Supervised learning
A person gives the right answer, a label. It works, but every example is somebody’s work, and labels do not scale.
Session №002
World models: how machines learn to predict the future, and why that is worth billions
AI learned by being told, by trying and by reading the internet, and that opened up knowledge work. World models, video models that predict what happens when you act, could give machines a place to practise for the far bigger physical economy. Matthias Dantone on what they are, what they change for robots and software, and where the value will be captured.
Clip from the recording
1 / 27
0:00 / –:––
Click a chapter, a slide or a transcript line and the video jumps there. When a clip plays in the talk, the recording takes the stage. Arrow keys move between slides, space plays and pauses.
Transcript follows the video
Four ways to learn
2012 · Told
2016 · Trying
Reinforcement learning
The environment keeps the score. Trying needs a place to try and a score to keep: games and code have both, the physical world has neither.
2020 · Observing
Self-supervised learning
Hide part of the data, predict it, check. Reading gives a plausible first draft; trying on top of it makes it correct.
Now · Imagining
World models
A model of how the world responds to actions: state and action in, next state out. A world you can try things in before you act.
Where the value is
Not our game
The frontier models
Billion-dollar rounds into the general world models, against Google and NVIDIA. A big risk, and not where a pre-seed fund plays.
Where we look
New verticals
Worlds the internet’s video does not show: physics, biotech. Areas where the data out there is not enough.
Where we look
Picks and shovels
If world models become the standard: how they get trained and served, and what gets built on top of them.
Take it with you
Also watch
