Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Wed 2 Sep · 16 stories · 5 sections

Frontier labs 4 items

Anthropic ships Claude Fable 5.1 (GA) and partner-gated Mythos 5.1: cache reads cut 75%, first Anthropic models to watermark text output.
OpenAI says Astra is its first model past the 'critical' cyber threshold; a public version comes 'soon' with cyber tools gated to partners.
Runway previews Solaris, an 'interface world model' rendering apps as live video: an LLM reads each click, Gen-4.5 draws the next frames.
Fei-Fei Li's World Labs unveils Atlas, a multimodal world model with pixel-perfect camera control that rebuilds generated scenes in 3D.

Agents & tooling 3 items

Perplexity rolls out Hybrid Compute to all Mac app users: tasks split between a frontier cloud model and a local LLM for sensitive data.
Gemini API adds agentic video understanding on 3.7/3.6 Flash: DeepMind cites up to 88% fewer tokens, 66% lower cost, 7% quality gains.
Simon Willison finds the ChatGPT/Codex desktop app quietly bundles a 1.7GB runtime: full Python, Node.js, Poppler, git, and LibreOffice.

Research & papers 3 items

Anthropic let Claude agent teams run alignment research end to end: fixes for 10 failure types averaged ~4x better than human experts.
An Imperial College AI reads a routine ECG in under two seconds, catching 81% of heart failure and 90% of valve disease cases in tests.
A from-scratch transformer trained for 1.5 hours on one RTX 5090 scores 44% on ARC-AGI-1 for $0.67 of compute, beating many larger LLMs.

Business & policy 4 items

Apple alleges OpenAI destroyed evidence in their trade-secrets fight; OpenAI's filing calls the case 'a mess of Apple's own making'.
SoftBank-backed SB Energy files for an IPO: 8.8GW of data-center capacity built or contracted and a $430B backlog, heavily tied to OpenAI.
CNBC's LLM token index: average cost per 1M tokens fell to $0.97, extending a months-long slide from a $2.07 peak on May 28.
The Pentagon launches Starshield AI's Grok for Government on GenAI.mil, opening non-classified use to roughly 3M military employees.

Also notable 2 items

Dan Luu audits how Ed Zitron's AI-skeptic predictions actually panned out, grading the concrete claims one by one against outcomes.
Google launches Pics, a Gemini and Nano Banana design suite inside Workspace, taking aim at Canva and Adobe Express.
Back to top ↑