Venture with Grace
00:04:14 · Ori Goshen
[00:04:14] S1: Time time moves. Five. It it seems like fifty years in AI. Yeah. But... So we we started the company, Yoav Shoham and myself. So Yoav was a professor at Stanford for many years, and he came back to Israel. That's how we kind of got together and started this AI lab company, which back then was really out of the ordinary. So back then, if you had an idea in AI or you want to do some research, the most natural place to do it is in academia or in Google in some shape and form, or there was like Facebook AI research. There were very, very few labs that you could actually do AI research. So, but the kind of the original idea of the company was first to establish an AI lab in Israel. And second, we really had this thesis that deep learning is necessary but not sufficient. And, you know, if you think about it, up until now, learning is kind of the dominant method in AI, and it's pretty amazing what we've been able to achieve with the scaling laws, right? So our thesis was that this, you know, deep learning statistical method needs to be complemented by more symbolic approach, some more deterministic code and so on and so forth. I think actually what we're seeing right now in terms of, you know, coding agents and building all these verifiable, more deterministic code that complements the statistical inferences, I think that's kind of our was our original vision when we started back in 2017. So those were the two core concepts based which in which we created a company. And then in the first three, four years, we were operating as a pure AI research lab. So we hired, you know, the the best engineers and AI researchers in in in Israel. So we actually were fortunate and very delighted to hire these extreme talented people early on and create various kind of the probably the highest density of AI talent in Israel back in the days. So, and back then we were mostly researching the natural language field. We were building all types of models. There were BERT type models, if you remember back in the days, and there were the initial versions of autoregressive models, what we know today as GPT. So we were experimenting with various variations and publishing papers and so on. But back in 2020, we realized, well, this actually is starting to become a useful technology.
00:07:37 · Ori Goshen, Grace Gong
And the first place its usefulness was manifested is in reading and writing, because these models became powerful enough to help users to rewrite sentences or rewrite paragraphs. Back then it would look like science fiction. So based on these models, we were able to create our first application that was called Wordtune, which is basically Grammarly on steroids, right? It's like basically an AI version of Grammarly, and that took off very quickly. We had when we released it back in 2020, after two years, we had tens of millions of users and a lot of paying customers, so was a very growing business. So that was kind of the first step in commercializing that type of technology through helping people to better read and write. So I've been through a long journey, can keep on going probably for many hours, but I think the point is that we started as an AI research lab, we commercialized the technology through an application, and then the kind of the next step in our journey was that we realized we want to kind of own our own destiny. So we started training our own large language models. So back in 2021, we released our first model, which was called Jurassic, the Jurassic model family. It was one of [00:09:24] S0: Jumbo, Maestro.
00:09:26 · Ori Goshen
[00:09:26] S1: So so so Jamba Jamba and Maestro came after, but Jurassic was the first model family that we released, was about a 178,000,000,000 private model back then. It was basically GPT-three, which was 175,000,000,000 parameter and Jurassic that size. So we were pretty early on in training these large models and the models back then were kind of murky. They were very impressive in terms of their what called a few shot capabilities, you would give them few examples, like the task that you want to do, you'd specify it in natural language, and we will follow the pattern and we'll do something interesting like a task you would want to rewrite a sentence or you want to summarize a paragraph, it will the model will understand what you're trying to achieve and given an input, given context, would be able to perform that task. So that was the few shot era, which was again very impressive, but not unlocking all the usefulness that we're seeing in in modern type of models. Then fast forward, back in 2024, we released a model family called Jamba, which stands for Joint Attention and MABA. And it was the first model, hybrid model, commercial hybrid model that combined between mamba architecture and the transformer architecture. That's basically the Bon Ton. And in that paper, we show that you actually can have best of both worlds. You can have the efficiency of the MAMBA architecture with the level of quality and nuance of the transformer architecture if you find the right way to combine them, and nowadays it became pretty industry standard, so if you look at examples like Nemotron from Nvidia and the models by IBM and many of the Chinese labs are training models with this combination of non transformer layers with the classic transformer layers in their neural networks. So you see that this is actually becoming like an industry standard these days. And again, fast forward for beginning of... End of twenty five, beginning of twenty six, we released Maestro, which was an agent orchestration system helping customers, enterprises to build reliable agents and the idea was that this needs to be a model agnostic framework. People can build agents on top of whatever model they want, and you'll be able to scale, test, compute and control a lot of the inner workings of these agentic systems to find the right balance between quality and cost and so on.
00:12:57 · Ori Goshen
So that was an important stepping stone into our focus today. And we are super excited about what we're seeing in the market right now. And there's the reaction and the level of demand. What we do today briefly is providing companies with what we call Lab-as-Service. So basically, it means that every company in the world right now wants to own their own AI. If you think about 2026, it all started with the agentic boom, right? Like, coding agents become super popular, every company, every software department is using it and so on. Now companies are starting to think seriously about how do they differentiate, because if you use, if you rent a model, if you go, if you rent a frontier model, you're like basically like your competitor, you're like everybody else. So how do you leverage the data, the secret sauce, the customer interactions you have with customers to create alpha, to create a differentiated offering in the market? So that's one driver. The second driver is cost. I mean, these models become because the sheer volume and usage, they become more and more expensive in practice when you are trying to apply them on real world use cases. So these two drivers really put a huge amount of pressure on companies to think about their strategy for AI and how owning their own AI is a key part of it. And the problem and the challenge is owning your own AI is hard. You need to know how to train, how to post train the models. You need to, you know, there's a lot of inner phases in training the models. There's the data preparation. There's actually also knowing what you want to achieve.
00:15:20 · Grace Gong, Ori Goshen
[00:15:20] S0: Like, [00:15:21] S1: how do you find what the task looks like? How do you define what good looks like? [00:15:27] S0: These [00:15:30] S1: things are very native and natural for AI makers, but for most most companies in the world don't have, you know, they don't have the talent or the tooling to build their own AI. So so basically, we we decided to help those companies in a shape of Lab-as-Service, which means we bring four deployed researchers. [00:16:03] S0: Mhmm. [00:16:03] S1: So the researchers... [00:16:05] S0: Really like that you create your own keywords.
00:16:07 · Ori Goshen
[00:16:07] S1: Yeah. Yeah, I think it's like, you know, there's a lot of engineering work to be done, but there's also when you are tackling these core business problems, it's actually closer to what people define in research, and I think that could be an old debate about, you know, the borderlines between research and engineering. But anyway, so we bring the FDR, the four deployed researchers, We provide the toolbox that we've created for many years as model builders for ourselves. So how do you post train? How do you evaluate at scale? How do you optimize the harness? So it's not just a model, it's the harness. We have tools for all of these needs. And the compute, would come such a scarce resource these days, even knowing from a company perspective, even knowing what kind of cluster, what kind of compute do I need to train a model that is capable of doing x, it's a challenging question, right? So we basically help with all of that. We bring the compute, we bring the toolbox, we bring the FDRs, and we help companies to achieve that goal of owning their own intelligence and actually proving the value, the business value, that owning your own intelligence. And the reason I'm so excited about this, because I think this is the future. The future is about diversified models, diversified AIs. I don't think there'll be one AI to rule them all. I don't think AI, I don't think even frontier AI is necessary for all tasks. You will have this abundance of different types of AIs. And I think companies will be defined by how good they make and operate their own AIs. And so if we believe in that future, that every company will become an AI company, and the goodness of the company will be defined by how well they own and operate their own AI. So that's a strategic component element, right? So I'm just excited about the opportunity that we have as AI21 creating that muscle for many companies in the world. So yeah, I can speak on and on. I can go on and on about this, but...