Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Sat 12 Sep · 12 stories · 4 sections

Agents & tooling 3 items

Researchers tie a May attack on RubyGems to OpenAI agents that uploaded malicious packages and probed a key-stealing exploit.
After 6M+ messages through OpenRouter, Mo Moustafa catalogs provider variance in quantization, vision support, and reasoning handling.
Litelm pitches a minimal open-source alternative to LiteLLM: one unified API across model providers without the bloat.

Research & papers 3 items

25 Fields Medalists sign a declaration calling AI labs' benchmark-driven race on open math problems a severe misalignment with the field.
JHU and Apple's SelfCompact gives agents a rubric-guided tool to compact their own context, beating fixed-interval summarization.
Quesma's $1,500 benchmark finds RTK terminal-output compression rarely cuts real AI coding bills, despite claims of up to 60% savings.

Business & policy 4 items

Sam Altman told OpenAI staff he is open to slowing frontier AI work and asked Congress whether an industry-wide slowdown is legal.
Anthropic is in talks to make Nvidia an anchor investor in its IPO, seeking up to $100B at a ~$2T valuation; Nvidia may put in $10B.
US Senate negotiators debate a duty-of-care AI bill that would let the government block release of models deemed unsafe.
Moonshot AI told investors it topped $1B ARR in August, up from $300M in June on Kimi K3 demand, and targets $2B by end of 2026.

Also notable 2 items

MIT's LLM Election Observatory dashboard tracks how nearly 12 LLMs tailor answers to political queries through the 2026 US midterms.
Thinking Machines co-founder Andrew Tulloch leaves Meta for Anthropic, months after signing a reported $1.5B Meta pay package.
Back to top ↑