for those wanting to do something similar I'm using https://t.co/lbo8XVm2Nm
they have not released their fast version yet (the version that they claims rivals genie-3) so you have to post-train it in order to get an interactive/real-time.
> started @fdocinc artifact 5 weeks ago
> pivoted 1 week before demo day because old idea was not a problem space I was interested in
> waymo dropped how they're using genie-3 for their world model : https://t.co/UVP8acRofv
> also was heavily inspired by @hugothomel 's world model demo
i'm going to take a bet that world models will improve as fast as we went to gpt-2 -> gpt4.5 in the next 2 years or so. I believe that there will be many cases, such as driving, where we can create high-quality synthetic data via these world models as opposed to having to manually collect data. Now anyone can train a SOTA FSD system.
i made an open sourced google colab alternative
- local + customizable
- connect to gpus by lambda/vast.ai/runpod
- and more features coming soon
try using: uv tool install more-compute
or roast my code at: https://t.co/2BJCACLgSj