The video explores the intriguing discussion surrounding artificial intelligence's potential impact on major recognition awards, like the Nobel Prize. It centers around opinions from notable voices in the tech world, questioning if AI may soon be acknowledged with prestigious accolades in the humanities, such as literature or film. The talk underscores AI's significant achievements in science, particularly with recent Nobel Prizes in chemistry and physics awarded to pioneers of AI-driven projects.

The video draws attention to AI's evolving role across different domains, positing whether its advancements justify such high accolades or if they might be perceived as hype. Experts like CTO Chris Hay and VP Edward Kalvisbert discuss how AI is reshaping industries and its implications on human contributions to intellectual outputs. The conversation expands on AI's journey from early exploration by figures like Jeff Hinton to modern-day applications, shining a light on historical contributions and the debates surrounding AI's future responsibilities.

Main takeaways from the video:

Please remember to turn on the CC button to view the subtitles.

Key Vocabularies and Common Phrases:

1. distinguished [dɪˈstɪŋɡwɪʃt] - (adjective) - Used to describe someone or something that is recognized for excellence and is respected and admired. - Synonyms: (renowned, eminent, prestigious)

Chris Hay is a distinguished engineer and the CTO for customer transformation.

2. collaboration [kəˌlæbəˈreɪʃən] - (noun) - The action of working with someone to produce or create something. - Synonyms: (cooperation, teamwork, partnership)

So it's going to be AI in humans in collaboration.

3. hype [haɪp] - (noun) - Exaggerated publicity and the intense promotion of a product or idea. - Synonyms: (exaggeration, publicity, fanfare)

Are the Nobel Prizes basically scumming to AI hype, or is this the start of something way bigger?

4. theoretical [ˌθiːəˈrɛtɪkəl] - (adjective) - Involving the theory of a subject or area of study rather than its practical application. - Synonyms: (conceptual, speculative, academic)

Has been a little theoretical and not hitting with the real world for a while.

5. cornerstone [ˈkɔːrnərstoʊn] - (noun) - An essential or basic element; the foundation. - Synonyms: (basis, foundation, keystone)

Actually, I think it's a good thing because this is the cornerstone of the next few years.

6. nuanced [ˈnjuːɑːnst] - (adjective) - Characterized by subtle distinctions or variations. - Synonyms: (subtle, refined, intricate)

It's not just speed, it's scale and the capacity to consume more data and to have more nuanced relationships between that data.

7. commoditized [kəˈmɑːdəˌtaɪzd] - (verb) - Made into a commodity; turned into something that can be bought and sold. - Synonyms: (standardized, marketed, commercialized)

I think we go through cycles where hardware becomes totally commoditized.

8. synthesis [ˈsɪnθəsɪs] - (noun) - The combination of ideas to form a theory or system. - Synonyms: (integration, amalgamation, unification)

These are very different use cases, very different impact on individuals, on society, and they pose totally different risks, so it's really not just about that technology, it's really what the technology is being applied to that I think is, needs to be assessed.

9. inference [ˈɪnfərəns] - (noun) - The act of drawing a conclusion based on evidence and reasoning. - Synonyms: (deduction, conclusion, reasoning)

But actually, I think the bigger thing is on inference, right?

10. exponential [ˌɛkspəˈnɛnʃəl] - (adjective) - Increasing rapidly in amount, number, or degree. - Synonyms: (rapid, escalating, accelerating)

It's an exponential rise in both the training and inference capabilities.

AI in the Nobels, DGX B200 arrival, and Unstructureds $40M funding round

It's 2027. Has an AI generated work won the Nobel Prize for literature? Chris Hay is a distinguished. Enjoy. Let me start that again. Give me that pause. Man starts yelling even before I ask the question. I'll do the intro. 2027. Has an AI generated work won the Nobel Prize for Literature? Chris Hay is a distinguished engineer and the CTO for customer transformation. Chris, welcome to the show. What do you think? Absolutely. And while we're at it, AI is going to win a few oscars and an Emmy as well. Okay. All right. And next up, Edward Kalvisbert is vice president, product management for the Watson X platform. Edward, welcome to the show. What do you think? No way. I don't think the Noble institution will allow board. All right, cool. Well, with a difference of opinion like that, you know it's going to be a good show. All that and more on today's mixture of experts. I'm Tim Hwang, and it's Friday, which means that it's time again to take a deep dive with our experts into the week's news in AI. We're going to talk about OpenAI's new DGX B 200, the new round of funding for a company called Unstructured. But we're going to start today with the big news story of the week, which is basically that AI has been taking the Nobel Prizes by storm this year.

In the prize for chemistry, David Baker, Demis Hasabas, and John Jumper took the prize with Hasabus and Jumper winning in part for DeepMind's work on alphafold. And then in the Nobel Prize in physics, one of the founding fathers of modern AI, Jeff Hinton and John Hopfield, another leading light of the field, won for their work in neural networks. So I think, Chris, I want to start with you first, is what do we make of this? Are the Nobel Prizes basically scumming to AI hype, or is this the start of something way bigger? I love it, actually. I think. Well, I think the Nobel Prize uses, if I'm being completely honest, has been a little theoretical and not hitting with the real world for a while. So actually recognizing AI and the impact that it's going to have in multiple fields, such as physics and chemistry, and if we think about the big innovations that are going forward in the next few years, it is going to be more AI led, right? So it's going to be AI in humans in collaboration. And how do you distinguish that? Well, is it fair to say that the people who founded AI in the first place then don't get rewarded for that work course not so. Actually, I think it's a good thing because this is the cornerstone of the next few years where AI is going to massively help in these areas. So I'm all for it. Go for it. And while we're at it, I think my next AI means I'm going to become the MVP of the NFL soon as well. So, Aaron Rogers, watch out, here I come.

I guess, Edward, to bring you into the conversation, Chris has already taken a very strong stance that basically it's going to be a few years from now and it'll just be AI winning every single Nobel prize award. I guess, Edward, there's two questions. I think what you said in the opening question, you were like, well, I don't know if the Doebell institution will really allow it. Does that mean that you don't think it's deserved or basically that you think the institution won't really being into awarding and giving AI its due? Yeah, I mean, I think mainly the former. I mean, I do think that AI is going to be an incredible tool in almost every aspect of our lives, and it's going to do amazing good for society and well being and quality of life and basically all the things that the institution stands for. But I do think the human contribution to those outputs that deliverables is really essential. Right. So whether it's 10% AI or 90% AI, I think if it's 100% AI, maybe it goes a little bit too far. Right. I think, Chris, one thing I wanted to do is, I think I take a lot of pride in the fact that mixture of experts is a good way for people who might not be reading every single archive paper, working their way through every single machine learning textbook to learn about what's happening more deeply in the AI space. I think what's interesting is you can often get lost with all the stuff that's happening at the enterprise layer around AI. I think there's a lot of people listening to this who may actually not even know Jeff Hinton. I guess I'm curious if you feel comfortable. We want to give our listeners just kind of a quick explanation for why someone like Hinton is so kind of important to the field and what exactly he sort of, like, contributed here.

No, absolutely. So I think the first thing I would say is Hinton really is kind of, he's considered as the OG goat of godfather of AI, which is. I love that term in that sense. And he's been doing AI for a very long time, machine learning specifically. He has been there before. It was so he was doing that right, back in the sort of 1980s, right? So if we were actually really uncool, basically, right? Yeah, very uncool. Exactly. So, but if we think of the modern foundations of what we've got today, so we think of things like deep learning that all comes down into these really deep, massive neural networks with billions and even trillions of parameters these days. Right? And I'm not gonna go into the massive details of that, but if we look at the work that Hinton has done there, even as far back as sort of 2011, right, when Alexnet came out and Ilya Susque was part of that as well, then that was a time where really the deep learning revolution kicked off, which was the sort of first kind of CNN on a GPU for training against images. And if we go further back in time there as well. So if we look at the work that he did for things like backpropagation, which is a key cornerstone of what we even do today with deep learning. So all of this goes back to the eighties, and Jeff Hinton was doing this when it was uncool. Right? So if he hadn't done that work, we wouldn't be where we are today. So I think you have to sort of recognize that fact. And as I said earlier, the impact that AI is having and going to have in the future is going to be incredible. So I think actually the impact that it's having in these different fields, he should be recognized for the work that he's done, right.

Even in physics. Right. Because I know there's some physicists I saw on my Twitter feed grumpy about like, well, what's this computer science person doing it here? I guess kind of ultimately you're like, actually this is significant enough that, like, it actually should be recognized in this context. Absolutely. And I think it does open us up. Right? So I think it's a good thing for physics as well, right? You don't want to be seen as this kind of boring thing. Here's a bunch of formulas. Oh, look, here's another telescope in the sky. Do you know what I mean? It's like, this is stuff. Exactly. AI is impacting every field. And therefore, I think it's a really good move by the Nobel Institute. Edward, I'd love to bring you into this because I think one of the things I love about Jeff Hinton in particular is just how down to earth and open he is about his views. There's a great quote that the Nobel committee had posted on Twitter about how he was like, oh, yeah, I was just in this low rent hotel room when I got the news and I had to reschedule my medical appointments to go deal with this Nobel Prize win. I think one of the things that Hinton has been very strong on, I would say, in the last few years, is warning about the risks of AI. I think people have taken him very seriously because he's been at this, of course, for a very long time. As Chris explained, he's the goat hipster of neural nets. I guess I'm curious about how you think about those as a leading light in the fields, his kind of sort of dark warnings about where AI is going. Seriously, do you think he's on the right track? I'm curious about how you kind of think about those kind of risks. And he, I think, made it actually a center of, like some of his comments during his, some of his interviews around this prize. And so I did want to make sure that we talk about it before we move on to our next topic.

Yeah, I mean, I think he's raising the warning just to make sure that that voice is always considered right. That risk is always kind of part of the, part of the math that enterprises, individuals, institutions do when they're applying AI to the particular problem or use case that they're applying it to. And I really think that's what it comes down to when we think about risk assessments or AI governance. It's really in the intersection of the technology and the use case. It's not the same thing to apply AI, to do creative writing, as we mentioned this morning, than it is to do credit underwriting for a bank or to do a hiring decision for an organization. So these are very different use cases, very different impact on individuals, on society, and they pose totally different risks. So it's really not just about that technology. It's really what the technology is being applied to that I think is, needs to be assessed. You know, the more this technology makes its way into international security, into defense. Right. Obviously, it's a much different consideration than poetry. Yeah, for sure. And is this, are you hearing this like, you know, because you work very close to the metal in some ways, right? Like the Watson X platform is something that, like, customers are using and relying on. I mean, sometimes I think I feel like a lot of discussion about, like, oh, AI is really dangerous, kind of takes place in like a totally different domain. But I guess I'm kind of curious. I mean, it sounds like you're sort of suggesting that, like, even day to day you're sort of hearing from customers and the market that these kinds of risks and these kinds of concerns are things that people are thinking about. Yeah, I mean, they're not existential risks, right? No, but they're definitely risks to brands. Right. Their business risks, their regulatory compliance risks. And managing these risks is definitely one of the top considerations that enterprises are.

That's really acting as an inhibitor to more wide scaled adoption of the technology. And it's something you can't really do after the fact, because so much of managing this risk is the end to end lifecycle. It starts with the data that goes into the model and what model you selected and how you customize and tuned it all the way to monitoring and guardrails, separation of duties, right. Between development, deployment. So all these things that you kind of have to start thinking about from the beginning, because if you don't, then at the end they become a wall or a real obstacle to try to reconstitute post fact. So we've been working with clients in that perspective, with that approach, and that's what's leading to some of these use cases making their way into production. And I wasn't being hypothetical when I was talking about credit risk underwriting and hiring decisions. These are real world use cases where the risk is being assessed mitigated in order to implement these workflows. Yeah, for sure. Chris, do you want to get a final shot here? Curious about what you think about sort of, I guess Hinton's kind of late career turn as being sort of like a voice of warning around some of these technologies. I like to think of this as the difference between waterfall and agile, which is probably a weird way of putting it, which is go into that more if we think of waterfall projects.

Nobody does waterfall projects anymore because we realize that we are too dumb, and I mean this in the nicest possible way, to figure out every requirement in advance and be able to plan everything because the world is too complicated. I sort of feel that way about AI risks. I think we are too dumb to figure out every single risk and every exploitation and be able to get ahead of everything in advance and pre planden. So therefore, I kind of think like a software project, I think we need to be agile, which is you need to experiment, and then you need to discover in a safe and controlled fashion what those risks are and let them evolve. And that means we're going to do dumb things. We really are. Right. But then in the process of doing dumb things like sticking your fingers in a wall socket or whatever, you realize, oh, I better not do that. Right. And then you put safety things in there. Now, I'm not saying that we should go that far with AI, but I hope that history tells us that in human existence, we've done enough dumb things, that we shall do enough dumb things before they become catastrophic dumb things. So I think a little bit of agility will help us discover that stuff. We need to have control. But I don't think we are going to all blow ourselves up because I think we're going to do much dumber things much earlier. That is my opinion. I'm going to move us on to our next topic. There was an incredible photo that OpenAI put out onto social media. It's of the team celebrating their receipts of the new Nvidia DGX B 200. And it's a great photo because you can see clearly that everybody is so jazzed to be standing next to this fresh new piece of compute that it's like Christmas morning. It's like people are so thrilled to get this computer in their hands. And I think it's a nice cook to talk a little bit about this kind of next generation of platform that Nvidia is rolling out and is actually having a really material effect on the market for compute.

There's a great chart I saw earlier in the week about how the prices for Nvidia H 100s, which were last season's, gotta have it hardware, those compute costs are just dropping all of a sudden as these new boards are coming available and online. I think it's a nice hope to talk a little bit about what's happening in the hardware markets. And I think maybe, Edward, I'll turn to you first. Is what we're seeing here just more speed? Right? Like, I guess there's one kind of point of view, which is it's Christmas morning, because it's really cool to be standing next to what's basically like an f one racing car for compute. But, like, is what we're getting here largely just faster? And if not, you know, what's different about it? Yeah, I mean, I think it connects back to the first topic we talked about, right.

The evolution of this technology and really trying to build it in a way that somewhat models the way our brain works. Right. And this kind of almost infinity, I know that's a big word, but of nodes and connections and relative strengths between them. It's not just speed, it's scale and the capacity to consume more data and to have more nuanced relationships between that data. So I'm not a hardware expert, but I think it's a great time in technology. When hardware matters again, I think we go through cycles where hardware becomes totally commoditized and then it matters again and then eventually it becomes commoditized again. So we're definitely in a stage where it matters. I think that's a signal that the innovation, the innovation frontier is active and moving rapidly, and I think that's all very positive. Yeah, I mean, I think, Chris, it's stunning. I was talking to a friend of mine who is working on some of these clusters and he's basically like, the hardware is literally moving so quickly here that they can only really afford to do like one big training run on a cluster they've built and then almost immediately they start moving to building the next cluster that they're going to do training run on. I guess I'm kind of curious here as someone who kind of like thinks about this and researches in this space. Where is this all going? Is the cluster just going to get bigger and bigger and faster and faster? Does this top out at some point or what is the trend here in the next twelve to 24 months? I think there's a couple of trends going on, and I think I might have said this on another episode, but I'm going to say it again. It's almost like following the bitcoin trend, which is if you follow the bitcoin trend, everything started on CPU, then it moved from GPU's and then it moved to FPGA's and it moved to ASICs.

Basically you went from kind of cpu, you went from compute being CPU to being GPU bound, and then you were going to custom hardware. And we're kind of seeing the same thing again. Because people, you need to bring the cost of compute down, you need to bring the cost of training down. That is, you've got bigger and better models to train. But actually I think the bigger thing is on inference, right? So you got to run these models at a low cost and speed. Now, if there is one criticism I would say of Nvidia over anything, is that the speed of tokens per second and the cost on these GPU's are quite expensive. And you've seen this in the marketplace already. This is where folks like rock have been coming in and they've been sort of releasing these chips that go really, really fast and then IBM's got their north Pole chip as well, and then Google's got their TPU chip. So everybody's trying to bring down the cost of inference because if you're running these massive models on the cloud, everybody's consuming compute, you want that to be as cheap and as fast as possible. The big thing if you look at these new Nvidia boxes is, yes, the training speed was much faster, but actually, if you look at that chart, the cost of inference, the speed of inference came down massively. They've obviously put a focus on that as well because they know that if they don't improve the inference speed, if they don't improve that inference cost, then all of these other providers are going to start eating their lunch as well because everybody's going to go cheaper. But I think this push and pull between kind of general purpose gpu and sort of custom chips is really important. But again, in the training point of view, different from inference, everybody's just focused on, I need to get the biggest and fastest, I need to get my model out really quickly and therefore throw away your last card, put in the latest card, because I just need to get my model out all the time. So there's a different dynamic that's going on over time. You're going to get faster architectures, you're going to get different architects going to get cheaper, and these cost speed performance ratios are going to change over time.

Yeah. The architecture bit of this component, I think, is a really interesting part of the market. I think one theme that we've had pop up on a lot of mixture of experts episodes is customers want smaller models, they want faster models, they don't want the gigantic model that's really expensive. Right. And so there is that pressure there, but it feels like there's kind of two ways of getting there, right? One of them is, well, we start marketing just smaller models right, where we, like, lower our demand. The other one, which you're arguing is, well, the chips get good enough that the cost of inference finally falls for running larger and larger models. And the two are kind of like in a little bit of a race, it sort of seems like, and I don't know if you have predictions on kind of who wins that race in the end, because you can imagine, like, the market might eventually settle and say, hey, look, these models do 99% of what we need them to do. We don't need crazy near AGI models to do this work. So at a certain point, you just don't need the chunkier model. Right. I guess there's another point of view, which is, well, but if the cost is cheap enough, you would still go bigger. And I guess I'm kind of curious about how you think about that relationship. It's a little bit complex and it's unclear where it lands in the market today. I think it's just going to keep pushing and pulling because we are going to want to run our models on device if you think of things like Apple intelligence, et cetera. So I think smaller models and faster compute are just, you're going to need both for a while. Will one win it? Yeah. Will one out win out? I don't think so, because the smaller that you can make the models and the faster you can make and smaller you can make the chips, then the more you can put them on embedded devices which open up a whole set of other scenarios which are kind of low latency.

And again, you even see that like this week. So what llama three two was out last week and they released their 1 billion model and their 3 billion model. I think it was right. And again, just smaller models. And I think the big thing there is, folks are getting really good at taking these larger models and distilling them down into much smaller models, and that's going to continue. I think we're looking at 1 billion parameter models. But let's project forward maybe six months, a year. You're going to then start to be back into the million parameter models, and then the chips are going to get faster and we're just going to go back and forward, back and forward, and it sucks forever. Yeah. Edward, are you seeing that in the market? I mean, it feels like one interesting outcome of what Chris is talking about is that there's a lot of market pressure to have a lot of the models, just more on edge devices everywhere. And it strikes me that part of the idea of a platform is you're running it in the cloud and all the advantages of cloud. But it does seem like there's actually really powerful kind of economic incentives eventually pushing us to sort of all on device here, not really like in the model that we're familiar with. Do you think that's a real possibility going forward? I mean, I think it's going to be all of the above, and we're the hybrid cloud company, so edge to us is definitely a continuum. The data center compared to the hyperscaler cloud is effectively a type of edge, and then you go down to facilities and eventually devices. So yes, it's going to be all of the above. And finding the right balance is always very specific to the requirements of the use case. I think what we see a lot is clients to get started, use a big model, because that's a way of accommodating a very broad range of requirements. Use cases, languages, all sorts of things, right? So you kind of prove out the business case with a big model that's going to help you accelerate. But then when you're there, you're like, okay, how can I do this as cheaply and with the least latency as possible?

Right? And now you start to really kind of optimize and customize. Right? Once you've validated that business case and really want to want to scale it with real economics behind it. So it's like you use the swiss army knife, it's going to give you a lot of flexibility, but eventually you're going to want to use that fit for purpose tool to get the job done. Yeah, that's super interesting. I never really thought about it as this lifecycle, but it's very interesting that, well, just for the pilot, we use the biggest, baddest model because it gives us the most optionality. And then as an organization, kind of tunes in the use case, it gets much more discreet and smaller, and you're optimizing for cost and all these other sorts of things. It's very interesting. Is this the time to mention agents? I realize we haven't mentioned the word agents in this episode yet, so, I mean, we're not contractually obligated to talk about agents, but if you want to mention agents, Chris, go for it. You can do the final hot take. Before we move on to the last topic. This is needed for agents because you need. Your agents are going to be highly specialized, they're going to work together, and they need to have low latency etcetera. So actually, the smaller model and being able to run on device and being able to run, whether it's on data center, on device and run in different locations, that is 100% necessary for this agentic world that we're in. So it's a good thing. Agents, for sure, agents a lot more to get into there, for sure. So for our final story of today, I really wanted to make sure that we had a chance to talk about a company called Unstructured, which recently closed a $40 million round. And this round was led by IBM and Nvidia and a long list of kind of prominent companies and investors in the space.

What's most interesting about unstructured is that it's a company that focuses purely on transforming unstructured data into structured data, which is not something that you normally think of as being something that you'd invest $40 million in. So I wanted to make sure that we talked about it first. Edward, if I want to bring you in just like if you want to resolve that mystery for some of our listeners, like, why is unstructured data important? And why is structuring it incredibly, incredibly important for AI? Yeah. Well, unstructured data is most data nowadays, right? And I think the most relatable type of unstructured data, the most usable type of unstructured data today for LLMs is document data, right? So it could be the content on the Internet, or word docs, or PowerPoint presentations, right? But effectively document data, and that is enterprise knowledge. Right? That is, that is what runs the world, right? It's, it's these documents in this language information, and that's what large language models are built on, right? That's what they're trained on, and that's what they're excellent at processing, summarizing and making usable. Right? So bringing that data, bringing that enterprise and institutional knowledge to the models is really the way in which, in which an organization can make it their own, customize it to the knowledge of their business, the language of their business, the tone, the entities, the relationships, the values, all the things that you need to do to put a model in service of a business or a goal. You need to do that by effectively teaching it with your data. And that's what this company focuses on. I've met them, they're very focused. I think that's really been part of their strength and success. They're very focused on taking that unstructured data that relies in different locations and different formats, and then making available for the models, particularly inductor stores for retrieval, augmented generation as an initial use case, which is effectively universal at this point, but then beyond that, you know, identifying relationships in the data for a graph rag, taking the data and putting into structured format to really increase the precision and accuracy of some of those queries. So I think rag is, you know, very popular, really valuable, but already kind of running out of gas a little bit for the next evolution of use cases.

And that's really all about continuing to unlock the value of the data in those documents. Yeah, that's really interesting. Can you go into that a little bit more? Why is rag running out of steam? It's kind of like, again, it feels like twelve months ago is the new hotness, or people were still definitely leaning into it as the key strategy for doing retrieval. What's missing, where the cracks appearing? Yeah, I mean, I think it's a great starting point, and I think it's essential in most cases, right. But for example, graph rag is going to give you the ability to have richer contextualization right. By identifying non obvious relationships. If I prompt the model with a certain set of words, it's really only going to limit its ability to reason, right, including retrieving the, the knowledge base to that domain. And there may be hidden relationships. There may be. For example, if I'm going to search something about Facebook, but I don't get a response about Instagram, then I'm not really getting the whole picture. But the model is not necessarily going to know that Facebook and Instagram are related because those relationships could potentially be non obvious. So the graph rag pattern is going to give you strength in relationships that are non obvious and in doing so, provide you richer contextualization that will be more relevant, right, to the question being asked, even if it's not asked with those specific words, right. So it's, it's again, mimicking a little bit of how our brain works in identifying those relationships. That's, that's one example. And, but even that is not necessarily going to be perfectly accurate, right, because there's data about transactions that may have like a SKU, a SKU number or a particular id has no semantic value. It's just a bunch of characters. It's like your license place. It doesn't really mean anything, right. So you need to have that type of data in structured formats and really combine rag or semanticsearch with SQL with structured queries, and that's going to give you more accurate responses. The questions that have, you know, to do with transactions or other types of data that are, that are very, you know, very important to a particular business, very important to particular domain, but don't have semantic value in a conversational or language sense, right? So now you have to complement rag with a different dimension, which is structured data. So those are just two examples, right, of how you really need to complement kind of classic rag to make it more, more accurate. That's really helpful, Chris.

Again, I think there's another, I think this story made me think a little bit about the market for data structuring, which I think is really interesting. Whereas we normally think about the people who generate data, the people who do the training, the people who offer the models to the consumer as the supply chain. One link in that chain I haven't really thought about is just this layer that exists between the data that's out there and the data that's usable. I guess one question I want to ask of you is that it feels like there's lots of different potential ways you could go about doing that. There's companies like unstructured where we have a specialized service that does this structuring of data for you. You might imagine that the foundation models themselves become good enough that they can do the structuring out of the box. You don't actually have to do much additional post processing to make it happen. You could imagine that synthetic data gets good enough. We don't even need this unstructured data because we can just generate a purely out of nowhere. And it feels like there's a lot of contenders to the throne of getting data that's usable, I guess. How do you size that up? Do you think that at some point, say, synthetic data just gets good enough that you don't need to do this data structuring anymore? Or is there always going to be a niche for this kind of structuring business? Just curious about how, where you think this market is going.

Oh, goody. I get to say the word agents again. My favorite. Yes, please do. Yeah, we got to get a few more in before the episode's up. So actually, I think everything is going to move into a marketplace in the future. So I do think we're going to have a marketplace of data, we're going to have marketplaces of agents, and we're going to have marketplaces of models, and I think we are going to get more outcome focused. So specifically on the data. I think we're doing a lot of human work at the moment to curate that data. And even if you look at things like structure, et cetera, but they do great work because they're actually taking away a lot of the complexity to get your data into your vector databases, to follow rack, because it is really hard. You have to do things like chunking. You're constrained by the context of the model, that is the short term memory that it can work within. You have to work out which data is going to be associated with what. As Edward saying, you need to start building out things like relationships. Then you've got to understand, I've got to get this data from this format. I'm getting this from an s three bucket, getting this from where we, it's really complicated, but actually we are, even though that's a faster process, we are still humans who are figuring this out and doing transformations, et cetera, and doing these sort of ETL pipelines. If I project a little bit forward in the future, back to our earlier discussion, where the models are going to be smaller, they're going to have less latency, they're going to have faster tokens per second, you're then going to be able to train these smaller models to start to do that restructuring work. And therefore, I think you're going to be in this world where agents are going to help you get your data into a structured format. And once your data's into a structured format, you're going to be able to train your model, and then you're going to loop around and you're going to be in this nice, virtuous circle.

So will there be a marketplace for this? Absolutely. Because at the end of the day, people own data, right? So the publishing companies, the media companies, they're all sitting on gold mines right at the moment, because that's data that is highly valuable, highly creative. There are things that are probably can be synthetically generated, so things like all the math data, et cetera, you could probably argue that will just be commoditized over time, because that will just get generated and synthetically created, and that would be the same for anything that are puzzles, games, et cetera. So there will be this push and pull of who owns that data. And I think that human data, especially the creative spaces, will still be highly valued. So I don't see the record companies giving up their ownership or songwriters of. Yeah, exactly. So I think that's going to be the push and pull that we have over time. But we are going to be moving into this marketplace where sort of that soft ip is just going to be the big thing that distinguishes companies. Because one of the examples I like to give is if you have got a model trained and you have the data of all the kind of spanish legal texts, and you've got that structured etcetera, and your model can answer spanish legal queries better than any general purpose model. If I'm going into court, you know what? I want the model that's really good at spanish law, as opposed to the model that's got a vague understanding of spanish law, because that's the difference of me getting a large fine or going to jail, right? There's a huge value on that locality, and I think that will be one of the biggest trends as models are going to get more and more specialized. And we're just going to be like we've been having with the general purpose benchmarks, MML use and all that.

We're going to have a benchmark for everything. You can imagine, Tim. It's going to be, here's the spanish legal benchmark, here's the car parking benchmark, you name it. It's going to be benchmarks everywhere, and we're just going to be in this big mess. Of marketplace of specialization. I love the image of hiring an agent, AI agent, attorney. Right, to defend you in a case. I mean, I think that is a feature I can get behind. I used it myself. I did an insurance claim. I looked at the insurance document. I was like, I have no clue what any of this means. And it was a kind of medical condition thing. And I was like, I run through the LLM. It's like, tell me. Gave me the key points, wrote to the insurance company payout, and you're like, you know, this could go somewhere. Exactly. That's what you want from these things. So I'm. Yeah, but we are going to be in a wild ride. We are. We are going to be having like the kind of the uber style marketplaces where you're matching up AI's to people, AI's to AI's. It's going to be wild over the next few years.

Edward, do you want a final thought to close us out for the day? Agency agents? Agents, absolutely. I mean, some of the work we're doing at IBM with agents is super exciting, and it really is going to kind of. I think it's going to be a step function, right. In terms of the complexity of the workloads and the use cases, the creativity to solving problems potentially beyond even our approaches, the automation. Right. The fact that you're gonna have so much work happening 24 7365, a lot of stuff already works that way. But this is gonna take it to the next level. And I think it's exciting, it's productive. I think it's gonna level the playing field for consumers, in some cases for smaller institutions. Right. So we're excited to be part of this future and to really be co creating it with our clients and our community.

Artificial Intelligence, Innovation, Technology, Nobel Prize, Ai Risks, Machine Learning, Ibm Technology