Venture with Grace
00:37:17 · Grace Gong
[00:37:17] S0: Can we talk about... Let's say like, I'm just using deal as example, because it's like a very straightforward as like, I think most people have like used like deal or like equivalent companies before. And I guess when you guys go in, first of all, how do you guys think about... You mentioned forward deployed researcher, right? So at what point do you go into the company and how do you think about monetization? You mentioned instead of buying buying a service or buying the thing, there's a third way, which is partnering together. So can we talk about what does a partnership look like with the middle layer of company that's kind of like tech native, but not necessarily like the most like, you know, front and center of like AI.
00:38:05 · Ori Goshen
[00:38:05] S1: Right. Yeah. So it's, I think it's pretty straightforward. We, you know, we come with a very clear purpose of building a model or building an AI system, like optimizing an AI system to achieve certain goals. And then we work in different types of, I would say, mode of operations. We have one mode of operation which is called embedded, which we work shoulder to shoulder with the team, an existing team, it could be a machine learning team, it could be a data science team, it could be a team of engineers, we'll work together to achieve a task. So that's one mode of operation. Second mode of operation is what we call manage, where the company would essentially create a spec. We may help with that spec creation because in the AI world, creating a spec is non trivial. We will do that. We'll help with that as well. And once you have the spec, the spec means what level of quality you need to achieve, how the evals look like, the rubrics, how do you what the costs should be like, what the latency, all these relevant dimensions that needs to be specified. And then once we get that spec, we actually translate it into reality. So we train the model, we provide the optimized harness. This is a really interesting point because the harness and the model really interplay with each other, so you want to make sure they're compatible and are optimized towards the same function. We even optimized inference endpoint to that... Towards that model architecture, so you get a full managed endpoint Mhmm. For the workload that you're trying to serve. So that's the managed and we'll serve it, we'll do the inference for you. We'll keep training the model so it will improve over time. We'll do that kind of as a managed service. So we support both mode of operations and we price them accordingly. Of course, try to reach as much as alignment as possible. We have our customers, we're very value driven. We really care about the value that these companies unlock. Just to give you an example, I mean, the models that we build with Gong and Deal and all these other great partners achieved 50x less cost than the frontier, exceeding frontier quality. It's just a great, great value that is being unlocked by these specialized models. So we're trying to align interests, and by the end of the day, these are, I think, really important... A directional decisions for these companies.
00:41:28 · Ori Goshen, Grace Gong
So we are pricing it in a way that is we're trying to simplify things. We have like packages that are very straightforward. Are correlated to the amount of effort and complexity of doing that, but they are as aligned as possible with the value that is created to the customer. That's the way we structure it. [00:41:54] S0: We
00:41:54 · Grace Gong, Ori Goshen
[00:41:54] S0: We [00:41:59] S1: really believe in our partners, like the partners we work with. The most important aspect is their growth. So if they grow, we grow. And that's, I think, a very healthy way to grow a business. And it's just the sheer size. That's what makes me excited. I'm always thinking, I wake up every morning and I'm thinking about the amount of companies that would need to build their own AI, that would want to fulfill themselves, that would want to create something of value, a new type of intelligence, is just incredible to me. I speak with customers every day, every industry you can think of, this touches a nerve. So I'm just very excited about the opportunity. And just to give you a sense, I think that the market for specialized custom models about twelve, fourteen weeks ago barely exists. Now it's everywhere. And I don't think it's a fad. I think it's here to stay. I think that's as more... As open source, open weight models become more and more popular, and they are. I mean, also another interesting stat. Back in June, open source token consumption was about 1520%. Now, this was June, now November, it's about 80%. 80% of the tokens in the world are open source, are based on open weight models. So that gives you a perspective of where the market is heading and the opportunity to create these specialized AI systems and models.
00:44:02 · Grace Gong, Ori Goshen
[00:44:02] S0: I wonder, so when you're working with a deal, so how many people from your company will actually be forward deployed to their company? Obviously we don't have to say the specific, but just curious, and then how do you separate the company into different... Do you separate them into different sector? Do they work with similar ish companies? Or how do you make sure that it's not like you are doing something scalable? Because nowadays I feel like AI talent is more interesting. [00:44:36] S1: I think you you have a great question here. It's pretty industry agnostic. [00:44:42] S0: Mhmm. [00:44:43] S1: And we're seeing similar technical problems Across all industries. Across across industries. [00:44:50] S0: Exactly. The biggest problem that you feel like it's across the industry people have to solve?
00:44:56 · Ori Goshen
[00:44:56] S1: I think the process of defining what you want, defining kind of the rule of the game, what's valid, what's not, what's what good looks like, how do judge a good output, how do you judge a good outcome, is the bottleneck. And I think even if you have good conceptual understanding of that, translating it into an eval and an eval framework is hard. And and I think that's... And then and then you need to apply all the different training algorithms, right, and know which one to apply when and which would be effective when and so on and so forth. So that's kind of the third bottleneck. But what we've been... And that goes back to your scalability issue. What we've been building ourselves is actually a platform that identifies the work of a researcher in the company. So every FDR in the company, he has two responsibilities. First and foremost, make sure the customers are happy and are getting huge amount of value from their work. And second is that they're being educated because part of the value is just by doing this, the company is doing it and being involved in the process. Now they suddenly create that skill. The third job of FDRs is to identify themselves. How would they, when they go to the next deployment, are able to streamline to make that process much more, much faster, much cheaper? How they when they solved an issue, when they applied an algorithm in a very specific way, how do they create that skill, if you will, so that when they come to the next deployment, they can repurpose, they can reuse that skill. So we're doing this on an ongoing basis, actually what we've seen where, let's say, mid year projects of pretty complicated RL training, you know, would take several months with all the debugging and the reward hacking and all the things that you need to address there. Now it takes a week or two for our experts using those tools. So that's the way you make it scalable. And it's less a question of this industry or the other, because we see a lot of similarities in the types of problems. Definitely, if you kind of translate them into the technical language, it's more about, you know, how do you, a sense, platformize and how do you identify the process so you can actually scale. And that's what we're seeing in the field.