Venture with Grace
00:20:53 · Ori Goshen
[00:20:53] S1: Right. So many questions. A to unpack, but I'll start with the beginning about the open rates. I think something very important to realize and related to the different waves and shifts we're seeing in the market. So if you think about it, we had the chat GPT moment back in 2022. Then we had twenty twenty five ish, the first, I mean, '24, we had the reasoning models, 2025, we had the agentic models, let's call them. So I think what's interesting in this wave, it just unleashed so much demand, right, in the form of these agentic workflows and particularly in coding. I mean, you look at the majority of tokens in the consumption in the world right now, they're coding related tokens. So I think that what happened is there's this point in time you can actually identify. In insights, it's easier, but if think you about OPUS 4.5, I think that was like the true native agentic model release that unleashed a lot of the... Unlocked a lot of the demand that we're seeing right now. That was a point in time. Then if you look back then, I think it was end of twenty five, I think there was still a pretty significant gap between open weight models the frontier model at the time. Now that gap has narrowed significantly. If you look only six months into the future or so, mid twenty twenty six, May, June, we had significant releases like Kimi K3 and, you know, new DeepSeek and Qwen, all came about the same time. These were the first, I would say, agentic open weight models. And that's why we're seeing kind of that, I think explains the gap narrowing that we've seen in the last few months. That explains why there's such a strong movement toward open weight models, they suddenly became useful, they became a real alternative to the frontier. And so we're what? We're talking about, you know, May, June, it's like less than six months ago, right?
00:23:34 · Grace Gong, Ori Goshen
[00:23:34] S0: I have a question. So do you feel like open weight is like a different kind of... When people use open weight, does that cancel out their usage of inference or no? What inference? So obviously I feel like inference is essentially a layer of people tailor the model to their own company or needs. So when people using OpenWait, essentially they are just technically downloading like something, and then they are like, just like using it in house, but within their own like data and stuff. So assuming it's not like conflicting, no? No. [00:24:16] S1: I think it's it's it's not really conflicting. You can take open weight models, you can serve it in your own native environment, like you're in your cloud, you can serve it with your one of the neo clouds or one of the inference providers. So you can you can choose whatever works for you, and then you can optimize the the inference layer to the workload that that that you're... That you have in place. I think so it gives you a lot of flexibility. Right? I think the next layer is that you can start customizing the open weight model. That's a big benefit. [00:25:00] S0: So [00:25:01] S1: and customizing is worthwhile if you're, you have some proprietary data, something that is not achieved, is not accessible on the public domain. So if you have, let's take an example, you mentioned Deel, which are great partners of ours, we released [00:25:20] S0: them I'm for technically wearing a Deel socks. Got it from the Valkyrie team they hosted us.
00:25:27 · Ori Goshen
[00:25:27] S1: Oh, that's great. Shout out to Deel. So, yeah, big shout out, great partners of ours, and they... We released together Pangaea, their first model for payroll. So if you think about companies like Deel that have very special, unique, proprietary data that is hard to acquire in the public domain, that's something that they can leverage. So having a specialized model in payroll is actually better than the frontier on payroll questions or payroll related tasks because of that very unique data. So that's, I think, very important benefit of the open weight model. Now you can customize them and create your own version. You can start seeing the benefits. So it's not just data that is sitting there and is used as context, but something that you can actually leverage to create a competitive advantage in the market. So I think that's an important layer. The other layer that we related to is now that you can customize and leverage the data, in many cases, can actually use even smaller models and achieve same level of performance of general purpose models, only with really small models, much cheaper to serve, much more economical, then you can start scaling. So one of the... Another another partner of ours that you mentioned, a company called Gong, the leading... Of the leading revenue intelligence companies. These companies try to apply these agentic workflows on all the transactions, like in case of Gong, it's all the calls and emails that they process. It's just huge amounts. And so if you have a specialized small model that's very performant, you can now scale and use that agent on top of all of those transaction.
00:27:44 · Grace Gong, Ori Goshen
[00:27:44] S0: What's the difference between, like, small model and and, like, open weight? So I feel like small model is kind of, like, tailored towards, like, some sort of, like, specific tasks. Right? Yeah. And it's, like, trained for that particular task, but isn't that that, like, kinda similar to, like, if you just [00:28:01] S1: It's orthogonal. So even in frontier models, you have different sizes, right? You have like Luna Terasol, Haiku, Sonnetopolis, right? So you have different sizes. I mean, and if you think about open weight models, their families also come in different sizes. The unique thing in Open Weight Models that you can customize them, you can then leverage the data that you have in place. Now that you can customize them, you can find the right model size that is most suitable for what you're trying to achieve. Now if you have that muscle of training models, you have the recipe, how to train models, you can find the model size that you need to achieve your task. And of course, obviously, the more data you have, the more data you collect, then you can do it with smaller and smaller models and achieve high level of cost effectiveness. So that's a very important cost and scale. Basically, different sides of the same coin are really important in this game of agents.
00:29:30 · Grace Gong, Ori Goshen
[00:29:30] S0: Speaking of agents, do you feel like, where's agents going in the business universe? Whenever I chat with executives from bigger companies, especially the traditional companies, people are very slow at adopting agent. I think the most advanced agent would be people go on Slack and then it's like a Slack bot. And then for a traditional company that's outside of Silicon Valley. But obviously, you guys work with some really good, just AI company or pretty, I guess, tech native companies. So people have different adoption level of agents. So maybe we could talk about, since you mentioned about, you call them forward deployed researchers. What's the difference between forward deployed researchers versus engineers? And in the agent first universe, how do you think like, you know, AI21 Labs like perform in this kind of like, like the AI kind of like landscape? [00:30:34] S1: Right. So I would segment the market broadly speaking into three types of organizations and companies. And there is some correlation to the level of maturity and the pace of adoption. So you have the AI natives. These are companies that were born, roughly speaking, post chat GPT, [00:31:00] S0: and
00:31:01 · Ori Goshen
[00:31:01] S1: their core product is AI based. So, know, we can name the names of the Harvey's and Rogos and Cognition and so on and so forth. So these companies, relatively speaking, they have the talent and they build that training, model training muscle almost from the get go, some more than others, but roughly speaking, they are pretty proficient in the crafts of training models and deploying them and scaling and so on. Had to, because that's the only way to survive, right? If you want to be competitive, you want to have the right economics, if you want to have the right edge, you have to own your own model. That's a very natural flow. So that's one segment of the market. The other segment is digital native. These companies that were born in the cloud era and have huge user bases, right? They've built amazing SaaS businesses. Now they transition into the agentic era. Now they are becoming AI natives. If you come to think about all of these products, now AI is not like an afterthought, it's becoming core to what they offer to their customers. So that's the main segment we are focusing on, helping those companies to actually scale, to actually become AI natives. And I think one of one of their challenges is how do you build the team? How do you build the framework? How do you bring the right amount of compute, and we really help accelerate these companies. And we are happy. We don't call them customers, we call them partners. And the reason is we don't believe in this new world of AI, there's build versus buy. It's not necessarily build versus buy, there's a third option of partner. As a company, you can work with partners like us that can augment, accelerate your AI development, becoming an AI native. So that's second segment. The third segment is a more traditional enterprise as you think about, you know, finance and insurance and healthcare and all these more established businesses, but they have, you know, very sophisticated machine learning and engineering operations, very, very sophisticated. They're also... They've done the transition to the cloud for many years. Most of them have done it, you know, successfully.