Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Fri 9 Oct · 13 stories · 5 sections

Frontier labs 1 item

StepFun's Step 5 Preview, a sparse mixture-of-experts flagship for agentic work with a 1M-token context, surfaced on OpenRouter.

Agents & tooling 3 items

SpaceXAI's Grok Bot will route tasks to rival models, including Claude and Midjourney, picking whatever is likeliest to do the job best.
Uzu, Mirai's open-source Apple-silicon engine, decodes 4-bit Qwen3.5 9B at 92-117 tok/s on a base M5, vs 22-25 for llama.cpp and MLX.
Anthropic's OSS Scanner gives open-source projects free opt-in security audits, sending AI-generated vulnerability reports unreviewed.

Research & papers 3 items

Biohub expanded its virtual-cell initiative to $1.8B with NIH, the Energy Dept, DeepMind, Meta and Isomorphic to train AI cell simulators.
Scott Aaronson: AI labs are testing whether internal models can break cryptographic protocols; Ethereum researchers urge key migration.
OpenAI withdrew three of its math papers over a sign error; the Association for Human Mathematics urged ending collaboration with the lab.

Business & policy 4 items

OpenAI told investors annualized revenue neared $50B in September, ~$20B below earlier reports; Nvidia, Oracle and CoreWeave fell on it.
Three fired OpenAI safety researchers say they were ousted for prioritizing safety, and urge labs to preserve reasoning monitorability.
OpenRouter: business spending on OpenAI vs Anthropic models was roughly even in September; Anthropic held a 75%+ share in January.
Anthropic's usage policy now bans 'sustained and needless abusive or cruel behavior' toward Claude; ending chats is the main enforcement.

Also notable 2 items

Cactus released Whistle, an open 16.9MB on-device speech-to-text model: seven languages, CPU-only, 11ms to first token.
Google's AI Edge Foresight, a macOS note-taking app, transcribes meetings entirely offline with EmbeddingGemma 2 and a Gemma 4 assistant.
Back to top ↑