Jensen Huang の NVIDIA CEO 辞任と、Claude Fable 5.2 は AGI なのか

I cannot think of any company anywhere that is going to have more of a culture shock to losing their leader than NVIDIA will. All that and more on this week's Mixture of Experts. Hello and welcome to Mixture of Experts, IBM's weekly show, which we unpack the week in AI news. Later in this show, I will be joined for a special appearance from Sean Falconer, who leads AI product and strategy at Confluent. He'll be talking about the Streamhouse data architecture. But before that, we're going to have two stories with my esteemed panel of three guests. Brie Kopecki is AI engineer and a regular MOE guest. Q Gabriel is field CTO of data at IBM. Derek Robertson is a senior writer at IBM Think. I'm your host, David Zacks, also a writer at IBM Think. In just a moment, we'll be talking about Claude's release of Fable 5.2, but we're going to start with late breaking news, the resignation of NVIDIA CEO Jensen Huang.

Breaking news as we sit down to record this, Jensen Huang, the CEO of NVIDIA, the embattled CEO, has announced that he will resign. We're going to focus a little bit less on what happened and more about what we think this means for the AI industry. This is the rare CEO resignation that takes the very top slot on NewYorkTimes.com, for instance. This is not just a matter of business importance, but arguably world and geopolitical importance. Let's begin with you, Q. What are you reacting to in the news today, and what are you speculating about the future? Yeah, I think a lot of people, you know, NVIDIA is almost synonymous with AI now, And so I think people are like, oh, that's just another tech company that's, you know, big in AI. NVIDIA has been around 33 years and Jensen founded it. And literally that whole company has been built around him. He is sort of that cult of personality of NVIDIA. And he has enjoyed a very long tenure, which has allowed him to place a lot of really long term bets that very few CEOs in tech get that sort of liberty to make those kinds of investments.

and a lot of those were like, for example, CUDA, which most people will say, oh, yeah, that's NVIDIA's moat. That's what's given them this position in AI. For 10 years, people criticized Jensen's decision to continue investing in that. They're like, oh, it's a loser. You need to cut that off, and yet that is what put NVIDIA where it is today. CUDA is correct. Correct me if I'm wrong. It's like the coding language that people must use if they want to use NVIDIA chips. Is that right? That's exactly right. It's a set of frameworks and libraries that make doing the kinds of work we want to do with either objects or math or video or whatever and making it easy for programmers to leverage this supercharged GPU hardware. Because without that, using GPUs is actually very, very difficult. And that's why you can see some other companies with GPUs far, far behind or only used in more narrow cases like just inference now. So, yeah, thanks for clarifying that.

But I cannot think of any company anywhere that is going to have more of a culture shock to losing their leader than NVIDIA will. I mean, I have a lot of friends at NVIDIA. My heart kind of goes out to them because, you know, they're very proud of him as a leader and everybody follows him. And so now losing him is going to be very disorienting, I think. Bree, what's your reaction today and what are you looking out for? Yeah, I agree with Q. I think this is a monumental shock to the world of AI, but also to NVIDIA. I mean, Jensen is tied to this company out of his umbilical cord. He founded it over 30 years ago, and he has poured every single day of his life. He was recently interviewed. He always works. There isn't a day he takes off, not on Christmas, not on New Year's Eve. He loves what he does. He's a visionary, and he's passionate about what he does. So replacing him will be very difficult because you might put a very competent NVIDIA employee in his seat,

but they will not be able to match that. So it's going to be a huge challenge for NVIDIA to continue in the short term, especially in the long run. There may be people that step forward, but in the short term it's going to be very, very difficult. If we look to the world that has consumed NVIDIA chips and will continue to consume NVIDIA chips, I think there, too, there will be re-diversification, I think, happening. I think huge disruption in the market. It's going to have political effects as well because AI is a political – it has political implications, not just economical implications. So I think none of us can say exactly what has happened and what will happen, but there will be changes. Derek, you have considerable expertise in sort of tracking policy and looking at the big picture of the political environment around companies.

When you look at a company like NVIDIA, which for me is like a 90s kid, for me it was synonymous with video games for the longest time. Then all of a sudden it became this backbone of the future of technology writ large. And when you think about this incredible concentration of power, I think they're set to control about 90% of the world's AI-intensive GPU chips. What do you think, looking at this really with a wide lens? It's an open-ended question, but what are you reacting to today? Long way, since, you know, NVIDIA was what I associated with playing Half-Life 2 or something like that when George W. Bush was still president. But to piggyback off what Bree was saying, you know, as you mentioned, you know, I have background as a. political reporter and a policy reporter. And I have been thinking about, you know, this is also piggybacking off what Q said, you know, Jensen Long had such personalized leadership of this massively important organization. So as we've seen with many other,

you know, tech founders, people like Steve Jobs, Mark Zuckerberg, et cetera, et cetera, Their personal beliefs, their personal preferences are extremely important in ways that are simply not the case for other corporations that have more traditional forms of governance. Jensen Huang was someone who has always been very much in favor of doing business with China, having more of a traditional kind of free trade approach to chip manufacturing and chip sales, which is a disputed position at the political and economic, the highest levels of politics and economics, to say the least, right now. So I don't want to speculate, but whenever you have someone with that level of personalized control over an important organization leave the picture, even if only by one degree or however this ends up shaking out, we could see this could have real geopolitical ramifications.

And like I said, that all remains to be seen, but that's what I am going to be monitoring most closely. But also style implications, because I don't know that a CEO ever wears a bomber jacket again, which I think is a loss for, you know, cartorial diversity in the C-suite. You know, even Zuckerberg ditched the hoodie eventually and, you know, never to be seen again, seemingly. So you're right. We could be moving into a new phase here. Yeah, I think that I also weigh in with this, you know, to sort of the other side of the argument, which is some of the long-term bets that we see that Jensen has placed, they're high bandwidth networking. You know, that's already highly profitable for them. The things they're doing, like DPUs, the investments they've made in physical AI, which is, you know, broadly robotics and automotive, edge computing. They've really done a great job of pivoting this success that they have achieved around AI and setting themselves, building a hardware moat, mostly, around the company to make them very strong to be the continued, you know, dominant force in AI.

Because AI, a lot of people don't realize, it's not just the GPU. A lot of it's actually I.O., just moving the data around. And the infrastructure necessary to make that happen and feed these hungry GPUs, which can consume ever more data, that is very specialized hardware If you look back at Grace Hopper so this is a combination of a Grace CPU and at that time the H100 their front top of the line AI chip and the bandwidth the interconnect between those two chips, it's like 900 gigabytes per second compared to PCIe, which was, I think at the time it came out, 56 gigabytes per second or something. So massively different. And yet when it came out, everybody was like, what are we going to use this for? I don't know what we're going to use that much bandwidth for, but now we're starting to see that that set them up for refining that technology, and now in the Vera Rubin line, that technology is dramatically more potent,

and they've kind of worked out the kinks. So just because people say, oh, we're going to come up with a competing product for this thing that NVIDIA has, don't just think that the Rev1 is going to work well, because it usually doesn't. So they've got a really good lead in a lot of key areas. I'm going to move us along to our next story, this one, a product release. So Anthropic is out with Fable 5.2, also Mythos 5.2. The two are kind of synonymous. Mythos is the more cybersecurity-minded and has lax certain safeguards that Fable, which mere civilians like me, use. But Fable 5.2 came out, and of course, it's inevitably become a conversation about whether we are at that famed threshold of artificial general intelligence, AGI, which is variously defined, but something like it could do anything that a pretty smart human could do. But I thought one way to just get quickly introduced, there's a wonderful academic named Ethan Mollick, I believe at Wharton, who's I think like one of the best AI users in the world.

And he published on his blog, oneusefulthing.org, and I encourage you to check this out directly, this amazing video that he generated with Claude Fable 5.2. Remarkably, it's doubly remarkable because as Fable notes in the video itself, Fable doesn't generate video. So how did it do this? Well, let's take a look at the video and get a sense of what's going on here. The End

Amen. Derek's going to weigh in on Schopenhauer and Schlock, but for me, I've succumbed to it. I have a little bit of a chill in my spine. To think that what one prompt resulted in, who knows how many hours later, that output, a tremendous amount of reasoning and planning and thinking, what is really new about me, Fable 5.2? What is distinct about me?

What would impress an audience? Well, what would probably impress an audience would be a video, but I don't even make video. Wait, let's make that the point of the video, the fact that I can't make a video, but that I can write a video with text through, we were talking about this earlier, but it was probably generated code in a vector language that renders as a video if presented in a certain way. So inevitably, I think that Ethan Mullick needed to do something to actually make it become the video that we saw. for it to decide and maybe consult a library of product launch videos of the sort that are common because this feels like a product launch video that you would get directly from an Anthropic or an OpenAI to decide that it needs a twist ending and that the cute twist ending would be to have its very prompt output be what's it called, an acrostic, that is the name of the blog. So much happened between that one prompt and the output. We don't know how long. We don't know how many tokens. but it's, I'm feeling, I'm feeling a little bit, my nervous system is reacting. How are you guys reacting? What do you think is we're looking at here? You said your nervous system is reacting.

And I think that is so great to point that out. We're, we feel something when we watch this video. We feel something and then we think, wait a minute, who are we? Are we that simple? Can we just be reduced to numbers and letters? Is it really that simple or is it not? Is there more to it? I get very pensive when I'm facing this kind of marketing because it is marketing. So I think we have to understand whenever a new product and a new model is a new product. When we're looking at an AI company that releases a new model, it's a new product. And for people to want to use it, there has to be marketing around it. I think it may be a tiny step closer to AGI, but as someone who builds AI systems every day with some of the most advanced companies in this nation, I can tell you a model alone is still not everything you need to make it work.

You need a harness. You need a ton of scaffolding for your AI agents to do the very thing it's supposed to do. So it's way more performant. I have tested it. It is faster, and it's incredibly smart going through the weeds of tons of data. But it is not AGI at all to me. Kiyu, do you think it's AGI? Do you think that's the question we should be asking? Yeah. Yeah, we just keep moving the goalposts on what is AGI. Yes, it is absolutely AGI by some standard of where the goalposts were, I don't know, a month, two months, five months, a year ago. Two years ago, let's just say three years ago. Three years ago, this is unimaginable science fiction. Literally no person on the planet would have looked at this and said this is not AGI. Nobody. I don't care what field they were in. And so the fact is we just keep, you know, saying, OK, increasing the, you know, like I said, moving the goalposts.

Yeah, I think the thing that I that I love to see, you know, talking about this video to some degree. And like Bree said, it makes you feel something. One of my generally very heartfelt topics on this is the story for creators, people who, whether you're an artist, painter, writer, creative profession. And I'm going to actually include software engineers in that. If you haven't really lived in the software engineering world, you've never built software, you don't understand that it is every bit the same level of art as any other expressive art, and that it can be taken to that level. And when I write software, I feel I get into flow. I feel that creativity flow. Same thing I would with almost any other creative practice. and yet I will probably never write software again and I think a lot of people had to mourn that and when I get into every new level of model I get into like Fable

you start to see how it's unlocking new creativity new ways to be creative I can express my creativity my desire to build things and show things and teach people with the things that I want to express, I can do that now in a completely new way. I'm not saying I didn't have to learn new things, but it's a new unlock for those who step forward and really embrace it and say okay I going to learn to create with this And Fable really takes that scale way up Like I think that for small tasks I was able to do the same things I could do with Opus, even earlier versions of Opus. But in terms of if I want to do some really, like, crazy ambitious work, like one of the cases that Anthropic Reference was they built an operating system with it, that's a very sobering level up.

And to say that even for my work where I'm not going to spend $10,000 worth of tokens, it's still opening doors to me that were out of reach even with Opus 5.5 or Astra models. Piggybacking on the theme of creatives, Derek, you and I share an interest in filmmaking. I know you outside your IBM world, you wrote a review recently of a biography of was it Spielberg and Coppola and Lucas and I'm a obvious filmmaker. Derek, if you turn up at Sundance Film Festival in five years and there is an AI produced film in the lineup, what is your reaction? Are you going to see it? How are you going to review it? How are you going to think about it? Well, I would certainly see it. I think that to some point this question is already complicated by the fact that I'm absolutely certain that animators are, you know, using the AI tools for certain things in the context of animated films. I think that

to, and I was thinking about this a bit as Q was talking about software development, there is an, as a creative person, I think that there is an irreducibly human element of creativity to some extent, And, you know, I've been thinking about this a lot in the book world recently because there have been a string of high profile scandals of authors whose many times award winning work have been found to be generated by AI in part or sometimes in whole. and I do not think that you can legitimately say that something like a book or a film that is entirely generated by AI is a work of art because I think that art requires something that is irreducible and to the point of this video that cannot be reduced to a formula or a proof

or a description, something that is in code. Does that mean that these tools have no place in artistic endeavor? I think that's just an interesting conversation that will be developed as years pass and the tools become more sophisticated. But to sum up, what is a very long answer to your very simple, I think, yes or no question? If I were to sit down in front of an entirely AI-generated movie, I would view it as a curio and a technical proof and not necessarily something that I am enthusiastic about engaging with as like, you know, part of the human endeavor that also brought us Ovid's Metamorphoses and the film Chinatown and, you know, Tom Jones and Moby Dick. So that, yeah. But what happens if, like Bree said, what happens if that movie makes you feel something the same way that a better or worse movie may? How do you reconcile that? I'm curious, generally, how do you reconcile that? It's a good question.

I mean, there's a lot of human-generated art that I think makes me feel something that also makes me feel resentful because I feel that I am being manipulated to feel something by somebody who is being sentimental or, you know, otherwise manipulative of my emotions. And that is something that AI is really good at. I think that even if I had that reaction, I would still feel the same way because I think there is a difference between having a connection between myself and another human being through the proxy of the art form and myself and an engine that was fed a prompt. I, you know, just as an essay, as a critic, not to be pretentious about it, I am not convinced yet that a prompt is a high enough level of engagement for me to connect with you in that manner. I haven't seen anything that has changed my mind about that. I strongly agree with that. And I'd like to use this panel as an example.

Imagine the four of us were speaking to other AI agents, right, like an AI that looks like us or looks like a person. I think we'd be differently engaged. I certainly wouldn't be as excited to hear about Q's opinion. I'd be like, okay, it's an AI talking to me that knows a lot of stuff. But I'm interested in Q's opinion. I'm interested in Derek's opinion. And I'm curious what David is going to ask me next because you're human beings. And I would even argue if we all sat in the same room today talking to each other, the conversation would also be even more invigorating because we'd be feeding off each other's nonverbal cues. We all know more than 90% of communication is nonverbal. That, too, would be introduced into the mix and we'd be even more excited. So I think consuming a piece of art and something can definitely be artistic that was made by AI, but we all know behind it is a prompt and probably an extremely good prompt.

And we all know prompt engineering is an art, and there is even a job that is called prompt engineer. Well, it used to be, now it's context engineer. Anyway, it keeps evolving. But we know what's behind it, and it is not a complex, messy human being that has been through the depths of the valleys of life and then the most joyous moments of riding a horse or, you know, holding your newborn baby. I think that's what makes the difference. Thank you for that. Yeah, just as a last question and a bit of a backtrack off of that, what would be a perfect flourish and end to the episode and maybe will be in the edit. But, Kiyu, I am curious. You include coders and software engineers in the creative professions. And, you know, I'm thinking about the things that automation replaces and the things that it does not. No one decides to run from New York to LAX.

They that technologies have have replaced the the sort of motor function of getting, you know, it was a horse, then it was the train, then it was the airplane. Nevertheless, we have the Olympic events where people run. We have marathons where people run, where they celebrate individual achievement. The same thing might unfold. The same thing has unfolded in chess. Right. No one really cares. Well, probably a handful of niche people are interested in seeing AI models can play chess against one or the other. But if you follow the chess world at all, which I do lightly, all the focus is on Magnus Carlsen and on the human personalities playing chess, all of whom can be trounced by the, you know, stockfish, right? So, Q, what is the... I'm trying to figure out what are the criteria by which these domains of human creativity persist, flourish, should continue to have attention share, market share, and what are the criteria by which they die away? Relatedly, the simple question I wanted to ask you is,

are there human coding competitions that exist that people track where they want to see human coders compete at human coding tasks, or is that just dying away? Yeah, predicting the future on that is a tough... I can see that persisting, you know, those kinds of things persisting in what I'll call analog creative practices, that is painting, sculpting, those kinds of things that produce physical things we can touch, more so than in software. I mean, although I will say that one of the coolest evolutions of these, just the thing we demonstrated in that video, is like I'm doing work with video and rendering 3D user interfaces that are completely out of reach for my skill set, my software engineering skills. And more importantly, they're completely out of reach for most people's willingness to invest time into the product. I can use LLMs to build extremely creative and sophisticated software on my laptop

that can help me do the work that I want to do in the way that I want to do it Like for example I work well with faces and I don do as well with names So like when I looking through my email and my Slack I want to see faces So I built an app that helps present that to me in faces on a timeline, which is the way my brain works. And so I think we're going to see some really cool creativity there. Whether or not, you know, back to your original question, you know, whether or not we're going to see that continue in, like, competitions, I don't know. But what I do know is that the bar for amazing software is going through the roof. We're going to see some of the coolest software exploiting this hardware, which has been available for a long time over the next, you know, it's already happening now. On a daily basis, I'm seeing like, OK, that's awesome. And it's going to go more and more as more people who do not have the ability to write software directly discover that they can now express their ideas and exploit this hardware in a way that they could never do before.

Which brings us to the question that Derek pointed us to, which is that some animators are already using this stuff. So the question of human-machine hybrid creativity is a very tangled and gray one that we will have to address on a future episode. I really enjoyed this talk today. We're going to keep it to two stories today. We went deep on two. I appreciate that. Kiubri, Derek, thank you so much for joining. I'm going to move us to our third segment, which is an interview with Sean Falconer of Consulant. So I'm joined by Sean Falconer. Sean, thank you so much for joining today. A few questions. Let's just begin with this one. What is Streamhouse? Well, that's a great question. So really, it's the name that we in a working group have given to an architectural pattern. And the way I think about this architectural pattern is it's a way for businesses to really be able to deliver what the current state of the business is. So imagine you're some enterprise, you've got data all over the place. It's very hard to answer sometimes simple questions in that scenario.

Like if you're an airline, what is the current state of a flight? Can I push back or not? because you end up with these kind of fragmented systems of record that are all traveling at different speeds. So they have different sort of state of the current truth. But we are to trust AI agents and AI systems to run key parts of our business, take on operational parts of the business. They really need to be reasoning and taking action on what that current truth is. So that's fundamentally the problem that the Streamhouse is designed to solve is to be able to make all your business data current. Now, how does that happen is through this pattern of essentially having change be the fundamental primitive of the data architecture. So a lot of times when we think of data, we think of tables, rows, and columns. But the thing that happens in businesses is change is essentially the only thing that we rely on. Your change is happening all the time. So we want to be able to capture change incrementally from all these different source systems, hundreds of systems, if that might be the case,

and then be able to use technology to essentially transform and reshape that data to create universal representations of that information. So you think of like a customer 360 or a product representation, latest invoice, and then make that data also instantly queriable. So you're still able to operate with the data in a table structure, but it's change that essentially is the fuel that keeps everything up to date. Your use of the word state struck me because when I hear people talk about state, Often it's in the context of an application, right? Is this application stateful or not? Is it remembering where it was? Was there even one state of a business prior to the era that we're living in? To answer your question, could a business have an accurate view of all their state in the past? I don't think that we've really had data architecture that allows us to answer that question easily. You know, people have kind of stitched things together. And I think very sophisticated companies have been able to put together a kind of collection of technologies together to be able to solve that.

So they may be using, you know, some sort of, you know, fast database for serving. They're maybe using Kafka to incrementally capture change. They're using stream processing to sort of stitch that stuff together, sync it down into this fast, you know, cache or fast representation in the database to be able to serve this need. But you're stitching a lot of like open source technologies together that are kind of on you to make them work together. It's on you to be able to scale and operate them. The other paths have been centralization of data. So bring all our data together in a warehouse or lake house. The challenge there is that even though that in many businesses, that's probably the most accurate view of their data in one centralized place. it's never going to be the current truth because of the fundamental nature of how those businesses work. Like those are data architectures work. The proliferation of data is much faster than our ability to centralize it. So it's always going to be trailing what the current truth is. So if I had an agent that is reading from the centralized data store to understand which hotel rooms are available in a hotel block right now and I'm booking somebody.

well, while there's going to be some delay between the actual system of record where hotel rights are happening about which hotel rooms are available and the source of truth that my agent's reading from. So that could be a multi-minute delay and then I could actually end up booking a hotel that's already been taken but it hasn't been reflected down to the centralized data because it happened essentially in the transactional system. So that's a bad world to live in if you're trying to build AI systems that can actually make you more efficient as a business or operate end-to-end scenarios without a human essentially babysitting it. I feel like as a non-technical person, we're moving into a realm where the metaphors are getting more difficult to grok, right? So back when it was a data lake, I could envision a lake, you dump stuff in it. Back when it was a data warehouse, I could envision that there were rows and columns, as it were. Now that we have a stream house, how can I visualize this? What do you think should be the governing visual metaphor people have now as they conceive of this new architecture? You could think of it as the, you know, even the world of like lakes and streams, you know, to stick with the water metaphor, you know, typically the streams is the fuel that goes into the lake, right?

And the ideal scenario is that essentially all the data or all the water that's going to end up in the lake start originating in these streams. A good way to think about it is the value of data or some information that you have. A lot of times the value of that data changes depending on the time delay from when you can actually look at it. So to use the surveillance scenario, imagine we're security guards and we're watching a videotape of a break-in. We could be watching that break-in happen right now, and that has one value to us. Like we could call the police. We could probably prevent the break in. Maybe we stop the robbers, arrest them in real time, prevent loss of goods and so forth. I could also look at the tape 18 hours later. And that doesn't have zero value, but it's a different value. So the value there is more retrospective. It's more, you know, I can now think about, OK, well, if this happens in the future, like maybe we need a different set of security protocols and so forth.

And I could prevent future break ins. It's harder for me to go and potentially find those people that were responsible, get the goods back. By now they probably sold it or it's gone. So it has a different value attribution. It's not zero value. It's just a different value. Both are important in the business world, just as they are in this hypothetical surveillance world. But they have a different value. And I think what businesses have really struggled with had not had really the fundamental data systems to solve the first part of that of I'm watching this videotape in real time. How do I prevent the problem from actually happening? And how do I reason and take action right now? But if we're going to have AI systems that are actually operating key parts of the business, they need to know what this truth is. If there's one thing I know about AI agents, they work quickly. They can work quickly, right? so you might as well get them that data in real time. Thank you, Sean, for this introduction to ScreamHouse, and congrats on the launch and articulating this vision for an architecture for the future.

Absolutely. Thanks for having me, David. And thank you to all our viewers and listeners. If you're watching this on YouTube and you wouldn't mind hitting like, comment, or even subscribe, that really does help us grow the show. I'll look forward to welcoming a new panel next week, and thank you so much for listening. you