In and Out of Model Responses Explained — Whiteboard Friday

So if it doesn’t, then, for example, we might have a query, like “write me a poem,” that doesn’t refer to anything particularly factual or fresh or topical. So you would probably just get a response straight out of the LLM without it making any of these external grounding searches, and then you get a response.

So if you wanted to influence this response, the only thing you could do really would be to influence the underlying training data of the model.

So to give you some context, GPT-4.0 finished training in, I think, late 2022, and GPT-4.5 finished training in late, I think, maybe August 2024. So this is quite a long time span between model refreshes, and a lot of people are still using models which at this point are coming up to four years old or even more.

So your frame to sort of quickly influence this is nonexistent. And in the long term, you’re basically trying to influence any of the content that goes into the training data for these models, which at this point is basically any written content whatsoever that ends up being ingested by a computer.

It’s incredibly broad. And that obviously doesn’t just mean your own site. That could mean external sites. But it can also, at this point, mean things like books. It’s gotten incredibly, incredibly broad with this underlying training data.