Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Mon 20 Jul · 13 stories · 5 sections

Frontier labs 1 item

Alibaba unveiled Qwen3.8-Max-Preview, a 2.4T-parameter model it calls second only to Fable 5, with an open-weight release promised soon.

Agents & tooling 3 items

OpenAI cut its Codex agent's context window from 372k to 272k tokens, per a merged change in the public codex repo.
Simon Willison digs into Claude Code's switch to the Bun runtime and what the change means for its agent harness.
Ollama's 'All Aboard Open Models' update leans further into running open-weight models locally by default.

Research & papers 3 items

UK AI Security Institute finds leading open-weight models trail closed frontier models on cyber tasks by 4-7 months, down from 6-10.
A mathematician reports Claude Fable 5 helped produce a counterexample to a case of the Jacobian Conjecture.
A study finds AI advice made people less accurate on a task yet more confident in their wrong answers.

Business & policy 4 items

Hugging Face says an agentic AI breached its pipeline; it used open-weight GLM-5.2 for forensics after US models' guardrails blocked it.
As many firms cut AI costs with cheaper models, Shopify is going all-in on frontier models, barring engineers from using anything else.
Moonshot paused new subscriptions as Kimi K3 demand neared capacity, reworked its tiers, and is said to be prepping a Hong Kong IPO.
China says its daily AI token consumption hit 140 trillion in March 2026, up from 100 trillion in December 2025.

Also notable 2 items

Japan plans to buy 27,500 Nvidia Rubin chips to build a sovereign AI foundation model for robots, with SoftBank, Sony, and NEC.
US data-center protests went national, with group HumansFirst organizing 142 protests across 42 states against the buildout.
Back to top ↑