AI labs are shifting from crowdsourcing to expert data

Transcript

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

1. 13:58

   kind of differentiating from each other? It seemed like scale was really in a position to run away but then now there's been a bunch of kind of new players in the landscape like how do you how do you think about that world yeah i think like the key thing that most people don't understand and the like data annotation evaluation landscape is just the shift in the market and how dramatically different it is from what it used to be two years ago because when chat gpt came out it was the models like weren't that good it was easy to trip them up they were making mistakes left and right. Even a high school student or college undergrad could do a lot of completions or evals to help improve the models in this crowdsourcing fashion, where they run these huge pipelines to get hundreds of thousands of pieces of SFT or RLHF, SFT being inputs, outputs,

2. 14:44

   RLHF being choosing between a bunch of different preference options like you would see in ChatGPT. But as the models got really good, that crowdsourcing model started to break, Because you needed really high quality people that would work directly with researchers to help them understand why is this model doing well? Why is it not doing well? How can we create this really complicated data that helps to trip up the model and actually reflects the real world things that we want to automate?

3. 15:12

   And so our platform of finding exceptional people that you would want to work with was perfectly positioned for that, and that we can hire these really high quality people super quickly. And that caused us to take off and have all of the traction working with the big labs. And I think that trend will continue. And that like the companies that are stuck in these like super high volume crowdsourcing pipelines are probably are certainly going to see a lot of churn.

4. 15:38

   And it's going to be the new players that understand the direction the market is headed and lean into really high quality talent underpinning it that are going to continue taking a lot of market share. Do you think there'll always be demand for, I guess, humans in the data labeling process? there is you know obviously more and more that can be done with these models or big model gets really get one task and then can train smaller model like how do you see that evolving over time yeah the way I think about it is that so long as there are things in the economy that humans can do that models can't do. We will need to create evals or RL environments so that models can learn how to do those things. So I think there's certain domains where that's just going to get solved sooner than others, right? Like within math or even many parts of code, like you

5. 16:27

   don't need that much data. It's super verifiable. The models will solve those problems. But then there's other domains that are like much, much more open-ended. What makes a good founder when we're assessing them. Honestly, a large chunk of knowledge work domains, maybe a majority of them are these open-ended problems that are really difficult to verify and understand what good looks like. And you just need to get all of that understanding that the models don't have into the models. And that's why I expect orders of magnitude increase in the human data and evaluations market over time. If I understand correctly, you guys,