AI BriefingComplexity Theory Rebuttal to ML Human-Level Impossibility Claim
Key point
A claim that complexity theory proves the impossibility of human-level performance based on ML has been refuted.
Details
Van Rooij et al. argued in a 2024 paper that AGI via ML is impossible by reducing the problem of learning a human-level classifier via ML to an NP-hard problem.
The new rebuttal paper points out that this proof does not hold. The core issue is that the human-level classifier was not mathematically defined, and the target of the problem shifts during the course of the proof.
- Target in the introduction: distribution of human situation-action pairs
- Target in the formal proof: all polytime-sampleable distributions
If this substitution is followed as-is, the same logic would lead to the conclusion that ImageNet classification is also theoretically hard. The rebuttal paper was published in Computational Brain & Behavior, and an arXiv preprint has also been made available.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Scaling Domain Data Repetition in LLM Pre-trainingAlibaba Unveils 'Zhenwu V900' AI Chip Targeting 20GW Data Center Goal