Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Fri 11 Sep · 13 stories · 6 sections

Frontier labs 1 item

Cognition's SWE-2, RL-post-trained from the 2.8T Kimi K3, hits 50% on FrontierCode 1.1, a point off Fable 5.1 at 64% lower cost.

Agents & tooling 2 items

OpenAI's Agents API exposes the Codex harness as a managed service: sessions, orchestration, sandboxed execution, and MCP tools.
OpenAI adds a Data agent to ChatGPT Work that queries Redshift, BigQuery, Snowflake, and more, then builds shareable dashboards.

Research & papers 3 items

Anthropic's threat report details disrupted misuse: Iranian surveillance of US targets, bio-risk research, and query distillation.
OpenAI claims progress on a second Millennium Prize problem while mathematicians press for proof its Navier-Stokes work is original.
An OpenAI case study shows a researcher using Codex and ChatGPT to screen candidate antimicrobial molecules end to end.

Business & policy 4 items

Newsom signs California bills limiting minors' chatbot interactions and regulating third-party AI safety evaluations.
Microsoft aims to more than triple its data centers to 38GW+ of capacity by 2032, a third of it built for AI-specific chips.
Ant International, Visa, and Mastercard plan a shared standard for AI-agent payments, citing $3-5T projected agentic commerce by 2030.
Huawei, Cambricon, MetaX, and Iluvatar raise AI chip prices 20-50% as the high-bandwidth memory shortage bites, Reuters reports.

Tutorials 1 item

Build multi-turn agents with CrewAI conversational flows: session state, per-turn execution resets, intent routing, and tracing.

Also notable 2 items

HN users report ChatGPT's 'allow training' privacy toggle repeatedly re-enabling itself; the thread drew 448 points in a day.
Apple's iPhone 18 Pro adds Reference Image, a cryptographic check that verifies a photo as camera-captured rather than AI-made.
Back to top ↑