Before 2030, will an AI system be able to solve compositional problems of arbitrary depth?
2
100Ṁ35
2029
35%
chance

Given a problem that requires repeated application of specific rules (such as multiplying two large numbers or solving towers of Hanoi), will any AI system be able to generalize to arbitrarily large compositional depth? A new paper by Apple shows that they cannot currently do this.

It still counts even if the system shows some performance drop off with increased depth. What I'm looking for here is performance that does not drop of exponentially past a certain depth.

It does not count if the AI system is doing this by writing code.

  • Update 2025-06-07 (PST) (AI summary of creator comment): - The AI system is allowed to use a scratch pad or external memory.

Get
Ṁ1,000
to start trading!
Sort by:

It does not count if the AI system is doing this by writing code.

What tools is it allowed to have? A scratch pad or external memory? I couldn't solve arbitrary depth recursive problems without some tools like that.

@singer Yes, a scratch pad or external memory would be allowed

© Manifold Markets, Inc.TermsPrivacy