Evals tell you what AI product to build

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 19:19

   do you think this is going to be so important for people building products in the future? Yeah, sure. I think the easiest way to think about it is almost like a quiz for a model, a test to gauge how well it knows a certain set of subject material or how good it is at responding to a certain set of questions. So in the same way, you take a calculus class and then you have calculus tests that see if you've learned what you're supposed to learn. You have evals that test how good is the model at creative writing? How good is the model at graduate level science?

2. 19:52

   how good is the model at competitive coding and so you have these set of evals that basically perform as benchmarks for how smart or capable the model is. Is it like a simple way to think about it, like unit tests for model? Yeah, unit tests, tests in general for models, totally. Great, great. Okay, and then why is this so important for people that don't totally understand what the hell's going on here with evals?

3. 20:17

   Why is this so key to building AI products? Well, it gets back to what I was saying. You need to know whether your model is going to, there are certain things that models will get right 99.95% of the time, and you can just be confident. it. There are things that they're going to be 95% right on and things they're going to be 60% right on. If the model is 60% right on something, you're going to need to build your product totally differently. And by the way, these things aren't static either. So a big part of evals is if you know you're building for some use case. So let's say, let's take our deep research product, which is one of my favorite things that we've released maybe ever. The idea is with deep research for people who haven't used it you can give chat gpt now a an arbitrarily complex query like it's not

4. 21:06

   about returning you an answer from you know a search query which we can also do it's it's here's a thing that if you were going to answer it yourself you'd go off and do you know two hours of reading on the web and then you might need to read some papers and then you would come back and start writing up your thoughts and realize you had some gaps in your thinking so you go out and do more research you might it might take you a week to write some like 20 page answer to this question you can let chat gpt just like chug for you for 25 30 minutes you know it's not the immediate answers you're used to but it might go work for 25 30 minutes and do work that would have taken you a week so as we were building that product we were designing evals uh sort of it at the same time as we were thinking about how this product was going to work and we were trying to go through

5. 21:53

   like hero use cases you know here's a question you want to be able to ask here's an amazing answer for that question and and then turning those into evals and and then hill climbing on those evals so it's not just that the model is static and we hope it does okay on a certain set of things you can teach the model you can make this a continuous learning process and so as we were fine tuning our model for deep research to be able to answer these things, we were able to test, is it getting better on these evals that we said were important measures of how the product was working? And it's when you start seeing that and you start seeing performance on evals going up, you start saying, okay, I think we have a product here. You made a kind of a comment along these same lines around evals that AI is almost like capped in how amazing it can be by how good

6. 22:43

   we are at eBells. Does that resonate? Any more thoughts along those lines? I mean, these models are, they're intelligences, and intelligence is so fundamentally multidimensional. So you can talk about a model being amazing at competitive coding, which may not be the same as that model being great at front-end coding, or back-end coding, or taking a whole bunch of code that's written in COBOL and turning it into Python, you know? And that's just within the software engineering world. And so I think there's a sense in which you can think of these models as incredibly smart, very factually aware intelligences, but still most of the world's data knowledge process is not public. It's behind the walls of companies or governments or other things. And same way, if you were going to join a company, you would spend your first two weeks

7. 23:40

   onboarding you'd be learning the company specific processes you'd get access to company specific data it's you can teach these the models are smart enough you can teach them anything but they need to have the sort of the raw data uh to to learn from and so there's a there's a sense in which Yeah, I think the future is really going to be incredibly smart, broad-based models that are fine-tuned and tailored with company-specific or use-case-specific data so that they perform really well on company-specific or use-case-specific things.

8. 24:22

   And you're going to measure that with custom evals. And so, you know, what I was referring to is just like these models are really smart. you need to still teach them things if the data is not in their training set. And there's a huge amount of use cases that are not going to be in their training set because they're relevant to one industry or one company. I'm just going to keep following the thread that you're leading