The problem is that the overwhelming majority of input it has in-fact seen somewhere in the corpus it was trained on. Certainly not one for one but easily an 98% match. This is the whole point of what the other person is trying to comment on i think. The reality is most of language is regurgitating 99% to communicate an internal state in a very compressed form. That 1% tho maybe is the magic that makes us human. We create net new information unseen in the corpus.
> the overwhelming majority of input it has in-fact seen somewhere in the corpus it was trained on.
But it thinks just great on stuff it wasn't trained on.
I give it code I wrote that is not in its training data, using new concepts I've come up with in an academic paper I'm writing, and ask it to extend the code in a certain way in accordance with those concepts, and it does a great job.
This isn't regurgitation. Even if a lot of LLM usage is, the whole point is that it does fantastically with stuff that is brand new too. It's genuinely creating new, valuable stuff it's never seen before. Assembling it in ways that require thinking.
I think you're missing the point. This is my own paper and these are my own new concepts. It doesn't matter if the definition of the new concepts are only 1% of the paper, the point is they are the concepts I'm asking the LLM to use, and are not in its training data.
That would be a weird default assumption. It's not hard to come up with new ideas. In fact, it's trivial.
And if you want to know if a specific concept is known by the LLM, you can literally ask it. It generally does a great job of telling you what it is and is not familiar with.
I think it would be hard to prove that it's truly so novel that nothing similar is present in the training data. I've certainly seen in research that it's quite easy to miss related work even with extensive searching.
Except it's more than capable of solving novel problems that aren't in the training set and aren't a close match to anything in the training set. I've done it multiple times across multiple domains.
Creating complex Excel spreadsheet structures comes to mind, I just did that earlier today - and with plain GPT-5, not even -Thinking. Sure, maybe the Excel formulas themselves are a "98% match" to training data, but it takes real cognition (or whatever you want to call it) to figure out which ones to use and how to use them appropriately for a given situation, and how to structure the spreadsheet etc.
