Introducing Mistral Large 4: Le chonk

• <p><strong><a href="https://mistral.ai/news/mistral-large-4/">Introducing Mistral Large 4: Le chonk</a></strong></p> Mistral are back in the game. Today they're releasing a preview of Mistral Large 4, a 1 trillion parameter, 49 billion active parameter model trained on their own cluster of 3,800 NVI

Reflection AI's Beam proves sparse MoE models can match frontier performance at a fraction of the compute cost, while OpenAI quietly makes watermarking mandatory infrastructure for EU users.

• Reflection AI released Beam: 501B total params but only 23B active (MoE), Apache 2.0 licensed, FP8/NVFP4 formats, 1M token context, 3-4x cheaper to run than GLM-5.2 despite GLM having 250B fewer active params — full weights drop this month
• OpenAI launched textGrain, an invisible watermark for ChatGPT/Codex outputs to comply with EU AI Act; auto-enabled for EU users, opt-in elsewhere; detector access restricted to approved researchers; editing ~25% of words drops detection to 17%
• Vals AI ran 90 Claude agents for 3 days to simulate materials science and surfaced two room-temperature magnetic semiconductor candidates, including one (KV[Cr(CN)6]) synthesized in 1999 but never tested for this property
• DeepSeek open-sourced its GPU math library, claiming a 98% cut in token costs for compatible workloads
• Multiverse Computing released an optimized Whisper Large V3 Turbo claimed to be 15x cheaper than comparable speech-to-text APIs
• FlashAttention-3 has a reported hidden BF16 bug that silently corrupts training after ~25B tokens — worth checking if you're fine-tuning at scale

A single Claude Code skill now turns one photo into a navigable, sound-equipped 3D world in under 5 minutes, ready for Unity, Unreal, or Blender.

• image-blaster (open-source): one photo in, full 3D world out — Claude Code orchestrates object segmentation, Hunyuan 3D mesh generation, and ElevenLabs audio, exporting .glb/.obj with physics colliders for Unity/Unreal/Godot/Blender/Three.js in under 5 min.
• AgentCraft: open-source MIT tool lets Claude coding agents work inside a Minecraft world — parallel worker agents code isolated repo copies, walk over to ask permission for risky commands, nothing pushed without in-game diff review.
• Aleph Alpha released Kolibri-1, a 78B MoE model (only 3.46B active params/token) under Apache 2.0, with 1M token context, agentic tool-calling, and strong German/English support — runs on 2x A100 80GB.
• Redis Agent Memory hit 86.5% accuracy on LongMemEval at ~$0.07/session, combining raw conversation evidence with extracted facts for persistent agent memory.
• Wavestone audit of 11 coding agents found a consistent 7-part architectural blueprint across the industry.

Scrimshaw Jukebox

• <p><strong>Tool:</strong> <a href="https://tools.simonwillison.net/scrimshaw-jukebox">Scrimshaw Jukebox</a></p> <p>I wanted to see if Claude Opus 5.5 could compose music, so <a href="https://claude.ai/share/1f721c20-2499-4d23-b368-3ab57146d956">I tried this</a>:</p> <blockquote> <p><code>I want you

Quoting Felix Rieseberg

• <blockquote cite="https://twitter.com/felixrieseberg/status/2107206431376334975"><p>The "old" version of Cowork runs model inference in the cloud, executing tool calls in an Anthropic-provided VM we shipped to your computer. We added the VM for capability, safety, and security reasons - mapping in j

MCP for agent-to-agent comms may be the riskiest protocol you've never heard of

Microsoft ships domain-specific voice/transcription models that beat rivals on accuracy and latency, while Suno pushes into combined voice+music generation.

• Microsoft launched MAI-Transcribe-2-Streaming: tops the Artificial Analysis Word-Error-Rate leaderboard, beating Grok, Meta, and OpenAI, with text returned in as little as 100ms across 60 languages.
• Microsoft also released MAI-Voice-2.1 (strongest multilingual TTS, 23 languages/26 locales) and MAI-Voice-2.1-Flash (optimized for high-volume, latency-sensitive use)
• Suno launched 'Speech' in beta: generates voice and background music in one cohesive track from text + style description, though accents and dramatic pauses still misfire
• Microsoft's strategy is now explicitly domain-specific models (MAI-Thinking-1, MAI-Code-1.1-Flash, MAI-Image-2.6, plus voice/transcription) rather than one general frontier model, targeting enterprise workflows where focused performance beats breadth
• OneTrust's Blake Brannon argues for 'guardian agents' that police other AI agents in real time, since human review can't keep pace and agents inherit access without inheriting trust
• Amazon committed $1B over 5 years ('Built Together') to win community support for data centers amid 7-in-10 American opposition, though this may not address deeper public anxiety about AI itself

llm-openai-decisions 0.1a0

• <p><strong>Release:</strong> <a href="https://github.com/simonw/llm-openai-decisions/releases/tag/0.1a0">llm-openai-decisions 0.1a0</a></p> <p>OpenAI released their new Jev-style <a href="https://developers.openai.com/api/docs/guides/decisions">Decisions API</a>, as previously announced at last week

llm-mistral 0.16

• <p><strong>Release:</strong> <a href="https://github.com/simonw/llm-mistral/releases/tag/0.16">llm-mistral 0.16</a></p> <p>Adds support for reasoning models, such as the newly released <a href="https://simonwillison.net/2026/Oct/6/le-chonk/">Mistral Large 4</a>.</p> <p>Tags: <a href="https://simonwi

EmbeddingGemma 2

• <p><a href="https://news.ycombinator.com/item?id=49980487#49983751">My comment</a> on <a href="https://news.ycombinator.com/item?id=49980487">EmbeddingGemma 2</a> &mdash; Hacker News.</p><p>I really appreciate that <a href="https://blog.google/innovation-and-ai/technology/developers-tools/embeddingg

Anthropic's hardware standard is quietly becoming the 'MCP for physical devices' while Meta Muse's relationship-mapping feature reveals how deep agent context-gathering can go.

• Anthropic's Model Hardware Standard (MHS), announced quietly in August as 'MCP for hardware,' received 30x more organization sign-ups than planned, spanning semiconductors, manufacturing, aerospace, and lab automation — early results include QuEra stabilizing quantum computer lasers (695/700 success) and Carnegie Mellon running autonomous iterative lab experiments
• Meta Muse builds detailed 'pages' on every person in a user's life (facts, history, relationship-strengthening tips) via hourly background processes — Meta says it's opt-in and sandboxed per user, but researchers call the depth of relationship-mapping 'creepy'
• Sam Altman publicly distances OpenAI from Anthropic's risk posture, stating OpenAI accepts 'bad things happening' for AI's benefits and rejects the idea of 'catastrophic risk' or loss of control — a clearer articulation of the two labs' diverging regulatory philosophies
• Both OpenAI and Anthropic still signed Trump's non-binding 'superintelligence' safety accord despite this philosophical split, showing alignment on optics even as risk tolerance diverges
• GPT-6 Astra took the #1 spot across 4 Design Arena leaderboards and ChatGPT ranks #1 in a16z's Top 100 consumer AI apps — competitive signal for anyone benchmarking creative/design model quality
• Cohere shipped its biggest update yet (North 2) with 15+ new enterprise features, continuing the push toward agentic enterprise tooling