The Exchanges

Q back now to the other end of the spectrum, like the very beginning of the company. So I was, as I was prepping for this actually, uh, read that the company started in 2017, which for, for the world of NLP was, was a big year, uh, but you were very much at the beginning of that, that, that big NLP wave. So like, how did it all come about?

A Sure. So, uh, it all started, uh, so we started a company, uh, Yoav and myself, um, I had a technical background. Um, this was my, uh, second company. The first one was, uh, analytics company, uh, in the networking space that was, uh, basically acquired by another Israeli company called Cellwise that, uh, was eventually acquired by Qualcomm. And, um, and I, I was, um, I was extremely curious about, uh, AI. I didn't have any background in AI, but I was, I was very curious about this space and actually I had a few ideas. Um, and, um, I kind of randomly, uh, it's an interesting story by itself, but I met, uh, Yoav, who's, uh, my partner and, um, and Yoav, his background, he was a professor at Stanford for almost, uh, 30 years. Um, he ran the AI lab and, um, Actually started. This is his fifth company and, um, and, and, uh, all of his previous, uh, companies were acquired. And when, um, when his last company, uh, acquired in 2015, they decided the whole family to move back to Israel. And, and that's how we met. And, um, and we joined forces to start a 21. And, uh, shortly after, uh, Amnon Shashua joined us as the third co-founder.

AI assessment note: “we joined forces to start a 21.”

Q a few minutes ago, which I, I loved, uh, which is that the industry has gone from sporadic exploration to massive experimentation, I believe you said, which is a wonderful way of putting it. Um, from your perspective as somebody who's in the trenches every day, what are some of the top use cases and maybe, um, talk about some of your customers and some of the case studies there?

A Sure. So it's really across industries and, you know, from, um, kind of cool startups to, uh, uh, healthcare, uh, companies, um, And, um, and it's really, it's really diverse and interesting, uh, ways of using these models. Um, I think I'm I'm most excited by just to see how these are employed to, uh, enhance productivity of, um, of, of knowledge workers, of, you know, of all of us, basically. So we work with, um, a few financial, uh, institutions and they essentially are using these tools to, um, Um, make the information. So a lot of these, uh, financial institutions have repositories of research, uh, huge repositories. Um, it could be. Millions of documents. And, and, and then there are all sorts of, uh, roles in the organizations that didn't access to that information and, um, and kind of typical search won't do it. Um, sometimes you have a question and the information resides in cross, you know, many documents and you need, um, a system that can pull together all the pieces of information and, Stitches, stitches together into a coherent, uh, and relevant, uh, answer. And it is quite magical. I mean, you'd see, we've seen some of the analysts that, uh, said, wow, this would take us, uh, like three or four hours. Now with this tool, we're doing it in just a couple of minutes. And We've never had the chance to, cause we haven't never had the chance to think about these specifi…

AI assessment note: “we work with, um, a few financial, uh, institutions and they essentially are using these”

Q of application, the, the, the task specific model. So just, um, again, reviewing my notes, uh, talk about text editors, retail software, knowledge and support platforms, content management systems. So those are our, our discrete business applications. So how does that, how does that, um, how does that work? And, uh, I mean, presumably they, they, they're fed by Jurassic two models underneath, maybe just Go through the high doors.

A Yeah. So we're, well, you know, the market has just shifted from, um, sporadic exploration to massive experimentation in the last nine months. And we've, you know, we, we learned a lot from the market and what we, what we saw is that 95% of the use cases in the enterprise are actually, uh, pretty specific. I mean, there's barely areas where you'd want an open-ended chat system that can, um, uh, you know, that can address any query that you have in mind and be that generalized. Most enterprise, when you think about How to incorporate this in a workflow, how to incorporate this in an application. They, they have a much narrower use case in mind, very powerful, very valuable, but, but narrower. Um, and, and then we came to the conclusion that this is maybe using the general purpose model as is, it may be an overkill. And, uh, we, we should take a different approach and, and we, and we think about it as a metrics. You have the tasks. As the columns and the, um, the industries or domains as a rose. And then you have like summarization of financial reports or, um, tax generation. We have certain types of constraints in the retail industry for producing, uh, product descriptions, right? So you have these Lego blocks that each piece is really good at a particular task. It's also a customizable piece. It's not kind of a closed box. You can then customize it to a certain, you know, styli…

AI assessment note: “we actually wrap these models into systems that really take care of the reliability part”

Q you, you still got better results with just like super powerful deep learning, uh, even when you try to build in the, the, the stuff, in other words, the hybrid system performed, uh, not as, not as well as the, just the pure deep learning. Um, I guess what, what do you say to, to, to, to that, to the extent you can talk about, um, any, any of this?

A Yeah, to be, to be modest here, I think the jury is still out, but, um, uh, I have a sense, uh, that if you truly want to deal with the, uh, long tail and the problems in life, like, and if you didn't want to deal with natural language, uh, robustly, the long tail is extremely long and diverse. And um, And, and my sense is that, uh, the way to move forward to make the way to make these systems more reliable and transparent in a way is not going to be by creating a more powerful model, like, you know, more parameters or, um, tweak the data. Uh, I mean, there's definitely kind of performance gains that you can keep on squeezing, but, uh, it will eventually fall off the cliff. And we'll face the same, the same fundamental problems we're facing with the strong models today. And we're already start seeing diminishing returns. I mean, you take the architecture, you, you, um, increase the size of the models or, and you see that, um, At these levels of scale where, you know, we are dealing with, uh, things, diminishing returns, we start, we start seeing it. So I think it's a, um, it shows that, um, in order to fundamentally resolve the problem, there is, there needs to be kind of, uh, we need to take a different approach. I think there, there were actually quite recently, uh, success cases where the hybrid systems, uh, actually reached better results. Uh, from the kind of, you know, en…

AI assessment note: “there were actually quite recently, success cases where the hybrid systems, actually reached better results.”

Q um, you know, what should we look up, uh, research, what, uh, who should we be familiar with in terms of companies or universities or researchers, um, you know, and or anything you're excited about in the ecosystem, any concern that you have, so whichever way you want to take it, but, uh, Would love to be educated on the, on the sort of one-on-one of the Israeli AI ecosystem?

A Sure. I think, I think generally speaking, Israel is just an amazing talent pool, uh, technical talent pool, um, very creative, very, um, uh, you know, chutzpah mentality and, um, And that serves us very well. I think if you look at, you kind of, uh, Israel tech ecosystem from a technology, technological perspective, you see that it adopted many, many times and very fast. Like, you know, back in the nineties, it was all about chips and hardware, and then thereafter networking and telecommunications. And then it's just, it's, um, it's making these transformation. And I think You know, if you pinpoint right now, Israel is very well known for its cybersecurity talent, very unique and good at cybersecurity, but we, we are feeling that there is a, you know, a movement towards, um, the next wave, which is basically AI. And to be fair, there's also, uh, an AI talent here for, for a while now. There are a lot of, um, computer vision, Uh, companies, you know, and Mobileye is a great example. Um, there, of course, there are other computer vision, uh, companies here that was the, basically that were built in the last decade or so. Um, so, uh, it's, it's the DNA and the core, you know, the expertise in deep learning and so forth is definitely, is definitely here. Another, uh, kind of phenomenon, uh, that, uh, I think people should Be aware of is that, you know, when you look at NLP researc…

AI assessment note: “Mobileye is a great example.”

Q access to the actual model itself. I just call it, um, And then the, so just to play back with what, uh, you just said, uh, depending on my needs, uh, so how do, how do I know which one to use? I mean, obviously the parameters around latency and price, but, um, are there certain tasks where I know I'm gonna need the bigger one versus the small one?

A Yeah, I guess, uh, typically, um, the light is great at kind of sort of extractive, uh, or classification tasks. A mid is for, um, uh, short form generation or, um, uh, rag type of use cases. And, um, we'll try is one that, um, is very good at, you know, longer form generation, keeping, taking long context and kind of referring them and, Having the, you know, the, the ability to attend to, um, uh, more information and be, uh, consistently, um, uh, cohesive with, um, so, and I think basically people should try. I mean, when they have the use case, my, my first recommendation to customers is put, um, even on a small scale, make, uh, Do your best effort to create an evaluation set that represent the, the problem or the distribution of the problem you're trying to solve. Once you have that in place, uh, you can get a sense for our models and also other models in the market and, and, and see what works best, uh, best for you. Um, another, another thing I should mention is that we're taking a very neutral approach. So we're, um, delivering our models, Uh, in many different ways. So we're, we're working, uh, with, uh, we have our SaaS platform that is, is actually multi-cloud, and we also serve our models through various types of integrations. So, uh, with, with Google Cloud, you can basically consume our models through the marketplace and in the future through Vertex. And, and, um, a…

AI assessment note: “the light is great at kind of sort of extractive, uh, or classification tasks.”

Q important is, uh, fundraising strategy? Do you think at this moment in the general AI market, obviously headlines have been dominated by like massive fundraisers, and this is certainly a very large one, but is that is a little bit of like whoever raises the most wins because you need to spend money on GPUs and, and, and, and attract super expensive talent. Like, how do you think about it?

A Yeah, I think, um, One should think about it strategically. Um, um, you know, there's, there's a lot of excitement and I think the excitement is for a reason. Uh, what we're experiencing right now, it's, it's just a huge platform. It's a huge technology shift. So it's gonna have impact and it's going to have, um, you know, economical gains. Uh, but I think, uh, also we should be, I mean, we as company leaders, we should be responsible and think how do we grow the company healthily and, um, and, and build business around it. Like we build technologies that are, you know, transformative, but we also build businesses and we need to constantly calibrate Uh, the progress of the, uh, business side, uh, to the valuations and the amounts that we raised because each step in this road, you know, just creates, um, an expectation for the next step. So I think, um, again, as, as, uh, Uh, uh, young companies in this space and young, like, I mean, relatively to the incumbent, um, uh, we should be very responsible here and think about how we grow these companies and, uh, in terms of GPUs and compute resources, and I'll speak about it later. I think there's, there's a more, um, practical approach. Um, I mean, we can, you know, build these massive models that are, Uh, very, very generalizable and can do many things, but, uh, we can also be, uh, more thoughtful about how do we train even the larg…

AI assessment note: “calibrate Uh, the progress of the, uh, business side, uh, to the valuations”