ainch

My problem with it is the shift in focus. The original Genie was pitched as a simulator of the world that an agent can interact with to learn so that it could be deployed into the real world. This new version looks like a good, controllable video generator. The fact that you're comparing it to the will smith video, rather than a simulator, is indicative of the fact they've strayed from that original goal. I think that's a shame

if you want GPU-backed/compiled numpy code then JAX is also a good option.

I think you're exactly right. This was already a problem with Genie 3, but they've appear to have gone all in on video generation over controllability.

The original Genie was a great piece of work with a real sense of purpose - to enable agents to learn about the real world in the way Dreamer or Schmidhuber's paper sketched. But now most of the people that drove that mission have left Deepmind (Jack Parker-Holder, Tim Rocktaschel, Demis) and it's become this instead.

Not sure what what specific claim you're making. Do you think gay romance 'soiled' Mass Effect?

I use MonoLisa for the same reason

Not available for me on Pro in the UK.

I think that's quite a strong claim, especially since chain-of-though reasoning means a modern LLM has its own private scratchpad to workshop sentence structure in, if it were a significant problem.

That's true, but I think scientific writing often veers too hard towards impenetrability. At least in machine learning, many of the hallmark papers are remarkably clear and simple in a way that most papers I read day-to-day aren't. Even something as simple as the problem setting can be written insightfully in a strong piece of work.

I'd guess it's named for a Bernoulli distribution, which is how you model a single trial with a binary outcome, like a coin flip.

There are some groups doing it for LLMs - like the Allen Institute for AI's Olmo models and Eluether AI's Pythia.

I know someone that's mucking about with it in a hierarchical control setup. I think the awkward thing is that you need to have prespecified options for it to choose from, rather than have a model that can specify arbitrary joint positions or target poses.

Curious to see how they get on with VLAs. They sound great till you have to sit and record hundreds of teleop demos to teach them how to solve your task...

I believe he is! His lecture series at Stanford on large-scale training was excellent.

It's a good read but I believe you're mistaken. Ilya and Alex Krizhevsky worked with Hinton.

You might be thinking of Andrej Karpathy as one of Li's more famous PhD students?

That makes sense, thank you! Although it's telling that, with language understanding, a human player is able to discover all the goals for themself, rather than needing to have them specified and shaped up-front. Seems like we still have a long way to go on that front :)

And in terms of what I'd try, I think it's a really hard question haha! I'm interested in the idea of hierarchical world models, where you have something doing prediction on a long-horizon, abstract task level (if I beat this gym I can fight the next one) as well as a short horizon model predicting over individual inputs moment-to-moment. I believe LeCun was actually attached to a hierarchical JEPA paper for robotic control - but it's still very early days.