Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Sat 15 Aug · 13 stories · 5 sections

Frontier labs 2 items

Alibaba releases Qwen3.8 weights under Apache 2.0; testers report the 27B posts big coding-benchmark jumps and runs on one consumer GPU.
MiniMax releases Music 3.0, an open-weights model turning lyrics into full produced songs, extending its open-release run beyond video.

Agents & tooling 3 items

Mixedbread launches Toast 1, a specialized search agent it says matches Opus 5 and GPT-5.6 Sol on retrieval at up to 10x lower cost.
Willison highlights a retrieval trick from Doug Turnbull: let the model hallucinate tags, then snap them to a real vocabulary by embedding.
Daily Dose of DS maps 7 agent-harness design decisions, arguing scaffolds now swing agentic benchmarks more than frontier model swaps.

Research & papers 2 items

Google engineers detail making homomorphic encryption practical for private AI, running inference on queries it never sees in plaintext.
RCT: one 18-turn session with a belief-reframing chatbot on Claude Sonnet 4.5 cut breakup distress at 7 days vs assessment-only controls.

Business & policy 4 items

WSJ: Nvidia reworked its financing of OpenAI's Ohio data-center campus to initially guarantee only half of the planned $250B backstop.
Investor docs: Anthropic generated $11.5B+ revenue with positive adjusted operating income in Q2, up from $787M in Q2 2025.
Reuters: Apple trained a China-specific LLM with Alibaba's support — the first foreign company to offer a proprietary model in China.
Anthropic raises its misalignment risk estimate from very low to low and says it does not plan to release a stronger internal 'Model 2'.

Also notable 2 items

Google adds a Gemini/Flow toggle removing visible watermarks from AI images, video, and music; SynthID and C2PA metadata remain.
Andrew Ng publishes an AI Engineering Skills Map: building AI apps, software fundamentals, coding agents, and shaping the build.
Back to top ↑