Along these lines I’m curious how a scratch trained model with a few built-in tools (memory, calculator, ?) available during a single generation would behave. Not tool calls as turns.
replya single generation would still allow looping in the "CPU" because the instruction set can do it. The model is almost a Turing machine (except for the limit on memory)
replyAs I understand, the problem of continual learning is that if you continue training, it will "overwrite" previous knowledge and this CPU will be lost
replyThere is nothing in theory preventing you from training while keeping a subset of weights locked.
replyHow do you ensure that subset of the weights will continue to be accessible through the others?
replyI don't understand your question.
reply[flagged]
reply