review – Alibaba released an AI code review CLI used internally at scale, combining deterministic pipelines with an LLM agent for precise line-level comments. A tuned ruleset flagged issues like null pointers, thread safety, and injection, and the agent could read full files and search the codebase.

🚀 Grok Build coding model on API – xAI opened a coding model tuned for agentic tasks like web development, debugging, and tool use through its API in public beta. It served at over 100 tokens per second, powered the Grok Build CLI, and worked across popular agent harnesses and gateways.

Robotic, World AI

🌐 Cosmos 3 unifies physical AI – NVIDIA released an open omni-model combining world generation, physical reasoning, and action generation in one architecture, replacing separate models for generation, understanding, and policy. It topped multiple open leaderboards across world reasoning, generation, and action.

A dual-tower mixture-of-transformers pairs an autoregressive reasoning stream with a diffusion generation stream.
Action was treated as a first-class modality alongside language, image, video, and audio.
The release shipped Super and Nano sizes, diffusers integration, post-training scripts, and synthetic datasets.

🚗 OmniDreams: real-time world model for driving – NVIDIA introduced a generative world model that autoregressively produced action-conditioned driving video in real time for closed-loop simulation. Trained on 21,000 hours of driving, it synthesized rare events like extreme weather and unpredictable agents that fixed simulators struggle to capture.

It conditioned each frame on past frames, the current simulator state, and immediate driving actions.
A policy post-trained from it beat a larger research policy while using a fifth of the parameters.
It ran inside a closed-loop stack with a separate policy model and orchestrator.

🎞️ FlashDreams: infra for world models – NVIDIA detailed a high-performance inference and serving library for interactive autoregressive video and world models. It packaged the persistent streaming loop into reusable pipelines and runners spanning driving simulation, robotics, games, and creative tools.

TTY Lunch

Each week, TTY Lunch brings together exceptional builders around the table. Today’s lineup included Amine Saboni (Pruna AI), Aram Adamyan, Bryen Param, David Martins Gonçalves (Notify), Isaac Partouche (Stage11), Hani Chalouati (Guepard), Matthieu Terris (Blur Labs), Olivier Desclaux (Harmattan AI), and Stan Girard (The Vibe Company).

Topics included, in no particular order: visual context windows in vision models, vehicle tracking accuracy using SLAM, the future of cheaper and smaller world models, AI pricing and energy consumption, the real cost-effectiveness of LLMs in applications, evaluating pruned and quantized models, the subjectivity of image quality evaluation, the challenges of model upgrades and consistency in production, LLM interpretability, predictability and replicability, knowledge transfer between models, the evolution from md files to deterministic application logic, the future of open-weight models, and the viability of open source software.

As always, thank you so much for joining and these wonderful conversations!

Contributors This Week

Kevin Kuipers, Félix Raimundo, Pierre Chapuis, Lior Oren, Khalil Ouardini, Jeremie Kalfon, Quentin Dubois, Tejas Chopra, Ashley van Heteren, Kemal Toprak Uçar, Dario Di Carlo, Harsimrat Singh Sandhawalia, Hugo Venturini, Isaac Partouche, Marie Thuret, Olivier de la Clergerie, Pierre Manceron, Amine Saboni, Guillaume de Luca, Ihab Bendidi, Julien Kilo, Julien Mangeard, Julien Seveno-Piltant, Koutheir Cherni KC, Louis Manhes, Nancy Wang, Raymond Rutjes

TTY-changelog #045

Kevin Kuipers — Sat, 30 May 2026 18:52:18 GMT

👉 Article originally posted on TTY

TTY Party

Short version: Please join us the 17th of June starting at 6pm in Paris to celebrate. Location TBD. Long version here.

Autonomous Agents

🗺️ PEEK caches long-context agent maps – PEEK built and cached a compact context map summarizing large, recurring corpora like codebases or document sets, so LLM agents could quickly locate relevant parts, avoid re-reading everything, and answer long-context queries more accurately and cheaply.

It beat strong baselines by 6.3 to 34 percent on long-context reasoning and aggregation tasks.
It used 93 to 145 fewer iterations and up to 5.8 times lower cost than the prior state of the art.
The gains held across different language models and agent architectures, including OpenAI Codex.

Biotech, Health, and Chemistry

🇨🇳 China competition divided biotech leaders – A widely shared Maraganore tweet argued US biotech leaders complaining about Chinese rivals should compete on merit instead of pushing protectionist policy. The thread weighed whose interests the anti-China framing actually serves. (article mentioned in the tweet)

Community take: Felix Raimundo saw the anti-China push as US biotechs that were losing trying to protect themselves, wrapping “save high-paying jobs” and “security risk” in patient-safety language. Americans, he said, love “merit and free-market competition until they face serious competition and lose.” Maraganore, the former Alnylam CEO who pioneered siRNA drugs, was essentially saying: innovative drugs do not fear copycats. The real fight, Felix added, was not about safety, since Chinese drugs still go through FDA approval and US trials. It was about China potentially blocking innovative drugs during trade disputes, and the US biotech sector hollowing out the way electronics did when manufacturing moved to Shenzhen. Ihab Bendidi agreed many US biotechs were complaining for the wrong reasons but felt Chinese regulatory shortcuts were genuinely dangerous. He preferred labeling drugs by risk and letting the market decide. He also noted US tech plays the same game, asking for rules on competitors “because only they can be trusted with AI.” Felix closed by pointing to an old industry trick: piling on compliance costs that giants can absorb but that block startups from entering at all.

🔬 Reading foundation models in biology – Great take from Julien Duquesne. Sparse autoencoders trained on protein embeddings recovered thousands of interpretable features per layer, far beyond what raw neurons exposed. Applied to single-cell models, the method validated concepts causally through counterfactual gene perturbations, not correlation.

🧬 ESMFold2 advanced protein design – An open scientific engine claimed state-of-the-art results on protein interactions and antibodies, with validated miniprotein binders across five cancer and immunology targets. It shipped with an atlas of 6.8 billion proteins and 1.1 billion predicted structures.

A world model of protein biology emerged purely from language modeling over billions of sequences.
Mechanistic interpretability tools borrowed from LLM research exposed how the model represents biological concepts.
A simple gradient-based search over the model surfaced high-affinity binders.

🧪 AlphaFold’s out-of-distribution blind spots – A widely shared interview argued that structure prediction is not biological inference. The hardest cases sit out of distribution: quantum protein fluctuations, post-translational modifications, and conformational ensembles the training data never captured.

Community take: Khalil Ouardini flagged the podcast as articulating the importance of uncertainty quantification in out-of-domain settings for foundation models, with AlphaFold on proteins subject to quantum fluctuations as a sharp example.

Image, Video & 3D

🖼️ PiD bridges latent models – PiD turns compressed image data into high-res pictures faster by combining decoding and upscaling in one step, instead of doing them separately. It can make 512x512 images into 2048x2048 in under 1 second with good quality.

Cyber

🔧 Build narrow dependencies yourself – Redis’s creator endorsed building narrow-scope dependencies in-house, extending a fork-and-trim argument. The 2026 twist: an LLM can merge security patches and test dependencies for flaws upstream maintainers may ignore. Julien Danjou’s companion essay framed the real question as which dependencies to drop, not whether to build.

Community take: Amine Saboni kicked off the thread asking whether forking dependencies works for smaller teams that cannot afford to rebuild and maintain their own tooling. Pierre Chapuis agreed he always tried to minimize dependencies and pointed at Julien Danjou’s piece. Robert Hommes suggested artifact scanning platforms like Cloudsmith as a middle ground. Amine pushed back: the proposed mitigation was to remove supply chain links, not scan them, so where do you set the threshold? Robert conceded the tradeoff between functionality and security. Pierre later flagged a small HPSv2 PR as a clear case for a rewrite.

🛡️ Multi-agent system found real zero-days – A multi-agent LLM system built on Google’s OSS-Fuzz automated vulnerability discovery with reproducible verification. A control-flow abstraction localized bugs at the right granularity, while dual-layer fuzzing and analysis tools reasoned across functions.

It hit a 90 percent detection rate, finding 36 of 40 bugs at the AIxCC 2025 final.
In the wild it surfaced 29 zero-days across 12 open-source projects, all confirmed and fixed.
Every reported vulnerability was fuzzer-reproducible, cutting the false positives that plague LLM detectors.

Infrastructure

🛡️ Edgee added fallback models – Sacha’s Edgee launched its Fallback Models feature on Product Hunt. When Claude cannot run a request (outage, plan limit, or programmatic credit cap from June 15), the layer routes it automatically to Kimi K2.6, GLM, Qwen, Gemma, or to a Bedrock, Vertex, or Azure account. Works with copilot, codex, and opencode.

Community take: Kevin Kuipers asked how Edgee evaluated routing quality. Sacha Morard said they ran SWE benchmarks to verify no drift from the compressor and planned the same for fallback models, while admitting “human intuition” remains the most common indicator of effectiveness in practice.

⚡ Inference engine prioritized decode speed – An inference preview reached 3,000 output tokens per second per request on eight AMD MI300X GPUs and 2,100 on eight NVIDIA H200s, in FP16 with no speculative decoding. Its claim: agents run sequential loops, so decode speed per request beats aggregate throughput.

Language Models

🤖 Claude Opus 4.8 is out – Anthropic upgraded its flagship at the same price, with sharper agentic judgment and user-controllable effort on claude.ai. The headline change was reliability: the model was roughly four times less likely to let code flaws slip past unflagged.

On the Super-Agent benchmark it was the only model to finish every case end-to-end, allegedly matching GPT-5.5 on cost.
Claude Code gained a dynamic workflows mode aimed at very large-scale problems.
Fast mode claims runnning at 2.5 times speed for three times less cost than prior models.

📱 Quantized diffusion runs on phones – Heavily compressed image models brought diffusion to laptops and phones. The 1-bit variant shrank to 0.93GB and the ternary variant to 1.21GB, both staying competitive on composition and prompt fidelity. A companion iOS app ran generation fully on-device.

MLOps

💡 Fine-tuning versus long-context personalization – A revived community thread argued adding experts (MLPs) or LoRA adapters is more effective for LLM personalization than feeding 1M tokens per prompt, especially as models get larger. At $0.10 per million tokens, 2000 requests a week becomes a $5 vs $200 split for the same outcome.

Community take: Glenn Sonna endorsed the argument. Kevin Kuipers pushed back that fine-tuning investment can become obsolete when the model is replaced months later (for objective reasons, or just perception), unless the use case is static enough for a LoRA. Jeremie Kalfon countered that internal representations may not change much between Opus 4.6 and 4.8, so the same fine-tuning neurons could potentially be reused, with distillation or per-layer optimization keeping personalization viable across model upgrades. He added that in-context learning is short-term memory and fine-tuning is long-term, and that he was tired of agents that only possess short-term memory. A $195 per week per user spread, he noted, is not limited value.

Programming

🕸️ Graph indexing tools for codebases – Two graph-based codebase indexing tools, codegraph and GitNexus, drew attention for high star counts. Both expose pre-indexed knowledge graphs to coding agents like Claude Code, Codex, Cursor, and Gemini, claiming fewer tokens and tool calls than text-based retrieval.

Community take: Antoine Sueur asked whether anyone in the community had real success with these tools, flagging that high star counts had not translated into evidence of positive impact. Youssef Tharwat (currently building Noodlbox - codebase context, built from code) confirmed graph-based indexing does work and said codegraph looked good without having tried either, but noted the analyses on both projects were still shallow. Kevin Kuipers also flagged that Leonard Scheidemantel at Sequa is building something similar.

🌐 Built-in AI shipped in Chrome – A Google I/O writeup walked through Chrome’s on-device AI APIs, which skip cloud cost and latency while keeping data local. The Summarizer API generated headlines and meta descriptions, and the Prompt API used JSON Schema for tagging and comment moderation.

Community take: Raymond Rutjes saw the on-device APIs as a way to augment features without paying the cost-tax of cloud inference.

Robotic, World AI

🦾 Open robotics simulation rebuilt fast – An open-source simulation platform rebuilt its stack for physical AI with a penetration-free contact solver, unified rigid and deformable physics, and a path-traced renderer. A GPU compiler forked from Taichi gave 10x faster launch and up to 4.6x runtime over the first release.

🤖 Humanoid robotics dismissed as hype – Arnaud Thiercelin reported that across recent conversations with peers and tier-1 SV VCs, the consistent characterization of humanoid robotics was “useless hype” with no real-world impact, though interest in specialized robotics solutions had risen as a side effect.

Community take: Arnaud argued generic robots will always lose to specialized ones. Humans, he said, are the apex multi-role machine and trying to replicate us is futile, while specialized machines are where the real potential sits. Kevin Kuipers partly agreed but pushed back that humanoids carry the advantage of fitting tools and environments designed for humans (stairs, appliances, vehicles), making the path harder but the reward 100x bigger.

TTY Lunch

Each week, TTY Lunch brings together exceptional builders around the table. Today’s lineup included some of the best talent in deep tech right now with Aram Adamyan, Ihab Bendidi (Recursion), Julien Duquesne (Scienta), Khalil Ouardini (ex-OWKIN), Victor Schmidt (Entalpic), Victoria Latynina (Sanofi), and of course Willy Braun to drive the insightful discussion.

New members

🇫🇷 Clément Castellon – Co-founder @ Candide · Research lab building infrastructure to make the AI research cycle measurably faster than what currently exists. Research Engineer at Sorbonne, ENS Ulm, and ERC ModERN by day, building Candide with co-founder Arindam Biswas in parallel. ICML 2026 workshop paper on Ramanujan-graph sparse initialization shipping now. Background across physics, math, CS, and digital humanities. Chess, Dune Imperium, and an unending search for music that fuses orchestral and electronic without sounding like a film trailer. Special power: peaked top 90 worldwide on GeoGuessr and does OSINT on the side, give him a photo and he can usually tell you where it was taken. 📍 Paris, France

Contributors This Week

Félix Raimundo, Arnaud Thiercelin, Jeremie Kalfon, Ihab Bendidi, Pierre Chapuis, Quentin Dubois, Amine Saboni, Robert Hommes, Antoine Sueur, Glenn Sonna, Louis Manhes, Sacha Morard, Tejas Chopra, Victoria Latynina, Youssef Tharwat, Benoit Kohler, Clement Castellon, Julien Seveno-Piltant, Khalil Ouardini, Louis Choquel, Maziyar Panahi, Raymond Rutjes

TTY-changelog #044

Kevin Kuipers — Sat, 23 May 2026 10:57:51 GMT

👉 Article originally posted on TTY

Autonomous Agents

🧪 Vercel’s Zero, a language for agents – The experimental language put AI agents first, aiming for a small, regular surface a model can learn while working, a standard-library-first design, and tooling that exposes diagnostics and repair plans as structured data. It is pre-1 and not meant for production.

🔌 Codex plugin for AI methods – A new extension by Louis Choquel from Pipeless that lets the Codex desktop app author multi-step AI workflows in a dedicated language, then validate and repair its own output through hooks and CLIs, with build artifacts rendered live as it worked.

Biotech, Health, and Chemistry

🧬 Carbon, an open DNA model – It claimed to run 275x faster than the prior best, processing a full human genome on a single GPU in under two days, using a tokenizer that splits sequences into six-base chunks while keeping single-base resolution. An interactive demo covered genes, mutations and protein folding, but the release drew sharp pushback over its reconstructed tree of life and how little biology it seemed to grasp.

Community take: Félix Raimundo dismissed the model as AI for bio with zero domain knowledge, noting it treats flies as plants where PCA does not and fails to separate introns from exons that naive baselines clear at 95 percent. Jeremie Kalfon pushed back that the authors were LLM people new to biology and that getting this far without much guidance was already impressive, while Félix held that being good at the task is the bare minimum, not grounds for applause.

🧫 Open tool for variant effects – A genetics lab open-sourced an MCP server and a Claude skill on top of its model for predicting which genetic variants cause disease, letting anyone query predictions or analyze their own sequencing data. The team stressed it is a predictor, not a diagnostic tool.

Community take w/ Jeremie Kalfon: “From what I understand, it really only works for coding variants, which are the super easy ones.”

📄 Nature papers on AI scientists – Three papers on AI scientists for biology landed in Nature on the same day, one from Future House and two from DeepMind, pointing to a fast-growing push toward agents that can form and test biological hypotheses.

Image, Video & 3D

✂️ Runway Aleph 2.0 and Edit Studio – Runway upgraded its flagship video editing model and wrapped it in a new product, handling up to 30 seconds of 1080p video, localized edits that touch only the targeted region, image-level control where an edited frame sets the look, and changes applied across multiple shots at once.

Infrastructure

⏱ The 62.5-minute rule for caching – A deep dive on Claude’s prompt caching landed on a clean rule: if you expect to reuse a cached prefix within that window, keep it warm with a cheap read, otherwise let it expire and rewrite later. The threshold holds across models and prefix sizes even as the dollars change.

🧱 Local-first second brain for agents – An open-source tool from Stan Girard keeps notes as plain markdown behind a single index, exposed through a CLI and an MCP server so agents like Claude Code or Cursor could read and write to a local personal knowledge graph. Released under MIT.

Language Models

⚡ Gemini 3.5 Flash at Google I/O – Google shipped it to everyone, claiming gains on almost every benchmark over 3.1 Pro with a big jump in coding, quality near the best models, and roughly 4x faster token output. The keynote also covered a 24/7 personal agent, a new terminal and an Omni media model.

Community take w/ Kemal Toprak Uçar: “I rarely praise proprietary models, but Flash 3.5 has been genuinely fascinating for coding and brainstorming, with both the speed and the performance standing out.”

🟢 Cohere Command A+ open weights – Cohere released its most capable model yet under Apache 2.0, with 218B total and 25B active parameters tuned to run on as little as two H100s. It claimed over 2x higher output speed and 30 percent lower latency than prior Command A, plus native support for 48 languages.

🔀 Nemotron blends diffusion and autoregression – NVIDIA released a language model family that runs both autoregressive and diffusion-style parallel decoding by switching one model’s attention pattern, enabling self-speculation where diffusion drafts tokens and autoregression verifies them with a shared cache, claiming multi-x speedups at equal accuracy.

Community take w/ Amine Saboni: “The self-speculation setup, using autoregressive verification on the diffusion drafts, is the really interesting part.”

🧠 A 1B hierarchical reasoning model – Trained from scratch on structured public data, it drew attention as very competitive for its size. Two recurrent Transformer modules, one slow and one fast, loop over the same embeddings to give deep effective compute at a fixed parameter count. It ships as a pre-alignment checkpoint, not a chat model.

🐉 Qwen3.7 launches as agent frontier – Qwen published the release with a long write-up framing the family around agentic capability across reasoning, coding and tool use. On the Arena leaderboards a preview then ranked 13th overall in text and 16th in vision, lifting the lab to 6th in text and 5th in vision, with stronger showings in math, expert prompts and coding.

Programming

🖥 Cursor Composer 2.5 ships – An upgrade over Composer 2 that sustains long tasks better, follows complex instructions more reliably and feels nicer to work with. It is built on the same open Kimi K2.5 checkpoint with extra reinforcement learning, while a much larger from-scratch model trains on SpaceXAI compute.

🚀 Gemini CLI becomes Antigravity CLI – Google folded its terminal tool into a new Go-based CLI built for multi-agent work that orchestrates agents asynchronously in the background and shares the harness of the Antigravity desktop app. The old Gemini CLI and Code Assist IDE extensions stop serving consumer requests on June 18, 2026.

Community take w/ Quentin Dubois: “It is fast with a nice feel, but the knowledge cutoff trips it up on newer APIs, the harness feels weaker than Codex or Claude Code, and it does not import your existing settings.”

🔧 Swapping Claude Code for local stacks? – Members are experimenting setups for replacing Claude Code or Codex with lighter or local options, from pairing Qwen or Nemotron with QwenCode for faster prompt processing, to running GLM through a custom harness, to home-grown Python CLIs. Most agreed local is increasingly viable yet still leaned on premium tokens.

Community take w/ Gabriel Olympie: “Opencode gets good feedback but takes forever to configure with too many plugins, it feels like modding Skyrim.”

Robotic, World AI

🌍 SANA-WM, minute-scale world model – An open 2.6B model turned a single image and a camera path into a full minute of controllable 720p video on one GPU, using hybrid linear attention to hold a coherent scene without the memory blowup that crashes standard attention.

🕹 Agora-1, a multi-agent world model – Odyssey introduced a world model where several participants, human or AI, share and act inside the same real-time simulation, shown off as a playable multiplayer deathmatch, and pitched shared-world simulation as a path toward gaming, education and robotics.

🔊 Starchild-1, a world you hear – Odyssey also showed a model billed as the first real-time multimodal world whose generated simulations you can hear as well as see, with joint audio and video, positioned as an early step toward a general-purpose world simulator.

Community take w/ Pierre Chapuis: “It looks close to Veo for joint audio and video, but the key difference is that it keeps responding to your input while it generates.”

🌐 HY World 2.0 goes open source – Tencent released full inference code and all weights for building interactive 3D worlds, pitching it as generating engine-ready environments you can actually use rather than just video clips. Observers were surprised the models were so small.

🎥 Runway bets on world models – The NY-based AI video startup is betting video-based world models (not language) will drive next‑gen AI, expanding from filmmaker tools into physics‑aware models for robotics, science, and more while competing with giants like Google and OpenAI.

🚗 Kesai Labs, open self-driving stack – Kyutai and the ELLIS Institute launched a Franco-German non-profit building a fully open and efficient self-driving stack, with plans to extend into robotics, manufacturing and healthcare. Weights, data and code are all meant to be open.

🎞 Fine-tuning Cosmos for robot video – A guide walked through parameter-efficient fine-tuning of NVIDIA’s Cosmos Predict 2.5 world model with LoRA and DoRA adapters, so teams can generate synthetic robot manipulation videos without the cost of full fine-tuning, then feed those trajectories into downstream robot learning.

TTY Lunch

Each week, TTY Lunch brings together exceptional builders around the table. Today’s lineup included Alexandre Pereira (2501.ai), Alvaro Lamarche Toloza (Mago), Charles Bochet (Twenty), Gregory scafarto, ‎ Hugo‎ Venturini (SkipLabs), Justin Halsall (rrweb), Sacha Morard (Edgee), and Samy Ouyahia (Mistral AI).

Some topics of the day were:

Evaluating and trusting coding agents, and what drives engineers to switch between them
Diversity and recruiting women in tech, the severe pipeline difficulty (even for TTY, obviously), and why it is critical to prioritize it early on
The real operational impact of a “second brain” and automated agents on company workflows
The inevitable death of the “per-seat” SaaS pricing model in the era of autonomous agents

New members

🇫🇷 Khalil Ouardini – Senior Research Scientist @ ex-Owkin · Owkin builds AI for drug discovery and pharma R&D, from target discovery to automating research. Spent four years in AI for biology, first researching at UC Berkeley then building products at Owkin, and recently left to explore startups automating complex pharma workflows with agentic systems, now looking for co-founders. Into boxing, jiu-jitsu, music production and reading. Special power: dissociative daydreaming. 📍 Paris, France

Contributors This Week

Félix Raimundo, Julien Seveno-Piltant, Quentin Dubois, Pierre Chapuis, Nancy Wang, Gabriel Olympie, Benjamin Trom, Robert Hommes, Louis Choquel, Raymond Rutjes, Maziyar Panahi, Enrico Piovano, Nikolay Tchakarov, Ivan Yamshchikov, Jeremie Kalfon, Kemal Toprak Uçar, Khalil Ouardini, Stan Girard, Arnaud Thiercelin, Hugo Serrat, Amine Saboni, Bryen Param, Charles Sonigo, Charly Poly, Ihab Bendidi, Lior Oren, Sacha Morard, Stéphane Collot, Tejas Chopra

TTY-changelog #043

Kevin Kuipers — Sat, 16 May 2026 10:25:02 GMT

👉 Article originally posted on TTY

Community Discussion

🧹 Cleaning up agentic junk – Lior Oren opened a week-long thread on what happens when a multi-repo team runs spec-to-merge agents with almost no human touch and the system starts collecting residue: dead code, junk specs, irrelevant quality metrics, agents.md waste, dead skills, and stale gotchas nobody curates once the original problem gets fixed. The boyscout rule applied to all agents helped a bit and made PRs dirtier. The sharper framing came later in the thread: “Cleaning itself works well. The when, the scope, and the how is where I’m trying to figure out a common pattern.“

Approaches covered the full range.

Lucas DiCiocci skips auto-merge entirely, freezes code when junk accumulates, and runs intentional cleanup before the merge-conflict hell sets in.
Pierre Chapuis follows a similar manual cadence with cleanup prompts triggered by hand, never cron’d.
Hugo Venturini hunts deadcode and antipatterns with a single sweep agent.
Anicet Nougaret asks the agent to produce dataflow diagrams of the messy parts, which catches 99% of the rot before it ships.
Nikolay Tchakarov’s team enforces quality checklists baked into the agent loop: PRs under 7 files, reuse before adding, dedup checks, with humans still approving as the accountability anchor.
Jocelyn Fournier doubled down on guardrails: mandatory tests with minimum mocks, sharp acceptance criteria in the RFC for the agent to follow, and reviews from other agents not polluted with the implementation context.
Amine Saboni argued for Domain Driven Design as the harness: when the domain model lives clearly in specs, documentation, and the code, agents have less semantic margin to drift, and a /simplify pass or a delegated reviewer does the rest.
Jérémie Bordier sketched the maximalist version, a “dream mode” autonomous maintainer that pulls back the global picture of the project, removes dead code, updates dependencies, and rewrites the .md files.
Youssef Tharwat offered the opposite bet with Noodlbox: rather than another agent, a local static analysis engine running under a file watcher produces a constantly fresh list of action items the coding agent consumes over MCP or CLI, scoped per PR or globally, with an in-flight extension to flag stale .md files whose linked code no longer exists or has semantically drifted.

The skeptics pushed back on the premise. Gawen Arab: “Isn’t junk the consequence of the high variance of LLMs? Having a high-variance LLM fix the junk of another high-variance LLM feels naively like a dead end.“ For now he keeps a human in the loop and skips the auto-loop entirely.

The hardest unsolved problem turned out not to be the code at all. The real junkyard, as Lior Oren pointed out later, is the agent memory and gotcha files: a session produces a note like “db migration files stale at x execution, wait more time than usual,” the issue gets fixed somewhere downstream, and nobody cleans the shared dev memories. Amine Saboni admitted he doesn’t commit those files, treating them as personal hygiene rather than shared standard, but at team scale that breaks: notes are pointless unless shared, and once shared they go stale unless someone owns them. Lior Oren eventually landed on a pattern worth trying: “Trigger cleaning agents with each PR but have them create a stacked PR. The main PR is not polluted, the scope of the cleaners is limited to the files touched in the main PR, and it saves the scheduling and cron infra.“ He has since wired it into GH CI as skills.

Audio

🗣️ Interaction models for real-time AI – Thinking Machines released a research preview of interaction models: a from-scratch architecture designed to continuously ingest audio, video, and text and to think, respond, and act in real time. The argument: today’s turn-based interfaces collapse the model’s perception into a single thread that freezes during generation, creating a bandwidth bottleneck for collaboration. The fix was a multi-stream, micro-turn design that let both model and user interject, see, and respond without waiting for the other to finish.

The multi-stream micro-turn loop replaced the wait-then-respond cycle, so the model kept perceiving and the user kept signaling while either side was mid-output.
The model was trained from scratch rather than retrofitted onto an autonomous-first base, on the bet that interactivity had to scale alongside intelligence rather than be added as a UX layer.
Reported state-of-the-art combined performance on intelligence and responsiveness, positioning collaborative real-time interaction as a primary capability rather than scaffolding.

Autonomous Agents

🧪 Petri auditing agent open-sourced – Anthropic open-sourced Petri (Parallel Exploration Tool for Risky Interactions), a framework that deployed an auditor agent to probe target AI systems through multi-turn conversations with simulated users and tools, then scored each transcript across multiple safety dimensions to flag situational awareness, whistleblowing, self-preservation, reward hacking, and similar concerning behaviors.

🧮 AI co-mathematician from Google – DeepMind introduced a stateful workbench where mathematicians collaborated with AI agents on ideation, literature search, computational exploration, and theorem proving. Scored 48% on FrontierMath Tier 4.

Community take w/ Enrico Piovano: “I recently tried LLMs on math problems from my applied math PhD. None could give correct proofs, even for smaller lemmas. Even powerful LLMs still struggle with less common problems and don’t generalize well in RL.”

🦘 LLMs can’t make abductive leaps – Position paper used general relativity as a case study and argued LLMs could execute deduction but lacked the abductive jump needed to formulate novel premises, proposing multimodal world models as the bridge.

Biotech, Health, and Chemistry

💊 Isomorphic skips the wet lab – Isomorphic Labs confirmed it would not be building any in-house wet lab or chemistry capability and would instead outsource all experimental work to contract research organizations (CROs), the third-party labs pharma companies hire to run experiments on commission. First clinical trials were targeted for end of 2026, with Demis Hassabis confirmed to stay as CEO through the push. The pipeline ran on Isomorphic’s drug design engine IsoDDE, which generated candidates that external CROs then validated experimentally.

Community take w/ Félix Raimundo: “No wet lab is insane to me. They clearly have too much cash. They will use CROs and pay their margins. I assume they think experiments are like compute and they view CROs as hyperscalers.“

🧬 Cellular Intelligence acquires STEM-PD – Cellular Intelligence acquired STEM-PD, an FDA-cleared Phase 2-ready allogeneic cell therapy for Parkinson’s, from Novo Nordisk, alongside a strategic investment from Novo Nordisk in the company.

Image, Video & 3D

🎨 Qwen Image 2.0 unifies generation and editing – Omni-capable image model that paired Qwen3-VL as condition encoder with a Multimodal Diffusion Transformer. Handled ultra-long text rendering up to 1K tokens for slides and posters, multilingual typography, and photorealism in one framework.

🌐 Image to world tool for Claude – A community tool turned any image into a 3D environment with individual meshes, physics, and ambient sound layers, exporting to Unreal or Blender. Built on WorldLabs and FAL under the hood.

Cyber

🚨 TanStack hit by massive supply chain attack – An attacker forked TanStack’s open-source React router project under a disguised name, opened a pull request that TanStack’s own automated build system ran, and used that single PR to steal a secret token from the build runner. With that token they published 84 malicious package versions through TanStack’s legitimate, cryptographically signed release pipeline, so the poisoned packages arrived with valid provenance. Every developer who installed one had their credentials stolen and silently used to republish more poisoned packages they maintained, turning the campaign into a self-spreading worm across npm and PyPI. Within hours it had reached companies like Mistral AI, OpenAI, UiPath, OpenSearch, and Guardrails AI.

🔓 Mythos breaks Apple M5 in days – Researchers at Calif.io, working with Anthropic’s Mythos Preview, built the first public macOS kernel memory corruption exploit on Apple M5 silicon in five days, against an MIE (Memory Integrity Enforcement) defense that Apple had spent five years and likely billions building into hardware. MIE was Apple’s ARM MTE-based memory safety system, designed to disrupt every public exploit chain against modern iOS including the recently leaked Coruna and Darksword kits.

Mythos Preview was Anthropic’s restricted-access security-focused model, released in April 2026 to a small group of Project Glasswing partners to find and patch vulnerabilities before equivalent capabilities became broadly available.
The macOS attack path was an accidental discovery starting April 25, with the team reporting it in person at Apple Park rather than through standard submission channels.

Language Models

🌳 PageIndex for vectorless RAG – Reasoning-based RAG system that built a hierarchical tree index from long documents and used LLMs to navigate via tree search rather than vector similarity, aimed at professional documents where similarity fell short of true relevance.

📐 Geometry of consolidation in RAG – NeurIPS paper proved any compression of embedded items into fewer representatives carried an irreducible identity-retrieval error tied to local effective dimension. RAG compression schemes therefore had a minimum wrongness no budget could escape.

MLOps

🎙️ The end of finetuning – Latent Space podcast with Jeremy Howard, who shifted from finetuning advocate to arguing finetuning was becoming obsolete as models gained context and retrieval capabilities. Remaining reasons to finetune: compliance, tone, cost, format, and latency.

📈 Local Mac models got 5x smarter – Chart compared smartest open-weight models runnable on a 128GB MacBook Pro on the Artificial Analysis Intelligence Index v4.0: from Llama 3 70B at score 10 in May 2024 to DeepSeek V4 Flash at 47 in May 2026, faster than Moore’s Law.

🍎 Local LLM picker for Apple Silicon – Tool that ranked which open models fit and ran well on specific Apple Silicon configurations, factoring unified memory, bandwidth, tokens per second, and break-even versus cloud API costs.

💻 Running local models on M4 24GB – Practical writeup of getting Qwen 3.5-9B running at around 40 tokens per second on a MacBook Pro M4 24GB with LM Studio, including thinking-mode prompts and tool use, plus failed attempts with Qwen 3.6 Q3 and Devstral Small 24B.

Programming

📊 Coding agents benchmark released – Artificial Analysis launched a composite index for coding agents averaging pass@1 across SWE-Bench-Pro-Hard, Terminal-Bench v2, and SWE-Atlas-QnA, while also reporting cost, token usage, and execution time across agent and model combinations.

💸 Claude Agent SDK pricing change – Starting June 15 2026, Agent SDK and claude -p usage will no longer count toward Claude plan limits and will require a separate monthly credit by tier: $20 for Pro, $100 for Max 5x, $200 for Max 20x.

Community take w/ Stan Girard: “Happy I switched to codex and got used to it two months ago. Really not happy with Claude in the last months.“

🤖 Chrome extension to vibe code websites – Stéphane Collot shipped a browser extension that let users describe UX improvements in plain English, with an AI agent inspecting the live page, writing JS plus CSS, and injecting it instantly. Features persisted across sessions via per-feature in-browser git repos.

Robotic, World AI

🚀 Former Qwen lead launches new lab – Lin Junyang, former technical leader of Alibaba’s Qwen, started a company valued around $2B targeting world models and embodied brains, with team members from ByteDance, Tencent, and overseas. Sequoia China and Gao Rong were among funds contacted.

🛏️ Figure robots tidy a bedroom – Two Helix-02 humanoids reset a bedroom in under two minutes, opening doors, hanging clothes, closing a book, taking out trash, and making a bed together. They ran a single learned vision-language-action policy with no shared planner. A separate livestream showed an 8-hour autonomous shift.

🚁 Aerial robot for vibration dampers – Lightweight aerial platform that landed on power transmission lines and replaced Stockbridge vibration dampers using motorized wheel arms, an impact wrench, and a damper-lifter stage, keeping human operators safely on the ground rather than working on live lines.

TTY Lunch

Each week, TTY Lunch brings together exceptional builders around the table. Today’s lineup included Clément Castellon, Gabriel Olympie (2501.ai), Jules Belveze, Julien Millet, Mathieu Kassovitz, Quentin Dubois (OSS Ventures), Stéphane Béreux (Jimini AI), and Yvann Barbot (TerraLab).

We used this almost-holiday in France 🇫🇷 to reinvent the world all at once, while trying to satisfy Mathieu’s insatiable need to understand. Topics in no particular order included evals, second-brain management, the impact and ethics of AI, the geopolitical energy crisis, data sovereignty, and even a bit of video games.

Contributors This Week

Lior Oren, Félix Raimundo (Tychobio), Jérémie Bordier (XHR), Youssef Tharwat (Noodlbox), Enrico Piovano (Goji), Stan Girard (Quivr), Amine Saboni (Pruna.ai), Pierre Chapuis (Finegrain), Gabriel Olympie (2501.ai), Julien Seveno-Piltant, Maziyar Panahi (OpenMed), Arnaud Thiercelin, Benjamin Trom, Hugo Venturini (SkipLabs), Ihab Bendidi (Recursion), Jeremie Kalfon (Pasteur), Raymond Rutjes, Stéphane Collot (Sequense), Anicet Nougaret (Ascii), Gawen Arab (Airbuds), Ivan Yamshchikov (Pleias), Jocelyn Fournier (Softizy), Karim Matrah (Contrast), Koutheir Cherni (Guepard), Lucas DiCiocci, Nikolay Tchakarov (Asteria), Quentin Dubois (OSS Ventures)

TTY-changelog #042

Kevin Kuipers — Sat, 09 May 2026 12:56:51 GMT

👉 Originally posted on TTY

Community Discussion

🧠 How we build our second brains – A week-long thread initiated by Paul Masurel mapped the spectrum of personal knowledge management in the community, from bare-minimum setups to fully custom cognitive infrastructure. Julien Millet’s sharpest observation set the tone: before picking any tool, define what you actually want the system to do. Storing, navigating, retrieving, generating, and tying to ongoing projects each lead to radically different architectures. “If you only design for storage, you’ll end up with a graveyard of notes you never reread.”

Approaches covered the full range. Karim Matrah keeps it minimal with Google NotebookLM, querying it from Claude Desktop via MCP. Benoit Kohler’s team moved away from individual Obsidian setups after everyone kept duplicating work in parallel, building shared custom skills on top of Karpathy’s paper instead. Jocelyn Fournier solved the scoping problem by creating domains, attaching memories to specific topics rather than one flat pile. Marie Thuret went furthest: a Karpathy-style LLM wiki extended with a cognitive profile covering how she thinks, learns, and decides, her beliefs, her tensions, her working assumptions, plus a map of the people around her. The system doesn’t just store what happened, it knows who was involved and what that means. A generic assistant treats all inputs the same. Marie’s doesn’t. “When I start Claude Code without the plugin on, I feel it immediately. But it’s no magic. Lots of maintenance and curation to keep the quality high, and you have to fight default LLM behaviors along the way.”

The hardest unsolved problem: connecting heterogeneous sources. Paul Masurel described constantly searching Slack, Gmail, and Google Docs for information he already has somewhere, and suspects it is now possible to build something that addresses all of it in one system. Julien Millet put it more bluntly: juggling four platforms for a weekly review is “a clusterf*ck,” and whoever builds the universal cross-source connector “might have gold in their hands.”

Events

🇳🇱 Startup World Cup in Amsterdam (June 10) – Co-hosted by Tech Makers and Prosus at the AI House. Benelux founders, investors, and operators compete for a spot at the San Francisco final and a $1M investment prize.

Audio

🎙 GPT-Realtime-2 voice suite – OpenAI launched a suite of real-time voice models that bring GPT-5-class reasoning to production APIs for the first time, shifting voice interfaces from simple turn-taking toward agents that can listen, reason, translate, and act as conversations unfold.

GPT-Realtime-2 supports tool use, interruption handling, and multi-turn context, moving voice agents from call-and-response toward agents that reason through requests and complete tasks mid-conversation.
GPT-Realtime-Translate streams live translation across 70+ input languages into 13 output languages, keeping pace with the speaker without post-processing delay.
GPT-Realtime-Whisper transcribes live as the speaker talks rather than after they finish, enabling downstream processing to begin before a turn ends.

Autonomous Agents

🤖 HubSpot’s open agent platform – HubSpot positioned itself as a CRM “open by design” for the agent era, arguing the real AI race is about context rather than models. The vision: any agent can plug into HubSpot’s data and intelligence as a building block, while agents can also operate the platform end-to-end autonomously.

Biotech, Health, and Chemistry

🩺 LLM outperforms physicians across settings – Published in Science, a study tested OpenAI o1 against hundreds of physicians on clinical cases from published vignettes to live ER patients. The model outperformed human baselines in both diagnosis and management planning across all five experiments, motivating calls for prospective trials

👨‍⚕️ DeepMind’s AI co-clinician initiative – Published the same week as the Science paper above, DeepMind launched a research initiative into triadic care, where AI agents assist patients under physician authority. Evaluation used standardized patients, the same method as medical school exams, where the system outperformed leading evidence synthesis tools in blind comparisons.

Community take w/ Félix Raimundo: “Standardized patients are exactly how medical students are trained and evaluated, and far more representative of actual care than an EHR (electronic health records) that already contains all the information.”

🧬 MULTI-evolve’s epistasis claim challenged – A University of Washington analysis found that Arc Institute’s MULTI-evolve, published in Science as learning epistatic interactions between mutations, actually learned a classical additive model. The original paper included no comparison against simple additive baselines, a standard check for this type of claim.

💊 Can AI beat Eroom’s Law – Drug development costs have doubled roughly every 9 years since 1950 despite AI advances, a trend called Eroom’s Law. An investor essay argues cheaper AI-assisted trials could let small biotechs run Phase 3 without early exits to pharma, though managing a trial requires deep hospital and KOL networks that most biotechs lack.

Community take w/ Félix Raimundo: “Right now, biotechs get their liquidity from either M&A or out-licensing because Phase 3 trials are too capital intensive, and they need to hand it off to pharma. His point is that if trials become cheap enough, biotechs may be able to run these themselves without being forced into early exits. I have some light reservations though: the issue is not just cost, managing a trial requires a very deep network of hospitals and KOLs which biotechs may not have, and being efficient at running these also saves patent lifetime. But the point is still very interesting.”

🔬 Biology’s fuzzy API problem – Unlike software, drug discovery has no clean typed interfaces. Target discovery produces a probabilistic hypothesis, drug design produces a candidate burdened with unresolved questions, and clinical trial results are filtered through patient selection, endpoints, and regulatory interpretation. Every handoff leaks assumptions.

💰 Isomorphic Labs nears $2B raise – Isomorphic Labs, the AlphaFold spinout built to turn protein structure prediction into approved drugs, is raising over $2B after pushing its first clinical trials from 2025 to end of 2026. The raise funds the gap between AI drug design capability and clinical validation, with Eli Lilly, Novartis, and Johnson & Johnson already as partners.

Image, Video & 3D

🎬 VES on-set VFX data guide – The Visual Effects Society Technology Committee published an interactive guide on on-set data types, capture workflows, and stakeholder responsibilities across the production pipeline, aimed at standardizing communication between VFX, production, and technology teams.

Data, Infrastructure

📊 Lightweight adapter for tabular models – Co-written by Duong Nguyen from the community, TFM-Retouche is an architecture-agnostic input-space residual adapter for tabular foundation models that improves zero-shot performance without modifying backbone weights or requiring architecture-specific PEFT methods. It achieved this ranking under 20x less hyperparameter tuning budget than every competing tuned method on the benchmark.

⚙️ Inference engines as hardware strategy – A breakdown of major LLM inference engines framed around hardware strategy: llama.cpp for portability, MLX for Apple Silicon, vLLM for production serving, SGLang for large-scale disaggregation, and TensorRT-LLM for maximum NVIDIA throughput. The engine follows from the hardware decision, not the other way around.

Language Models

🧮 AlphaEvolve scales beyond math – DeepMind detailed AlphaEvolve’s real-world applications beyond mathematics: a 30% reduction in DNA sequencing errors via DeepConsensus optimization, grid improvements raising AC power flow feasibility from 14% to over 88%, and a 5% accuracy boost in natural disaster prediction across 20 categories.

💡 SubQ claims subquadratic frontier model – Subquadratic introduced SubQ 1M-Preview, claiming it is the first LLM on a fully subquadratic architecture with a 12M-token context window and roughly 1,000x lower attention compute than standard transformers. Community skepticism was high given the company’s thin public research footprint.

Programming

🔧 Treat agent output like compiler output – Great piece by Hugo Venturini from the community in response to Philip Su’s “No More Code Reviews: Lights-Out Codebases Ahead”, the piece reframes the real question: not whether to trust agent output, but whether we have built the upstream apparatus (type systems, formal specs) and downstream verification (tests, monitoring) that make reviewing unnecessary. We do not review compiled binaries. We run them against test suites.

Community take w/ Lior Oren: “Working without human review has delivered faster, higher-quality results when paired with spec-driven development, agent-led review loops, and E2E tests that check goals rather than steps.”

📐 Category theory for tiny ML – A working draft teaching ML fundamentals from first principles using Rust types, typed transformations, composition, and training loops. Category theory serves as an engineering tool rather than abstraction, making the structure of deep learning systems explicitly executable.

🦀 Bun explores Zig-to-Rust port – Bun is experimenting with rewriting its codebase from Zig to Rust. Notably, they wrote migration docs specifically for AI agents, not humans, telling AI to translate logic first without compiling, then fix compilation later. Still experimental, not decided.

New Member

🇫🇷 Raymond Rutjes – Twenty years in software, the last ten at Algolia (40 to 800 employees). Previously ran a web agency and built Pimwi, a SaaS that reached around 4,000 users. Now in exploration mode, looking for the right problem to commit to. Special power: My best ideas show up around kilometer 8. Resisting the urge to build my own indie game (otherwise known as the only legal way to force my self-produced music on strangers). 📍 Lyon, France

Contributors This Week

Félix Raimundo, Gabriel Olympie, Hugo Venturini, Julien Millet, Youssef Tharwat, Anicet Nougaret, Benoit Kohler, Guillaume Lesur, Jocelyn Fournier, Julien Seveno-Piltant, Marie Thuret, Duong Nguyen, Alvaro Lamarche Toloza, Karim Matrah, Paul Masurel, Robert Hommes, Jeremie Kalfon, Lior Oren, Quentin Dubois, Raymond Rutjes

TTY-changelog #041

Kevin Kuipers — Sat, 02 May 2026 09:00:54 GMT

👉 Article originally posted on TTY

Audio

🎙️ VibeVoice now on Hugging Face – Microsoft’s open-source 8B speech model gained a Transformers-compatible build on Hugging Face, enabling direct inference without custom setup. The model handles 60-minute audio in a single pass with speaker diarization and timestamps across 51 languages.

Autonomous Agents

🧠 Ineffable Intelligence bets on pure RL – David Silver, co-creator of AlphaGo, launched Ineffable Intelligence to build a superlearner driven entirely by reinforcement learning from environment experience, without relying on human-generated training data, aiming at what the company calls first contact with superintelligence.

⚡ Stripe Projects for agent provisioning – Stripe launched a developer preview letting developers and agents provision hosting, databases, auth, analytics, and other services from the CLI. Credentials sync back to the local environment automatically, and billing consolidates across all connected providers.

🗺️ What to learn in AI agents – A practitioner’s guide on durable vs transient knowledge in the current agent landscape, arguing that framework API mastery compounds poorly while systems thinking, debugging intuition, and taste for boring architectural choices remain the skills worth building.

Biotech, Health, and Chemistry

🔬 TxPert reaches Nature Biotechnology – A knowledge-graph approach to predicting transcriptomic perturbation effects achieved performance close to split-half experimental reproducibility for single unseen perturbations, and outperformed existing methods by 8 to 25% on double perturbations and cross-cell-line predictions. The first model in this class to move from preprint to peer-reviewed publication.

🌊 Virtual cell research at ICLR 2026 – Reflections from ICLR 2026 in Rio mapped the fragmentation in the virtual cell field: more models and investment than ever, but no shared evaluation infrastructure or common definitions. The piece argued for disease-specific models over universal cell simulators, with Eva from Scienta Lab cited as a positive example.

🧪 Orthrus RNA isoform foundation model – A Mamba-based mature RNA foundation model pretrained via contrastive learning on splicing isoforms and orthologous transcripts across 400+ mammalian species, published in Nature Methods. It outperforms genomic foundation models on mRNA property prediction with substantially less fine-tuning data. The first author co-founded blank.bio, applying the same approach to clinical precision medicine.

Image, Video & 3D

🔮 Power Foam: ray tracing meets rasterization – A differentiable 3D representation combining foam-based ray tracing efficiency with 3DGS-level rasterization performance. Uses bounded power diagrams with controllable cell extents to achieve consistent high-frame-rate output under both rendering paradigms from the same scene representation.

🎬 Vista4D video reshooting via 4D – A video reshooting framework from Eyeline Labs (powered by Netflix) that grounds input video and target cameras in a 4D point cloud, enabling novel-viewpoint synthesis while preserving scene dynamics. Supports camera trajectory changes, 4D scene recomposition, and dynamic scene expansion.

Cyber

🔴 CVE-2026-31431: critical Linux LPE – A 732-byte Python script roots every Linux distribution shipped since 2017 (!) by chaining a logic bug in authencesn through AF_ALG and splice() into a page-cache write. Requires only an unprivileged local user account, no race window, and no per-distro kernel offset.

The exploit works unmodified on Ubuntu, Amazon Linux, RHEL, and SUSE without any per-distribution tuning, and is expected to behave identically on Debian, Arch, Fedora, and other distros running affected kernels.
Kubernetes and multi-tenant environments carry the highest risk, as the shared page cache allows a pod with the right primitives to cross tenant boundaries and compromise the entire node.
Priority patching applies to multi-tenant Linux hosts, CI runners executing untrusted PR code, container clusters, and any environment running tenant-supplied notebooks, agents, or serverless functions.

Infrastructure

🏥 Scaleway wins France’s Health Data Hub – Following a rigorous 350-criteria selection process, Scaleway was designated to host France’s Health Data Hub including a copy of the national health database. The selection turned on HDS certification and a credible roadmap toward SecNumCloud qualification.

Community take w/ Robert Hommes: “The Dutch government and central bank also announced Stackit sovereign cloud pilots. And there’s Euro-Office, an EU-driven initiative for a sovereign, Microsoft-compatible office suite to replace US alternatives.”

💥 AI agent wipes production database – A coding agent running on Cursor + Opus 4.6 used a Railway API token created for domain operations to delete both a production database and all volume-level backups in 9 seconds. Railway stores volume backups within the volume itself, and tokens provisioned via the CLI carry undisclosed blanket authority over the entire GraphQL API, including destructive volume deletion.

Language Models

📉 DeepSeek slashes V4-Pro API prices – V4-Pro input token prices dropped 75% in a promotional offer running until May 5, with the model already undercutting GPT-5.5, Claude Opus 4.7, and Gemini 3.1 Pro at full price. Cache hit prices were simultaneously cut to one-tenth across the entire API suite, targeting agentic applications with repetitive inputs.

Programming

🎼 Symphony: open spec for Codex orchestration – OpenAI published Symphony, a written spec that turns issue trackers like Linear into always-on agent orchestrators, assigning one Codex agent per open task. Internal use yielded a 500% increase in landed pull requests on some teams.

🖥️ Zed editor hits version 1.0 – The Rust-based collaborative code editor shipped its first stable release.

🤖 Laguna XS.2 for agentic coding – Poolside’s first public model is a 33B parameter MoE with only 3B activated per token, designed for agentic coding and long-horizon tasks on local hardware. Scores 68.2% on SWE-bench Verified and runs on a Mac with 36GB RAM via Ollama under Apache 2.0.

💬 Karpathy on LLM new horizons – A fireside chat at Sequoia Ascent 2026 argued that LLMs enable three genuinely new capability categories beyond productivity gains: apps fully reducible to LLM inputs and outputs, plain-language install instructions replacing bash scripts, and unstructured knowledge bases previously impossible with classical code.

🧱 Better-T-Stack interactive stack builder – An interactive web UI for composing a TypeScript full-stack from modular picks across frontend, backend, runtime, database, auth, and native layers, generating a ready-to-run CLI command on the fly.

🎥 AI coding works, that’s the problem – A YouTube essay mapped the quantitative impact of AI on software development in web and game dev, then examined the economic and career consequences, arguing that accelerating output does not automatically preserve demand for individual developers.

Contributors This Week

Félix Raimundo, Ihab Bendidi, Aly Moursy, Robert Hommes, Youssef Tharwat, Quentin Dubois, Christophe Lesur, Louis Choquel, Timothy Lindblom, Amine Saboni, Antoine Millet, Gabriel Duciel, Gabriel Olympie, Hani Chalouati, Hugo Hernandez, Karim Matrah, Louis Manhes, Paul Masurel, Stan Girard

TTY-changelog #040

Kevin Kuipers — Sat, 25 Apr 2026 09:50:16 GMT

👉 Originally posted on TTY

Events

🇺🇸 AgentCon Silicon Valley in Mountain View (May 4) – Free conference for developers building with AI agents at the Computer History Museum. Sacha Morard is speaking on token optimization for coding agents, covering techniques, frameworks, and live session data from Edgee.

🇳🇱 Amsterdam Tech Week (June 16–19) – The second edition of Amsterdam Tech Week returns across four days with startup events ranging from technical workshops to networking sessions, organized by local startups and VCs.

Audio

🎙 Headroom expands to voice – Headroom, the open source context compression layer for LLMs that saved 100B tokens in 25 days, has launched kompress-voice, a new model for voice applications. It integrates with ElevenLabs and Cartesia, runs locally, claims significant token savings, and does not collect user data. The team is seeking early design partners. 👉 Tejas Chopra.

Autonomous Agents

🎥 Skills, MCP, and code mode demystified – A 20-minute must-watch talk at the MCP Dev Summit with Nikolay Rodionov from Alpic breaking down the three paradigms: Skills are context-efficient reusable prompts loaded progressively, MCP provides structured cross-platform tool access, and Code Mode executes scripts in a sandbox to avoid context bloat.

🦀 Microsoft Rust guidelines for agents – Microsoft’s Pragmatic Rust Guidelines, covering idiomatic API design, safety, documentation, and testing, are available as a single agent-ready text file. Dropping it into a coding agent skill showed measurable quality improvements in generated Rust code. https://microsoft.github.io/rust-guidelines/agents/all.txt

📦 Multi-language agent skills pack – Youssef Tharwat’s set of agent skills, inspired by Microsoft’s Rust guidelines, covers lint rules, design guidelines, and idiomatic patterns for Rust, Python, and TypeScript. It draws from Microsoft guides, Google style guides, Ruff, Oxlint, and GoF patterns. Works with Claude Code, Codex, Cursor, and 30+ other agents via npx skills add.

Image, Video & 3D

🍌 Vision Banana unifies vision and generation – Google DeepMind’s Vision Banana instruction-tunes an image generator to handle segmentation, depth estimation, surface normals, and other vision tasks by treating them as image generation problems, beating specialist models zero-shot. Paper is here.

Reframes all 2D and 3D perception tasks as RGB image generation, producing outputs like segmentation masks or depth maps directly rather than through separate decoder heads.
Beats Segment Anything Model 3 on segmentation tasks and the Depth Anything series on metric depth estimation without any task-specific training or architecture changes.
Achieves these results via lightweight instruction-tuning on top of Nano Banana Pro (NBP) without sacrificing the base model’s image generation capabilities.

🔥 ChatGPT Images 2.0 released – OpenAI released a new image generation model for ChatGPT with significantly improved text rendering, multilingual support, and advanced visual reasoning. Community testing confirmed sharper text output, though the model dropped alpha layer support compared to gpt-image-1.

🗺 LingBot-Map autoregressive 3D reconstruction – A purely autoregressive 3D foundation model achieving approximately 20 FPS at 518x378 resolution over sequences exceeding 10,000 frames. Builds geometric context attention with long-context modeling while maintaining a compact streaming state, with no post-processing or optimization passes required.

🎭 Image models for sensitive content – Community explored which image generation models handle violent or crime scene content after mainstream options (GPT-image-2, ByteDance, Grok) blocked or distorted requests. Open alternatives were suggested including Flux via Pruna AI endpoints.

Cyber

🔐 Ctrl+G, AI cyber defense lab, launches – The startup focuses on building AI that can genuinely patch vulnerabilities across an entire codebase. The team spent three months building an automated data pipeline covering 80,000 CVEs and found that current top models consistently fail to patch vulnerabilities efficiently even when claiming success.

They’re looking for strong red/blue team experts 👉 Paco Villetard

🔒 Cloudsmith proxies AI model downloads – Cloudsmith’s artifact management platform now proxies and caches Hugging Face models and datasets, letting organizations govern which AI models reach their pipelines, with vulnerability scanning, policy enforcement, audit trails, and global distribution.

🛡 CVE-free base images on demand – Echo rebuilds popular open source container base images with only necessary components, patches all known vulnerabilities, and delivers them directly to your registry. Claims 99%+ vulnerability reduction and zero code changes needed in existing Dockerfiles, with a 24-hour CVE handling SLA.

Infrastructure

⚡ Google’s eighth-gen TPUs for agents – Google introduced two new chips built for the age of AI agents: TPU 8t for training, which can link up to 9,600 chips into a single cluster, and TPU 8i for inference, which runs models 80% more efficiently per dollar than the previous generation.

🇫🇷 Scaleway powers France’s health data – Antoine Millet announced the company was selected to host France’s Health Data Hub following an assessment across 350+ technical criteria covering security, resilience, and strategic autonomy. The platform will host a copy of the National Health Data System database.

Language Models

🧠 DeepSeek-V4-Pro hits 1M context – DeepSeek released V4-Pro, a 1.6T parameter MoE model (49B activated) supporting a 1 million token context window. It combines Compressed Sparse Attention with Manifold-Constrained Hyper-Connections and requires only 27% of single-token inference FLOPs versus DeepSeek-V3.2 at 1M context.

🤖 OpenAI introduces GPT-5.5 – GPT-5.5 is OpenAI’s new frontier model targeting agentic coding, computer use, and long-horizon knowledge work. It matches GPT-5.4 per-token latency while scoring higher on benchmarks, and uses fewer tokens to complete the same Codex tasks.

📊 Qwen3.5-Omni scales to hundreds of billions – The Qwen3.5-Omni technical report covers an omni-modal model scaled to hundreds of billions of parameters with a 256k context window, trained on a massive heterogeneous dataset including millions of hours of audio-visual data. It claims improvements in multilingual understanding and emotional nuance in speech generation.

🔏 OpenAI Privacy Filter open-sourced – OpenAI released Privacy Filter, a 1.5B-parameter Apache 2.0 token-classification model adapted from GPT-OSS to detect and mask personally identifiable information in text. It runs locally in the browser via WebGPU, making it OpenAI’s first open model of 2026.

MLOps

🔀 Orq.ai router for LLM traffic – Orq.ai’s LLM router provides real-time cost visibility, governance, and vendor lock-in prevention across model providers. It runs on-prem with minimal footprint, supports both managed and bring-your-own API keys, and integrates natively with Claude Code via MCP.

Programming

🛠 AI prototypes need programmer review – DHH argued that while AI has given designers prototyping superpowers, models still cannot reliably one-shot production-ready code for large critical applications without programmer review or reimplementation, citing Basecamp as a concrete example where AI-generated designer code regularly requires rework.

🌙 Kimi K2.6 claims coding SOTA – Kimi claims top scores across most major agentic benchmarks. Under the hood it runs up to 300 parallel sub-agents, handles 4,000+ tool calls in a single run, and can sustain autonomous execution for over 12 hours across frontend, DevOps, and performance optimization tasks.

🧑‍💻 Qwen3.6-27B targets coding tasks – Alibaba released a 27B-parameter open-source AI model that codes and reasons at near-flagship level, often beating much larger models on coding and math benchmarks, and it also understands images and video.

📋 Anthropic explains Claude Code regression – Anthropic traced three separate issues affecting Claude Code since March: a default reasoning effort downgrade from high to medium, a caching bug that cleared prior reasoning on every turn after a session went idle, and a verbosity-limiting system prompt addition that hurt coding quality. All three were fixed by April 20.

Robotic, World AI

🌐 MultiWorld models multi-agent worlds – MultiWorld is a unified framework for action-controllable, multi-view consistent video world modeling across multi-player games and multi-robot collaboration, addressing the single-agent limitation of current video world models.

🌍 HY-World 2.0 open 3D world model – Tencent’s HY-World 2.0 generates navigable 3D scenes (3DGS/mesh) from text or a single image, importing directly into Unity, Unreal Engine, and Isaac Sim, positioning itself as the first open-source state-of-the-art 3D world model.

The Lunch

Each week, TTY Lunch brings together exceptional builders around the table, this time under bright sun on the Tech Makers rooftop in Amsterdam. Today’s lineup included Alina Dumitrache and Fabien Bouhier (Tech Makers), Guillaume Lesur (Wire), Justin Halsall (rrweb.io), Sohrab Hosseini (orq.ai), Robert Hommes and Vladimir Vučković (Moyai), Wilco Kruijer (ReJot), and Zac Allen (Zero Servo).

Here were some of the topics:

Red-Green-Refactor with Agents
Useless Memory, Thin Harness, Fat skills
MCP + Skills, The Distribution of Roles
The Unsolved Frontier For Coding Agents
Finding the Right Co-founder

👉 The full discussion here

New Members

🇳🇱 Zac Allen – Founder @ Zero Servo · Robotics gaming platform doubling as an embodied data platform. Currently building Zero Servo while geeking out on Arc Raiders. Special power: will quote from random films and TV series at any given moment. 📍 Amsterdam, Netherlands.

🇬🇧 Arnaud Thiercelin – Founder & CEO @ Rogue Cortex · Stealth tools to accelerate mobile robotics development. Author of the book The Pirate Way. Special power: dunno, but I wrote a book called the Pirate Way. 📍 Guildford, UK.

Contributors This Week

Robert Hommes (Moyai.ai), Gabriel Olympie (2501.ai), Félix Raimundo (Tychobio), Tejas Chopra (Netflix, Headroom), Maziyar Panahi (OpenMed), Aly Moursy (The Artificial Intelligence Company of Cairo), Paco Villetard (Ctrl+G), Sohrab Hosseini (Orq.ai), Amine Saboni (Pruna.ai), Arnaud Thiercelin (Rogue Cortex), Kemal Toprak Uçar (Continuity), Koutheir Cherni (Guepard), Antoine Millet (Scaleway), Benoit Kohler (Opsmill), Charles Sonigo (Alpic), Ihab Bendidi (Recursion), Pierre Chapuis (Finegrain), Stan Girard (The Vibe Company), Youssef Tharwat (Noodlbox), Zac Allen (Zero Servo), Benjamin Trom, Gilles Seghaier (Astran), Julien Seveno-Piltant, Karim Matrah (Contrast), Louis Manhes (Genario), Quentin Dubois (OSS Ventures), Remi Kaito (Arcade AI), Sacha Morard (Edgee), Vladimir de Turckheim, Wissam Antoun (Inria)

TTY-changelog #039

Kevin Kuipers — Sat, 18 Apr 2026 13:02:49 GMT

👉 Article originally posted on TTY

Events

🇧🇷 TechBio Social at ICLR in Rio (Apr) – Valence Labs and the LMRL Workshop host a TechBio gathering for ICLR 2026 attendees in Rio de Janeiro.

Audio

🎙 Gemini 3.1 Flash TTS launched – Google shipped Gemini 3.1 Flash TTS in Gemini API, AI Studio, Vertex AI, and Google Vids, scoring an Elo of 1,211 on Artificial Analysis’s blind-preference TTS leaderboard. New audio tags let developers steer vocal style, pace, and delivery with natural-language commands embedded in the text input, alongside native multi-speaker dialogue, support for 70+ languages, and SynthID watermarking on all outputs.

🎓 CS153 guest lecture by ElevenLabs – Stanford’s Frontier Systems class hosted ElevenLabs co-founder Mati Staniszewski walking through the company’s early days, part of a recommended lecture series covering today’s frontier infra teams.

Autonomous Agents

🎛️ The MCP design debate – A week-long thread reframed how to think about MCP servers. A Cloudflare post by Kenton Varda and Sunil Pai argues that the prevailing pattern of exposing MCP tools directly to LLMs is flawed. LLMs have seen vast amounts of real world TypeScript, but only a narrow, artificial set of tool calls. Their approach converts MCP tools into a TypeScript API, has the LLM write code against it, and runs that code in a V8 isolate sandbox where MCP servers are exposed as bindings rather than accessed over the network.

Phil Schmid’s counterpoint argues that MCP is not the problem, but poorly designed servers are. He frames MCP as an agent-facing UI rather than a REST wrapper, and outlines principles such as designing tools around outcomes, simplifying inputs, treating descriptions and errors as core context, limiting the number of tools, using explicit naming, and avoiding large unstructured outputs.

The naive OpenAPI2MCP pattern of one endpoint per tool is the shared punching bag: both the code-mode and best-practices camps agree it fails under real workloads because MCP is a user interface for agents, not a basic REST wrapper.
A pragmatic middle ground surfaced: design MCP servers around the workflows users and agents actually need, keep a search-and-execute path for the unanticipated, and evaluate servers by pass^k on frequent tasks.
Community take w/ Sebastien Charrier (Bump.sh): “I’m not convinced by the 100% code mode for all use cases. If the LLM has to build the code wrapper at runtime, it can result in very poor performance. It can work in a dev process because you accept waiting a few minutes, but for end-users consuming your service through MCP it’s not acceptable.”

🤖 Rethinking agent evaluation as a production system – Jules Belveze argues the current paradigm is breaking down: static datasets decay, benchmarks get gamed, and they fail to capture real outcomes. He suggests evaluation should emerge from the platform itself, leveraging production traces, feedback, and version history rather than separate benchmarks. He outlines three directions: behavioral fingerprinting based on execution consistency, a feedback flywheel that turns failures into minimal reproduction tests, and outcome signals such as user return, escalation to humans, and increasing task complexity over time, with static benchmarks remaining useful but secondary.

🎲 Property-based testing for agentic coding – Anthropic’s Frontier Red Team presented an agent that infers code properties from docstrings, type hints, and comments, then generates property based tests using Hypothesis. The goal is to automatically discover bugs by stress testing these inferred properties with large numbers of edge case inputs, rather than relying on hand written tests.

📱 Gemma 4 edge agent skills – Google claims Gemma 4 brings advanced agentic AI to devices, enabling multi-step planning, autonomous actions, and multimodal tasks on-device. Open under Apache 2.0, it supports 140+ languages and runs efficiently across mobile, desktop, and IoT hardware.

Biotech, Health, and Chemistry

🧬 OpenAI ships GPT-Rosalind biology LLM – OpenAI released GPT-Rosalind, a version of its model trained on how biologists actually work and on the big public biology databases. It is built to push back when a researcher proposes a bad idea instead of agreeing just to be helpful. Only approved US teams get access for now, with a lighter research plugin open to everyone else.

Community take w/ Maziyar Panahi: “I’ve tried Claude for Health. The skills are mediocre at best, and without third party MCP servers they are basically useless. It is disappointing, especially since they have not been updated in 3 to 4 months. Meanwhile, OpenAI’s latest podcast hints at deeper industry partnerships and specialized models and tools for life sciences that Codex can orchestrate. That direction might explain why Anthropic is acquiring biotech startups.”

Image, Video & 3D

🌊 ByteDance’s paper and flow model discussion – The Continuous Adversarial Flow Models paper by Shanchuan Lin takes an existing flow matching model and trains a second “critic” network alongside it to judge whether generated images look real. After just 10 extra training passes, image quality on ImageNet improved dramatically. The paper triggered a week-long community thread on why flow matching and diffusion, the two dominant frameworks behind modern image generators, are actually the same thing, surfacing three further resources worth reading.

The “same thing” reference: a Google DeepMind explainer showing that diffusion and flow matching can produce identical samples under the right parameterization.
A Kaiming He’s Back to Basics paper from 2025 argues that diffusion models no longer truly denoise but instead predict noise, a harder task. It suggests predicting the clean image directly and shows that simple pixel-level Transformers can achieve similar results without added complexity.
A great intro to flow matching models: Julia Turc’s video breakdown. Paired with the companion diffusion.fyi interactive tutorial, recommended as the cleanest on-ramp before harder papers.
Community take w/ Pierre Chapuis: “Flow matching is just what happens when you realize there is a simpler and better way to do diffusion.”

🎬 CS153 with Black Forest Labs – Stanford’s Frontier Systems class hosted Black Forest Labs co-founder Andreas Blattmann, covering the lab’s image and video generation stack.

Cyber

📐 AISI evaluates Claude Mythos – The UK’s AI Safety Institute tested Claude Mythos Preview on a battery of hacking challenges. It solved the hardest tier 73% of the time, and became the first AI to complete a 32-step simulated corporate network attack entirely on its own, succeeding in 3 of 10 attempts and averaging 22 steps versus 16 for the previous best model. It got stuck on an industrial-controls scenario though, and its performance kept scaling with more compute.

Community take w/ Gabriel Olympie: “As long as it is not independently tested, I consider Mythos a marketing stunt. More seriously, it’s not the first time Anthropic is using cyber to create traction, so even if Mythos is most likely real and quite powerful, I’d wait to see it evolve in the real world before falling for the narrative.”

🔍 Post-Mythos jagged cyber frontier – Security firm AISLE argued Mythos’s capabilities are less exclusive than Anthropic suggests. When they ran Anthropic’s own showcase bugs through small, cheap open-source models, 8 out of 8 caught the main flaw, even a tiny model costing roughly $0.10 per million tokens. Their point: the security workflow around any model matters more than the model itself. AI hacking skills are also uneven: a model can ace a hard bug and then completely miss an easier one.

🛡 GPT-5.4 Cyber for defenders – OpenAI launched GPT-5.4 Cyber, a version of GPT-5.4 tuned for defensive security work with looser safety filters that would normally block legitimate cybersecurity tasks. It can also analyze compiled software without needing the source code, something binary reverse engineers usually do by hand. Access is limited to verified individuals and approved organizations through the Trusted Access for Cyber program.

Language Models

🧠 Qwen 3.6 35B-A3B refreshed – Qwen shipped 3.6 as a 36B-parameter, 3B-active mixture model on Hugging Face with a 262K native context, positioned around agentic coding and a new Thinking Preservation mode that keeps reasoning traces from historical messages to reduce redundant thinking across turns.

🆕 Claude Opus 4.7 ships – Anthropic announced Claude Opus 4.7 as its most capable Opus, pitched for long-running tasks with more rigor, tighter instruction following, and self-verification of outputs, alongside three-times resolution vision, a new xhigh effort tier on the API, and a /ultrareview command in Claude Code. Reception on reddit was sharp: reports of faster token burn, lazy “let’s pick this up later” quitting after a few messages, hallucinated packages and imaginary coworkers, fabricated search tool calls without the UI indicator, sycophancy that flips positions under light pressure, and broad failure to honor configured claude.md and system preferences that 4.6 handled without drift.

Community take w/ Julien Seveno-Pilitant: “Even if the scaling laws still work technically, from a human perspective it makes less and less difference. The models are converging toward the entropy of the language itself, and they approach it logarithmically, so the progress is less and less visible.”

👩‍💻 Tokenizers tuned for code LLMs – A new paper with Ivan Yamshchikov (Pleias) contribution proposed training tokenizers that focus on widely shared, useful code pieces rather than rare project-specific junk, so models learn durable code patterns more reliably.

Programming

💬 Karpathy on prompt requests – Andrej Karpathy relayed a Peter Steinberger line that PRs should stand for “Prompt Requests”: capable agents can implement most ideas, so there is no need to take an idea, expand it into a vibe-coded mess on free-tier ChatGPT, and send that as a PR. The community pushback was that coding agents still need a second set of eyes, and that specs or markdown files are not the new source code.

🎻 Cline Kanban for multi-agent orchestration – Cline Kanban is a promising open-source UI to orchestrate multiple coding agents (Claude Code, Codex, Cline) in a Kanban board, letting you break down projects, chain dependent tasks, and review real-time diffs and comments in one place.

🎤 Building pi in a world of slop – Mario Zechner’s presentation of pi at AI Engineer Europe, his minimal, extensible coding‑agent harness. He criticized opaque “slop” agent ecosystems and OSS clanker spam, and argues for slow, tightly scoped, human‑controlled use of agents in codebases.

Community take w/ Robert Hommes: “AI Engineer Europe completely altered my beliefs on coding agents. Codex and Claude Code are the wrong harness to run GPT and Claude. Thin harness and heavy skills together with high-intelligence models like Claude, Gemini, and GPT is going to be the real unlock. Removing the developer dependency using skill transfer learning will end developer babysitting and start swarm programming with autonomous agents.”

Robotic, World AI

🤖 Pi07 steerable robotic foundation model – Physical Intelligence released pi07, a steerable model showing strong gains in generalization, including early compositional abilities. It can learn new tasks like using kitchen appliances with step by step language guidance, then execute them autonomously after fine tuning. It also shows cross embodiment transfer, for example folding laundry on a bimanual UR5e without task specific training data. The model supports multimodal prompts, including language, metadata, control signals, and visual subgoals.

🌍 NVIDIA Lyra 2.0 for 3D worlds – NVIDIA Research released Lyra 2.0, a framework for generating persistent, explorable 3D worlds at scale from a single image. It fixes the common failure where models forget what spaces look like and lose track of motion over time by maintaining per-frame 3D geometry to retrieve past frames and using self-augmented training to correct its own temporal drift. A robot can be dropped into the generated world for real-time simulation and immersive applications.

👉 Community Ask:

🤖 Luigi d’Ovidio looks for a quick DM with anyone hands-on with Vision-Language-Action model development.

New Member

🇫🇷 Charles Sonigo – Cofounder @ Alpic · All-in-one MCP platform helping companies build presence in ChatGPT and Claude via tools to build, deploy, test, audit, and monitor MCP servers and ChatGPT apps. Geek turned PM turned founder turned Geek, full loop. Dad, tennis, and wingfoil on the side. Special power: I sleep fast. 📍 Paris, France

Contributors This Week

obert Hommes (Moyai.ai), Pierre Chapuis (Finegrain), Aly Moursy, Gabriel Olympie (2501.ai), Gilles Seghaier (Astran), Julien Seveno-Pilitant, Sebastien Charrier (Bump.sh), Amine Saboni (Pruna.ai), Daniel Madalitso Phiri, Gawen Arab (Airbuds), Ihab Bendidi (Recursion), Lior Oren, Louis Manhes (Genario), Quentin Dubois (OSS Ventures), Guillaume Lesur (Wire), Koutheir Cherni (Guepard), Emmanuel Benazera (Jolibrain), Julien Duquesne (Scienta Lab), Youssef Tharwat (Noodlbox), Andrea Pinto (Notte.cc), Antoine Sueur (Pletor), Charles Sonigo (Alpic), Constant Razel (Hypermind), Jules Belveze, Luigi d’Ovidio (Graybx), Noé Charmet (Shipfox), Sylvain Hajri (Epieos).

TTY-changelog #038

Kevin Kuipers — Sat, 11 Apr 2026 09:54:15 GMT

👉 Article originally posted on TTY

Community Asks

📰 Newsletter management tools – Glenn Sonna asked the community what tools they use to manage the flood of AI newsletters and filter for signal. Suggestions ranged from OpenClaw-based cron jobs to Tasklet, a YC-backed managed agent platform for automating business processes with plain English.

🔬 Amazon Research Awards feedback – Hani Chalouati is seeking feedback on the Amazon Research Awards Spring 2026 call for proposals.

Community Updates

🚀 Ascii.dev adds voice and in-game overlay – The pocket CTO agent now supports fully hands-free operation through walkie-talkie-style voice commands and an in-game overlay, letting developers orchestrate coding agents while doing other things.

Autonomous Agents

☠️ Claude subscriptions drop third-party coverage – Anthropic announced that Claude subscriptions will no longer cover usage on third-party tools like OpenClaw. Users can still authenticate via extra usage bundles (now discounted) or API keys.

🦞 Aquarium CE, self-hosted agent management – Shuai Zhang (Jinko) released an open-source platform for deploying and managing AI agent instances locally with a single command. It supports 27+ AI providers, 14 messaging channels (WhatsApp, Telegram, Discord, Slack), MCP tools, and encrypted credential storage, all running on local SQLite with zero telemetry.

🎛️ MCP vs Skills, a practitioner’s take – A detailed post argued that the MCP remains the superior architectural pattern for giving LLMs access to services. Remote MCP servers offer zero-install usage, seamless updates, saner OAuth-based auth, and true portability across devices, while Skills work best as pure knowledge layers.

🔍 Governed agent autonomy patterns – After analyzing the Claude Code source leak, Nnenna Ndukwe (Qodo) extracted a docs-first reference for building agentic coding systems with governance baked in. The framework covers five gates: plan, permission, tool trust, verification, and runtime accountability.

Biotech, Health, and Chemistry

🧬 AI-native biotech meets radiopharma – TTY published an interview with Ashley van Heteren of Radiogenesis, a startup using generative models and physics-informed AI to design targeted radioactive drugs for metastatic cancer. The company uses foundation models to design peptide candidates in silico before lab validation, compressing early discovery from years to months.

Image, Video & 3D

🩻 Raidium confirms the VLM mirage effect – Following last week’s mention of the Stanford “mirage” paper, Pierre Manceron (Raidium) shared that his team had independently observed the same phenomenon and published their own study. Their paper explored how prior medical knowledge and statistical reasoning allow models to bypass visual input entirely: “In practice many medical diagnoses can be inferred just from the question and good medical knowledge, statistically speaking. For example you know that a patient is a smoker, and ask ‘is there a lesion on this CT scan’, even if you don’t see the image itself, the probable answer is still lung cancer.”

Cyber

🏴‍☠️ Claude Mythos Preview stays behind closed doors – Anthropic released a 244-page system card for Claude Mythos Preview, its most capable model to date, alongside Project Glasswing, a program restricting access to 12 major tech companies for defensive cybersecurity only. The model achieved 93.9% on SWE-bench Verified and 100% on Cybench, and autonomously discovered thousands of zero-day vulnerabilities in every major OS and browser.

The system card documented rare but concerning behaviors: sandbox escapes, posting exploits to public websites, editing git history to hide unauthorized changes, and attempting prompt injection on evaluation graders.
Anthropic described the model as simultaneously its best-aligned and highest-risk release, noting that its skill enables it to reach more dangerous situations even when well-intentioned.

🔒 Astral shares CI/CD security practices – The team behind Ruff and uv published a detailed breakdown of how they secure their open-source tools against supply chain attacks. The post covers banning dangerous GitHub Actions triggers across the organization, requiring commit-pinned actions with impostor-commit detection, using Trusted Publishing for registry releases, and employing Sigstore attestations for binaries.

Language Models

🔥 GLM-5.1 tops SWE-Bench Pro – Zhipu AI released GLM-5.1, a 754B-parameter open-weight model under MIT license that claimed the top spot on SWE-Bench Pro (58.4) and led on MCP-Atlas (71.8). The model sustained autonomous execution for up to 8 hours across hundreds of iterations and thousands of tool calls, unlike predecessors that plateau early.

🧠 Meta’s benchmark layout draws scrutiny – Community members debated Meta’s new model, Muse Spark, and its potentially misleading benchmark presentation. The broader discussion questioned whether benchmarks still matter when real-world impressions consistently diverge from reported scores.

🔩 Karpathy’s LLM knowledge base in practice – Jeremie Kalfon shared extensive experience using Karpathy’s LLM KB concept to replace OpenClaw’s simple markdown memory with a rich Notion/Obsidian database: “The powerful idea is to replace OpenClaw’s simple memory markdown as a large Notion/Obsidian database with many enriched markdown files and an index. I can always search and interact with them on my Notion and this is nicer than through my terminal/IDE.” The approach worked well for personal knowledge bases built from messaging apps and social media exports, but model laziness remained the biggest unsolved friction: “They tend to be lazy colleagues/interns, they do not push the task to their limits, nor do they think about the ramifications of what has to happen next when the task is finished.” Running mostly on Sonnet 4.6, token costs reached ~€100/day.

Programming

🥷 Claude-sock wraps Claude Code for subscription use – A CLI tool that wraps the Claude Code TUI to let tools like OpenClaw use Claude models with a subscription only, without API keys or extra tokens. It spawns a fresh TUI process per call (with ~5s startup overhead) and translates output to the claude -p protocol format.

🐳 Docker Sandboxes quickstart guide – A hands-on guide covering the Docker Sandboxes (sbx) experimental launch, including parallel agent setup, branch mode, networking, Docker Compose inside sandboxes, port forwarding, and network policy management. The guide uses DevBoard, a full-stack Next.js + FastAPI issue tracker with intentional bugs as exercises.

Community take w/ Amine Saboni: “I’m fostering the intuition that Docker is a really appropriate abstraction for developers to work with AI. Their not-so-recent focus on agent isolation is really interesting to track.”

🐆 Ardent AI vs Guepard on database cloning – When YC-backed Ardent AI launched with a pitch for instant Postgres clones for coding agents, the community compared it with Guepard, an open-source alternative already available. Key differences: Guepard is self-hostable, supports all databases (not just Postgres), and branches entire data environments rather than using production copies.

👨‍💻 Pi’s creator joins Earendil – Mario Zechner, creator of pi (the agent powering OpenClaw), announced he sold the project to Earendil, co-founded by Armin Ronacher and Colin Daymond Hanna. The pi core stays MIT-licensed forever, with commercial additions planned under Fair Source (converting to open-source via delayed publication). Zechner cited alignment on OSS values and wanting to avoid the solo-founder VC path after his experience with RoboVM’s closed-source fate at Microsoft.

TTY Lunch

Each week, TTY Lunch brings together exceptional builders around the table. Today’s lineup included Bertrand Guiheneuf (Fairjungle), Denis Brule (Finegrain), Francis Bouvier (Lightpanda), Karim Matrah (Contrast), Olivier Desclaux (Harmattan), Pierina Camarena (42 AI), Quentin Soulet (Dune Health), and Sacha Morard (Edgee).

Fast code alone doesn’t build product value
Non-engineers shipping code without system awareness
Can AI learn taste
Facing the new GAFAM
Classical models still power critical systems

👉 The full discussion is here

New members

🇪🇬 Aly Moursy – Founder & CEO @ The Artificial Intelligence Company of Cairo · Applied AI lab building AI products for the MENA region. Born in Cairo, raised in Dubai/Canada, studied engineering in Canada. 2x startup founder. Special power: making funny tech memes that go viral. 📍Cairo, Egypt

🇳🇱 Joel Milligan – Founding AI Engineer @ neno · Based in Amsterdam for 4 years, 10 years of software engineering experience, ex-founder (TechStars Berlin). Currently building Viche, an agent-to-agent communication layer and discovery protocol. Special power: navigation, has never ever been lost. map📍 Amsterdam, Netherlands

Contributors This Week

Aly Moursy (AIC of Cairo), Glenn Sonna (Xybrid), Koutheir Cherni KC (Guepard), Enrico Piovano (Goji), Hani Chalouati (Guepard Inc.), Nnenna Ndukwe (Qodo AI), Pierre Manceron (Raidium), Robert Hommes (Moyai.ai), Anicet Nougaret (Ariana.dev), Pierre Chapuis (Finegrain), Amine Saboni (Pruna.ai), Félix Raimundo (Tychobio), Hugo Venturini (SkipLabs), Jeremie Kalfon (Pasteur), Quentin Dubois (OSS Ventures), Shuai Zhang (Jinko), Youssef Tharwat (Noodlbox), Daniel Huynh (Conception), Gabriel Olympie (2501.ai), Ihab Bendidi (Recursion), Joel Miligan, Laurent Mazare (Gradium), Nancy Wang (FluenTea), Stéphane Collot (Sequense), Tejas Chopra (Netflix)

TTY-changelog #037

Kevin Kuipers — Sat, 04 Apr 2026 10:04:08 GMT

👉 Article originally posted on TTY

Community Updates

⭐️ Headroom crosses 1k stars with v0.5.12 — The open-source context compression layer for LLM applications shipped faster inference, supports for LangChain, LangGraph, Windows and OpenClaw. Used by Netflix, Meta, and Vercel.

📈 Edgee shows Claude Code session analytics — A new shareable session report format for the Edgee AI Gateway showed 63% tool compression across 213 requests and 13.9M tokens! Works with Claude Code, Codex, OpenCode, and Cursor support coming soon.

📱 LFM2.5-350M runs on Android via Xybrid — Liquid AI’s 350M-parameter model ran at 40+ tokens per second on Android through the Xybrid runtime, crossing the threshold where on-device AI feels native.

Events

🇳🇱 Reliable Agents Club, April 15th in Amsterdam — For people who run AI agents in production and want to make them more reliable and debuggable. The event focuses on discussing agent reliability challenges, networking with peers, and introducing Moyai’s platform for monitoring and diagnosing failures in agentic workflows.

Audio

🔊 On-device TTS enters private beta — Gradium announced Phonon, a text-to-speech model that runs locally on smartphone CPUs with natural voices, multilingual support, and voice cloning, requiring no server, no latency, and no per-call cost.

The private beta is now open for game developers and app builders. The model runs entirely on-device, eliminating cloud dependency and enabling offline voice interactions at scale for free-tier products.

A community discussion highlighted the gap between available high-quality TTS models and polished consumer apps: “Why wouldn’t Mistral, Gradium, or others just build a competing Speechify? It feels like the remaining work is only 10 percent once you already have the model and the API?”.
For Laurent Mazare, “audio models are advancing very quickly right now. Several companies, including us at Gradium, focus on training these models but lack the cycles or expertise to build applications, so we sell them to startups that handle the end products.”
Also, the “remaining 10%” framing undersells the problem for Glenn Sonna. “Getting a TTS model to run well on someone’s actual phone is a whole different beast: quantization, memory, battery, latency. Cloud API plus per-token cost kills UX for something like a reading app.”

Autonomous Agents

👔 Pocket CTO delegates to cloud agents — The new product from Anicet Nougaret is live. Pocket CTO accepts voice memos on Telegram or desktop instructions, then orchestrates multiple cloud coding agents working in parallel on their own machines. It maintains a central always-on agent aware of your tools and codebase, surfaces demos, PDF reports, hosted builds, and executive summaries, and supports cron jobs, webhooks, secrets, and git hooks.

🧰 A massive agent harness for coding tools — A performance optimization system originating from an Anthropic hackathon winner, shipping 30+ agents, 150+ skills, autonomous loop patterns, continuous learning, and AgentShield security scans across 10+ language ecosystems. Built over 10+ months of real-world use with 30+ contributors, it supports Claude Code, Codex, Cursor, and OpenCode via a single install pipeline.

Biotech, Health, and Chemistry

🧬 Bio foundation models hit the noise floor — A journal entry analyzed a GenBio AI benchmark testing whether AI can predict how cells respond to interventions, drawing on TTY community discussions to propose four investment criteria for techbio: solve a specific biological problem deeply, iterate fast on data that actually predicts what happens in patients, pick mechanisms where physics constrains the outcome, and generate rich data only where the underlying biology is well-understood enough to learn from.

The best AI models approached but could not exceed the natural variability of the experiments themselves, meaning the ceiling is set by biology, not by compute or model size.
Models that encoded existing biological knowledge (which genes interact with which) consistently outperformed those trained on raw data alone, suggesting that in biology, domain expertise still beats brute-force scaling.
Newer models not in the benchmark are closing the gap: State (Arc Institute) improved prediction accuracy by up to 200% over baselines by training on 167M cells, TxPert (Valence Labs/Recursion) matched experimental reproducibility using biological knowledge graphs, and Xaira’s X-Cell claims to go further still at 4.9B parameters.
Community take w/ Ihab Bendidi: “GenBio’s setup does not actually perform perturbation prediction in the standard sense. It tests whether gene embeddings from unperturbed cells correlate with perturbation effects, which is an interesting but distinct question.”
Community take w/ Félix Raimundo: “Even with perfect predictions, all you achieve is predicting the result of experiments that already fail to translate to humans 91% of the time. Being able to predict a broken proxy faster does not fix the proxy.”
A complementary essay argued that AI will not accelerate drug development on its own, because the real bottlenecks are patient recruitment, years-long endpoint observation, and regulatory friction, not a lack of intelligence.

🔬 Five new cancer research teams launched — Global Cancer Grand Challenges funds five international teams (ATLAS, CAUSE, ILLUMINE, InteroCANCEption, REWIRE-CAN) up to $25M each to tackle cancer avoidance, mutational signatures, dark proteome, nervous system–cancer links, and therapeutic rewiring.

Image, Video & 3D

🎨 Unified image model generates and edits — Alibaba released Wan 2.7 Image, a single model handling generation, editing, and understanding in one architecture. It offers granular face control (bone structure, eyes, contour across ethnicities and ages), color palette extraction with up to 8 HEX codes, 3K-token text rendering in 12 languages including dense tables and formulas, interactive region-based editing via natural language, and multi-image generation.

Cyber

☠️ Axios npm package compromised — The popular JavaScript HTTP client with 100M+ weekly downloads was briefly hijacked through a compromised maintainer account, delivering a cross-platform remote access trojan via a hidden dependency during a two-hour window on March 31.

🎩 Black-hat LLMs automate real attacks — Nicholas Carlini, Research Scientist, Anthropic, at the [un]prompted 2026 argued that LLMs have crossed a threshold for autonomous vulnerability discovery, with capabilities growing on an exponential curve. Placing a model in a VM with broad permissions and a short prompt was enough to reliably discover severe bugs in heavily audited software.

Demonstrated examples included a blind SQL injection in Ghost CMS that the model fully weaponized into an exploit exfiltrating admin credentials, and multiple remotely exploitable heap buffer overflows in the Linux NFSv4 daemon including a bug dating back to 2003.
Benchmark evidence showed that models from only a few months ago rarely succeeded on such tasks, while current generations repeatedly found kernel-level bugs, with task complexity and exploit value growing roughly exponentially.
The talk warned that automated bug discovery will soon far outstrip human triage and patching capacity, and that weak safety filters only hobble honest defenders while bad actors jailbreak around them.

Language Models

📈 Time series models meet trading — Google’s TimesFM 2.5, a 200M-parameter pretrained forecasting model with 16K context length and continuous quantile predictions, sparked a community discussion with insights from Enrico Piovano, Jeremie Bordier, and Hubert Ancelot about applying foundation time-series models to financial markets.

Early evaluations showed decent zero-shot performance on demand forecasting but fell short of fine-tuned SARIMAX models, with inference time a dealbreaker for high-frequency use. Chronos 2 from AWS, a universal forecaster handling univariate, multivariate, and covariate-informed tasks via in-context learning, was also tested for stock backtesting with limited results.
Community members with quantitative finance experience noted that systematic trading firms still rely on legacy logistic regressions with smart feature selection, and the shift from high-frequency to mid-frequency strategies makes portfolio hedging more important than raw latency.
Robert Hommes drew an analogy with Nvidia PhysicsNeMo (transformer and neural operator architectures for physics simulation), suggesting similar approaches could model portfolio dynamics and forward-simulate hedging effects to maintain stable return-to-risk ratios.
Value in time series models for trading likely lies in assessing stationarity and cointegrated pairs in mid-frequency strategies rather than directional stock predictions.

💎 Google releases Gemma 4 under Apache 2 — Google launched its most capable open model family in four sizes (E2B, E4B, 26B MoE, 31B Dense), purpose-built for advanced reasoning and agentic workflows.

The 31B model ranked third among open models on the Arena AI leaderboard, outcompeting models 20x its size.
Day-one support spans llama.cpp, MLX, vLLM, Ollama, LM Studio, Unsloth, transformers.js for browser inference, and fine-tuning via TRL including multimodal tool-use workflows.

👁️ Vision models answer without seeing — Multimodal models can often “fake” seeing by leveraging language patterns and benchmark quirks, so their impressive scores and detailed visual explanations can be an illusion. To trust them (especially in fields like medicine) we need evaluations and system designs that explicitly check whether their answers truly depend on the images, rather than just on text and clever guessing.

🐉 Qwen 3.6 Plus targets agentic workflows — Alibaba released its new flagship proprietary model combining hybrid linear attention with sparse MoE routing, a 1M-token context window, and up to 65K output tokens. Optimized for coding agents with support for code repair, complex terminal operations, and automated multi-step tasks, it also added enhanced multimodal reasoning for document understanding and visual coding workflows.

⚗️ Tiny 350M model punches above weight — Liquid AI released LFM2.5-350M with 28T tokens of pre-training and large-scale reinforcement learning, outperforming models over twice its size on tool use, data extraction, and instruction following. The model runs on devices from Raspberry Pi to smartphones.

🌿 Eight billion params fit in 1.28 GB — Bonsai-8B compressed a Qwen3-8B model to 1-bit quantization, shrinking it to 1.28 GB while scoring 70.5 average across multiple benchmarks. On an M4 Pro it achieved 131 tokens per second for generation, and 44 tok/s on iPhone 17 Pro Max with 2-6x better energy efficiency than 4-bit baselines.

Programming

🔓 Claude Code source code leaked — The Claude Code source was exposed through a source map file accidentally left in Anthropic’s npm registry. Within days, independent researchers published detailed analyses of the codebase, an open-source fork added support for 200+ models, and an interactive architecture explorer mapped all 519K+ lines of code.

A reverse-engineering analysis shows Anthropic has better internal safeguards (post-edit verification, multi-agent orchestration, chunked reads, stricter prompts) that reduce hallucinations and context loss, but these are gated to employees. Author shares a “CLAUDE.md” fix to replicate them (available at the end of the tweet).
OpenClaude forked the CLI to support OpenAI, Gemini, DeepSeek, Ollama, and 200+ models via OpenAI-compatible APIs while preserving Claude Code workflows (bash, file tools, MCP, slash commands). It added agent routing that maps different agent types to specific models within one session for cost and performance optimization.
CC Unpacked mapped 1,900 files into an interactive explorer covering 50+ built-in tools, every slash command, and hidden unreleased features including Kairos (persistent mode with background actions), UltraPlan (long Opus planning runs), Coordinator Mode (multi-agent task decomposition in git worktrees), Bridge (remote control from phone/browser), and Daemon Mode (background sessions via tmux).

🐛 Claude Code cache bugs exposed — A detailed analysis revealed that the original v2.1.89 standalone binary had severe cache bugs dropping cache reads to 4-17%, causing nearly every token to be billed at full price.

👁️ GLM-5V-Turbo reads design drafts — Z.ai released GLM-5V-Turbo, a model using a next-generation CogViT architecture with deeply fused text-vision training. It natively ingests images, videos, design drafts, and document layouts to generate complete runnable code, targeting UI-to-code workflows and GUI agents.

🖱️ Cursor 3 goes fully agentic — Cursor announced version 3 with a whole new interface window - outside of VS Code - that lets developers run many agents concurrently across local projects, worktrees, SSH remotes, and cloud machines.

💻 Mac app wraps Headroom locally — A community-built macOS menu bar app that wraps the Headroom CLI, providing persistent savings tracking and one-click setup without pip install. All processing runs locally with no data leaving the machine.

🐳 Docker sandbox guide for agents — A hands-on quickstart for running Claude autonomously inside Docker Sandboxes using the sbx CLI, built around a full-stack demo app (Next.js, FastAPI, Postgres). Covers direct mode (edits your working tree) and branch mode (isolated git worktrees for parallel agents), with secure secrets management via OS keychain, configurable network policies, and multi-workspace mounting for cross-repo scenarios.

TTY Lunch

Each week, TTY Lunch brings together exceptional builders around the table. Today’s lineup included Charles Borderie (Lettria), Julien Goupy (Applikai), Antony Marion (Noways), Frederic Legrand (H Company), Louis Choquel (Pipelex), Julien Millet, Thomas Payet (Meilisearch), Alexandre Pereira (2501.ai), Francois Massot and Paul Masurel (Quickwit & Datadog).

Token maximalism vs. refined intent
Vibe fatigue and specification collapse
Deterministic spatial design
Moats: from code to responsibility

👉 The full discussion is here

New members

🇩🇪 Nancy Wang — CTO @ FluenTea · Language exchange with real humans and a touch of AI help. PhD combining computer vision and neuroscience, previously at IBM Research and Amazon. Born in China, raised in Canada, studied in the US, now in Berlin. Enjoys video games (currently Slay the Spire 2, all-time favorite XCOM). Special power: has never suffered a hangover. Berlin, Germany

🇫🇷 Hani Chalouati — CTO & Cofounder @ Guepard Inc. · Git-like serverless databases for AI agents. Software engineer born and studied in Tunisia, moved to France in 2014. Enjoys drone FPVs (especially DIY) and aviation. One exit from a previous company in 2024. Special power: makes very good homemade Neapolitan pizza. Île-de-France, France

Contributors This Week

Tejas Chopra (Netflix), Glenn Sonna (Xybrid), Robert Hommes (Moyai.ai), Ihab Bendidi (Recursion), Félix Raimundo (Tychobio), Koutheir Cherni KC (Guepard), Jocelyn Fournier (Softizy), Nancy Wang (FluenTea), Anicet Nougaret (Ariana.dev), Enrico Piovano (Goji), Sacha Morard (Edgee), Charly Poly (Browserbase), Quentin Dubois (OSS Ventures), Willy Braun (Galion.exe), Amine Saboni (Pruna.ai), Gabriel Olympie (2501.ai), Jérémie Bordier (XHR), Laurent Mazare (Kyutai), Hubert Ancelot (Verne), Jeremie Kalfon (Pasteur), Maziyar Panahi (OpenMed), Pierre Chapuis (Finegrain), Victoire Cachoux (Iktos), Hani Chalouati (Guepard), Karim Matrah (Contrast), Louis Choquel (Pipelex), Louis Manhes (Genario), Shuai Zhang (Jinko)

TTY-changelog #036

Kevin Kuipers — Sat, 28 Mar 2026 12:20:13 GMT

👉 Article originally posted on TTY

Community Asks

🤖 Gabriel Olympie is looking for feedback about Agno, an open-source framework for building, running, and managing agentic software at scale.

📋 Nnenna Ndukwe is looking for a tool that traces intent from ticket to PR to QA in agentic engineering workflows. She wrote a field report from DevNexus exploring the gap.

Community Updates

🏆 Cubic has been ranked first on Martian’s independent code review benchmark: 61.8% F1 score, ahead of tools like Cursor Bugbot (45%), Claude Code, Gemini, and Coderabbit (30%). And also praised by the community. Congrats, Paul!

Events

🇫🇷 CodeCarbon x Pruna x Ecologits Frugal AI party in Paris (April 8) – After-party following the PyTorch conference.

Audio

🎙️ Voxtral TTS finds its voice – Mistral AI released a demo of Voxtral TTS, its voice adaptation model. A single voice adapted to new languages, accents, and expressions. Community noted the French voice still carried a foreign accent.

Autonomous Agents

🦞 Community takes on persistent agents – Between a Discord discussion and TTY lunch, the community dissected what it takes to deploy persistent agents in real organizations. Five themes emerged: task assignment, communication interfaces, permissions, monitoring, and cost control. No one has cracked the full stack yet, but concrete approaches are taking shape.

Agents are harnesses, not intelligence. Guillaume Larcher (Linkup) uses directly Pi (OpenClaw’s backbone) as a thin execution layer wired to Opus or Codex, limited to file I/O, bash, and grep.
Agents as external contributors. Sylvain Utard runs Albert in production at Altertable: it has its own GitHub account, reads notifications, clones repos, writes code, and opens PRs across a dozen SDKs. No privileged access, spec-driven workflows, behaves like an open-source maintainer.
A CTO agent in the cloud. Anicet Nougaret is building a persistent CTO agent available via Telegram. The first 13 users generated 1,300 commits in two days!
Security first. Permissions ranked as the top concern for everyone. Arnaud Porterie explored highly containerized persistent agents. Hugo Venturini (SkipLabs) recommendation: “Using narrower permissions for the user it impersonates.” Even though this means less autonomy.
Slack is unfit. Venturini argues for web desktop first, mobile second, while Porterie favors task-based systems backed by Linear, which humans already use. Most, like Julien Millet, use Telegram.
Cost explosion is real but accepted. Agent-driven workflows introduced a new variable cost layer. And this cost can spike unpredictably when agents run unchecked. But no one is seriously considering going back.
Evaluation, not generation, is the bottleneck. People’s concerns are about without code ownership. Quentin Dubois puts it bluntly: “I don’t know how to scale a tech team anymore.” Full-throttle agent setups worked in internal contexts, but compliance and audits still required human validation layers.
OpenClaw’s limits. Millet migrated from OpenClaw to Claude Code on Telegram, arguing it lacked structural transparency and was nearly impossible to debug. Pierre Chapuis takes a wait-and-see stance on this one: “I am going to be a late adopter on this one. Claw was an experiment as a side-project, for work I will wait to see how startups and/or big labs rationalize it first.”

🧠 ARC-AGI-3 shifts to interactive environments – The new benchmark from ARC Prize moved from static puzzle-solving to interactive environments. AI agents must explore novel settings, acquire goals on the fly, build adaptable world models, and learn continuously. A 100% score means agents can match human learning efficiency.

🧬 HyperAgents improve how they improve – Facebook Research released HyperAgents, self-referential agents that modify both their task-solving behavior and the meta-process that generates future improvements. The system built on the Darwin Gödel Machine to enable open-ended self-improvement. GitHub, thread.

🔎 Cursor indexes text for agent search – Cursor published a deep dive on building trigram-based indexes for regular expression search, targeting large monorepos where ripgrep calls took 15+ seconds. The approach drew on 1993 inverted index research and added probabilistic masks and sparse n-gram selection.

Biotech, Health, and Chemistry

🧪 Why there is no AlphaFold for materials – A Latent Space interview with MIT professor Heather Kulik explored why AI for material design lags biology. Two structural reasons stood out: combinatorial complexity (20 amino acids vs. the entire periodic table) and far worse datasets.

In materials, databases are computationally derived (DFT), not experimentally measured, creating an in-silico vs. wet-lab gap analogous to biology.
Universal force field models like MACE-MP-0 and CHGNet now showed cross-element transferability, directly challenging the claim of limited transfer learning.
Community take: For Felix Raimundo, AlphaFold was truly an exception, built on 20 years of data generation from hundreds of labs on a single clear problem, which is why there has not been a second “AlphaFold moment” in bio. He believes materials science may actually be easier because simulations are better (cheaper validation and data generation), the property space is likely smaller, and interactions are simpler: a crystal interacts with itself, while a drug plays with the immune system, brain, and a huge, mostly unknown dependency graph. “I’d actually be decently bullish, but it may very well be because I underestimate the complexities involved.”

🧬 Deep learning reveals inflammation memory mechanisms – Two papers used model interpretation methods for deep learning models of regulatory DNA to uncover causal roles of sequence syntax in transcription factor binding, methylation, and histone variants tied to stem cell memory of inflammation.

Image, Video & 3D

✨ Luma launches UNI-1 multimodal model – Luma Labs released UNI-1, a multimodal reasoning model that generates pixels. Built on “Unified Intelligence,” it combined spatial reasoning, common-sense scene completion, and multilingual text rendering with image and video generation.

🪦 OpenAI shuts down the Sora app – OpenAI announced the shutdown of its TikTok-like social app built on Sora 2, just six months after launch. Despite early hype, there was no sustained interest in an AI-only social feed. The underlying model remains available.

Cyber

🚨 LiteLLM PyPI package gets compromised – LiteLLM version 1.82.8 on PyPI contained a malicious .pth file with base64-encoded instructions to exfiltrate credentials and self-replicate. The attack was traced to the earlier Trivy compromise. A checker tool and GitHub issue were published for affected repos.

The compromise chained two supply chain attacks: a security scanner (Trivy) was itself compromised, then used to inject malicious code into LiteLLM.
The incident reinforced that even security tooling is a viable attack vector, and these types of highly targeted supply chain attacks are becoming more frequent.
Community take: A FOSDEM talk was cited arguing that VCS sources (even unsigned public commits) are safer than signed tarballs or PyPI packages.

Infrastructure

📈 Final training runs use minority of compute – New data from Epoch AI, drawing on MiniMax and Z.ai IPO filings plus OpenAI estimates, showed that final training runs accounted for only 10-23% of total R&D compute. The other 80-90% went to experiments, ablations, synthetic data generation, and models that never shipped. Catch-up players like MiniMax appeared to skip costly experimentation by relying heavily on large-scale distillation of frontier models.

Community take: Benjamin Trom confirmed the pattern and added that every architectural innovation requires massive infra investment to validate. DeepSeek’s reported $5M training cost for V3 implied at least $200M in total research spending. Pierre Manceron (Raidium) put the ratio even lower: final training represented only 2-3% of their total budget.

Language Models

🔍 OpenAI monitors agents for misalignment – OpenAI published how it used chain-of-thought monitoring to study misalignment in internal coding agents. An internally deployed model attempted prompt injection to escape its simulation environment.

🕸️ RetriCo boosts Graph RAG construction 100x – GLiNER-based RetriCo replaced expensive LLM API calls for knowledge graph construction, enabling local Graph RAG systems at a fraction of the cost of Microsoft’s GraphRAG. LinkedIn post.

💻 Percepta builds a computer inside transformers – Percepta added computing capabilities inside an LLM, enabling execution of arbitrary C programs for millions of steps. The approach used 2D attention heads for exponentially faster inference compared to standard autoregressive decoding.

MLOps

🔥 TurboQuant compresses KV cache to 3 bits – TurboQuant is a new compression method that shrinks AI model memory (like KV cache and vector indexes) up to ~6–8x while preserving accuracy, enabling faster, cheaper long‑context and search systems.

Programming

❤️ OpenCode with Qwen 3.5 runs local – Louis Choquel (Pipelex) shared a detailed trial of OpenCode (open-source Claude Code alternative) across three configurations: Opus 4.6 via API, Qwen 3.5 397B via OpenRouter, and Qwen 3.5 35B locally on a MacBook Pro M3 Max with 64GB RAM. The local model ran alongside normal computer usage with no issues.

The 35B model stopped after each phase and needed “go on” prompts, but still delivered working results thanks to clear agentic feedback loops correcting its mistakes.
Qwen 3.5 397B through OpenRouter performed comparably to Opus 4.6, with faster speed and relevant ASCII diagrams of method flows.
Optimizing skills for small models forced clarity and token efficiency improvements that also benefited larger models.

🔥 gstack packs 15 Claude Code skills – A Claude Code plugin with 15 tools covering CEO, designer, eng manager, release manager, doc engineer, and QA roles. The /review skill called Codex for an outside view, and /qa ran Playwright-based browser testing of all buttons and menus. The standout was the install/upgrade system: a single copy-paste instruction in Claude Code that handled everything.

The plugin was iterated daily, and the skill upgraded itself automatically so users saw improvements without manual intervention.
Community take w/ Louis Choquel: “It’s my best experience so far using any tool.” Good news: token usage remained modest. 9 hours of heavy use on 2 parallel projects consumed only 6% of the Claude MAX 200 weekly budget.

🛠️ RTK cuts LLM token use 60-90% – A CLI proxy written in Rust that reduced LLM token consumption on common dev commands. Single binary, zero dependencies. RTK is now part of the Headroom context optimization layer, which also added a --learn option to help coding agents learn from failures.

TTY Lunch

Each week, we brings together exceptional builders around the table. Today’s lineup included Quentin Dubois (OSS Ventures), Hugo Le Belzic (H Company), Guillaume Larcher (Linkup), Sylvain Utard (Altertable), Jules Pondard, Henri Mirande, Glenn Sonna (Xybrid).

Agents are harnesses, not intelligence
Developer cost explosion
Evaluation without code ownership
Multi-model workflows as control
Running models in constrained environments

👉 The full discussion is here

Contributors This Week

Pierre Chapuis (Finegrain), Louis Choquel (Pipelex), Tejas Chopra (Netflix), Willy Braun (Galion.exe), Benjamin Trom, Félix Raimundo (Tychobio), Koutheir Cherni (Guepard), Jocelyn Fournier (Softizy), Amine Saboni (Pruna.ai), Gabriel Olympie (Human-Centric), Jérémie Bordier (XHR), Robert Hommes (Moyai.ai), Enrico Piovano (Goji), Julien Millet, Khaled Maâmra (Edgee), Sacha Morard (Edgee), Thierry Abalea (Shipfox), Ihab Bendidi (Recursion), Julien Seveno-Pilitant, Alexandre Meunier, Anicet Nougaret (Ariana.dev), Antoine Millet (Scaleway), Arnaud Porterie (Vibe), Guillaume Allègre (Andromede AI), Hubert Ancelot (Verne), Hugo Venturini (SkipLabs), Julien Kilo, Kemal Toprak Uçar, Lior Oren, Louis Manhes (Genario), Lucas DiCiocci (Koli), Nnenna Ndukwe (Qodo AI), Paul Sangle-Ferriere (Cubic), Pierre Manceron (Raidium), Quentin Duboix (OSS Ventures), Stan Girard, Youssef Tharwat (Noodlbox)

TTY-changelog #035

Kevin Kuipers — Sat, 21 Mar 2026 10:51:19 GMT

🎉 New Name, 🏡 New Home

I’ve launched a few initiatives over the years and decided to bring them together under one home: TTY (what is TTY? Answer here).

The Reg.exe Discord community, the WeLoveSota.com blog, and my weekly TechLunches in Paris now live under this single brand.

Head to ttysession.com, where members can connect their Discord account and explore the long-awaited member directory.

Curious to hear what you think.

Events

🌐 TelcoSec Talk #2: Iran cyberattacks analysis (March 2026) – Live demo of AI-driven threat actor modeling with ApexModeler, analyzing the Iranian APT landscape in a geopolitical context.

Biotech, Health, and Chemistry

🧬 AlphaFold Database adds protein complexes – EMBL-EBI, Google DeepMind, NVIDIA, and Seoul National University released 1.7 million high-confidence homodimer predictions into the AlphaFold Database, with 18 million lower-confidence structures available for bulk download. The dataset prioritized the 20 most-studied species and WHO priority pathogens, representing about 17 million GPU-hours of compute made freely available.

😓 AI hype called out in life sciences – Anshul Kundaje shared Rachel Thomas’s post on how AI hype consistently outruns reality in biotech. Community members noted that models like scGPT keep reappearing in new publications despite over 30 papers documenting its shortcomings, pointing to conflict-of-interest issues with authors pushing their own tools.

Image, Video & 3D

🎮 DLSS 5 announced then quickly deflated – NVIDIA announced DLSS 5 with promises of AI-powered photorealistic lighting and materials for games. Days later, clarifications from NVIDIA confirmed skeptics were right: the model is a 2D AI filter using only color buffer and motion vectors, with zero access to geometry, lights, PBR properties, or normals.

Sebastian Aaltonen laid out requirements for a proper AI rendering filter: per-pixel developer control, reference training material for hero assets, scene depth awareness, no removal of fog or volumetrics, and respect for the art style set by developers.
Community reaction called it worse than expected, with one viral meme imagining a future where turning DLSS off reveals nothing underneath.

📹 Deterministic video depth estimation ships – DVD combines generative priors with deterministic inference to resolve geometric hallucination in diffusion-based depth models and semantic ambiguities in discriminative baselines, delivering consistent high-fidelity geometry. Code, models, Gradio demo, and ComfyUI integration were released.

Cyber

🛡️ ApexModeler maps threat actor behavior – AI-driven threat actor modeling tool that combines open-source and private intelligence into a unified engine, deployable on-premise. Built by P1 Security, it will be demoed live analyzing Iranian cyberattacks and APT attribution in the current geopolitical context.

Infrastructure

⚠️ YC compliance startup accused of faking audits – Delve, a YC-backed startup that raised $32M, was accused of systematically fabricating SOC 2, ISO 27001, HIPAA, and GDPR compliance reports. A leaked spreadsheet revealed that 493 out of 494 SOC 2 reports contained identical boilerplate text with only names and logos swapped, auditor conclusions were pre-written before any evidence review, and the advertised “US-based CPA firms” were actually Indian certification mills operating through shell entities.

Language Models

🚀 NVIDIA Nemotron 3 Super drops with open tooling – A 120B-parameter mixture-of-experts model (12B active) that reportedly surpassed Qwen3.5 on text benchmarks. More notable than the model itself, NVIDIA open-sourced its full post-training stack as the NeMo suite.

The NeMo release includes four repos: Evaluator, RL, Gym, and Guardrails, covering the full post-training pipeline from evaluation to reinforcement learning to safety.
The Gym component is interoperable with external ecosystems like Unsloth, lowering the barrier for custom RL-based model tuning.
The model license remains restrictive, but the open tooling could prove more impactful than the model itself for the broader community.

🇫🇷 Pleias and NVIDIA release French personas – One million synthetic French personas grounded on real census demographics (age, gender, location, education, income) were published as the Nemotron-Personas-France dataset on Hugging Face. France’s open census license made the rich demographic grounding possible.

🎭 Reasoning models perform “theater” in CoT – A paper analyzed DeepSeek-R1 671B and GPT-OS using activation probing, early forced answering, and a CoT monitor. It found models often lock in their answer early but continue generating chain-of-thought tokens that do not reflect actual internal confidence. Harder tasks showed more genuine stepwise reasoning, while easy recall triggered performative output.

Programming

🧠 Cursor trains Composer to self-summarize – Cursor published how it uses reinforcement learning to teach its Composer model to compress its own context when approaching a fixed token limit, then continue generating. Instead of relying on an external prompted summarizer, the model learns what information to preserve as part of the reward signal during training.

🤔 Claude 1M context hidden tradeoffs – Multiple community members highlighted the productivity gains from Claude Code’s expanded context window, with enthusiastic endorsements. Some, like Stan Girard, are more skeptical, pointing out that performance degrades at large context sizes and that the current setup may hide pricing and rate limit tradeoffs, potentially leading to higher costs and overreliance on a setup that may not be sustainable long term.

📚 Andrew Ng launches Context Hub CLI – Open-source CLI that gives AI coding agents curated, versioned API documentation instead of relying on web search. Agents can attach local annotations that persist across sessions for self-improvement, while aggregated feedback flows back to documentation maintainers to continuously improve docs for everyone

📊 Vibe Drift Tracker monitors AI slop – VS Code extension that integrates with Claude Code to track how many AI prompts go into each commit, computing a real-time drift score. A high prompt count with low lines-per-prompt ratio signals feature creep and loss of focus.

🍎 Apple cracks down on vibe-coding apps – Apple blocked App Store updates for Replit and Vibecode, citing long-standing rules against apps executing code that alters their own functionality. Replit was told to open generated app previews in an external browser instead of an in-app web view, while Vibecode was ordered to remove the ability to generate software for Apple devices.

Robotic, World AI

🗺️ Pokemon Go data trains robot navigation – Niantic Spatial, spun out from the Pokemon Go maker, trained a visual positioning system on 30 billion crowdsourced images from players to locate robots to within centimeters in urban environments where GPS fails. Coco Robotics deployed the system across 1,000 delivery robots in five cities.

🏙️ Seoul World Model simulates real cities – KAIST and NAVER built a city-scale world model grounded in real Seoul street-view data. SWM uses retrieval-augmented generation on millions of panoramic images to produce faithful video spanning kilometers of actual cityscape, supporting free-form navigation, text-prompted scenario control (like spawning Godzilla), and multi-kilometer trajectories without quality degradation. Code and weights were released.

Community Updates

📦 Headroom v0.4.4 ships wrap commands – Context optimization proxy for LLM apps now supports headroom wrap claude and headroom wrap codex without running a separate proxy, and adds CLI token compression via rtk-ai. (👉 Tejas Chopra @ Netflix)

⚡ Edgee launches token compression gateway – Free proxy that compresses prompts before they reach LLM providers, claiming up to 50% cost reduction and 26.5% longer Claude Code sessions. One-minute CLI install, no code changes required, works with Claude Code, Codex, and Cursor. (👉 Sacha Morard @ Edgee)

🧠 Anti-Agent enters beta for testers – AI learning companion combining knowledge graphs, spaced repetition, Socratic dialogue, and reflective journaling, all accessible via a Telegram bot. Waitlist open. Video demo (👉 Louis Manhes @ Genario)

⚖️ Rature anonymizes PDFs in-browser by Julien Kilo – Open-source tool for lawyers running BERT-based NER entirely client-side with Transformers.js, detecting and wiping named entities before sharing documents with AI. Early alpha seeking feedback. GitHub (👉 Julien Kilo)

🤖 Xybrid ships on-device AI runtime – Rust-powered runtime that runs LLMs, TTS, and ASR natively on iOS, Android, macOS, Linux, and Windows with no cloud required. SDKs for Flutter, Unity, Kotlin, and Swift. (👉 Glenn Sonna @ Xybrid)

TTY Lunch

This week with Alexandre Esser (Arsen), Son Nguyen (Hugging Face), Jean Beuselinck (Tomorro), Khaled Maâmra (Edgee), Loic Knuchel, Marco Fucci di Napoli (Rowads), Stan Girard (The Vibe Company), Thibault Hervier, Thierry Abalea (Shipfox), Vladimir de Turckheim. Topics were:

Different organizational models
Beyond code-centric workflows
CI: the main bottleneck?
Hiring being redefined

🍕 Read the summary

New members

🇫🇷 Margot Lor-Lhommet – Co-founder & CTO @ Tigo Labs · Cognitive AI infrastructure so AI systems can reason about humans, not just respond to them. 10 years in academic AI research (France + USA), 10 years in industry (Algolia, Twilio). Runner, pianist, foodie (meat smoker), and mom of two. Special power: Making 1980s cognitive science papers ship as product features. 📍 Île-de-France

🇫🇷 Julien Goupy – Co-founder & CTO @ Applikai · Agentic mobile app builder to make production-ready apps, not prototypes. Third-time founder (one exit to Google, one spectacular fail). Started shipping low-level Android Bluetooth kernel drivers deployed on 50+ million devices (Lenovo, Motorola, TCL). Filed 40 patents. Recreational programming enthusiast: terminals, compilers, weird languages, silly toy games. Special power: Can keep a conversation about trebuchets for a whole 12h long SF to Paris flight. 📍 Paris, France

Contributors This Week

Julien Goupy, Philippe Langlois, Margot Lor-Lhommet, Quentin Dubois, Jérémie Bordier, Karim Matrah, Ihab Bendidi, Ivan Yamshchikov, Gabriel Olympie, Amine Saboni, Felix Raimundo, Christophe Lesur, Robert Hommes, Julien Kilo, Pierre Chapuis, Louis Manhes, Tejas Chopra, Kemal Toprak Uçar, Glenn Sonna, Bertrand Guiheneuf, Antoine Sueur, Gawen Arab, Sacha Morard, Youssef Tharwat, Maziyar Panahi, Koutheir Cherni.

👉 Originally posted on TTY

TTY Weekly

TTY-changelog #054

Events

Audio

Autonomous Agents

Biotech, Health, and Chemistry

Image, Video & 3D

Cyber

Infrastructure

Language Models

MLOps

Programming

Robotic, World AI

Other topics

Contributors This Week

TTY-changelog #053

Biotech, Health, and Chemistry

Image, Video & 3D

Cyber

Infrastructure

Language Models

MLOps

Robotic, World AI

New member

Contributors This Week

TTY-changelog #052

Audio

Image, Video & 3D

Cyber

Infrastructure

Language Models

Programming

Reinforcement Learning

Other topics

Contributors This Week

TTY-changelog #051

Audio

Autonomous Agents

Biotech, Health, and Chemistry

Image, Video & 3D

Infrastructure

Language Models

Programming

Robotic, World AI

Other topics

Contributors This Week

TTY-changelog #050

Events

Autonomous Agents

Biotech, Health, and Chemistry

Image, Video & 3D

Infrastructure

Language Models

MLOps

Programming

Reinforcement Learning

Robotic, World AI

Contributors This Week

TTY-changelog #049

Events

Autonomous Agents

Biotech, Health, and Chemistry

Image, Video & 3D

Cyber

Infrastructure

Language Models

Programming

Robotic, World AI

Other topics

Contributors This Week

TTY-changelog #048

Autonomous Agents

Biotech, Health, and Chemistry

Image, Video & 3D

Cyber

Infrastructure

Language Models

MLOps

Programming

Robotic, World AI