#244 Yoav Shoham on Jamba Models, Maestro and The Future of Enterprise AI

This episode is sponsored by the DFINITY Foundation. DFINITY Foundation's mission is to develop and contribute technology that enables the Internet Computer (ICP) blockchain and its ecosystem, aiming to shift cloud computing into a fully decentralized state. Find out more at https://internetcomputer.org/ In this episode of Eye on AI, Yoav Shoham, co-founder of AI21 Labs, shares his insights on the evolution of AI, touching on key advancements such as Jamba and Maestro. From the early days of his career to the latest developments in AI systems, Yoav offers a comprehensive look into the future of artificial intelligence. Yoav opens up about his journey in AI, beginning with his academic roots in game theory and logic, followed by his entrepreneurial ventures that led to the creation of AI21 Labs. He explains the founding of AI21 Labs and the company's mission to combine traditional AI approaches with modern deep learning methods, leading to innovations like Jamba—a highly efficient hybrid AI model that's disrupting the traditional transformer architecture. He also introduces Maestro, AI21's orchestrator that works with multiple large language models (LLMs) and AI tools to create more reliable, predictable, and efficient systems for enterprises. Yoav discusses how Maestro is tackling real-world challenges in enterprise AI, moving beyond flashy demos to practical, scalable solutions. Throughout the conversation, Yoav emphasizes the limitations of current large language models (LLMs), even those with reasoning capabilities, and explains how AI systems, rather than just pure language models, are becoming the future of AI. He also delves into the philosophical side of AI, discussing whether models truly "understand" and what that means for the future of artificial intelligence. Whether you're deeply invested in AI research or curious about its applications in business, this episode is filled with valuable insights into the current and future landscape of artificial intelligence. Stay Updated: Craig Smith Twitter: https://twitter.com/craigss Eye on A. I. Twitter: https://twitter.com/EyeOn_AI (00:00) Introduction: The Future of AI Systems (02:33) Yoav's Journey: From Academia to AI21 Labs (05:57) The Evolution of AI: Symbolic AI and Deep Learning (07:38) Jurassic One: AI21 Labs' First Language Model (10:39) Jamba: Revolutionizing AI Model Architecture (16:11) Benchmarking AI Models: Challenges and Criticisms (22:18) Reinforcement Learning in AI Models (24:33) The Future of AI: Is Jamba the End of Larger Models? (27:31) Applications of Jamba: Real-World Use Cases in Enterprise (29:56) The Transition to Mass AI Deployment in Enterprises (33:47) Maestro: The Orchestrator of AI Tools and Language Models (36:03) GPT-4.5 and Reasoning Models: Are They the Future of AI? (38:09) Yoav's Pet Project: The Philosophical Side of AI Understanding (41:27) The Philosophy of AI Understanding (45:32) Explanations and Competence in AI (48:59) Where to Access Jamba and Maestro

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

• 0:00Speaker 00

  I think language models, even the latest so-called reasoning models, they have their limits of what they can do. And really, the world is going towards AI systems now away from just pure language models, no matter what the specific architecture is. So language models aren't going away, but they really are becoming part of a bigger picture.

• 0:19Craig S. Smith?

  This episode of Eye on AI is sponsored by the Definity Foundation. Definity Foundation is a Swiss not-for-profit that is home to some of the world's leading cryptographers, computer scientists, and experts in distributed computing. Their mission is to shift cloud computing toward a fully decentralized state by supporting the Internet Computer, also known as ICP. If you don't understand anything about the Internet Computer, you can see my episode with Dominic Williams. ICP's vision is that most of the world's software

• 1:02Craig S. Smith?

  will be replaced by network resident software that's an evolution of traditional smart contracts. To achieve this vision, ICP is designed to make smart contracts as powerful as traditional software, all while remaining tamper-proof, unstoppable, transparent, and verifiable. Since Definity's launch in 2016, they stand as Switzerland's most extensive blockchain research and development initiative and have been awarded more than 500 research grants worldwide. Definity remains steadfast in its mission to drive the advancement of the decentralized internet.

• 1:47Craig S. Smith?

  Network resident software can now be used to run AI models and RAG infrastructure, preventing them from becoming "hot wallets" from where data can be stolen, and increasing resilience. Furthermore, the technology has been designed to allow AI models to spin up and modify

• 2:10Craig S. Smith?

  Running web applications and internet services solo by addressing several key challenges. If you're interested in reading more about the internet computer, visit internetcomputer.org. I also encourage you to listen to my episode with Dominic Williams, in which he explains the internet computer.

• 2:33Speaker 00

  I got interested in AI when I was in college and went to study in the state, I went to Yale. And then I was offered the job at Stanford as an assistant professor there. And we're talking back a long time ago, nineteen eighty seven. I'm an old guy, but I honestly didn't know that that's what I wanted to do. But I figured it was a good place and I'll have a good time until I figure out what I really want to do. And twenty eight years later, I retired as a full professor. And along the way, I ran the AI lab there for a while.

• 3:15Speaker 00

  And my work there was largely theoretical, not-for-profit research. I worked on logic and philosophy and game theory as it pertains to computer science and AI. But in fact, I have an online game theory course together with some colleagues. That's that's that more than a million people view it already. This is for not particularly useful topic, but pretty, but not useful. But I do have an applied side, so I started several companies, AI twenty one being the last one. That's me briefly.

• 3:55Craig S. Smith?

  And how did you start AI twenty one? And and then we can talk about about what you guys are doing there.

• 4:06Speaker 00

  AI twenty one is seven years old now. And the idea germinated. I'd sold my last company to Google and and was thinking that AI started to put all the eggs in one basket, which was deep learning. You know, L L M weren't a thing, then large language models weren't a thing. But deep learning was and and machine learning in general. And and it sort of worked. I mean, there's a reason why people gravitated to it. But my feeling was that this is statistics, basically.

• 4:51Speaker 00

  Statistics will never give you the robust reasoning of the kind that that one needs and that you saw in AI early on. And we should kind of put together a good old fashioned AI with modern day deep learning based AI. So that that's kind of an odd reason to start a company. But truthfully, that's the reason we started the company. This is me and my co-founder, Ori Gosin, who's half my age, twice my brains. And another guy who's not that younger than me, but somewhat younger, but certainly very smart. Amr Shahshua was well known for having started and running Mobileye of the companies. But he's also quite an accomplished computer science professor.

• 5:35Speaker 00

  And I knew him from the academic side of things. That's why we started the company.

• 5:40Craig S. Smith?

  And you were combining, going back to symbolic AI and expert systems and infusing them with deep learning. Is that what you were doing in the beginning?

• 5:53Speaker 00

  So we did a lot of experimentation. And once, in fact, I remember early on, the first twenty employees, I'd recorded

• 6:05Speaker 00

  lectures on non representation. And I forced a poor guy to learn about logic and about temporal reasoning and about frame system and so on. Because they knew modern AI well, but they didn't know that.

• 6:23Speaker 00

  But it wasn't so much to take exercise and glue machine learning onto them, but to think more deeply. Uh, what's missing from modern AI and how best to imbue it then? For about three years, we did nothing but technology building.

• 6:41Speaker 00

  And coincidentally, we started the company right around the LLM revolution because that's the year that the Transformer paper came out. And it was clear to us that the the work was in language, not in vision. Vision is quote unquote an easy problem. So the arbitrage recognition, it's obviously not easy and but to know that this is a mug. I don't really care what this pixel over on the side is, but there's nothing local in language. Take a sentence and change a word over here, the whole, you can't get away from semantics. And so it's clear that that's where the work was. And then coincidentally, Transformers came out and

• 7:26Speaker 00

  as we we we felt we had to be very good at language models. And we, if I can say so myself, became very good at that. When GPT three came out, we were, I think, the heaviest, most demanding users in terms of the use cases and how we banked on it. And and at some point, we decided we we need to build our own and we built our first model called Jurassic One. It's not a most innovative model. It's a standard, sort of left to right autoregressive model, GPT like. It was slightly bigger, had five kinds of vocabulary. People who tested us find us sound, found us a little better than GPT three.

• 8:10Speaker 00

  But, you know, it was really at the time a novel thing to do, maybe even a dairy thing to do, and a small company. Doing and putting out this and at the time we so we were playing around with technology and building this language model, but at that point, this is already three years into our existence, or there about two and a half years, maybe two years. Uh, we didn't want to just be a research lab. And we had and still have really an amazing collection of talent. Honestly, I never I've had the luxury of working with very smart people, but such a big kind of dense density of of brain. I'm always the dumb person in the room.

• 8:53Speaker 00

  Uh, but we didn't want to just be, uh, and I, I, I, I always mentioned, uh, DeepMind as an example of a group I really admire, but early on, they really didn't have, they they're not a big business in solving Atari games. So the question of what business should we be in? And it was clear to us, we wanted to be in sort of the enterprise kind of B two B world. But there was no business there then. And so we had built, uh, we built our own business. It was an application which today is a very, it's a very busy space of writing assistants. But at the time, it was quite novel. There's nothing like it.

• 9:34Speaker 00

  We had an L L M powered AI writing assistant called WordTune, and it really was amazing. Still is. It's it crossed, it very quickly crossed 10 million users. It was a freemium model generating kind of a lot of revenues. But it was never meant to be our core business. And it's still there, but of course, the business became the enterprise. And it was somewhat audacious, you know, we, there we are, we're three years old. We have this application called WordTune. We have our own model. So we're, let's say, competing with Grammarly on, on, on, on, you know, writing assistants. We're competing with OpenAI on L L Ms.

• 10:18Speaker 00

  And we're like 30 people, you know, it is, uh, and then we raised, raised some money, but, uh, but less than OpenAI. So it was a little audacious, but that's what we did. And in time, you know, uh, we focused more and more on the enterprise. That's kind of roughly our arc.

• 10:39Craig S. Smith?

  Yeah, when you say focus on the enterprise, I mean, there are a couple of things I'm interested in. I was doing some reading about Jurassic 2, actually. And I read that you guys used Amazon SageMaker to build that. Is that right?

• 10:55Speaker 00

  Our Jurassic models were trained, we actually trained on both AWS and GCP. We tended to use the bare metal, not use much of the infrastructure on top of it. We were quite portable. Our models, starting with about a year ago, were trained only on Amazon. Because they're broke new grounds, there were no longer pure transformer architects. The thing with transformers, I mean, why did transformers take is because up until then, there were

• 11:34Speaker 00

  RNNs and, you know, things that look at the input and didn't have a long context to look back into. And like I said, you know, vision is kind of local that worked, but language isn't. So the attention making transformer allowed you to relate different, disparate parts of the input. The thing is, you pay a high price for that. It's suddenly becoming quadratic complexity. And if your input is like 1,000, 1,000 squared is fine. But now we're pushing a million, a million squared is not fine. And

• 12:13Speaker 00

  so our guys created this new architecture, a model called Jamba, the Jamba family. In fact, we've just now released Jamba 1.6. It's amazing. And the way the model kind of is architected, it's mostly a so-called state space model, specifically based on the Mamba.

• 12:35Speaker 00

  Model that came out of academia. We were the first ones to really to scale it to a large model. So the advantage of state based models is that they're more efficient. You're back to sort of linearish complexity. But you're not and you're not as bad because you try to remember the path and the state you carry along. That's why they call state based models. But that's not as effective. And the attention mechanism or performance. What our guys did every so often, every eight layers in that case, put a attention layer. And there were a lot of ablations about exactly how to do it. But the end of it was a model that whose performance in terms of the quality of the answers was competitive with the best model of similar size.

• 13:20Speaker 00

  And in terms of latency and throughput and memory requirements was just, there's no comparison. Way, way more efficient. This is a mixture of expert models. So you have a total number of parameters and then a smaller number of active at any moment. So our Jamba small was fifty two billion parameters, the total of which, oh my God, twelve billion, I think, were active. And the large model was three hundred ninety eight billion, of which ninety four billion were active. So the small model would fit on a single GPU, which is mind boggling. And the large model would fit on a single A GPU pod.

• 14:04Speaker 00

  And the throughput and latency are just, there's no comparison, they're almost linear. And not quite linear because you have a little attention, but almost linear. So those are the models that we're putting out now. And like I said, we've just released Jumbo one point six, which is our latest.

• 14:21Craig S. Smith?

  Can you talk a little bit about Mamba? I haven't done an episode on that. And a lot of people, a lot of the listeners have heard about it, but don't really understand it about its development and how you integrated it.

• 14:43Speaker 00

  So first of all, I really encourage you to speak with the creators of Mamba. They're nice, smart guys, academics.

• 14:52Speaker 00

  I'm sure they'll be able to go into as much detail as you want. But like I said, it's a state-based model. So it's order aggressive, goes left to right. And looking at the input token by token and predicting the next token and doing the stochastic gradient descent to update the the the weights. The difference is, as it goes along, it carries state and updates the state. And now the state can't, doesn't capture the, you know, entire history, it can't, but it captures enough of it. So you have this great, it's like a memory aid.

• 15:35Speaker 00

  That the model has. So that's how it operates. And like I said, it has certain advantages of speed and an advantage over, you know, more vanilla left to right kind of R N N type of mechanism. But but not quite competitive with explicitly every time looking at the entire history with the attention heads.

• 16:02Craig S. Smith?

  I had Sep Huckreiter. I can never pronounce his name the way he did. The creator of the L S T M on fairly recently. And he's pursuing that and

• 16:21Craig S. Smith?

  of is building a company around it. Around sort of an improved L S T M. Is Mamba nothing like L S T M? Or is it

• 16:35Craig S. Smith?

  a completely new architecture?

• 16:38Speaker 00

  Completely different, it's hard to say. And I think what Mamba did, it opened the door for people to experiment, to be brave enough to either experiment with.

• 16:51Speaker 00

  Totally new kind of model or tweaks. Mamba itself had had several versions. We ourselves didn't take Mamba as is, but we did a lot of adaptations. People have looked at now, taken a second look at bidirectional models. Not these left to right auto-regressive ones, but you know, with masking. So I think, I think something very healthy that people aren't intimidated by successive transformers. And it'll be interesting to see what comes up. I have to say, I think language models, even the latest so-called reasoning models, you know, the ones and R, R ones and so on.

• 17:34Speaker 00

  They have their limits of what they can do and and really the world is going towards AI systems now away from just pure language model. No matter what the specific architecture is. So language models aren't going away, but they really are becoming part of a bigger picture.

• 17:52Craig S. Smith?

  So in this hybrid model, you have, you were saying you have a Mamba layer and then several layers above that of a transformer.

• 18:03Speaker 00

  Exactly.

• 18:04Craig S. Smith?

  You stacked them with Mamba transformer, Mamba transformer, is that right?

• 18:11Speaker 00

  Yes, except it's not Mamba transformer. It's, there's a Jamba block. And by the way, I really encourage people to go and read because we actually gave quite explicit description of Jamba, the architecture and how it was built. We, we actually open sourced the, or open weighted the model. We felt it's especially novel for, for the community to help kind of improve things. But in the Jamba block, there were eight layers, one of which is transformer, the rest are Mamba layers. And there are also a mixture of experts, the layers there. Yeah, that's essentially what it looks like.

• 18:49Craig S. Smith?

  And so how does, how is it performing on all the various benchmarks?

• 18:57Craig S. Smith?

  That everyone is watching these days?

• 19:01Speaker 00

  So let me first make a general kind of pontificate generally on this topic of benchmarking.

• 19:10Speaker 00

  You know, we are

• 19:13Speaker 00

  somewhat cynical about the benchmarks. They're well intentioned and they give you some signal about the quality of the model. But, but it's a very weak signal.

• 19:26Speaker 00

  For two reasons. One, which is a sort of a unavoidable one, is the correlation between the tasks that are encoded in the benchmark. And what you see in the real, the real world is, is not strong. And so you may perform great on GSMA K or an MMLU or Scorehigh on, you know, LMCS or Arena Hard or something. And that doesn't necessarily mean that when you go and actually try to work on a real problem in the real world, the model will be better than the model that scored less. That's the one reason we take it with a grain of salt.

• 20:07Speaker 00

  The other is unfortunately, these are really easy to game.

• 20:14Speaker 00

  And now we've been, so by the way, to answer your question, we score well on these, despite my cynicism. They are a measure, they're good, they, they are an aid in optimizing your training. We never cheat. Um, but we take it with a big grain of salt. And I can tell you that our customers also take it with a great, a big grain of salt. There's You know, the, the, especially the sophisticated, the sort of, uh, you know, people, they'll, they'll always make a face when you discuss these benchmarks. But I mean, so that, I'm saying that we're okay.

• 20:58Craig S. Smith?

  Yeah, I mean, you can, you can train to those benchmarks, can't you?

• 21:03Speaker 00

  First of all, yes. And I do know about some others where people explicitly gave instruction, you know, we need to win this benchmark. And, um, and you could do it, even without cheating, not, I mean, the most blatant thing you could do is take the test data and put in the training data. But if you don't do that, you just spend more cycles in that area of mathematics, logic, you know, whatever it is, whether it's natural data or produce a lot of synthetic data in that area, you'll do better there, which doesn't necessarily mean you'll be better in real application. So, so people do that.

• 21:47Speaker 00

  So if you throw enough flaps at it and you throw enough data at it, um, you'll get good performance of the benchmarks.

• 21:56Craig S. Smith?

  Back to DeepMind's AlphaGo. People have been using reinforcement learning directly to train the models, not to refine or fine tune the models. What do you think about that? Did you use that approach with Jamba?

• 22:14Speaker 00

  By the way, this year's Turing Award winners are the two pioneers in AI of reinforcement learning, Andy Barto and Rich Son. Certainly RL having its moment now for a variety of reasons. RLHF is a misnomer, it's really not reinforcement learning. It's sort of reward modeling, in this case by humans. But RL is playing a role. In the Jamba models that are out there, we, we didn't use it. We didn't find the need there. Recently, we have.

• 22:52Speaker 00

  But when you're referring to...

• 22:56Speaker 00

  R1 or O1, those quote unquote chain of thought models.

• 23:05Speaker 00

  The RL plays a particularly important role there.

• 23:11Speaker 00

  As it happens, when you look closely, it's not enough. But that's really the primary way in which at training time, not at inter time, but at training time, you try to guide the model towards these chains of tokens, the correct way to say it, that are more likely to be productive than others. And that's the role of RL in those models.

• 23:39Craig S. Smith?

  But you guys, how did you train Jamba?

• 23:42Speaker 00

  We did, I'll call it standard pre-training with a lot of quality data, with natural and synthetic. We then did a lot of alignment tuning of various kinds.

• 23:57Speaker 00

  And using really, I mean, SFT supervised fine-tuning there. And that's primarily how we train Jamba.

• 24:11Craig S. Smith?

  So from your point of view, Jamba's this new hybrid architecture or hybrid model.

• 24:22Craig S. Smith?

  Are you going to continue building larger models? What's the goal? You said at the outset that your interest was in working with enterprise.

• 24:33Speaker 00

  Right. So I'm not in position to sort of announce what we'll do in the future, but I will say that we believe language models are and will continue to be important. It's very important for a company like ours to be very good at them. So I don't think that's going to go away. But increasingly, the emphasis is on AI systems.

• 25:00Speaker 00

  And in particular, we've just announced Maestro, which is a planning-based orchestrator that works with multiple L L Ns, as well as other tools and code and what have you, and orchestrated to work in a way that really will serve the enterprise, to give the reliability and predictability. And increasingly, we think that will be the emphasis in the industry, and certainly for us.

• 25:33Craig S. Smith?

  So are you going to continue

• 25:37Craig S. Smith?

  coming out with larger and larger Jambas? Or is

• 25:46Craig S. Smith?

  size of both compute and training not your goal here?

• 25:54Speaker 00

  Our goal is to produce the technology that serves the enterprise reliably. That's our goal.

• 26:05Speaker 00

  Language models are part of it. And the quality of the model is a function of multiple things.

• 26:13Speaker 00

  Size, unlike a sandwich, in language models, bigger is not always better. And it's not only the amount of training data. A lot of it is the quality of the training data and the and the ability to specialize the model to quickly, efficiently, reliably insert domains. And like I said, the efficiency of it, the way to serve it efficiently.

• 26:42Speaker 00

  That'll continue to be our focus. But like I said, I can't speak specifically to what kind of models we'll be training.

• 26:51Craig S. Smith?

  Yeah. And you were saying with Jamba, it is open weight.

• 27:02Craig S. Smith?

  And it has a particularly

• 27:06Craig S. Smith?

  large context window, is that right? And very high

• 27:13Craig S. Smith?

  throughput, token throughput compared to transformers. You can, you were saying you can fit a model on a single GPU. Uh, so what are some of the applications that you imagine with this model?

• 27:31Speaker 00

  Well, I don't need to imagine, but we have customers using it for a variety of of, uh, you know, the, the, the type of applications aren't that mysterious. You see it across the board. Uh, there, we have working with a bank, uh, to, on, in a call center setting to do question answering and troubleshooting. We're working with the big retailer to do product descriptions. But these online retailers have millions of products and writing product description is time consuming, costly, error prone. So we're helping with that.

• 28:12Speaker 00

  We are working with financial institution to query financial documents. And so the applications really vary. And and what we're finding is that

• 28:28Speaker 00

  enterprises realizes a big gap between a, a Fashi demo and something that actually works. And what we see in the industry is that we're in what we like to, just to think of as the second phase of the modern AI revolution. So up until maybe two and a half years ago, so you had sporadic experimentation. The CIOs or the CEOs had just gone, finished their migration to cloud and they didn't have energy and go ahead and play AI what you want. And then suddenly somebody turned on the switch and there's today, there's no CEO in the world doesn't say I'm an AI first company or I want to be one.

• 29:14Speaker 00

  And see how we're in the mass, in the era of mass experimentation. We have kind of customers, partners with hundreds of use cases. But the drop off between an experiment, a PLC to prototype to actual deployment is precipitous. There's one study by AWS that shows the drop of 94%, 6% of the projects see it to production. And we think we're on the cusp of the transition to the third phase, phase of mass deployment.

• 29:56Speaker 00

  But for that to happen, you need to deal with these two issues of have the model be efficient. Because the economics of LLM is, are not the economics of traditional software. It can be extremely expensive. Certainly, we use quote unquote reasoning models. And the second is accuracy. Because these models are probabilistic machines. And sometimes they give you wonderful, creative answers. And then they give you total garbage. And so if you're a high school student, you know, writing an essay, it's okay if you occasionally wrong, maybe you have time to correct. And if you don't, that's okay. Worst case, you'll get a poor grade.

• 30:41Speaker 00

  But if you're, you know, writing a summary of a patient checkup and you make a mistake. Or if you're writing a memo to your client or your boss and you make a mistake, you can't recover for that. And that, tackling that is required to get to the same phase of mass deployment. And this is where AI system will come in because no language model, no matter how good, how many parameters, how much data, how many guardrails you put there, they will never, we call it prompt and pray. So long as you're in the prompt and pray world, enterprise won't adopt it at scale.

• 31:26Craig S. Smith?

  So most of your deployments on prem where people have their own GPUs? Or is this primarily?

• 31:39Craig S. Smith?

  Being accessed through the cloud. And then what do you build around the Jamba model to protect against hallucinations?

• 31:50Speaker 00

  So first of all, our models can be accessed really anywhere. You certainly can get on the cloud, we're on all the hyperscalers, we have our own SaaS. But we also, in fact, we're viewed as the best on-prem solution. And we have it installed in some really significant companies. So it's all of the above.

• 32:17Speaker 00

  In terms of how do you get to high accuracy?

• 32:23Speaker 00

  You, again, it's, it's a fool's errand to try to, to, to force the L M to be deterministic. When you call it L M twice in a row, you'll get different answers. This is the way it works. And so you need to build a system around it. Also, an L M is not right, you need to call tools. If I'm going to access my proprietary data, you need some rag-like system to, to access that.

• 32:52Speaker 00

  Certain things, God did not put neural nets on earth to do arithmetic. H P gave us a calculator in 1970. We don't need to reinvent that wheel and get a crooked one of that. So you have these pieces of code that do stuff reliably, you want to use those. And so what you want is a, it's a system, not an A I, an L M, an A I system that knows to work with these L M, by the way, not just Jamba. So Maestro, our new sort of orchestrator, will work with Jamba, work with, with others as well. And is just smart about when to use what, which for what,

• 33:35Speaker 00

  how to coordinate them, how to parallelize things, how to plan ahead and give some performance guarantees about how much, how much will it cost me? And right now, if you go to the, the so-called reasoning models, they'll go and, you know, think, generate these sequences of so-called thinking tokens and come back after half an hour with an answer, maybe a bad hour, an answer, a good one. But you've just spent a dollar. You can't have that, not in the enterprise, not at scale. And so that's what the world we're looking at now of these orchestration systems that can

• 34:19Speaker 00

  work and in a production environment and really juggle everything in a way that transplants users, give the user control when they need it. And so that's the world where, and this is what I think certainly where Maestro is, but I think where the world is going in general. So what I was gonna say is what you see now happening in the enterprise before such an orchestration tool became available is that people do manual coding of static chains. So they would go and write a script, write a program, we'll call the L M here, check

• 35:04Speaker 00

  its output, call another L M, want the L M again, call some custom code, somebody went and wrote that code. And that's, again, that manual, it's a one-off for one use case. You have a different use case, a different workflow, you need to start it from scratch. And so that, that was the way until now, people controlled for the unruly behavior of L L Ms. But now we have A I to help control this through a very deliberate planning process.

• 35:36Craig S. Smith?

  Yeah. What do you think of this G P T four point five? That you let it run long enough and it, it comes up with an answer that's more accurate, less likely to be hallucination. Is that a system like as you're talking about, or is that a model?

• 36:03Speaker 00

  First of all, I have to say, to caveat it by, they haven't really shared the, the deal behind any of their models, but it clearly is a model.

• 36:17Speaker 00

  But it's a model of those chain of thought variety, where they've trained the model

• 36:26Speaker 00

  to predict not just the very next token, but a long sequence of tokens that hopefully represents some useful thinking pattern or reasoning behavior, culminating the right answer. And sometimes it works. The thing is, it's still a problem-pray regime. It can improve the answers. It's costly, like I said. It'll cost you much more than a single call, even a single call to an LLM, which itself can be expensive. And you get no guarantees. And it's something of a misnomer to call these reasoning models, because

• 37:09Speaker 00

  the thing that you see along the way sometimes corresponds to what you might think of logical or coherent reasoning. And sometimes it's just stuff. And so, you know, I call them large-musing models. The... One of my colleagues, Ra, who, well, a professor, or also one of our consultants,

• 37:38Speaker 00

  calls them chains of thoughtlessness.

• 37:42Speaker 00

  And

• 37:44Speaker 00

  it's a little unkind, because they do service a useful purpose. But they will not give you the reliability and cost control and controllability in general that the enterprise needs.

• 37:58Craig S. Smith?

  Maestro is out.

• 38:02Craig S. Smith?

  Jamba is out. What's next in your... In your... You, personally, in your work?

• 38:09Speaker 00

  Me, personally, it's complicated. I try to not interfere with smart people who are doing useful work.

• 38:19Speaker 00

  I will tell you, I have a pet project, which is I do in my kind of spare time of the weekends, together with my colleague Kevin Leighton Brown, who's a professor from Canada, brilliant, and also he happened to be working after the company, but this is a totally academic work on, we call it understanding, understanding, what does it mean to understand something? And do LLMs really understand? It gets to interesting philosophical questions, but in my spare time, that's what I do. But in terms of stuff that actually matters to the world and the company,

• 38:59Speaker 00

  we think this world of AI systems that are based on explicit planning is in very early stages, there's a ton of work to do there. What we've just released is a limited version of our system. It allows you, here you are, you could call an LLM, Jamba, but you could call Opus, you can call Grok, you could call GPT, 4O even, or instead, you could call Maestro,

• 39:39Speaker 00

  who would use those respective systems, but amplify the performance dramatically. On average, it amplified the accuracy by 50%.

• 39:51Speaker 00

  But that's the early functionalities we're making available. There's much more work to do in an AI system, how to plan things ahead, both at training time, at infant time, and so there's a ton of work to be done there.

• 40:10Craig S. Smith?

  I wanted to ask, you mentioned understanding, and as you said, this is kind of a bottomless philosophical question, partly because no one has defined understanding in any kind of a scientific way. But Jeff Hinton certainly believes that LLMs understand and think, what's your view? And where does understanding?

• 40:46Craig S. Smith?

  There's a gradient or a spectrum of understanding of and that that ends in awareness.

• 40:59Craig S. Smith?

  Where do you stand and where do you think the current L L Ms are on that spectrum and or not L L Ms, but A I models? And where do you think and do you think that it'll continue to climb that ladder? Or we've created something that has something that we call understanding and that's as far as it's gonna go?

• 41:26Speaker 00

  A lot of, a lot of stuff packed into that. Let me first say that

• 41:32Speaker 00

  um that I have the utmost regard for Jeff. Um, and for another guy, Andrew Ng, is my colleague at Stanford.

• 41:43Speaker 00

  And both brilliant people, but I mention Andrew because Andrew and Jeff had an online conversation about whether L L Ms understand. And agreed between them, they understand something to some extent.

• 42:00Speaker 00

  I think that is an ungrounded conversation because they haven't defined understanding. And what Kevin and I did to start out by saying, what does it mean to understand and how you, how would you evaluate it? And then we can meaningfully speak about. And it's a, it's actually a technical paper with lots of kind of equations, but I'll tell you, give you the gist of it.

• 42:30Speaker 00

  In order to demonstrate understanding, first of all, you don't just understand, you understand something, a domain. You understand arithmetic. You understand human biology, something you understand, I don't, you understand motorcycle maintenance.

• 42:50Speaker 00

  Okay, now, what does it mean to understand within that scope of that domain? So, there are really two or three things that you have to kind of keep in mind. One is, you've got to be competent. In other words, if I ask you, "Does the system or the person understand arithmetic?" You're going to ask the question, get answers. If by and large the answers are wrong, I don't care what, that person or that system doesn't understand arithmetic. That's kind of one baseline requirement of general competence. Let's call it the passing grade.

• 43:31Speaker 00

  You don't need to be 100 to be qualified in understanding arithmetic, but you can't get 15. That's one thing. The other thing is, you can't be ridiculously wrong. And what's ridiculous is a matter of context here. But let's say I want to know if the system understands arithmetic or multiplication. And I'll give it some multiplication problems, and it'll occasionally fail, and maybe long ones, but then I ask it to multiply 2 by 2, and it'll say 5. And I'll say, "Are you sure?" And it'll say, "Oh, yes."

• 44:16Speaker 00

  And it'll give me an explanation of why 2 plus 2 equals 5, and maybe some recent chatbot do try to gaslight me into believing that 2 plus 2 equals 5. That system does not understand mathematics. So, you've got to have general competence. You can't be ridiculous. And the third element is really important, which is explanations. You've got to give good explanations, and that's interesting. It took so long to figure, to understand why explanations are relevant here, and you see in high school, in school in general, in exam, a teacher will say, "Explain your answer." And why do they do that?

• 45:01Speaker 00

  You ask the student a small number of questions, three or five or 10. The set of questions you could have asked is huge, not infinite, but it's many millions of problems in arithmetic you could have given them. So what you would like to know is to... that the answers they gave represented. If you had answered others, you want to have confidence you would have gotten good answers there also. And one way is to ask enough questions that the statistics are such that it's unlikely that they would have lucked into a correct answer.

• 45:42Speaker 00

  But the way to amplify that is to actually give you an explanation. So if I tell you, show me one... solve this one multiplication problem, and tell me how you solve it. And I give you my multiplication procedure. And you believe that... I didn't just memorize it, I actually applied it. That's a good question, why should you believe it? But let's say you believe it. You don't need to ask me any more questions. Now you could have asked me any others and I would have applied the same procedure, or you believe I would have. So those are some of the elements in ghost who wants me to really understand something.

• 46:26Speaker 00

  And when you look at language mal today and take a domain that's not trivial, then you can show that they really don't understand the domain. What would it take to understand them? That's a longer discussion, but yeah. Now you asked another question and folded it. You said, "Will that give us... if we did understand, will that get us to awareness?" And now I'll really take you really far afield, if I may. Stop me if you... When I was a professor at Stanford, I had a freshman seminar that I called "Can Confusers Think? Can They Feel?"

• 47:11Speaker 00

  And this is kids who haven't been corrupted yet, the smart kids. And I would start with six questions. Can a computer... let's see, I can reconstruct the... computer...

• 47:26Speaker 00

  Think, can they understand? Can they be creative? Can they feel? Can they have free will? And can they be conscious? There, I got to six. And, and it was, you have to vote. No depend, no hedging, yes or no? And at the end of course, as we spoke about AI and machine learning, everything, they vote again, and the answers will be different. But I think the most interesting thing came out of it is that people realize that it's not so much, I don't know whether machines could in principle say, be aware, like you asked.

• 48:11Speaker 00

  If they start to question, what does it mean for people to be aware? And that's the thing, what's fascinating about AI, it forces us to think about ourselves, not just about machines. And I think that may be the most fun thing about AI.

• 48:28Craig S. Smith?

  That's fascinating. I should have you for another episode, just to talk about that, because I have more questions. But if somebody wants to try Jamba or Maestro, you say it's available everywhere, but the most direct way, I imagine, is to go to AI21. So what's the URL?

• 48:59Speaker 00

  So if you go to our landing page, ai21.com, all the information will be there. I just wanted to maybe clarify something. Jamba 1.6 is widely available.

• 49:15Speaker 00

  It's open waits unless you're a big company with more revenues and some threshold, then you know, you should pay us because you know, we paid a lot of money to develop it. So we need to recoup that money somehow. Maestro is in a closed preview right now. And so we're working with a small number of companies and there's a waitlist and we encourage people to sign up. We'll gradually bring more people in and at some point it'll be an open preview and we'll announce that.

• 49:49Craig S. Smith?

  Okay. And it's IA, I'm sorry, AI21.com. Is that the URL?

• 49:57Speaker 00

  That's the landing page and from there you'll get, you'll see exactly where to go.

• 50:01Craig S. Smith?

  This episode of Eye on AI is sponsored by the Definity Foundation. Definity Foundation is a Swiss not-for-profit that is home to some of the world's leading cryptographers, computer scientists, and experts in distributed computing. Their mission is to shift cloud computing toward a fully decentralized state by supporting the Internet Computer, also known as ICP. If you don't understand anything about the Internet Computer, you can see my episode with Dominic Williams. ICP's vision is that most of the world's software

• 50:43Craig S. Smith?

  will be replaced by network resident software that's an evolution of traditional smart contracts. To achieve this vision, ICP is designed to make smart contracts as powerful as traditional software, all while remaining tamper-proof, unstoppable, transparent, and verifiable. Since Definity's launch in 2016, they stand as Switzerland's most extensive blockchain research and development initiative and have been awarded more than 500 research grants worldwide. Definity remains steadfast in its mission.

• 51:24Craig S. Smith?

  To drive the advancement of the decentralized internet. Network resident software can now be used to run AI models and RAG infrastructure, preventing them from becoming quote-unquote "hot wallets", from where data can be stolen, and increasing resilience. Furthermore, the technology has been designed to allow AI models to spin up and modify running web applications and internet services solo by addressing several key challenges. If you're interested in reading more about the internet computer, visit internetcomputer.org.

• 52:06Craig S. Smith?

  I also encourage you to listen to my episode with Dominic Williams in which he explains the internet computer.