In the AI Progress Essay Contest, we opened the floor for investigations of predictions regarding the future of AI, especially as it relates to the prospects, timing, and impacts of potentially transformative advanced AI systems. We are honored to have received over 20 fantastic entries and we are very impressed with the results. Here are our winners!
First prize: Hyperbolic takeoff
ege_erdil uses a hyperbolic growth model to investigate how quickly the world economy might be transformed after the widespread deployment of advanced AI. Assuming some reasonable distributions over key parameters of the model, he finds that it’d take around three months to go from widespread deployment of AI to a radical transformation. Extending the model to consider smoother transitions, and using some data on the field of AI, ege_erdil argues that we will see AGI coming a year or two in advance. He assigns it a 50% chance we’ll see Gross World Product double in one year before it ever doubles within four years.
Second prize: Are We Surprised by AI Milestones?
Here, AABoyles investigates whether recent progress in AI surprised the Metaculus community, which has been trying to anticipate it since 2016. Metaculus was positively surprised by how quickly AI mastered Starcraft and achieved code-generation abilities (e.g. Codex and AlphaCode), and somewhat surprised by image generation (e.g. DALL·E 2). On the other hand, progress in protein folding and poker was anticipated, while progress on Winograd problems may have fallen short of expectations.
steven0461 sketches a dozen plausible scenarios that don’t feature transformative AI by 2050 and assesses their likelihoods. His top three most plausible scenarios, conditional on there not being transformative AI, are:
Hard AI Creating a transformative model takes much more compute than the baseline expects (21%)
Stabilized World — Relevant actors like regulators coordinate to prevent or delay major investment in projects to develop advanced AI (16%)
Ruined World — A catastrophe occurs before transformative AI would have been attained (14%)
Fourth prize: Will transformative AI come with a bang?
marius.hobbhahn investigates whether Metaculus expects a fast takeoff for transformative AI. He finds that Metaculus is uncertain—or perhaps inconsistent: The community predicts a slow takeoff when asked about broad economic or scientific questions, and a fast takeoff when asked about transformative AI directly.
Fifth prize: Phase transitions and AGI
ege_erdil uses historical data on economic growth to predict when we might have accelerated economic growth—and thus potentially transformative AI. He uses a Bayesian method for inferring when our economy will reach a phase transition, based on historical examples of such transitions. His preferred model predicts that the economy has a median of ~10 doublings left before it transitions into a new phase with supernormal economic growth. He points out that to suppose that economic growth might accelerate within ~20 years, you would need evidence amounting to a Bayes factor of about 6. He argues that the field of AI—although impressive—has not provided sufficient evidence to update that far in favor of imminent economic acceleration.
Congratulations to ege_erdil, marius.hobbhahn, AABoyles, and steven0461!