TTY-changelog #035
Cursor trains Composer to self-summarize past context limits, NVIDIA drops Nemotron 3 with a full open post-training stack, and DLSS5 turns out to be a plain 2D AI filter with no geometry awareness.
🎉 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
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