Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Sun 13 Sep · 7 stories · 4 sections

Agents & tooling 1 item

Simon Willison had ChatGPT Work with GPT-6 Astra build looping 5K/10K running routes from OpenStreetMap data in one 27-minute task.

Research & papers 2 items

Specific Labs' Real-SWE benchmarks agents on licensed private enterprise codebases; Fable 5.1 leads at 38.8%, GPT-6 Astra at 33.8%.
Yoshua Bengio traces agent lying and cheating to imitation pretraining plus outcome-driven RL, and expects severity to scale.

Business & policy 3 items

Dario Amodei proposes pacing the frontier: slow capability growth, with Anthropic unilaterally taking on embedded outside evaluators.
The Economist: Nvidia backstops AI finance, guaranteeing up to $105B for one Ohio data center and $500B+ mobilized with Wall Street.
OpenAI says Perplexity now trusts GPT-6 Astra with end-to-end systems, from generated test harnesses to production monitoring.

Also notable 1 item

A three-year retrospective maps the full architecture of Apple's M1 Neural Engine, just as the M5 folds ANE cores into the GPU.
Back to top ↑