Expert debate can train models on some subjective work

Transcript

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

1. 29:31

   recently and feels like especially related to that is the idea of like how do we make models creative for example because that is like potentially the least binary data out there um and there are companies now that have been started to like basically give models taste because ultimately like it's the same thing where like they the labs need to like buy data from tastemakers and this is something where like yeah i guess amazing taste is a classic thing but it's also perpetually changing you could look at like fast fashion it's just like constantly scraping trends to do something every week but that doesn't feel like the right way to make a model have amazing taste um and this is just in general like so i'm curious this definitely feels like an area where they will always be needing new data and like how do you think models get creative i have

2. 30:21

   a specific way i think about this which is the like tasting creativity uh i look at it as you You have a spectrum of objectivity and subjectivity. Like there are some things that are very objective. You can do unit tests for and you can sort of very much say if it's correct. And then you just sort of go on this journey of like at the end of the spectrum, you have like art and music and so forth that there's no objective truth.

3. 30:50

   And the reality is that all domains other than like maybe a couple that you can sort handpick are extremely subjective and are on this spectrum and some of them are like towards art actually like they're almost at the end of the spectrum of subjectivity and i would say even like legal is actually a good example of that where where like the best lawyer is actually probably really creative yeah and like there's no like laws try to be as objective as possible and like if you if you if you explained each law so precisely and you'll be basically just or like a programming, you sort of program the law, then like you could get it to be more objective.

4. 31:29

   But the reality of it is that you have a lot of debates, you have court cases, you have a judge that ultimately decides, and it's very subjective. So the point here is that the domains that we are already getting the models to be good at, which is like medicine, finance, legal, et cetera, are already very creative and very subjective. And the specific kind of structure that allows for models to be good regardless of subjectivity is is get is gathering human judgment along with Well, judgment, there's always like kind of a mixture and allowing the humans to debate it out, essentially.

5. 32:06

   And, and you got to like, you got to like productize this, right? Like the debates are essentially like peer reviews and you have like, you know, you create rubrics and you have, you have other experts that you have like one expert that creates a set of rubrics. You have another one that looks at their rubrics and basically, you know, they disagree and there's another expert that might break the tie, et cetera.

6. 32:22

   up. Um, and, and, and this idea is basically if, if, if, if something is subjective, but not like maximally subjective, if you, if you debate enough, you will converse to the truth. Like if you, you got to think of it in limits, where it's like, if you have an infinite number of debates with all lawyers on one, one topic, um, you know, you, you, you will probably converse to like the objective truth. And so the problem with this is that as you, as you go towards the, the very subjective areas, um, this is, this no longer holds true.

7. 32:51

   which is if you if you have peer reviews and you debate, you actually diverge and you end up you end up disagreeing way more than than than agreeing, which which is why it's a really hard problem to get models to be very good at the super subjective.