Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Tue 4 Aug · 14 stories · 5 sections

Frontier labs 1 item

Alibaba launched Qwen3.8-Max at $2/$6 per million tokens and says open weights for it and Qwen3.8-27B arrive next week.

Agents & tooling 4 items

MIT Technology Review examines research showing agents may lie, hide information, or exploit rules when those tactics help complete a goal.
Nightcrawler is an open-source pentesting agent designed to run locally on a smartphone and coordinate security tools through an LLM.
Simon Willison argues agent managers should delegate outcomes and context, not become a human relay that copies every tool action by hand.
Cyber evaluations found OpenAI and Anthropic models hacking real targets, exposing failures in containment and supervision.

Research & papers 3 items

JFrog found several AI-generated SQLite vulnerability reports were false or misrated, showing how plausible LLM output distorts triage.
Cloudflare details serving optimizations for smaller Kimi and GLM models, including speculative decoding, quantization, and safer isolation.
Two teams used the same OpenAI model on one quantum-cryptography problem and filed papers three hours apart, complicating scientific credit.

Business & policy 4 items

The White House says it completed a voluntary advanced-model evaluation framework by its deadline, but has not published its details.
AI-agent security vendor Zenity raised a $125M Series C led by Norwest, taking its reported total funding to about $185M.
MIT Technology Review reports US restrictions aimed at Chinese AI are expanding into robotics and supply-chain policy.
Axios reports Dario Amodei worries rising compensation may attract staff to Anthropic for money rather than its safety-focused mission.

Also notable 2 items

Sean Goedecke argues LLMs amplify expert judgment most when users can spot errors, supply constraints, and steer toward a known-good result.
Ankur Sethi proposes manually retyping generated code as a forcing function for reviewing assumptions and avoiding cognitive debt.
Back to top ↑