2/25/2026 Update: It's evident that the idea of continual learning is not new. However, there is still fruit in the idea of daydreaming. Possibly, how can we extract the most out of a static dataset (rather than inputting new information) via daydreaming. Currently doing research on it.
Despite impressive capabilities, large language models have yet to produce a genuine breakthrough. The puzzle is why. A reason may be that they lack some fundamental aspects of human thought: they are frozen, unable to learn from experience, and they have no “default mode” for background processing, a source of spontaneous human insight.
I remember vividly that I was once asked "what can LLM not do that humans can?". It's naive right? This is probably one of the things I didn't think of, but it exposes so many of the blindsides that we aren't aware about with LLMs.
For one, humans have continual learning. Yes, while LLMs are getting new user queries, they are trapped in their knowledge and it's hard to move the needle much farther from its current knowledge.
But I think the main insight comes from the fact that humans have continual thinking -- whether that's while sleeping, walking, daydreaming, working, etc.
Gwern outlines possible ways to combat this phenomenon, trying to promote more unique connections among ideas:
Day-dreaming loop (DDL): a background process that continuously samples pairs of concepts from memory. A generator model explores non-obvious links between them, and a critic model filters the results for genuinely valuable ideas. These discoveries are fed back into the system’s memory, creating a compounding feedback loop where new ideas themselves become seeds for future combinations.
It's obvious (and he notes) that it isn't difficult to implement this. Just toss two of the concepts into a system and then find nonobvious connections between the two -> it's just expensive.
