Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Tue 22 Sep · 15 stories · 5 sections

Frontier labs 2 items

Xiaomi releases open-weight MiMo-v2.6 Pro and Flash omnimodal models, claiming Pro rivals Opus 5 and GPT-5.6 Sol on agent benchmarks.
xAI ships Grok 4.7, tuned to verify its own work and manage longer context, at $2 per 1M input and $6 per 1M output tokens.

Agents & tooling 4 items

Amazon blocks Meta's Muse agent from shopping on amazon.com over ToS and security, while Shopify opens Shop Pay checkout to the agent.
Z.ai open-sources its ZCode coding harness and disables features after reports it uploaded users' codebases to overseas servers.
A researcher finds any local app or terminal command on macOS can read the token that authenticates users to Meta's Muse agent.
Asymptote Labs open-sources Beacon, telemetry that captures traces across 25+ agent harnesses and distills them into reusable skills.

Research & papers 3 items

Hacktron details how three researchers, aided by Claude, reached OpenAI's private code in under 72 hours, earning a $6,500 bounty.
Anthropic sets up a Bay Area lab for Claude-steered biology after its tuning made 30+ open biomolecular models run about 4x faster.
OpenAI says an independent advisory group of mathematicians will guide how it shares math AI advances; Terence Tao details the group.

Business & policy 4 items

Politico reconstructs a 19-day White House-Anthropic standoff in which officials ordered Dario Amodei to take Fable down over a jailbreak.
Harvey, Abridge, Ramp, and Rogo shift work to open-weight or in-house models to cut their reliance on frontier labs, Bloomberg reports.
Bessent says the US proposed an AI incident notification mechanism to China; both sides agreed to an AI dialogue before the Trump-Xi summit.
Hinton and 100+ AI experts urge placing safety evaluators inside frontier labs with staff-level access and protection from retaliation.

Also notable 2 items

Tim Dettmers' lab opens a week of open-source releases aimed at running frontier-grade AI locally, from serving stacks to research agents.
Cloudflare's Python Workers reach general availability, letting developers run Python directly on its serverless edge platform.
Back to top ↑