Pearl weighs in: JEPA world models can answer counterfactuals with two predictor calls
yudapearl · x · 2026-11-21
Judea Pearl amplifies Paolo Di Prodi's walkthrough of the new paper by Balestriero, LeCun and Varaku, "Counterfactual World Models: JEPAs That Reason About 'What If'", and asks how causal inference researchers should respond.
- Key claim: a learned world model can answer a counterfactual question with two predictor calls plus one addition.
- It works only if nature makes no leaps — dynamics must fit a structural causal model S{t+1}=F(St, At, Ut) with unobserved exogenous noise.
- Framed via Pearl's Ladder of Causation: ordinary world models sit on rung two (intervention); physics, not more data, lifts them to rung three (counterfactuals).
- The boat-and-current example shows why counterfactuals matter: separating bad decisions from bad luck for blame assignment and debugging.
More from AGI Musings
- Riemann Hypothesis proof passes independent Lean kernel checks: 17M lines, zero rejections — flngr · 2026-11-21
- "RH is not even a sliver of what is coming": investor says AI math breakthroughs validate his forecasts — burny_tech · 2026-11-21
- If AI proves the Riemann Hypothesis, it also proves verifiable tasks aren't general intelligence — bendee983 · 2026-11-21
- OpenAI's littmath: low p(doom), but p(gloom) higher than she'd like — mjdramstead · 2026-11-21
- Poster predicts Bitter Lesson will dominate everyone's minds in 2027 — akbirthko · 2026-11-21
- Truth-seeking vs community-seeking: why AI doomers and boomers can't understand each other — gleech · 2026-11-21