OpenAI and Tesla Need More AI Compute to Scale Realworld Video Generation

Tesla has been able to generate realistic video with real world physics using AI for over a year. Dr Know It All, John Gibb, had theorized this a year ago. Elon Musk confirmed this to be the case.

OpenAI Sora is much more visually impressive but FSD 12 is in many ways more amazing especially given that it had to inference at 36fps.

Tesla has been short on compute for FSD training so they have not been using this capability that much. In a training compute abundant situation, then self driving improvement will increase by using generating new training.

2 thoughts on “OpenAI and Tesla Need More AI Compute to Scale Realworld Video Generation”

  1. But we have already seen this kind of simulation from Tesla, and it was just classical simulation of roads. That’s why Elon is stating that it has accurate physics… Of course it has.. it is based on a physics engine and straight forward simulation just like a computer game.

    This is very different from Dora that can produce short clips using a text prompt only

  2. These simulations are high quality synthetic data for training all sorts of advanced NN robots. This will enable significant acceleration of training – which is a bottleneck. That means the new bottleneck is compute for creating this synthetic data and training NN – which is the big hardware shortage that Altman wants to throw Trillions of dollars in investment at

Comments are closed.