Meta engineer with diverse interests. Game Development, Virtual Reality, Human Computer Input, Artificial Intelligence, CNC Machining, Woodworking, 3d Printing
It's also related to your comments on distilling your working context back into weights as you sleep. My take on it distills it into a "lifetime context" (on disk) rather than weights, but similar thinking.
@karpathy re: Dwarkesh's youtube wrt pulling knowledge out of the weights. I've been noodling on a practical idea to do that exactly that.
https://t.co/HH4PUi6P6t
Texture => Context Window representing all of agents 'time'
Block of Pixels => Embedding at some LOD of time
Mipmap => incremental embedding summary tree
Virtual texture paging => Dynamically fitting all of 'time' into a fixed sized context window using the best embedding LODs
Some great talks at #UPPERBOUND2025.
I half expected @ID_AA_Carmack to unveil "Mega Embeddings". Mega Texture for Transformers. Streaming LOD. Infinite context! ๐ค
He could make it work.