[–] blovescoffee 1y ago ↗ More details on the actual arch would be nice. It seems like this autoregressive action world model space is a local max until the JEPA work takes over. [–] gunalx 1y ago ↗ Dont see how we could not hack on jepa instead of any other input encoders on a transformers llm .
[–] gunalx 1y ago ↗ Dont see how we could not hack on jepa instead of any other input encoders on a transformers llm .
2 comments
[ 2.6 ms ] story [ 18.4 ms ] thread