Comment by marcus_holmes
10 hours ago
> Nothing about the conception of knowledge as "justified true belief" implies that machines can't accrue knowledge.
They can't accrue knowledge because they stop learning once their training is finished.
10 hours ago
> Nothing about the conception of knowledge as "justified true belief" implies that machines can't accrue knowledge.
They can't accrue knowledge because they stop learning once their training is finished.
That's an imposed limitation, nothing structural. The models do accrue knowledge during training so they already check that box. The LLM companies simply set it up such that newly gained knowledge can't be accrued after that, they could allow it just as easily.
Humans stop learning once their training is finished too, it's called death.
The difference is that LLMs we can talk to them when they're in that dead state.
And of course for an LLM this is an implementation detail. One could just as well keep updating the weights of the same model over time. And its being done, it's called "continual learning". LoRA is also a way of doing that.
By this definition if they could instead buffer recent facts and then offload to their weights reliably and efficiently they would be able to accrue knowledge despite their substrate.
Yes, and they'd have to deal with the same problem that we do: how do we know what is true? what gets counted as a "recent fact" and what gets dismissed as mere opinion or disinformation?
In training they get feedback. Out of training, not so much.
Hard problem
You mean they don't accrue knowledge during the training?