The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch
00:13:13 · Harry Stebbings, Adarsh Hiremath
[00:13:13] S1: What does no one tell you about scaling that you wish they had told you? [00:13:16] S0: Scaling culture is harder than scaling software. When you're adding people to the team very, very quickly, there's this dynamic that the culture that you create with the first, you know, 20 people is in some ways the strongest the culture is ever gonna be. Ensuring that that culture stays strong as the company grows, does new things, and new people enter the company is really, really challenging, but in some ways, the most important part of building a legendary company. [00:13:42] S1: We mentioned scale earlier. One of your investors said to me that you are mostly doing data labeling for foundation models. Do you think that's fair? And is that a niche market or a wedge into a much larger market in your mind?
00:13:56 · Adarsh Hiremath
[00:13:56] S0: Actually, insight about the market is that human data and talent assessment have actually become the same thing. Where, you know, I can take you back five years where when we think of this data labeling or human data stuff, it's essentially a crowdsourcing problem. Let's say Waymo wants a bunch of their images labeled. You get a bunch of people across the world to draw boxes around stop signs to make the model better at classifying stop signs. Fast forward to today, and the nature of human data work has changed a lot. Now it's GPT four point or whatever model is not good in a particular domain, so we actually need an expert to make the model better in that domain. And figuring out who that expert should be is 100% a talent assessment problem and is a perfect application of the platform. With a lot of the labs that we work with, we're able to figure out who are the exceptional people in very, very specific domains how those people work with the labs. And the interesting thing about this is that it's essentially a forcing function on our long term objectives. Right? When you think about Mercor building this global, unified labor market, what do we need to make this happen? We need tons of smart people on the platform, and we need the ability to predict job performance and figure out what those people should be doing, which happens to be the exact set of problems that a lot of AI labs are having.
00:15:15 · Harry Stebbings, Adarsh Hiremath
[00:15:15] S1: Can I ask you then? When we think about, like, the AI labs today, I heard through the grapevine that as you mentioned that you work with some of the top AI labs. Merkle experts, how does that fit into these labs? What does that partnership look like? Just help me understand this. [00:15:29] S0: It looks exactly the same as placing someone to work at any company. So just like Merkor might work with startups making their first hires or companies hiring in a more traditional full time capacity, it's the exact same thing for, a lot of the large AI labs. They'll hire people through the Merkor platform to essentially help with post earning models. [00:15:48] S1: When you look at today, what is the satisfaction on a higher basis? Is 90% of hires successful? Is 60%? What are the metrics that you place? And what is the one core metric that you use for the success of the business? [00:16:01] S0: Customers keep growing their relationships with us. Net retention is over 100% by by a large margin. You know, as long as they keep expanding, it means that we're doing a good job [00:16:11] S1: at finding the right people. In terms of an infrastructure basis, what models are we sitting on top of today? [00:16:16] S0: It's interesting because the model landscape is just changing so so quickly. But we we leverage a variety of models and have been particularly thrilled with the OpenAI models. [00:16:28] S1: Have we always been on OpenAI predominantly? [00:16:30] S0: We've always used OpenAI in some capacity.
00:16:30 · Adarsh Hiremath, Harry Stebbings
[00:16:30] S0: We've always used OpenAI in some capacity. [00:16:33] S1: If improved in terms of any aspect of ModelAir, what would make the biggest improvement on the business and the product today for you with McCore? [00:16:41] S0: I think a concrete example would be the AI interviewer. We've built the product in such a way that whenever the models improve, the experience for applicants on our platform also improves pretty significantly. In general, this is something that's been, you know, on our mind. Right? There's like this huge wave of models getting better and better and better. And can we ride that wave to make our product better and better and better? So to summarize, well, we leverage LLMs and all these models throughout our product. The whole product gets better as the models get better. [00:17:09] S1: What do you think will be the next generation of models in terms of what they look like first before we get to training data? [00:17:16] S0: The whole market is shifting to reinforcement learning. You're already seeing this with o one, o three, the deep seek models. As a result, I think we're gonna see really, really powerful models in specific domains that can reason extremely well. And I think that'll be really, really exciting and will unlock just a huge number of use cases across a variety of different industries and domains. [00:17:43] S1: Do we live in a world of many, many specialized models, but very fragmented? Or do we live in a world of monoliths with one or two very horizontal platforms?
00:17:43 · Harry Stebbings, Adarsh Hiremath
[00:17:43] S1: Do we live in a world of many, many specialized models, but very fragmented? Or do we live in a world of monoliths with one or two very horizontal platforms? [00:17:53] S0: I think we'll live in a world with many, many models with different use cases. Right? We're already seeing this with a lot of application layer companies, right? Where they all have these specialized use cases for how they want to leverage the models. Right? For us, it's hiring and beating the expert hiring manager. For another company, it might be financial analysis in a specific domain. Across each of these use cases, I think these companies will need to make their models better for for their own purposes. [00:18:22] S1: How fair do you think the analogy is that the model landscape will be very much like the cloud landscape in terms of three or four juggernauts and it being very hard to switch out of? Do you agree that it's hard to switch out of them, or do you think given the model transients, it's actually much easier and much less defensible in that respect? [00:18:41] S0: There will only be a couple of companies that are able to build these foundation models that everyone builds off of. That analogy roughly holds. I don't anticipate there being 20 companies training foundation models that can all be be leveraged in the same way someone might leverage, OpenAI, for example. [00:18:58] S1: In terms of, like, the post training data side, how much will be human data versus how much will be synthetic data moving forward?
00:18:58 · Harry Stebbings, Adarsh Hiremath
[00:18:58] S1: In terms of, like, the post training data side, how much will be human data versus how much will be synthetic data moving forward? [00:19:06] S0: I think a lot of it will be human data going forward. I think a great example of this is Evals. Evals for models definitionally have to be outside of model capability. In order to see whether a model is doing well at a particular task, you need to have an Eval set created by humans that is better than the model at that particular task. And humans are gonna play a huge role in that, for example. And I think there are a whole set of other use cases, whether it be SFT, RLHF, RL environments, like how the models of tomorrow are being trained, that all require these expert humans to essentially teach the model how to get better. [00:19:42] S1: To what extent would you say that data is the bottleneck that prevents model improvement more than compute or algorithms? [00:19:49] S0: Data is the bottleneck. I think that would be an accurate statement. [00:19:52] S1: Why then does so many people tell me, including Jonathan at Grok, that actually synthetic data is often more high quality? It doesn't involve the drags of the Internet like Reddit, in a lot of cases being included. And actually, you'll see this exponential increase in model performance due to actually mostly using high quality synthetic data, not low quality human data. Why is that wrong?