Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Tue 28 Jul · 14 stories · 6 sections

Frontier labs 2 items

Moonshot releases Kimi K3's weights under a custom 'Kimi K3 License', with a technical report detailing the training recipe.
Microsoft debuts MAI-Cyber-1-Flash inside its MDASH harness, claiming top CyberGym results at 50% of the cost of leading models.

Agents & tooling 2 items

Independent audit questions Bun's $165K AI rewrite in Rust: six weeks after the merge, still no release tag and constant CI churn.
OpenWorker, an open-source desktop agent from Andrew Ng, works across files, Slack, and calendar, checking in before key actions.

Research & papers 3 items

Epoch and METR release MirrorCode: agents reimplement whole programs from CLI access; Opus 4.7 did a 2-17-week human task for $251.
OpenAI's study of 800K work messages finds 43.5% of occupation-specific use involves tasks traditionally from other jobs.
METR proposes the 'expenditure horizon': the spend level at which an AI agent stops being cheaper than a human on the same task.

Business & policy 4 items

Anthropic says it has never backed an open-weights ban, defends chip export limits to China, and calls for global model testing.
Nvidia forms the Open Secure AI Alliance with CrowdStrike, Hugging Face, and Dell to share tools for AI safety and cybersecurity.
Nvidia commits a reported $5B to Ilya Sutskever's SSI; the lab gets GPU access to raise its compute by an order of magnitude.
China vows 'all necessary measures' if the US sanctions its AI firms over distillation, calling the technique standard practice.

Tutorials 1 item

Hands-on: rebuild a Claude Code-style agent harness — planning, loop, subagents, sandboxing, memory, checkpoints — using CrewAI.

Also notable 2 items

Shared Claude chats surfaced in Google and Bing results; robots.txt disallowed crawling but the pages lacked a noindex tag.
Simon Willison's opinionated guide to which AI to use for what: models, tools, and workflows as of July 2026.
Back to top ↑