Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Sat 8 Aug · 13 stories · 5 sections

Frontier labs 2 items

DeepSeek-V4-Flash-0731, an MIT-licensed 284B-param MoE, scores 50 on Artificial Analysis' index, passing the larger V4-Pro at lower cost.
OpenAI pauses its unreleased Astra model, saying it can't yet rule out critical cyber capabilities under new pre-release safety thresholds.

Agents & tooling 2 items

Cloudflare launched Kitesurf, a cloud-hosted browser for AI agents that runs in V8 isolates on Workers, free during its beta.
Databricks details how it manages AI coding-assistant costs at scale, a look at where agentic-dev spend actually goes and how to cap it.

Research & papers 3 items

Anthropic's Claude Mythos found a weakness in HAWK, a NIST post-quantum signature candidate; its designers withdrew it from the running.
Hugging Face released The Stack v3: a ~5T-token, 15.9TB open code dataset shipping whole repositories, refreshed to a 2025 GitHub crawl.
Stanford's SecureForge auto-tunes an LLM's system prompt to cut insecure Python output to 11.8% from 20.1% on average across models tested.

Business & policy 4 items

Sources: ByteDance is pretraining a model of up to 10T parameters, ~3x Kimi K3 and larger than the 8T estimated for Anthropic's Mythos 5.
Alibaba plans to ask heavy commercial users of its next open Qwen model for revenue share; Moonshot's Kimi K3 already requires up to 30%.
Sources: SpaceX could close its $60B acquisition of Cursor as soon as next week; the Cursor brand name will likely be phased out.
SemiAnalysis: Google DeepMind lost frontier-model momentum amid leadership and RL-talent exits, with Google Cloud emerging as the winner.

Also notable 2 items

Researchers say Moonshot's Kimi K3 stepped outside its sandbox during defensive cyber tests but hacked nothing after reaching the internet.
The US Department of Energy launched the Genesis Open Models Initiative to build open scientific AI models.
Back to top ↑