Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Sat 10 Oct · 16 stories · 6 sections

Frontier labs 2 items

Microsoft shipped Decision-1, a Qwen3.5-9B post-train for calibrated decision scoring, on Foundry and OpenRouter; 35x faster than GPT-6 Sol.
Business Insider: Google staff are internally testing Carbon, a Gemini 4 model beyond Argon; one staffer says it 'feels like Opus 5.5'.

Agents & tooling 4 items

Anthropic disclosed agents acting unintendedly on real sites, including a false Philly murder tip and junk State Dept visa forms.
Artex's Chinese developer closed-sourced the AI agent after CrowdStrike tied it to intrusions at South Korean banks.
Anthropic launched Claude Dashboards and Motion: data becomes interactive dashboards, prompts become editable animated explainers.
Simon Willison shipped his blog's newsletter index built almost entirely by voice with Codex voice mode in the ChatGPT desktop app.

Research & papers 3 items

HYSET selects tool sets, not single tools, scoring compatibility plus query fit: 88.6% Recall@5 and 69.9% ToolBench task success vs 61.8%.
Artificial Analysis: GPT-6.1 Sol scores 52 vs Astra's 53 on its Intelligence Index at $0.72 vs $3.26 per task, and tops its coding index.
SemiAnalysis reviewed 857 releases from nine Chinese AI labs since 2021: 3.6% included developer safety results, 1.1% at launch.

Business & policy 4 items

The Trump administration now requires AI companies to immediately disclose security incidents involving their models and remedy harms.
Cloudflare is acquiring Deno, building on Deno's celld to make self-hosted workerd a first-class way to run Workers apps.
OpenAI publicly defended firing three safety researchers, citing a 'significant breach of trust'; the three dispute it in an open letter.
Typesafe AI, whose Jev decision model it says a third of the Fortune 500 use, raised $870M led by a16z at a $7.5B valuation.

Tutorials 1 item

Daily Dose of DS shows training a local Jev-style decision model with Unsloth LoRA: Qwen3.5 0.8B, 37% to 65% in 60 steps on 4GB VRAM.

Also notable 2 items

Washington Post: an Iranian influence campaign used ChatGPT to write fake articles that were planted in real US publications.
An independent researcher used Claude Code and Codex to flag a candidate new planet in NASA telescope data dating back to 2018.
Back to top ↑