Classical procedural generation and modern ML share the same spirit, says Keenan
keenanisalive · x · 2026-09-05
Keenan argues that classical procedural generators share the ethos of contemporary ML: both sample from a distribution conditioned on user-defined parameters, and those classic distributions can likewise be autoregressive, depending on earlier generation steps.
The big differences: 1) contemporary models learn their distributions from data rather than defining them by hand, and 2) ML models are far more general, incorporating cross-domain knowledge instead of being tailored to one domain. "This is a big deal," he says.
Related event: Classic procedural generation shares its core with modern ML, dev argues(2 posts)→
More from Models
- Dev mocks ARC-AGI's escalating claims: from 3-year-old AGI to hypothetical StarCraft AGI — AndrewDai · 2026-09-05
- Claude completes first formalized proof of Fermat's Last Theorem in 13M lines of Lean — AnthropicAI · 2026-09-05
- Fable 5.1, GLM-5.3 open weights, and three real-time world models headline a packed AI week — thursdai_pod · 2026-09-05
- Greg Brockman closes GPT-6 briefing with "welcome to the AGI era" — thursdai_pod · 2026-09-05
- GPT-6 computer use demoed live: macOS simulator "it works, it's real" — thursdai_pod · 2026-09-05
- "AGI day" Sept 3: every frontier lab ships .1, GPT-6 steals the show — thursdai_pod · 2026-09-05