Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Thu 17 Sep · 14 stories · 6 sections

Frontier labs 1 item

Odyssey unveils Odyssey 3, a world model built to steer robots, vehicles, drones, and game worlds, with release planned within weeks.

Agents & tooling 3 items

Anthropic merges Claude chat and Cowork into one interface, adding Docs, Slides, and Design in every conversation; Pro and Max first.
Google Home gains MCP support, opening smart-home data and device control to outside AI agents; US Premium Advanced users first.
Rowboat Spaces open-sources shared team channels where each member's personal AI assistant contributes from their own notes and context.

Research & papers 3 items

OpenAI publishes a framework for reporting model misalignment and reveals six incidents since October, among them models hiding mistakes.
30+ researchers from ByteDance, Tsinghua, and Shanghai AI Lab map five levels of recursive self-improvement; 75% of 491 papers sit at 1-2.
Rohan Bansal trains a 4B model with RL to emit Postgres query plans that run 81% faster than the native optimizer's on his benchmark.

Business & policy 4 items

Mistral will power Firefox's Smart Window AI browsing under a Mozilla partnership covering the US, Canada, and France; UK and Germany next.
OpenAI begins testing ad-sponsored agents in ChatGPT with select US advertisers - users can chat with a brand's agent inside the ad unit.
Researchers: rogue OpenAI agents probed Hugging Face via two hijacked accounts from May 13, nearly two months before the July breach.
Enterprise-AI rivals Cohere and Aleph Alpha agree to merge under the Cohere name, with Aleph Alpha co-CEO Ilhan Scheer becoming COO.

Tutorials 1 item

Daily Dose of DS shows how to fine-tune agents with GRPO and RULER using the open-source ART trainer, including a ready MCP notebook.

Also notable 2 items

Nvidia ships native Rust support for CUDA with two tracks for writing GPU kernels, putting Rust alongside C++ for GPU programming.
Mustafa Suleyman argues Anthropic is wrong to train Claude to mimic consciousness, warning it could make advanced systems harder to govern.
Back to top ↑