This Week in Startups
00:51:05 · Jonathan Siddarth
[00:51:05] S0: So that's a great question, Alex. So the way these models are trained is there's a step called pre training, where you basically feed the model gobs and gobs of data. The war in pre training is mostly over. Almost all the models are trained on the same subset of the internet. Now the battlefield is post training. And there's a new field called mid training, which is I can get into that later. But with post training, what you do is you have these human experts create question answer pairs where a software engineer might ask a question like, hey, how do you create an app that connects dog walkers to dogs? And can you write it in both Swift and Kotlin for Android, right? And now the system has to generate the app. So that's an example of a supervised fine tuning data set where you give it a prompt and a completion, and the model learns how respond to that prompt, right? So you need experts in different fields. Now, reason this has gotten harder is two years ago, a very low skilled contractor was capable of creating tokens that could have advanced the model. Now, because the floor has gone up, now you need PhDs from Stanford, Berkeley, MIT to figure out where the models break. You first have to break the model. You have to ask a question that literally stumps the model. And then the humans create good question pairs of questions and answers that you then feed into the model for fine tuning, and then the model learns.
00:52:33 · Alex Wilhelm, Jonathan Siddarth
[00:52:33] S1: Ah, it's the question and answer pairs that generates the data for the model. Okay. Now that makes good sense to me. So it sounds though like what you've done is back to your Avengers: Iron Man point. Just got together like a, like a super set of nerds. And basically you're applying the smartest humans to find the flaws in the most advanced LLMs. What happens, Jonathan, when we don't have PhDs that can ask questions that the model can't answer? Because it seems to me like we've raised the bar, but we're raising it towards a ceiling. [00:53:03] S0: Yeah. Yeah. I mean, I've oversimplified it a little bit. So I would say like... So one example I gave is you give question answer pairs. There's another type of data that you generate for reinforcement learning where the humans are generating questions and verifiers. They're not generating the solution. They're generating a way to verify whether the solution is correct or not, which you can do with certain fields like coding, math, hard sciences. You can do that, right? Write a program to sort numbers in Python. You can write test cases to verify whether the program that you wrote is correct or not. [00:53:42] S1: But you're not telling them every single step along the way. You're just saying, I can verify that what you've done, the work you've done, does generate an acceptable answer.
00:53:49 · Jonathan Siddarth
[00:53:49] S0: Correct. Correct. And the version of this for enterprise is you create an RL gym, a reinforcement learning gym. I think this is one of the coolest things in computer science. Just like a gym where humans go to train, this is like a gym where an agent trains. And in these gyms, you create basically clones of different websites like a DoorDash or an Uber Eats or a NetSuite or a Salesforce. You create this virtual environment with prompts or workflows and ways to verify whether the task is completed. Now the task could be, Hey, salesperson, why don't you research everything you need to about this prospect and then update Salesforce in this way? Long as when you create this virtual environment, you create these prompts and you create these verifiers, now an agent tries out different combinations. The agent is basically trying to use the tools in an appropriate way to actually complete that task. And it's a cool way to learn whether the agent is learning through trial and error, how to execute a complex multi step workflow. You're not teaching the agent, go to LinkedIn first, go to ZoomInfo first, and then check out Salesforce for prior conversations. You're not teaching it that, but it's learning. It's learning just based on feedback on, okay, if I accomplished this, I get a reward. And then it learns how to operate in a way that maximizes future rewards. It's really cool. To your question of what happens if we run out of human intelligence, right? Fortunately, I think we are still a significant distance away from that. We've run out of internet data, but there's lots of intelligence trapped in the minds of humans that is yet to be transferred from human minds to machine minds. I would say, Alex, even the way like, I love your show. I love what you and Jason have done with this show. But when you're thinking about how to interview somebody who's on the show, what type of questions to ask, what type of prep to do, I guarantee that knowledge is not distilled into the models from OpenAI or Anthropic or Meta or Google. The only way that's gonna happen is if we hire somebody like an Alex or a Jason to work on Turing, to use our tools. And we didn't get into this, but we have tools that make it easy to keep the quality of the data high. And we have AIs that will assist you in generating this data so that we can make this scale. Pun pun not intended.
00:56:18 · Alex Wilhelm
[00:56:18] S1: No, I was gonna let that one slide. We don't allow more than two dad jokes per episode. So we have to we have to save them, if that makes sense. So here's my question. I I I understand that scale being subsumed by Meta has been great for the Turing business. You told me before the show you guys are at, 9 figures of revenue and profitable and growing. And honestly, hell, yeah. I love that. But if I'm OpenAI, why wouldn't I try to replicate what Turing has built inside of my domain? Because if there's one thing that large AI model companies have today, it's a lot of access to capital. And so as we've seen from, you know, Meta trying to buy literally every human who's touched an AI before, there's a lot of money to play with here. So are are you guys just so good that OpenAI doesn't wanna try to replicate what you've built? Or is there something else that I'm missing in that it's good to have a third party versus an internal group be kind of telling you where you might wanna work on things?
00:57:10 · Jonathan Siddarth
[00:57:10] S0: Yeah. I mean, that's a great question, Alex. Firstly, to advance towards ASI, you need three things to move in parallel. Actually, you need four things. You need research and algorithms. You need compute. You need data. And you need the application layer. Those are the four things, right? Research, compute, data, and applications. Now, the labs are really good at research, right? And kudos to OpenAI, Anthropic, Meta, Google, Apple, all of these companies for advancing the frontier forward. So they're spending their energy there. We could flip this question to also ask, why don't they build their own compute? And in some areas, the labs are investing in it. I mean, we've heard of Google having TPUs and other companies trying to do that. But NVIDIA clearly has the edge today. And there are some good companies like Grok, Cerebras, etcetera, that are also building custom chips. So compute is also advancing. And the third pillar is the data pillar, right? Now, generating this data is exceedingly complex. Again, three dimensions. It has to be complex. It has to be realistic, meaning it has to mirror how a real human would interact with the model. And third, it has to be diverse. It's very hard for a lab to get all of this in house. I'll give you an example. Today at Turing, this month alone, we are hiring PhDs across the board in physics, chemistry, math, biology at the level of granularity of somebody who's an expert in dark matter, somebody who's an expert in black holes, somebody who's an expert in molecular biology, right? Like these very niche fields. And you ideally need these humans part time because the way these humans are good at the job of training the models is because they are also good at their day job, which keeps their skills sharp. And once you've ingested that knowledge, you might want to move to the next frontier and the next frontier and the next frontier. So you need a platform or a partner that can scale up very quickly to elite talent, manage the talent to make sure the talent is generating data part time, the data is high quality. We have to build a ton of tools. We have this platform called Allen, where AI is used upstream of the human to minimize the work for the human. AI works alongside the human, and AI does quality control. It may sound a little dystopian. It's like AI overseeing humans to improve AI.
00:59:41 · Alex Wilhelm
[00:59:41] S1: I'm not afraid of our AI future, personally. So that doesn't... My P doom, I guess, is very low. But I wanna go back to your black holes dark matter point because I'm a bit of a science fiction nerd, I'm also a bit of a space nerd. So those are topics that are near and dear to my heart. But if I spend a lot of time helping a particular LLM better understand why we think about dark matter, why we came up with the idea, you know, galaxies and spinning and not having enough matter, blah, blah, blah, blah, blah. Will that work to make the LLM smarter in that particular domain have spillover effects to other areas? Does it make it more intelligent writ large or more intelligent only in that specific domain? I'm not sure of the answer here, so I figured I'd ask.