Lex Fridman: uncertainty

Transcript context

…Yeah. Sometimes it just goes too far, it gives you this, “Oh, imagine you had this lemonade stand and you bought lemons.” I don’t want that level of analogy. Not everything’s a trivial metaphor. What do you think about the context window, this increasing length of the context window? Does that open up possibilities when you start getting to a hundred thousand tokens, a million tokens, 10 million tokens, a hundred million … I don’t know where you can go. Does that fundamentally change the whole set of possibilities? It does in some ways. It doesn’t matter in certain other ways. I think it lets you ingest a more detailed version of the Pages while answering a question, but note that there’s a trade-off between context size increase and the level of instruction following capability. So most people, when they advertise new context window increase, they talk a lot about finding the needle in the haystack sort of evaluation metrics and less about whether there’s any degradation in the instruction following performance. So I think that’s where you need to make sure that throwing more information at a model doesn’t actually make it more confused. It’s just having more entropy to deal with now and might even be worse. So I think that’s important. And in terms of what new things it can do, I feel like it can do internal search a lot better. And that’s an area that nobody’s really cracked, like searching over your own files, searching over your Google Drive or Dropbox. And the reason nobody cracked that is because the indexing that you need to build for that is a very different nature than web indexing. And instead, if you can just have the entire thing dumped into your prompt and ask it to find something, it’s probably going to be a lot more capable. And given that the existing solution is already so bad, I think this will feel much better even though it has its issues. And the other thing that will be possible is memory, though not in the way people are thinking where I’m going to give it all my data and it’s going to remember everything I did, but more that it feels like you don’t have to keep reminding it about yourself. And maybe it will be useful, maybe not so much as advertised, but it’s something that’s on the cards. But when you truly have systems that I think that’s where memory becomes an essential component, where it’s lifelong, it knows when to put it into a separate database or data structure. It knows when to keep it in the prompt. And I like more efficient things, so just systems that know when to take stuff in the prompt and put it somewhere else and retrieve when needed. I think that feels much more an efficient architecture than just constantly keeping increasing the context window. That feels like brute force, to me at least.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.