A source-first intelligence briefing
Tracking the evidence for the intelligence explosion.
The few developments that materially change the case for acceleration, explained without hype and linked to the evidence that supports them.
- 01
OpenAI tripled a reasoning benchmark score by changing two settings, not the model
- 02
OpenAI publishes cyber safeguards for a new model, Astra, before describing the model itself
- 03
Microsoft built practice worlds for AI agents, and a small model nearly doubled its score
- 04
An AI rebuilt software in 14 hours that would take a human up to 17 weeks
- 05
A tiny open model nearly matches systems over ten times its size on a key bug-fixing benchmark
How to read an AI benchmark claim without being fooled
A number went up. Six questions tell you whether that means anything.
Read the guide
Our evidence standard
Every claim should leave a trail.
Every reported fact links to the paper, lab, filing, or dataset it came from.
What happened is kept separate from our interpretation of why it matters.
People write and edit every issue and guide before publication.