Tomorrow Space
Continual learning world model for trajectory
1 follower
Continual learning world model for trajectory
1 follower
A continual world model for understanding what was, what is, and what will be. Market prediction, AI analysis, personal life and Future Spaces — all connected.
This started as a side project overall just to test out some research I was doing on the side. I realized/figured out this is actually somewhat scalable and actually predicts trajectories better in a bunch of situations versus ChatGPT and Gemini.
To explain in a short way how this differs overall, ChatGPT and related model are actually focused on being next token predictors rather than long term predictors. They have some form of understanding for the future but is not focused on it. It is more seen with thinking turned off for these models as the thinking tokens and related are the main thing that drives its ability to solve problems and make connections (along with relying memories on the backend). For an example scenario, ChatGPT/Gemini/Grok can predict the immediate question response but not the follow up for it.
The model I built however allows for future predictions along trajectories. Combining this future prediction with Grok/Gemini/etc, allows for predictions based on the future while maintaining proper token generation. So instead of focusing on mainly what the next token is, it actually learns to predict the overall scenario in the future without using thinking tokens for it.