Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Wed 9 Sep · 15 stories · 6 sections

Frontier labs 3 items

OpenAI ships ChatGPT Images 2.5: up to 50% lower latency than Images 2.0, a Sketch draw-to-prompt input, and new API model IDs.
Inception Labs' Mercury 2.5 diffusion LLM decodes at 1,107 tokens/sec and claims a 40% intelligence gain over Mercury 2.
OpenBMB open-sources MiniCPM5-2B, a 2B on-device agent model it says ranks first among sub-4B open models on AA's Intelligence Index.

Agents & tooling 2 items

Meta launches Muse, a personal agent on dedicated cloud VMs: Muse Spark 1.3 model, free to 100M tokens/week, $20 and $100 tiers.
Strix, an open-source AI pentesting agent (23k+ stars), crawls a running app, probes abuse paths, and returns PoC-backed findings.

Research & papers 3 items

OpenAI says an internal model solved Navier-Stokes using 10k agents over 88 hours; NYU's Buckmaster alleges his team was scooped.
DeepMind releases AlphaGenome Atlas, a 1PB predictive map covering every one of the ~9B single-base edits in a human genome.
Qwen3.8 27B quantization benchmark: 17GB Q4_K_M matches 55GB BF16 on Terminal-Bench 2.1 and fits a 24GB card; 1-bit collapses.

Business & policy 4 items

Devin-maker Cognition banks another $2B, nearly doubling its May valuation to $48B as run-rate revenue approaches $900M.
China puts its AI compute at 2,185 EFLOPS, up 177% YoY, and wants ~4x more by 2030 on the back of ~$532B in infrastructure plans.
Axios: Anthropic cuts ties with the ITI trade group after it opposed three export-control measures on advanced AI chips.
Bloomberg: Anthropic decides against acquiring Decart for ~$6B after due diligence; the companies may still collaborate.

Tutorials 1 item

Hands-on: turn Qwen3-4B into a reasoning model with GRPO via Unsloth and TRL - deterministic reward functions, no labeled data.

Also notable 2 items

Terence Tao warns good open problems are being mined 'in a non-renewable fashion': rumors of progress now trigger AI compute races.
Deltafin fork runs Kimi K3 (2.8T MoE, 1.45TB of experts) at 1 token/s on an M5 Max MacBook by streaming weights from four SSDs.
Back to top ↑