Orchestrating Smarter AI Systems with AI21 Labs' Yoav Shoham | AI Basics with Google Cloud

In this episode of AI Basics, Jason sits down with Yoav Shoham — Stanford professor emeritus and co-founder of AI21 Labs, creators of Jurassic-2, Wordtune, and the new orchestration system Maestro. They unpack: Why enterprise AI struggles with reliability What orchestration really means (and why LLMs alone aren't enough) The pitfalls of "agent-washing" Small vs large models, agent-to-agent protocols, and where real opportunities lie This one is for founders building with AI — if you're navigating hallucinations, chasing automation, or exploring multi-agent workflows, this episode is a must. * Timestamps: (0:00) Yoav Shoham joins Jason to discuss AI Basics. (1:05) What AI21 Labs is building — from Jurassic-2 to Maestro (5:49) The overuse and confusion of the seductive term "agent" (8:22) Small models vs large models: what's best for enterprise? (10:53) The verticalization of AI: legal, accounting, and beyond (12:17) The challenge of agent-to-agent communication and shared semantics (17:08) What's overhyped vs underhyped in AI — Yoav's advice for founders * Uncover more valuable insights from AI leaders in Google Cloud's 'Future of AI: Perspectives for Startups' report. https://goo.gle/futureofai * Explore Further Google Cloud's Report: The Future of AI Get insights from 23 leading experts on how startups can leverage AI for real business impact. 👉 Read the full report More AI Basics Episodes Founders, operators, and builders — catch up on our full Startup Basics series (Legal, Finance, Growth & AI): 👉 thisweekinstartups.com/basics Links from episde: Google Cloud — Build smarter with AI tools built for startups AI21 Labs — Makers of Wordtune, Jurassic-2, and Maestro Maestro — Orchestrate AI agents and tools across workflows Jamba — AI21's hybrid language model with breakthrough efficiency Learn About Google's A2A (Agent-to-Agent) Protocol The next evolution in agent interoperability and coordination: Official A2A Announcement Hands-on A2A Tutorial (Medium) * Follow Yoav: X: https://x.com/yshoham LinkedIn: https://www.linkedin.com/in/yoavshoham/ * Follow Jason: X: https://twitter.com/Jason LinkedIn: https://www.linkedin.com/in/jasoncalacanis * Follow TWiST: Twitter: https://twitter.com/TWiStartups YouTube: https://www.youtube.com/thisweekin Instagram: https://www.instagram.com/thisweekinstartups TikTok: https://www.tiktok.com/@thisweekinstartups Substack: https://twistartups.substack.com

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:04Jason Calacanis?

  All right, everybody. Welcome back to This Week in Startups. It's time again for our AI Basic Series. What is this? Ah, man, founders ask us all the time, the same questions over and over again. So we have done basics for legal, we've done it for marketing and growth tactics, obviously, accounting, and here we are in the age of AI. People need to understand the best practices and, hey, it's moving pretty quickly. If you want to get caught up on your AI basics, go ahead and download this fantastic report by our partner, Google Cloud. It's called The Future of AI Perspectives for Startups, featuring insights from 23 top AI experts. And today,

• 0:44Jason Calacanis?

  one of them is with us. Yoav Shom is here. He's a Stanford professor, emeritus, as you know, and he is the co-founder of AI21 Labs, the team behind Jurassic 2 and Word Tune, building large language models and tools for collaborative reasoning. Welcome to the show.

• 1:02Speaker 01

  It's fun to be here. Thanks for having me.

• 1:05Jason Calacanis?

  Let's talk a little bit about what we mean by reasoning, you know, here in 2025. Are these machines actually reasoning? And how do you get the best out of them? Right now, so many experiments going on in corporate America. So many people are testing and starting to deploy AI technology from large language models or with the basis of large language models. But there is some concern about the reasoning. Are these making the best decisions, hallucinations, et cetera? So what is the state today? And maybe tell us a little bit about your company, AI21 Labs.

• 1:39Speaker 01

  Maybe I'll start with the latter. AI21 Labs, it's about seven years old. We definitely, one of the main LLM builders, although we took a slightly different tack than most people recently with our Jamba family, which is not a pure transformer architecture for efficiency reasons, and we can speak about that. Most of our effort now is around orchestration and planning of these complex AI systems, in particular, a product we call Maestro, which we released. But that's directly relevant to your question, Jason. About, so you know, we in AI are guilty of using terms that are so ill-defined that they come back to bite us.

• 2:24Speaker 01

  And we can draw a long list from AGI to agents to reasoning. But if I step back from the actual terms, the issue, as you pointed out in the enterprise, is that there's a ton of experimentation. Like, three years ago, nobody paid attention. Maybe three years ago, you couldn't get a CEO or, you know, Chief Innovation Officer to pay attention. Now, everybody's on top of that. But for all the hundreds of use cases in a given company that you see, the number of deployments is very small. And the main reason is the issue of, well, there are many reasons. Issues of compliance and safety and, you know, use cases, new technology, it's all good, but the main reason is reliability.

• 3:07Speaker 01

  And again, the term hallucination may be not the best term, but these are probabilistic machines and they're, for sometimes, often, they'll give you brilliant output, but if you're brilliant 95% of the time, not just wrong, but total garbage 50% of the time, that may be okay in consumer land or not in the enterprise. And so there are many studies that show that the ratio of experiments in the enterprise, deployments, is like 10 to 1, 20 to 1.

• 3:38Jason Calacanis?

  And so that is super relevant. What it means is there's an enthusiasm for the technology, but when it comes to mission critical applications, if you're doing something in accounting, if you're doing customer support, obviously if you're a developer and you're pushing code to a server, it needs to be a lot more reliable. And that's why humans are in the loop. But with Maestro, I think the concept, and you'll correct me here if I'm wrong, is we have many different LLMs and having them, as crazy as it sounds, work in concert with each other, interesting, concert, Maestro, and having them check each other's work,

• 4:16Jason Calacanis?

  maybe trying to get them to justify their answer, as it were, can result in better output and more reliable systems, correct?

• 4:28Speaker 01

  Uh, largely yes. Let me do answers.

• 4:32Jason Calacanis?

  Please.

• 4:32Speaker 01

  So some would like to take the LLM, or as new version of them sometime called LRM, large reasoning models, which is really a misnomer, but system like, you know, O1, O3, R1, and try to get them to behave through guardrails and alignment efforts and everything. And all this is good to do, and we do that. But that will never iron out the variance in the models. And so what you really need to do is to put logic on the outside, to orchestrate. And it's not just a matter of routing between, you know, this LLM or that LLM. Sometimes you run code. Sometimes you use a tool. You know, you'll access the database.

• 5:17Speaker 01

  You'll call, you know, a weather API, what have you. And something needs to orchestrate all of that. And you're absolutely right that you want to do not only system testing, but unit testing. So every step of the way, you have an explicit plan. And every step of the way, you want to, as well as, as best you can, validate how well you're doing. And sometimes you'll do it with the language models, what's called judge language model. And often you'll just, for example, you want an output to be 600 to 800 words long. Just do the damn counting.

• 5:50Jason Calacanis?

  Let's talk about agents specifically. This promise has captured people's imagination because agents feel like a way for humans to stop doing chores. So much of what we do when we go to work, when we try to build a business, is the actual product and the service we're providing to people. But we all have to do our chores and cleaning stuff up, normalizing data. Just work that humans find monotonous, most humans, and they don't like to do, digging dishes. Like, the agents feel like the proper solution for those. How close are we to having agents at scale doing these repetitive tasks?

• 6:34Jason Calacanis?

  How often are people actually building an agent that is in 2025 when we're recording this? Getting the chore done reliably enough that humans can forget about that chore?

• 6:47Speaker 01

  Your mileage varies, depends on the store and depend on the quote unquote agent. The problem is that people have been using the term agent now. It's so seductive for anything that smacks of any kind of automation. And it'll come back to bite us. I call this the agent washing. And so you're absolutely right that the biggest bank for the buck is when you're trying to get the technology to take care of fairly simple, fairly mundane stuff, kind of like robotic process automation on steroids. So more and more stuff can be automated. Is that an agent or is this simply a program that you wrote that maybe is an LLM?

• 7:26Speaker 01

  We don't need to get anal about the definition, but typically when we speak about agents, what do we have in mind? We have in mind a system that's not the transactional calls in LLM. There's something that's more ongoing. There is the AI system, the agent can be proactive, not just respond to our prompt. It takes, it executes complicated flows, not just like a one step thing. It uses multiple tools. And as you do this, it gets dicey here because if a single call to an LLM carries some uncertainty, when you start to compose them, at some point you get more noise than signal.

• 8:10Speaker 01

  And that's where I think maybe some people are getting a little ahead of themselves. So simple road repetitive stuff, yes. More complicated stuff, we have work to do.

• 8:22Jason Calacanis?

  It seems like there's been an investment in smaller, more narrow models. And some debate about that. Some people believe the large models will figure it all out eventually. Other people think, hey, why not make a smaller model faster, cheaper, better and more constrained? Maybe you could talk about the state of. SMLs.

• 8:41Speaker 01

  I think the short answer is in consumer land, if you're looking for a very general purpose chat, let's call it call a spade a spade, a chat GPT like experience, there's probably no replacement for a very large language model that is being aligned tuned to cover a huge variety of cases because you can't anticipate the variety of inputs you get from consumers. As you go to the enterprise and your needs are much more narrower, there's several reasons to go narrow. First of all, is is just cost. It's it's cost not only in terms of dollars, but also latency. And so, you know, and as you know, we came up with this new family of models called Jamba.

• 9:26Speaker 01

  This is a hybrid state based model and transformer that in terms of quality, the answer you you get it's competitive with the most sized model. In terms of latency and memory footprint, there's just no comparison, especially as the input, as so called the context length increases, that kills you. The transformer architecture, this is what, you know, 2017, the famous paper from Google, thank you, Google, really moved the needle. Suddenly, stuff happened in the language that hadn't happened before, which did happen in vision. And the reason is that the attention mechanism transformer allows you to, as the term suggests, tend to very disparate part of the input.

• 10:08Speaker 01

  In vision, it doesn't so much matter if, you know, to know that this here is a, you know, is a phone, doesn't really matter what the pixel way over to the side is, but in language there's nothing local. So that made a difference. Problem is that it's expensive. It's a quadratic complexity in the input or the context length, as we call it. Now, when we had input of a thousand, thousand squared is fine, but we're now pushing a million, a million squared is not fine. And so you need to deal with that. Part of it is smaller language models that doesn't quite deal with the context length side of things. But then rethinking the architecture, so the state space model, which is inherently linear, not quadratic, and mixing it just a little bit transformer, it gives you the best of both worlds.

• 10:53Jason Calacanis?

  What about verticalization of knowledge? Is there a movement to say, hey, this? Small language model is going to focus really on accounting to just put it in business categories. This one is just really amazing at legal concepts and you know that when you throw this legal brief into it, or on the accounting side, you throw this RFP into it. These are huge context windows dumping cases of case law and trying to process them. Is it just about the speed and the cost? Or is it also about the accuracy? Because the model has been constrained to not have to worry about, oh, all the movies and songs ever written in the world and every blog post about those songs and music in the world.

• 11:38Jason Calacanis?

  Is it actually going to result in a higher fidelity of content?

• 11:41Speaker 01

  Short answer, yes. Long, I'm sorry, I'm always nuanced, but the longer answer is that there's some general common sense that the baseline model has learned that you want to retain, even if it's just, you know, mastering correct English, you know, language, you know, grammar. And so as you attain more domain-specific language, it's okay to forget certain things, but not others. So the art here is to just remember the right stuff.

• 12:16Jason Calacanis?

  Let's talk a little bit about agent-to-agent protocols. For people who don't know, A2A is a protocol for interoperability between agents. You know, technologists are always looking at what's around the corner. If you do get your agents working well, let's say the accounting department agent doing purchase orders and paying bills, eventually you might want that agent to interact between companies, maybe to put an RFP out to get five companies to bid for, I don't know, the new shed you're building or the new software that you want written. And agent-to-agent protocol is going to solve for that.

• 12:57Jason Calacanis?

  But this is, we're talking within the last 60, 90 days, this stuff is all starting to be publicly released by Google, other players, and we're starting to see some consensus.

• 13:10Jason Calacanis?

  From different technology companies and data sources that this is the next big thing. Is anybody actually got this in deployment now in your experience? What are the early results like? What are people thinking this will do?

• 13:23Speaker 01

  No, I think it's too early for anything to have been in production, even in the ideal scenario. So this is not a knock on H2A, just too early. So first of all, I think kudos to Google for shepherding this and a good start. But it's just a start and we have to realize the limitations. So the vision that there's several things that excite people when they hear about multiple agents coordinating. Part of it is the something from nothing. Oh, I don't need to think hard about the problem. I'll just build a bunch myself will build a bunch of agents. And then magic will happen when they come together. That historically has led to disappointment. Often the magic isn't the glue.

• 14:05Speaker 01

  It's good to compartmentalize and factor things out. That's always a good. But often the magic is how you put together things and what the algorithm is. But I think, as you said, the promise is that it's not only my agent speaking to my agents. It's my agents speaking to other agents in my company, but that I didn't build, but also outside my company. And here I think not so fast. There are two fundamental problems. One is, if you look at the protocol, it's, there's a, a, a, a part of what the agent communicates in JSON,

• 14:48Speaker 01

  which is its capabilities and other stuff where, you know, the contract it does with our agents. The problem is it specifies the syntax, but not the semantic, not the meaning. And that historically has been the pitfall of distributed object systems. That objects advertise their capabilities, but there's no reason for me to, for my agent to understand what you meant. When you put in language, I know how to find, you know, good flights. Well, what is good flights? Is it mean efficient time, you know? So we need to share semantics. It can be done, but it's a big undertaking. It's a community kind of activity. That's number one.

• 15:30Speaker 01

  Number two is shared incentives. If you go outside the boundaries of even my own unit and company, because we don't always share the same incentives, even if we're in the same company, let alone. So for example, if I'm looking, I have an agent that's trying to put together an itinerary for me and book a flight, and it's speaking with your agent and your, maybe, you know, an agent of one of the companies, we do not have the same incentive. And so you need to put in some control for that. And actually, I spent in my, wearing my academic hat, I spent a good fraction of my academic work on the area of multi-agent systems. In fact, we have

• 16:12Speaker 01

  a standard textbook in the area. And a lot of that has to do with crafting protocols for multiple agents to get them to play nice together, even though left to their own devices, they wouldn't. A lot of game theory and stuff like that.

• 16:27Jason Calacanis?

  I mean, if you think about what we went through when we tried to have a semantic web exist. Oh, hey, you're a chef and you have a really silly example, but you're putting your recipes online. We'd like you to make your recipe semantic. These are ingredients, these are steps, you know, and here's what the output looks like, and here's the origin, and here's the temperature. We want all this stuff to be semantic. It was like, okay, yeah, I'll do all that for you. And then, you know, a bunch of people just scrape all your recipes and you get less traffic to your website. And the promise was, oh, I would get more traffic. People would search for these three ingredients and, you know, my recipe might come up. So the devil is in the details there and, you know, thinking it through is critical. Let's end on this. What's hyped, what's overhyped? You got a lot of founders

• 17:11Jason Calacanis?

  listening here on this week in startups. They're building stuff. What's something they should be doing now that's obvious and going to pay dividends for their startup? What are things that, hey, maybe agent to agent falls into this category you could become aware of, there's an opportunity here, but it might be a bit too early to get significant gains from it? Where should they be focusing their time and effort in your mind?

• 17:35Speaker 01

  A little presumptuous for me to give a definitive answer because there's so many kind of degrees of freedom here. But I think that if you look for the maybe underhyped opportunities, maybe I'll mention two. One is the boring stuff, you know, getting workflows to be reliable is, and again, I'm thinking enterprise. This is kind of my lens.

• 17:58Jason Calacanis?

  Yeah, true.

• 17:59Speaker 01

  That is the biggest blocker in the enterprise right now. Getting the workflows to be reliable and customizing them per deployment. That's hard work, but that's where I think the real pain is. The thing is, it's not sexy. If you give a demo that did something amazing once or maybe many times, that's sexy. But it's not sexy to show that things don't fail. But that's where the real value is. So I think there's, that's where I, you know, one area I would focus in. Yeah, the area where I don't know that it's, you know, underhyped, but I think it's underserved is education.

• 18:38Speaker 01

  You know, I was there in the early days of online courses, you know, Coursera and Udacity started in my corridor at, at, at the Stanford. And, and, you know, I have a, an online course on game theory that's been seen by over a million people. But it doesn't begin to scratch the surface of what the real opportunity is in proactive, per student. Teaching, the issue is not how to get chat GPT out of the classroom so people don't cheat. The issue is how to get the technology in the classroom and rethink what education is really about. How we do it right with technology.

• 19:21Speaker 01

  So that's an area that I would really like to see kind of blossom.

• 19:25Jason Calacanis?

  It is super interesting. The first step was getting all those courses online and somebody who. Went to Fordham University, didn't quite hit the Ivies. I was always jealous, like, what's going on, you know, in at MIT, at Stanford, like, how different is it than my experience? And I was feeling particularly under-resourced in macroeconomics, right? And I was like, I really want to understand this. And I just went to YouTube and I found a courses at Stanford, MIT. And I watched the courses and I was like, wow, this is like alchemy. They're trying to figure out how macroeconomics works. But the fact that it was available for free on YouTube,

• 20:06Jason Calacanis?

  the same course that people were paying and had to qualify to be in the 0.1% of people on the planet to get there was available for free. And now you imagine if it was adaptive learning and you could answer some questions upfront. And then it said, yeah, you know, you should really start with this third video. Or actually, you're not ready for the first two videos. You should do this pre-calculus and maybe the statistics course first. Before you get in there, learn some basics about statistics. So you can actually understand the material better. And that would be so amazing to have it be adaptive and not leave any. As a professor, you wind up leaving some students behind. And at what fork in the road did they get disengaged is always the question, yeah?

• 20:50Speaker 01

  Absolutely. You know, it's, it's typical. When a new technology comes, comes on, you, you try to use it the way you use the old technology. So television initially was televised radio. And then you, over time, you understand what the medium was really good for. I think it's, the time is ripe now with, you know, the flat world and everybody has that having access to, you know, computers and networking to get AI to do, to rethink education.

• 21:17Jason Calacanis?

  Yeah, absolutely. And you, you think about the role of professors, creating courses, creating quizzes even. You can go into any LLM today, ask it to take Great Expectations, Charles Dickens, and give you a series of Q and A and be your coach. And you just give it that tiny prompt. And you could sit there with your phone and have a personalized tutor that you would have had to spend, you know, whatever, days or weeks to find them and pay hundreds of dollars and it's just available to everybody for free today. What an amazing discussion. Thanks so much, Yoav, for joining us.

• 21:49Speaker 01

  Thank you. It's really, really fun. Thanks for having me.

• 21:52Jason Calacanis?

  You can learn more in Google Cloud's Report: The Future of AI Perspectives for Startups. Go to goo.gle/futureofai. That's goo.gle/futureofai. Go check out Gemini Men. I love that deep research. And everybody, if you want to get more of our startup basics series from legal to accounting to marketing and now AI, go to thisweekinstartups.com/basics. Thanks again for listening. We'll see you next time.