Lex Fridman: belief

Transcript context

…I think that most of the time when I’m talking with Claude, I’m trying to map out its behavior in part. Obviously I’m getting helpful outputs from the model as well, but in some ways this is how you get to know a system, I think, is by probing it and then augmenting the message that you’re sending and then checking the response to that. So in some ways it’s like how I map out the model. I think that people focus a lot on these quantitative evaluations of models, and this is a thing that I said before, but I think in the case of language models, a lot of the time each interaction you have is actually quite high information. It’s very predictive of other interactions that you’ll have with the model. And so I guess I’m like, if you talk with a model hundreds or thousands of times, this is almost like a huge number of really high quality data points about what the model is like in a way that lots of very similar but lower quality conversations just aren’t, or questions that are just mildly augmented and you have thousands of them might be less relevant than a hundred really well-selected questions. Let’s see, you’re talking to somebody who as a hobby does a podcast. I agree with you 100%. If you’re able to ask the right questions and are able to hear, understand the depth and the flaws in the answer, you can get a lot of data from that. So your task is basically how to probe with questions. And you’re exploring the long tail, the edges, the edge cases, or are you looking for general behavior? I think it’s almost like everything. Because I want a full map of the model, I’m kind of trying to do the whole spectrum of possible interactions you could have with it. So one thing that’s interesting about Claude, and this might actually get to some interesting issues with RLHF, which is if you ask Claude for a poem, I think that a lot of models, if you ask them for a poem, the poem is fine, usually it rhymes. And so if you say, give me a poem about the sun, yeah, it’ll just be a certain length, it’ll rhyme, it’ll be fairly benign. And I’ve wondered before, is it the case that what you’re seeing is the average? It turns out, if you think about people who have to talk to a lot of people and be very charismatic, one of the weird things is that I’m like, well, they’re kind of incentivized to have these extremely boring views because if you have really interesting views, you’re divisive and a lot of people are not going to like you. So if you have very extreme policy positions, I think you’re just going to be less popular as a politician, for example. And it might be similar with creative work. If you produce creative work that is just trying to maximize the kind of number of people that like it, you’re probably not going to get as many people who just absolutely love it because it’s going to be a little bit, you’re like, oh, this is the out. Yeah, this is decent. And so you can do this thing where I have various prompting things that I’ll do to get Claude to… I’ll do a lot of this is your chance to be fully creative. I want you to just think about this for a long time. And I want you to create a poem about this topic that is really expressive of you both in terms of how you think poetry should be structured, et cetera. And you just give it this really long prompt. And it’s poems are just so much better. They’re really good. I think it got me interested in poetry, which I think was interesting. I would read these poems and just be like, I love the imagery. And it’s not trivial to get the models to produce work like that, but when they do, it’s really good. So I think that’s interesting that just encouraging creativity and for them to move away from the standard immediate reaction that might just be the aggregate of what most people think is fine, can actually produce things that at least to my mind are probably a little bit more divisive, but I like them.…

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