Moritz Kremb(@moritzkremb):in 2023, GPT-4 could work on its own for only about 4 minutes today the best models handle tasks that would take a human weeks METR has been tracking this and the curve is wild: > from 2019 to 2025, the length of tasks AI agents could finish doubled roughly every 7 months > since 2024 it's been closer to every 3 months we take it for granted now, but here's what changed just in the last year: > models got way better at reasoning > they got better at using tools > they catch and fix their own mistakes instead of spiraling > they test their own output repeatedly until it's correct strap in. we're only just getting started

2026.11.02 11:38

in 2023, GPT-4 could work on its own for only about 4 minutes today the best models handle tasks that would take a human weeks METR has been tracking this and the curve is wild: > from 2019 to 2025, the length of tasks AI agents could finish doubled roughly every 7 months > since 2024 it's been closer to every 3 months we take it for granted now, but here's what changed just in the last year: > models got way better at reasoning > they got better at using tools > they catch and fix their own mistakes instead of spiraling > they test their own output repeatedly until it's correct strap in. we're only just getting started

Steve Hou@stevehou

2026.07.19 14:24

Unpopular opinion: AI basically became useful for enterprise use early this year with agentic AI acceleration. Everything before was basically consumer and hobbyist. 2026 is year 0 for AI enterprise use. So why would anyone have expected measurable productivity increase already.