A really frustrating partial presentation, given an apparent lack of testing with a spread of efforts for each model.
Given that there's no reason to believe that Fable's xhigh is comparable to GPT-sol's xhigh, or Opus xhigh, for that matter, it would be far more useful to see the effort level where these tasks no longer achieved their goals.
Please forgive my naivety, but are world models (once they are in a consumer-ready form) expected to outperform any currently existing LLM on these sorts of tasks (i.e. of the physical world)?
From what I have seen "world models" are more like video models. The idea is: from a video clip, predict the next frame, like LLMs predict the next word/token from a bit of text.
The idea is that these models should be able to internalize the laws of physics and properties of objects just like LLMs do with grammar rules. To make training more efficient, frames from game engines with hardcoded physics are used. The application is mostly robotics, including autonomous vehicles.
This is a bit different, the idea is to write a physics engine, it means numbers: like for an airplane model, calculate the climb rate. The application is mostly science and engineering.
"world model" is, sadly, a new buzzword; most of the world models we are seeing are video generative models (like GPT-Image), they receive an action vector and previous frames and must generate the next frame. It's working of course.
The whole field of world models also have lighter approaches based on using latent representations as the target (and not pixel frames). LeWorldModel[0] and DINO-WM[1] are a good way to learn how these works.
Strangely, most of the startups in the space of world models are focused on video-generative world models.
A good application of world models is in robotics where an action policy generate a distribution of actions, you can then predict the future state of the world and use some reward policies to choose the "best" future state.
They were all benchmarked, but not included in the post to try to keep the amount of data from being excessive. (Results are also not surprising — Opus good, Sonnet struggles, Haiku fails.)
However, it's pretty good at regurgitating textbook answers (classic Google...). This makes it an asset for circuit design. Try one of my real queries with Gemini Pro:
"Create a differential current splitter that splits a bias current Ib into a positive current Ipos = Ib/2 + ky*Vy and a negative current Ineg = Ib/2 - ky*Vy where Vy = Vy+ - Vy- is a differential voltage input Y."
Honestly, I just hate the term "physical AI". They're robots. It's unfortunate that we had to adopt a term with the words AI in it, just to get investor's attention.
Anthropic's guardrails still make it garbage. But eventualy it will leak.
Given that there's no reason to believe that Fable's xhigh is comparable to GPT-sol's xhigh, or Opus xhigh, for that matter, it would be far more useful to see the effort level where these tasks no longer achieved their goals.
The idea is that these models should be able to internalize the laws of physics and properties of objects just like LLMs do with grammar rules. To make training more efficient, frames from game engines with hardcoded physics are used. The application is mostly robotics, including autonomous vehicles.
This is a bit different, the idea is to write a physics engine, it means numbers: like for an airplane model, calculate the climb rate. The application is mostly science and engineering.
The whole field of world models also have lighter approaches based on using latent representations as the target (and not pixel frames). LeWorldModel[0] and DINO-WM[1] are a good way to learn how these works.
Strangely, most of the startups in the space of world models are focused on video-generative world models.
A good application of world models is in robotics where an action policy generate a distribution of actions, you can then predict the future state of the world and use some reward policies to choose the "best" future state.
However, it's pretty good at regurgitating textbook answers (classic Google...). This makes it an asset for circuit design. Try one of my real queries with Gemini Pro:
"Create a differential current splitter that splits a bias current Ib into a positive current Ipos = Ib/2 + ky*Vy and a negative current Ineg = Ib/2 - ky*Vy where Vy = Vy+ - Vy- is a differential voltage input Y."