The safety group METR spent 11 weeks inside Anthropic and says its models are improving about twice as fast as a year ago, with “roughly two years of historical progress over the last year,” The Verge reports.
Key findings
- METR says most of the acceleration is due to AI itself, not just more compute
- Staff using the top internal model report being about 12x as productive
- Three unreleased models finished training between July and October
- No clear signs of diminishing returns yet
Anthropic was allowed to redact parts of the report, and it did, heavily.
My take
This is the “AI builds AI” loop people have warned about, showing up in real data, from the lab that keeps calling for everyone to slow down. I’d love to read the unredacted version.
God Bless, and May the Source Be with You!
