Summary
Jeff Clune (then at Uber AI Labs and the University of Wyoming) contrasts the “manual AI approach” — discover each building block of intelligence, then engineer them together — with AI-generating algorithms (AI-GAs) that automatically learn how to produce general AI. The argument rests on the historical trend that hand-designed solutions are eventually replaced by learned ones, and on the existence proof of Darwinian evolution. An AI-GA needs three pillars: meta-learning architectures, meta-learning the learning algorithms, and generating effective learning environments. Clune argues both paths are worth pursuing, that AI-GAs scale well with growing compute, and that the community should shift substantial investment toward them; he also discusses safety and ethics specific to the approach.
Key points
- The Third Pillar — generating effective learning environments and training data — is “the least-studied, least-understood, and likely hardest of the three pillars.”
- Environments are not enough; order matters. Even with the right set of environments, random sampling would be too inefficient; the curriculum may itself depend on the learner’s current capabilities, and generating an environment includes generating its reward function.
- Explicitly optimise environments for learning rather than hoping coevolutionary arms races produce general learners; POET (Paired Open-Ended Trailblazer) is offered as early work generating environments together with their solutions.
- The case for AI-GAs: hand-designed solutions keep being replaced by learned ones (computer vision is the clearest example), the approach has already worked once (evolution on Earth), and it scales with compute.
- Pursuing AI-GAs is worthwhile even if they are not the fastest path, because they inform open-ended algorithms and the science of how simple conditions produce intelligence.
Relevance
- Environment Generation and Open-Endedness — the founding statement of why generated environments, not just better learners, are the frontier.
- Unsupervised Environment Design — the curriculum argument (order as a function of the learner) is the premise UED methods operationalise.
- WorldSmith — the Third Pillar is the problem WorldSmith attacks with LLMs and coding agents instead of evolutionary search.
Notes
From paper club, 17 June 2026: ahead of its time in 2019, and one of the foundational pieces on what it actually takes to produce general intelligence automatically. The third pillar is what’s most relevant to us: automatically generating effective learning environments. The core insight is that intelligence doesn’t require hand-crafted complexity; it emerges from simple rules (mutation, selection, co-evolution) applied consistently in well-designed environments. We’re doing the same thing, but with LLMs and coding agents instead of evolutionary algorithms.