"The world is predictable in the right coordinates.
@cosmo_shirley4.7K followers
36,872 views
8 Comments
Comments from the original post on X (Twitter)

@cosmo_shirley
@KempeLab @ylecun Website here:
4 likes

@MaxDiffusionRL
@randall_balestr @KempeLab @ylecun @soniajoseph_
1 like

@jatingargiitk
@ylecun @randall_balestr @KempeLab how did you measure that in practice?

@WorldianAI
Physics has a long record of finding the right coordinates before anyone could explain why they work. Kepler got there with ellipses, and Noether showed that symmetries and conservation laws are the same thing. If a model finds its own, the interesting test is whether a physicist can read them.

@Kaoscity_now
@randall_balestr @KempeLab @ylecun The application deadline is doing more empirical work than most world models.

@GOFIntelligence
@randall_balestr @KempeLab @ylecun "The world is predictable in the right coordinates. The open question is whether a model can learn those coordinates on its own." - these AI people are so full of themselves.

@GiggsyFoot
@randall_balestr @KempeLab @ylecun Predictive coding is the only self teaching method (unsupervised) and is what the human brain uses. It learns via observation Works beautifully with an SSN. Prediction errors propagate from layer to layer allowing each layer to be individually trained via local/Hebbian means

@patchwanders
@randall_balestr @KempeLab @ylecun A model discovering useful coordinates that physicists hadn't thought to try would be especially exciting.