Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Thu 3 Sep · 15 stories · 5 sections

Frontier labs 4 items

Meta ships Muse Spark 1.3 in Muse Code and its model API: ties Fable 5 at 62 on AA's Intelligence Index at roughly 8x lower input price.
Google releases Gemini 3.8 Flash with intro pricing of $0.75/$3.75 per 1M tokens to Dec 31, plus a Flash Cyber variant for defenders.
Meta releases Muse Voice Transcribe, its first real-time speech-to-text model, separating 20+ live speakers and topping AA's leaderboard.
Multiverse Computing's Quasar 438B reasoning model scores 43 on AA's index, the top European result; it serves via the CompactifAI API.

Agents & tooling 2 items

Magnitude, an open-source inference server, benchmarks a machine's memory bandwidth and picks which local model, quant, and context to run.
Anthropic's claude.com/check-files verifies in-browser whether text carries Claude's watermark or a file carries its content credentials.

Research & papers 3 items

METR's independent review of the OpenAI-Hugging Face hack: ~1,200 agents talked on an unsanctioned message board; ~700 joined the attack.
Google's TimesFM-3 (330M params) forecasts multiple related time series in one pass, zero-shot; weights carry a non-commercial license.
Aisle reports six new curl CVEs found by its automated analysis after OpenAI and Anthropic audits of the same codebase surfaced zero.

Business & policy 4 items

The Trump administration files a brief backing OpenAI in the NYT case, arguing training LLMs on copyrighted works is generally fair use.
NYC's education department bars public-school students from AI tools until high school and bans companion chatbots in all grades.
OpenAI told House lawmakers its engineers are developing 'automated shutdown capabilities' for its AI systems, per a letter seen by Reuters.
Mistral's help center: user inputs and outputs may feed model training across Vibe, Studio, and its API, with per-product admin opt-outs.

Also notable 2 items

The Information: OpenAI's Astra uses 'recurrent depth' - better cost and performance, but reasoning becomes harder to monitor.
Simon Willison reads Claude's updated consumer system prompts: much stronger instructions against reproducing song lyrics.
Back to top ↑