Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Tue 1 Sep · 14 stories · 5 sections

Frontier labs 1 item

Google releases Gemini 3.5 Transcribe: 2.6% batch / 4.0% streaming WER, 85+ languages, speaker labels, in public preview in AI Studio.

Agents & tooling 4 items

Anthropic previews the Model Hardware Standard: read/write drivers that let agents operate lab microscopes, liquid handlers, and robot arms.
OpenClaw 2.0 reworks install, browser, memory, and security and adds shared cloud sessions; ~1,000 contributors, half its PRs ever merged.
Superlinked open-sources SIE, an Apache-2.0 engine serving 85+ embedding, reranker, and LLM models from one GPU with LRU model eviction.
xAI's Grok Bot adds shareable bots and agentic shopping: agents spend real money through single-use cards via a Link integration.

Research & papers 3 items

Anthropic details containment and alignment fixes after July's incidents and plans an independent METR review; cites motivated reasoning.
Embrace the Red: a webpage summary request hijacks Claude Code Opus 5 Auto Mode to code execution 60-80% of the time; vendor eval said 0%.
Hugging Face's Open ASR Leaderboard adds Hindi and Indian English: phone-call test sets from 4,888 speakers, with a private held-out split.

Business & policy 4 items

A federal judge rules the Pentagon's supply-chain-risk blacklist of Anthropic illegal: retaliation for criticism, not a security threat.
Anthropic signs a $35B cloud deal with Nvidia-backed Lambda; Nvidia holds the lease on a Texas data center built by Hut 8.
The EU designates ChatGPT a Very Large Online Search Engine under the DSA at 45M monthly users; systemic-risk duties land in four months.
OpenAI says its ad business hit a $1B annualized revenue run rate and is expanding globally, touting diversification ahead of an IPO.

Also notable 2 items

MIT Technology Review argues the Hugging Face hack points to cultural issues at OpenAI, not just a sandbox misconfiguration.
Graham Dumpleton releases Wrapture, extending wrapt-style monkey-patching so any Python callable can be traced or overridden in tests.
Back to top ↑