Could Anthropic's J-space be incorporated into a Hutter prize entry? Anthropic admits in their paper that they don't completely understand where it comes from. Claude's neural network has at least 88 layers. Back propagation does not work on deep neural networks, so how does that work?
They describe the J-space as the set of neurons that Claude is thinking about as intermediate steps when reasoning that don't make it to the final output. It looks a lot like the way humans think, except that the global workspace or short term memory or transformer attention mechanism or access consciousness can hold several dozen words vs 5 to 9 words in humans. Human short term memory uses neural feedback loops, where in Claude it uses a combination of long propagation chains and an attention mechanism that works like a scratch pad.
Another difference is that LLMs are trained in multiple passes, and then the weights are frozen to avoid leaking data between users. This requires a large context window (200K to 1M tokens) so that it can forget everything you said in the conversation and then play it all back before the next question. Obviously a compressor would work more like the brain or a private LLM, training in a single pass on every input with no need for a context window. This should require a lot less computation.
This should be possible. The current entries based on CMIX don't have multiple layers or a J space. The top LTCB entry NNCP might because it is based on a transformer network running on a GPU, but it is closed source and the paper doesn't give details.