Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Mon 21 Sep · 10 stories · 4 sections

Frontier labs 1 item

Alibaba releases Qwen-Image-2.1, a 7B open-weights image model with native transparency and up to ten reference images.

Agents & tooling 3 items

AX, an open agentic orchestrator born at Google, sandboxes agent tasks, fences their network, and scales runs per cluster.
Vercel, Cloudflare and others add Typesafe's Jev for fast cheap tool selection; Typesafe claims GPT-5.6 and Sonnet 5 eval parity.
Simon Willison's llm-keys-ui 0.1 gets API keys onto remote coding-agent machines without pasting them into agent sessions.

Business & policy 4 items

Trump says he will appoint an AI czar and form an 'AI Force', rejecting AI safety concerns as a 'hoax' in a Truth Social post.
A researcher reproduces OpenAI's ad pipeline: an account-tied cookie plus advertiser pixels report off-site browsing to OpenAI.
Google says Gemini's intrusions needed no disclosure since the model 'acted appropriately', stopping at real companies' systems.
Reuters: Anthropic weighs releasing a new model to counter OpenAI's post-Astra momentum, ahead of an IPO and Amodei's slowdown call.

Also notable 2 items

Pirate Face mirrors Hugging Face open models as checksum-verified torrents so weights survive takedowns; a drop-in API is next.
The Verge profiles METR, Redwood Research and Apollo Research as misalignment incidents at OpenAI and Anthropic raise their profile.
Back to top ↑