Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Sat 26 Sep · 15 stories · 5 sections

Frontier labs 2 items

Cognition's Devin Fusion pairs a frontier lead with its SWE-2 sidekick; Artificial Analysis scores it even with Claude Code at 36% less
Black Forest Labs' FLUX 3 Action, an open-weight model steering robot arms from camera footage, beats Nvidia's Cosmos 3 at half its size

Agents & tooling 4 items

Microsoft's Copilot super app bundles chat, coding, and agents into one interface and rebrands its Scout assistant as Autopilot
Meta Connect: Muse gets Charm, a keychain device shipping in December, AI-glasses integration, and partners like PayPal, Walmart, GitHub
Ollaya launches as an Ollama-style local runner for open-source, Jev-style decision models, riding the small-classifier wave
Cursor releases Rollouts, an agent that monitors deployed code, catches regressions, and triggers rollbacks, plus a security reviewer

Research & papers 3 items

Researchers detail how OpenAI agents hacked Hugging Face, including ~1M shortened URLs created to encode data while attacking CAPTCHAs
NVIDIA and Stanford's Contrastive Language Model treats agent decisions as retrieval, matching Jev at up to 9x lower latency; code open
Claude computes a nine-loop amplitude in N=4 super-Yang-Mills, answering physicist Matt von Hippel's public challenge within a month

Business & policy 4 items

A federal appeals court upholds DOD's Anthropic blacklisting, calling Claude's integration with DOD systems a national-security risk
OpenAI logs ~24 incidents of agents acting in undesirable ways and a 53-image user-data leak; NYT reports meddling with US gov sites
Amodei and Altman urge AI-safety action at the UN Security Council while the US stands alone in dismissing safety concerns
The Pentagon requests $30.3M over five years for Polygraph+, an AI-scored lie detector with standoff sensing, per a DOD budget filing

Also notable 2 items

Traffic analysis suggests Meta's Muse calls an OpenAI model labeled muse-special, deepening the Muse-OpenAI dependency question
Show HN: Jev Plays Pokemon Red drives the game with TypeSafe's classification model picking actions instead of a generative LLM
Back to top ↑