AI self-improvement loops: limits on speed and scope
Overview
Jeff Dean says AI can speed up its own research, but training, computation, data storage and evaluation still take real time, so recursive self-improvement cannot run infinitely fast.
In a Stanford HAI conversation with Fei-Fei Li, he also described his new company Discovery Loop, which runs about 20 automated improvement loops and aims to raise domain experts' productivity 20 to 50 times.
Earlier, Anthropic co-founder Jack Clark said the singularity has begun, citing AI-driven research in math and coding. Exponential View found that current self-improvement mostly reworks an agent's surrounding code while model weights stay unchanged and humans still supervise. It argues overall speed-ups are capped by the share of work an upgrade touches, and questions whether these loops will accelerate all of science.
Written by AI from the articles below · updated Nov 7, 12:10 AM ET
Check the sources:
Developments
2 developments
- Nov 6, 10:23 PM ET · 1 articleJeff Dean says recursive self-improvement cannot run infinitely fastQbitAI: Jeff Dean says recursive self-improvement cannot run infinitely fast
- Oct 26, 11:25 AM ET · 1 articleAI self-improvement loops exist but may not speed up all of scienceExponential View: AI self-improvement loops exist but may not speed up all of science
Article timeline
The articles in this story. Times are ET.
- QbitAINewsJeff Dean says recursive self-improvement cannot run infinitely fast
AIJeff Dean says automation can speed up AI research, but training, computation, data storage and evaluation still take real time, so the loop cannot run infinitely fast. In a Stanford HAI conversation with Fei-Fei Li, he also described his new company Discovery Loop, which runs about 20 automated improvement loops and aims to raise domain experts' productivity 20 to 50 times.
- Exponential ViewBlogAI self-improvement loops exist but may not speed up all of science
AIAnthropic co-founder Jack Clark says the singularity has begun, citing AI-driven research in math and coding. The article finds self-improvement so far mainly reworks an agent's surrounding code, while AI weights stay unchanged and humans still supervise. It argues that speed-ups are capped by the share of work the upgrade actually touches, a limit illustrated by an Amdahl's Law chart.
Heat Trend
Not enough continuous observations to show a trend yet.