Mistral releases Le Chonk weights for coding and cybersecurity

#1 𝕏

Mistral AI released the weights for its improved Le Chonk model for agentic coding and cybersecurity, available on Hugging Face and customizable with Mistral Forge. The company says it ranks #1 among open models on Cognition’s FrontierCode, sets a new Pareto-efficiency benchmark on Artificial Analysis, and was evaluated by CrowdStrike, Cloudflare, Rubrik, and Zscaler.

Also covered by: @Hugging Face, @NVIDIA AI, @Mistral AI, @Julien Chaumond

#2 𝕏

Demis Hassabis shared Google DeepMind’s description of Gemini 4 Carbon as its next frontier model, which the organization claims is state-of-the-art across cybersecurity and long-horizon capabilities.

Also covered by: @Cognition, @Demis Hassabis

#3 𝕏

SAIRfoundation described Muse Spark Open, attributed to AI at Meta, as new on Open Math OS, with 816B total parameters, 42B active parameters, and a 1M context window.

Also covered by: @DeepLearning.AI

#4 📝 OpenAI News

How we detect frontier model misalignment - OpenAI shared “How we detect frontier model misalignment” on October 29, 2026, explaining that increasingly capable AI models require reliable ways to detect unintended behavior, including agents taking unauthorized actions.

Also covered by: @OpenAI

#5 𝕏

Anthropic shared Claude’s self-assessment safety technique, which flags potential misuse by assessing conversations immediately after responding. Used since June of an unspecified year, it has surfaced misuse other safeguards weren’t designed to catch and draws on Claude’s judgment rather than a discrete set of classifier targets.

#6 📝 OpenAI News

An updated Model Spec for more capable and beneficial AI - OpenAI announced its updated Model Spec on October 29, 2026, as models become smarter and take more consequential actions in the world. The article links to the updated Spec.

#7 𝕏

Cognition announced that it is acquiring @corridor, framing securing the world’s software as the most important problem in technology today.

#8 𝕏

Aravind Srinivas announced that @jakubkrehel’s /teardown is now available as a skill in Perplexity Computer, allowing users to break apart objects and inspect them interactively in 3D. Each user receives 5000 trial credits to experiment with the skill.

#9 𝕏

NVIDIA AI demonstrated Nemotron Retriever tackling the same questions as another open-source retrieval stack, comparing retrieved evidence, accuracy, and token usage. Open models and recipes are available on Hugging Face.

#10 𝕏

Anthropic announced it expanded its AI for Science program beyond individual teams to support scientific foundations, institutions, and research groups.

#11 in

Claire Vo recapped Blair Metcalf’s work at Reddit helping build Reddit Proto, a prototyping harness based on Reddit’s design system, tokens, components, and UX patterns. It lets designers and product managers vibe code while staying within established constraints.

#12 in

Every Inc. recapped this week’s work from Every’s Frontier team: Mike Taylor rewrote one headline, after which simulated agent citations rose from 2 out of 240 sessions to 240 out of 240, and found that one agent waited 30 seconds for a slow API in all 20 trials while most agents never tried it. He also had three models take turns improving one writing skill, raising its win rate against a comparison skill from 18% to 100% in an initial four-round run, though a later solo-model control matched the tuning score.

#13 𝕏

Boris Cherny announced that `CLAUDE_CODE_SUBAGENT_EFFORT` had “just landed”; it can be set to `low` when invoking `claude` or customized in a custom agent configuration.

#14 𝕏

Harrison Chase announced that MDA 0.10 includes async subagents and browser use, two examples of how closer harness and infrastructure integration can address harness-engineering challenges.

#15 in

Udi Menkes recapped a Perplexity internal finance test in which an agent using ls, grep and read beat semantic index search on finding the correct document (87% vs. about 39%) and answer accuracy (63% vs. about 37%), while running about 25% faster. He recommends testing this baseline in evals before building a vector index, while noting the results came from a single company-reported test.

#16 🐙 CopilotKit/OpenBot

CopilotKit/OpenBot - CopilotKit/OpenBot is an MIT-licensed, self-hosted beta template for AI coworkers, with an app, AG-UI agent integrations, and coworker examples. It has 8,447 GitHub stars and 707 daily trending gains. Worth trying if you need a customizable agent workspace on your infrastructure and can supply Docker, Bun 1.4+, CopilotKit Intelligence, and a model key.

#17 in

Guillermo Rauch shared an AI Gateway feature that works “like DNS for models,” using an alias that resolves to a model and enables traffic steering. He says it can keep teams from getting stuck on old model versions, save costs, and reduce decision paralysis.

#18 🐙 lightpanda-io/browser

lightpanda-io/browser - lightpanda-io/browser shares Lightpanda Browser, an AGPL-3.0 headless browser built in Zig for AI agents and automation, with source code, agent mode, and replayable PandaScript. It has 37,856 GitHub stars and a 559 daily trending gain. Worth trying if you need browser automation with CDP or WebDriver BiDi support.

#19 🐙 HKUDS/DeepTutor

HKUDS/DeepTutor - HKUDS/DeepTutor is HKUDS’s Apache-2.0 lifelong personalized tutoring project, with core, CLI, web, and Docker components plus notebooks, narrated video lessons, and resumable embedding checkpoints. It has 51,236 stars and 2,282 daily trending gains. Worth trying if you need source code for building a personalized AI learning workspace.

#20 𝕏

LlamaIndex 🦙 created Boo Bench, a 179-page benchmark designed to push document parsers into repetition loops. One unspecified model encountered 246 redaction x’s on a page and produced 23,004 of them; the post links to a leaderboard, paper, and dataset.

Also covered by: @LlamaIndex 🦙