What would a language model trained on all of the literature present in 1500 look like? No doubt it would spout a lot about God and the geocentric model.
Would we expect it to ever develop the heliocentric model, Newtonian mechanics, or Ricardian economics?
Exponential growth is self-similar at all scales, so suggesting that LLMs can push the envelope of 21st century knowledge and beyond is akin to claiming that our hypothesised 1500s GPT could do the same in its own time.
I'm not really making a judgement about this question either way (although I certainly have my own view). I just think it's worth considering that belief in the former claim also requires belief in the latter claim, and that if you believe one but not the other, perhaps you're overindulging in scientistic mysticism.
In my view, I think it's far more reasonable to see LLMs as machines that 'fill in the gaps': both in the most literal sense, but also metaphorically. At their best they might allow us to infer knowledge, like mathematical proofs, that are already present in literature but are left unstated. I don't believe they have the means to push beyond into new territory, however, because I've not seen any indication that they possess world models.
Note that it's very easy to pretend one possesses a world model under the scrutiny of statistical tests by looking at the artifacts produced by other world models & imitating them. Humans have highly evolved intuitional discrimination, & so easily fall into the trap of seeing world models where none exist. But an LLM trained on the knowledge of the 16th century would be no more capable of identifying missing knowledge in a corpus of 1500 than an LLM of the 21st century could do the same for us.
Put another way: the dimensionality of an LLM's output is limited by the dimensionality of its own training data. Even the 'most perfect' LLM, that could exactly recreate the entire 15th century corpus purely through statistical inference, would - I believe - still be incapable of discovering Newtonian mechanics, because Newtonian mechanics sits outside of the dimensionality of the corpus and is knowledge that *requires* a world model to discover.
@jsbarretto This is one of my biggest fears about AI. Right now it looks incredible because it’s caught up with the latest human discoveries and creations. Over time, though, it may stagnate, and worse, it will reduce the incentive for humans to keep evolving themselves. It’s a complex dilemma.
What's more, this is a testable idea! I think it would be really interesting to collect a corpus of 500 year old knowledge, train a modern LLM on it with modern statistical methods, and see whether it can produce *any* modern knowledge. I expect it to not be capable of doing so, but I'm sufficiently unsure of myself that I'd love to see this done in practice.
@jsbarretto this just came across my feeds today. Looks like the exact model you'd like to test.
https://github.com/haykgrigo3/TimeCapsuleLLM
@bmitch Very cool! Unfortunately I suspect the corpus is limited enough that there's no good way to test things like the emergence of extra-dimensional concepts (concepts not already present in the corpus) but thanks for pointing me at this!
@jsbarretto most likely an LLM trained on 500 year old texts won't even seem remotely intelligent because there is not enough training data for it to imitate natural language sufficiently well. There is a reason why they had to scrape the whole internet including a large part of all books ever written to produce these things. But maybe I'm wrong on this. In any case I don't think they would produce anything really new, just slight variations, probably mostly inaccurate, of the information already present in those texts.
@jsbarretto I would actually love to see the results of such an experiment. I think it would be very helpful in shaping how we approach this new thing and what we should expect. The better we plan for it, the fewer disasters we’ll face.
@jsbarretto I'm not sure if it's actually all that testable. LLMs have their training steps include (actual) humans a lot and that introduces a modern information bias that, while maybe not impossible, would be extremely hard to account for and I don't see how one could even verify that it has been appropriately accounted for.
