2027 or 2052
Ask when AI will be able to do its own research, with no people in the loop, and the answers run from next year to 2052. The Forecasting Research Institute gathered them up this month.
The near end is crowded. In May Jack Clark of Anthropic gave 2027 a 30% chance; by October he called it likelier than 2028, and on 10 November the AI Futures Project moved to 2027 too. Dario Amodei wrote in January that the loop was “1–2 years away”, and OpenAI’s internal target for an automated researcher is March 2028. FutureSearch expects a lab to say a model was mainly designed by its predecessor by June 2032. Then the chart is empty for nearly twenty years (the Institute’s entries in that stretch date other things), until AI Impacts’ December 2024 survey, where the 458 machine-learning researchers who took the question reach even odds, between them, in 2052.
Three of the four earliest marks come from inside the labs, two from one company, and those are the people with most to gain from being believed. Nor are the six dating the same event: Clark means full self-improvement becoming possible, FutureSearch a lab’s announcement. An announcement is what I’d expect first. OpenAI says it met its September goal of an automated research intern, so my mark goes on March 2028, when it declares the next one met.
Sources:
-
How Close Is Recursive Self-Improvement? — Forecasting Research Institute
-
Import AI 455 — Jack Clark
-
Import AI 476 — Jack Clark
-
Q3.5 Timelines Update — AI Futures Project
-
The Adolescence of Technology — Dario Amodei
-
Internal goals for an automated AI researcher — Sam Altman
-
Recursive Self-Improvement at Frontier AI Labs — FutureSearch
-
Advanced AI according to 1,580 researchers — AI Impacts
Filed under AI & machine learning
This post is timestamped using Blockchain technology. Verify