But the question is how do you communicate between System and System?

[rifat28dddd]

If it’s the first idea, the model will be simpler, with only one model and one API to maintain. But it’s also harder to control because they are controlled at different frequencies. The “System” model has a lower control frequency, such as one decision per second (Hz), while the “System” model requires a higher control frequency, such as . decisions per second (. Hz), as I hold this glass of water at thousands of decisions per second.



This control of different frequencies is difficult to achieve cameroon phone numbers through a single model. So maybe a cascade approach would be better.Are they communicating through text or through some latent variable? It's not clear yet, and I think this is a very exciting new research direction. Sonia Huang Do you think we can break through System One by thinking through the expansion and the technology of Transformers? Or do you want to pray for luck and wait and see how things develop? Jim Fan I certainly hope that the data strategy that I described earlier will help us achieve this goal.



Because I feel like we haven't fully realized the potential of Transformers yet. Essentially, Transformers work by putting tokens in and out. Ultimately, the quality of the tokens determines the quality of these large Transformer models. For bots, as I mentioned, the data strategies are very complex. We have internet data, simulated data, and data about real robots. Once we are able to scale the data pipeline and get high-quality actionable data, we can tokenize this data and feed it into Transformer for compression.