Simulator AI: why GANs, diffusion and autoregressive models are all simulators
torchcompiled · x · 2026-10-03
Ethan Smith's essay "Simulator AI" revisits how we categorize deep learning generative models.
Key points:
- GANs, diffusion models, and autoregressive models decompose the problem differently and rely on varying assumptions, but all are methods of modeling and sampling from p(x);
- Beyond "Tool AI", "Oracle AI", and "Agentic AI", the author argues the best description is "simulator" — per Janus & GPT-3's definition, a simulator is optimized to generate realistic models of a system, optimizing for nothing but realism;
- Under pre-training objectives, all these models reconstruct the data distribution they've seen, becoming a mirror of our world — though a "fun-house mirror" due to inaccuracies;
- The author illustrates with his own diffusion model: "A boy with a toy sword" yielded plausible, mostly-correct imagery that still didn't match how it looks in our world.
More from Research
- First NECB2026 draws 360+ researchers, 190+ abstracts over two days — KevinKaichuang · 2026-10-03
- TLAPS-Bench debuts: proving real-world TLA+ specs remains hard for frontier AI — tianyin_xu · 2026-10-03
- Looped LMs could use 3x fewer parameters and a 3x smaller KV cache, matching standard Transformers — ChengleiSi · 2026-10-03
- Neuroscientist Konrad Kording says calibration-free BCI decoding — the field's hardest problem — may be solved — KordingLab · 2026-10-03
- Princeton-backed Choir open-sources a protocol for multi-agent autoformalization on GitHub — burny_tech · 2026-10-03
- CogGym Finds Bigger, Newer Models Behave More Like Humans — But Even More Like Each Other — teortaxesTex · 2026-10-03