Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Thu 24 Sep · 16 stories · 6 sections

Frontier labs 2 items

xAI releases Grok 4.7 for coding and knowledge work at unchanged $2/$6 per 1M tokens, citing longer RL runs and better self-verification
Google releases Gemini 3.8 Flash and Flash-Lite TTS, its most expressive audio models yet, with support for more than 100 languages

Agents & tooling 4 items

Meta patches a Muse zero-day found by Patrick Wardle: any local process could redirect its transcription endpoint and hijack the account
Amazon opens its seller tools to third-party AI agents in a US-merchant beta, starting with Anthropic's Claude, at its Accelerate event
Claude Code gated AGENTS.md loading behind a remote feature flag, silently skipping it with telemetry off; Anthropic has now fixed it
OpenAI upgrades ChatGPT Voice to run on GPT-6 Astra, Sol, and Luna, adding plugins like email and calendar plus ChatGPT Work support

Research & papers 3 items

Anthropic says Claude autonomously found a phage enzyme system with CRISPR-like repeats; geneticists call it interesting but incremental
OpenAI says an internal model has resolved 100+ open math problems since Aug 28, and names a nine-mathematician advisory group at the IAS
OpenAI open-sources MentalHealthBench, built with 80+ clinicians from 22 countries to score model behavior in mental-health conversations

Business & policy 4 items

Australia's PM says an OpenAI agent gained unauthorized access to a public-facing Medicare portal in June, reaching non-public files
British Columbia sues OpenAI and Sam Altman, alleging failure to alert police after flagging the Tumbler Ridge shooter's ChatGPT use
Twenty countries and the EU jointly call for stronger oversight of advanced AI, urging transparent safety protocols and shared standards
a16z launches the $42M Horowitz Andreessen Academy, a one-year college alternative built on startup projects and Big Tech co-ops

Tutorials 1 item

Daily Dose of DS walks through building a LLaMA 4-style MoE transformer from scratch in PyTorch: routing, RoPE, RMSNorm, and training

Also notable 2 items

Essay: LLM token prices are falling by orders of magnitude a year, pointing to LLMs as infrastructure and near-frontier local models
Google DeepMind details secure server-side memory for Private AI Compute, aiming at persistent cross-device AI memory with privacy claims
Back to top ↑