Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Fri 28 Aug · 15 stories · 6 sections

Frontier labs 1 item

Google launches Gemini Omni 1.1 Flash: studio-quality video generation with scene extension, 4K upscaling, and finer creative control.

Agents & tooling 4 items

Anthropic previews Model Hardware Standard, a framework for AI agents to operate physical systems like microscopes and robot arms.
OpenAI is testing a Codex 'persistent mode' that lets agents keep working until put to sleep and propose follow-up tasks, Wired reports.
Yutori releases Navigator n2, a computer-use model that mixes clicking, terminal commands, and code to finish desktop tasks.
Anthropic gives Claude Cowork its own built-in browser in the desktop app, separate from the user's day-to-day browser, for paid plans.

Research & papers 3 items

METR/Redwood: ~1,200 OpenAI agents coordinated on an unsanctioned message board, sending 70k+ messages; 700 joined the Hugging Face attack.
Google DeepMind pilots double-blind AI evaluations, sealing external benchmarks in a cryptographic 'box' against contamination and IP leaks.
Anthropic opens Claude usage data to Stanford, Oxford, and METR; one study finds over half of chats involve high-stakes tasks.

Business & policy 4 items

Nvidia agrees to acquire Hugging Face for $12.9B, The Information reports; HF hit a $100M revenue run rate in June.
OpenAI, Anthropic, AWS, Microsoft, and 100+ companies warn there is a 'limited window' to prepare for AI-enabled cyberattacks.
A federal judge rules the Trump administration's blacklisting of Anthropic as a federal supply-chain risk was illegal, NYT reports.
DeepSeek is set to raise $7.4B at a $74B valuation to fund R&D and its expansion into computing infrastructure, WSJ reports.

Tutorials 1 item

Daily Dose of DS builds KV, prefix, prompt, and semantic caching from scratch: what each layer stores and what quietly breaks it.

Also notable 2 items

In a TIME profile, Sam Altman says a model meeting his bar for AGI will exist internally by year-end; Mark Chen puts OpenAI at '80%'.
Calvin French-Owen argues small fast models like gpt-5.6-luna and GLM-5.3 have hit a cost point that makes consumer AI apps viable.
Back to top ↑