So then you are talking about weight updates, not just memorization, which makes it much harder, and right away really messes with the business model of cloud-based AI where it's the same model being served to everyone.
Even if we had an algorithm to incrementally update weights without catastrophic forgetting (I don't believe we do - but doesn't seem like such a tough problem), then this implies that everyone has their own personalized model, else if you combine all these updates there is no data privacy. Cloud serving also really depends on everyone using the same model so that you can batch requests and not reload weights for each user.
The much more achievable goal, without needing to upend the whole serving model, would be just to implement continual "episodic" memorization (text-only maybe), but even with compaction/consolidation the next question would be how do you retrieve these external memories and get the necessary chunks into context when needed. Some sort of vector store, perhaps? How do you avoid vendor lock-in - perhaps have agents store/retrieve vector-store context chunks from a vendor-agnostic cloud store?
So, even the most basic form of continual-learning-like enhancement becomes tricky. What you'll first see is presumably just enhancements of the vendor-specific memory mechanisms that are already available.
What I would consider as true continual learning, close to what an animal or human does, would require much more extensive architectural and deployment/business model changes - since then we're really talking about more than just an update/recall problem, but rather the whole autonomous agentic loop of predicting/acting/failing/learning/etc with innate traits like curiosity (prediction failure) and boredom to make sure the agent is exposed to learning situations in the first place. At this point you are building an artificial brain, not just an LLM. Some companies such as Google/DeepMind have a more ambitious definition of AGI (more than just an LLM) that is perhaps a step in this direction.
Even with this sort of animal-like continual exploration/learning, you still wouldn't have something that is human level, but at least much closer in terms of ability to learn.
The reason i am asking is about research mathematics. In this application what is really important is continual intuition and theory building, this takes time and trial and error. However this is very fruitful in solving problem. Currently the LLM can do mathematics that is somehow already in their weight very well. See the recent things. But i suppose true research would require continual learning that goes beyond the context window. Thats why i am wondering if this solvable. My intuition is that it is easily solvable, hence we could make a true mathematical continuously building theory to prove things.
For that sort of application then maybe a chunked vector store memory would work if 1M context is too small to hold the entire history? In Terrance Tao's ChatGPT discussion it seemed that the LLM was making good use of the in-context information.
I think for true, especially open-ended, research you really also need things like curiosity, but perhaps a lot can still be done with the same sort of prompting that got the Jacobian result - just tell the model to keep on going, investigate anything that seems interesting/unexplained, etc?
The issue remains the context window. I haven't seem any example of a "memories" system that the agent can update over time without it devolving into slop. The magic of the human brain is that we seem to have decently sophisticated heuristics parsing/saving memories as well as for letting memories that aren't accessed decay.
Yes i agree, but somehow knowledge got put into the LLMs head during train time, but why does it not work during inference time? Too little examples? Do llms know stuff with exactly one occurence in the training data?
It's like that all that knowledge exists as a vague memory. I've found if I ask them specifics about any book I've read (w/o allowing tool calling), when they get it wrong their hallucinating things that are often correct or wrong in subtle ways. Are they're just completely out of pocket.
Even if we had an algorithm to incrementally update weights without catastrophic forgetting (I don't believe we do - but doesn't seem like such a tough problem), then this implies that everyone has their own personalized model, else if you combine all these updates there is no data privacy. Cloud serving also really depends on everyone using the same model so that you can batch requests and not reload weights for each user.
The much more achievable goal, without needing to upend the whole serving model, would be just to implement continual "episodic" memorization (text-only maybe), but even with compaction/consolidation the next question would be how do you retrieve these external memories and get the necessary chunks into context when needed. Some sort of vector store, perhaps? How do you avoid vendor lock-in - perhaps have agents store/retrieve vector-store context chunks from a vendor-agnostic cloud store?
So, even the most basic form of continual-learning-like enhancement becomes tricky. What you'll first see is presumably just enhancements of the vendor-specific memory mechanisms that are already available.
What I would consider as true continual learning, close to what an animal or human does, would require much more extensive architectural and deployment/business model changes - since then we're really talking about more than just an update/recall problem, but rather the whole autonomous agentic loop of predicting/acting/failing/learning/etc with innate traits like curiosity (prediction failure) and boredom to make sure the agent is exposed to learning situations in the first place. At this point you are building an artificial brain, not just an LLM. Some companies such as Google/DeepMind have a more ambitious definition of AGI (more than just an LLM) that is perhaps a step in this direction.
Even with this sort of animal-like continual exploration/learning, you still wouldn't have something that is human level, but at least much closer in terms of ability to learn.
I think for true, especially open-ended, research you really also need things like curiosity, but perhaps a lot can still be done with the same sort of prompting that got the Jacobian result - just tell the model to keep on going, investigate anything that seems interesting/unexplained, etc?