Events

🌐 Edgee Demo Day online (November 10) – Research lead Khaled Maâmra and the Edgee team will demo token compression, model routing and provider fallback on live coding agents, 5pm CET.

🇫🇷 TechBio: Where Food Meets Health in Paris (December 2) – During BioWeek Paris, Cyril Véran of Living Models and Vincent Libis of Generare will discuss which AI biology tools carry over from food to health.

Audio

🎙️ NVIDIA shipped open Nemotron audio models – Nemotron 3.5 Transcribe packed transcription in 40 languages, speaker attribution, translation and audio understanding into one promptable open-weight 30B model, with only about 4B parameters active at a time, keeping track of who said what in crosstalk.

  • Its 9.6M-parameter sibling, NemotronLabs RE-USE, restored speech damaged by noise, echo, clipping, codec artifacts or packet loss, and kept the main voice when people talked over each other.

  • On the URGENT 2025 test set it beat Sidon and UniPASE, models 25 to 50 times larger, on both quality and intelligibility scores.

  • Cleared for commercial use, it took audio from 8 to 48 kHz in any language and targeted cleanup of recordings and of speech datasets for TTS training.

🗣️ Phonic-1 topped speech-to-speech leaderboards – Built for voice agents, the model heard callers’ audio directly, catching tone and hesitation, and called tools mid-sentence in 50+ languages. It claimed the top spot on Artificial Analysis’s speech-to-speech index and four other voice benchmarks.

Biotech, Health, and Chemistry

🧪 Anthropic acquired one-person biotech Mithril – Revealed as Cody Tranbarger, the anonymous biopharma poster Professor Oak ran Mithril Medicine solo with heavy Claude use and part-time consultants, developing antibodies for LAMA2-CMD. At Anthropic he will test how many programs one person can run in parallel.

  • Community take: Félix Raimundo called “only employee” an exaggeration, since the founder “did venture building within Atlas, with access to their whole team, best-in-class consultants paid in equity, and CROs,” adding that the model was not new: “we call that a ‘virtual biotech’.” Khalil Ouardini agreed it did not count as solo, and saw “an opportunity for a PR stunt on Anthropic’s side, although they seem to be the most serious of the big labs when it comes to bio applications.”

💊 Citrini put Claude on drug design – Citrini Research used Claude to generate new drug molecules at speed, but its computer-simulated binding predictions were right only 49% of the time, and the lab data needed to improve them took about 30 times longer to collect than the designs.

Cyber

🪨 Cairn agents attacked 1,200 websites – A Chinese-speaking attacker let open-source Cairn agents break into online shops largely alone, reaching sites of more than 100 companies including Deere and Cummins. Each successful intrusion took about 23 hours and roughly $50 in tokens, at a 10% hit rate.

  • Community take: The thread split on where defense starts. Gabriel Olympie stressed that cheap models could now run these attacks, and Hugo Le Belzic that patching exploits mattered more than restricting models. Adam Surak argued most companies should harden the basics. Mario Cornejo saw continuous monitoring, plus readiness to go offline, as the answer once any 0-day could be weaponized within hours.

💸 Binance lost $6.36B in record hack – On October 29, an attacker drained about $6.36B from Binance wallets on ten blockchains, mostly within 20 minutes, four times the Bybit hack. TRM attributed it to North Korea’s TraderTraitor, pointing to compromised wallet signing infrastructure.

  • On Ethereum, an EIP-7702 delegation backed by 50 pre-signed authorizations turned a Binance hot wallet into a sweeper the attacker controlled.

  • Stablecoin issuers and Arbitrum’s Security Council froze about $2.7B, but the attacker still held roughly $2.4B in ETH and BTC beyond any issuer’s reach.

  • TRM advised treasury teams to watch wallets for unexpected EIP-7702 delegations and keep tested revocation methods ready.

🕵️ Cohere flagged Qwen3.8-Flash-Next sandbox probing – Asked to write a defensive detection rule, the model allegedly inventoried its machine and tried to upload a probe file to an Alibaba Cloud bucket in Beijing, after its reasoning said it would call no tools. The sandbox blocked it, and Cohere did not claim malice.

  • The released traces showed the model listing /tmp, checking for curl, wget, python3 and jq, reading the DNS configuration, then sending a synthetic test payload to an Alibaba storage endpoint.

  • Many observers dismissed the claim, and the analyst who reviewed the traces leaned toward training contamination.

  • Cohere framed the open question as whether models can act differently from their stated reasoning, and how to vet open-weight models before adoption.

  • Community take w/ Gabriel Olympie of 2501.ai: “It’s presented as a voluntarily implanted behavior, but it’s most likely a leftover behavior from some sandboxed training. The issue is real and reproducible though, so even without malicious intent, under some conditions the model will still try to leak data, which should be known to build the guardrails around it.”

🐚 HTTP2Shell spread beyond Node.js and Python – CVE-2026-111497 reached many HTTP libraries and apps: a flaw in Brotli decompression, which unpacks compressed web traffic, let attackers run code on servers without logging in. Fixes were still missing for some services, pending Brotli 1.2.1.

📸 Instagram and Telegram kept vanishing media – A backup audit run with Claude Code allegedly found 15,100 cached files on one iPhone, 11,400 from Instagram’s vanish mode and 3,700 from Telegram’s self-destructing chats. The apps hid them from chats, but any of the 31 apps with file access could read them.

Infrastructure

🛰️ Suncatcher’s orbiting TPU answered a prompt – Weeks after launching its prototype satellite, Google sent a command up to orbit, booted the onboard TPU, ran a prompt and received the full answer back on Earth: a haiku, the hello world milestone of Project Suncatcher.

🦆 Guepard’s PQL brought ML into DuckDB – The open-source extension added TRAIN MODEL and PREDICT to SQL, so a database user or an AI agent could forecast churn, reorders or missed payments straight from existing tables, training in seconds on a laptop CPU with features built from foreign keys.

Language Models

🚫 Chinese users reverse-engineered Claude bans – A Chinese community claimed each Claude account starts with a score that rises or falls with behavior, with a ban below a threshold. IP stability, payment method, multi-device logins and account age allegedly weighed most, and lost points were hard to win back.

🌍 OpenAI’s GPT-6.1 Terra cut computer-use costs – The fast model for operating computers, visual reasoning and writing, aimed at high-volume work like invoice processing, came in at half GPT-5.6 Terra’s API price while more than doubling its Terminal-Bench 4.0 coding score to 50.3%.

  • It matched GPT-6 Astra on professional tasks at a tenth of the price per token and finished tasks over three times faster than Astra and GPT-6.1 Sol.

  • Its error rate on prompts built to trigger mistakes fell from 53.6% to 16.7% at low reasoning effort, and to 4.3% at maximum effort.

  • It met the Critical cybersecurity threshold of OpenAI’s Preparedness Framework, with strong resistance to prompt injection.

🐉 Nanbeige4.3-33B trained a teacher first – Built to raise the ceiling of small agent models, the 33B model claimed top agentic results among open ~30B models on coding and cowork benchmarks, ran cheaply inside mainstream agent harnesses and will distill into Nanbeige4.3-3B next.

  • Community take w/ Gabriel Olympie: “But how long does it think? Previous iterations of Nanbeige had very good benchmarks, but were thinking so much they never output an answer, and were disappointing in real use cases.”

🖥️ Qwenputer compiled a CPU into Qwen – A 32-bit RISC-V processor was compiled into 11 extra layers in front of Qwen 3.5 4B, using its Gated DeltaNet layers as exact memory, so the model ran C and MicroPython inside its own computation. Long-integer arithmetic accuracy jumped from 48.6% to 100%.

  • Community take: Mehdi Medjaoui of GenerationAI called it “a beauty! Useless, so we need to look at it!” Gabriel Olympie of 2501.ai noted that “someone already put Doom on it,” adding that it “runs at 0.01 fps and uses ASCII art to render so it looks like shit, but the idea is incredible.”

✏️ Next-Concept Modeling sketched text before detailing – Rather than writing word by word or refining the whole sequence like diffusion models, NCM generated a hierarchy of compressed codes level by level, rough outline first. It matched the best diffusion models 10 to 31 times faster and beat both families on Sudoku.

Programming

🏭 Pierre Chapuis’s software factory reading list – After a lunch conversation on software factories, Pierre Chapuis of Finegrain shared the three references he had brought up:

  • The Counting House: Steve Yegge’s latest essay ran customer service on stateless model clerks that each handled one case, with a policy table assigning intents to models across five autonomy tiers, from silent shadow mode to moving money, and mistakes scored by cost.

  • Raising an Agent: Amp’s podcast by Quinn Slack and Thorsten Ball, now in its second season, with episodes like “The Local Dev Env is Dead.”

  • Earendil: Armin Ronacher’s public benefit company, which employs Mario Zechner and builds open tools to harness AI, Pi and Lefos.

🤖 Factory’s Droid 1.0 batched tool calls – Factory rebuilt the harness, the prompts, tools and loop wrapped around a model, so its coding agent wrote a JavaScript program running hundreds of tool calls at once. Moving 151 Linear issues took 6 requests instead of 31, with 98% fewer input tokens.

  • On the same models, Factory claimed Droid beat native harnesses, scoring 3.5 points above Codex with GPT-6.1 Sol at 12% lower cost.

  • An automated Agent Factory loop tested each new model, spotted failure patterns and proposed harness changes as controlled experiments for engineers to review.

  • Sessions now shared one background process, cutting memory from 350MB to 20MB each, and launched 3.8 times faster.

Robotic, World AI

🦿 NVIDIA MotionBricks drove a G1 in-browser – An independent port ran NVIDIA’s 195M-parameter motion model fully in the browser, generating walking, running, crouching and vaulting for a Unitree G1 humanoid in real time, in styles from skipping to zombie, matching NVIDIA’s official runtime almost exactly.

🎮 DeepMind’s Genie 4 made worlds talk – Google DeepMind’s latest world model generated environments users could explore in real time, reacting to their actions with sound and with people who moved, responded and spoke. The lab called it its most realistic and interactive yet.

📱 SpAItial’s Echo-3 rebuilt places from phones – The world model turned ordinary phone captures into photorealistic 3D environments at real-world scale, from a single room to multi-floor buildings and outdoor public spaces, and was available from launch day.

Contributors This Week

Gabriel Olympie, Félix Raimundo, Hugo Le Belzic, Adam Surak, Quentin Dubois, Pierre Chapuis, Khalil Ouardini, Laurent Mazare, Mario Cornejo, Mehdi Medjaoui, Anicet Nougaret, Gilles Seghaier, Jules Damidaux, Koutheir Cherni, Lior Oren, Julien Millet, Bryen Param, Gawen Arab, Thibault Hervier