Bryan Casey: AI agents—what are we expecting next? How do we put planning and reasoning alongside this large representation of the world we have now? Are we going to have products that truly never incorporate generative AI? I think “never” is such a strong word. And what’s the most exciting thing happening in hardware today? It’s nice to see that finally we built big computers again.

I’m Bryan Casey, and welcome to this week’s episode of Mixture of Experts. We let Tim go on vacation this week, so you’re stuck with me. I’m joined by a distinguished panel of experts across product, research, and engineering: Volkmar Uhlig, who is the VP of AI Infrastructure; Chris Hay, who is the CTO of Customer Transformation; and Skyler Speakman, senior research scientist.

There’s been a lot of discussion in the market around reasoning and agents over the last six months or so. So the question to the panel is: do we think we’re going to get more progress in building reasoning capabilities through scaling compute—and this is just over the next year or so—scaling compute, algorithmic progress, or from good old-fashioned software engineering? Volkmar, over to you.

Volkmar Uhlig: Very clear: algorithmic progress.

Chris Hay: Software engineering.

Skyler Speakman: Algorithmic. That’s the next step.

Bryan Casey: I like it! We got some different opinions on this. This actually leads us into our first segment, which is a company called Milon released a new paper around “Agent Q.” This paper demonstrates improvements in reasoning and planning. The scenario they defined was using an agent to book restaurant reservations, using LLMs combined with other techniques like search, self-critique, and reinforcement learning. They demonstrated an order-of-magnitude improvement in the success rates of LLMs.

Maybe, Skyler, as a way of kicking us off, I’d love to hear why do LLMs struggle so much today with reasoning? And why is the work going on in this space, exploring other ways, so important to making progress?

Skyler Speakman: LLMs have this amazing ability to build a “world model.” I’ve seen that phrase popping up more. Sometimes it gets criticized—“oh, they’re just predicting the next word”—but in order to predict the next word as well as they do, they actually do have this... I’m not going to say “understanding,” that might be a stretch... but they have this model of the world. Up until these new recent advancements, they had no real reason, motivation, or agency to go out and explore that world, but they had created that model and could answer questions about it.

So I think LLMs did a very good job of creating the model of the world. I think these next steps are: now that we’ve got a representation of the world, which is pretty good at the next-token prediction problem, how do we actually execute actions or make decisions based on that representation? That’s the next step we’re seeing, not just from Agent Q, but from lots of research labs: figuring out how to put planning and reasoning alongside this large representation of the world we have now. These guys are off to a good start, one of the first to put something out there with a paper available. Lots of other companies are working on it, so I wouldn’t necessarily say they’re ahead of the pack.

Bryan Casey: Yeah. Maybe, Chris, I know we were talking about this: how indicative do you think the work this team did is of where everybody’s going in this space? Is this paper just another piece of data in a continuation of everybody exploring the same problems? Do we think this is pretty dialed in on where the problem space will be around agents over the next year or so?

Chris Hay: I think it is pretty dialed in. When I read the paper, it’s similar to some of the stuff we’re doing with agents ourselves, which is always good. If you really look at what’s going on, they’re not using the LLM for the hard bits; they’re using Monte Carlo Tree Search to actually work things out. One major thing they’re doing is using a web browser as a tool. If they’re trying to book a restaurant, they’re doing a Monte Carlo Tree Search and navigating using that tool to different spaces. They’re using the LLM to self-reflect and to create an initial plan, but they’re relying on outside tools and pieces like the tree search to work out where they’re going. The fact is, LLMs are not great at that. It’s more of a hybrid architecture.

Everybody’s doing the same thing with agents: bringing in tools, outside memory, things like graph searches—Graph RAG is becoming really popular. Everybody’s bringing in planning and reasoning. I think they’re doing interesting stuff with self-reflection and fine-tuning to create a virtuous circle. They’re probably further ahead than a lot of people in those spaces. But even open-source agent frameworks—we started with LangChain, now you see things like Glang, Graph is becoming popular, and then moving into multi-agent collaborations like CrewAI. Everybody’s on a slightly different slant, but they’re definitely on the right track. And by the way, back to my earlier argument: that is software engineering, my friend. That is not doing anything different with the LLM; it is engineering and putting stacks and frameworks around your toolset.

To that point, Bryan, I do want to hear Volkmar’s take on why “algorithmic” was his pick. You have to hold us to our answers. He’s going next.

Volkmar Uhlig: My background: I built self-driving cars for seven years, and we always had this decision between how much software engineering we can do and how much we can train into a model. Often, as Chris just said, it’s a packaging of different technologies together. Where we are right now is we have this really powerful tool, the LLM, which gives us a basic form of world understanding, a world model. Now we are trying to make it do stuff, which we haven’t seen; it’s not just predicting the next thing you do on OpenTable. You’re in an unknown, open world where you need to explore different choices.

I think the next step will be: you run this brute force, and once you have those choices, you will train a model. That’s my expectation, based on the path with driving. We always came up with some heuristic, a huge data corpus, tried something out, and in the end, it was always, “Oh, now that we figured out the underlying problem, let’s train a model to make this more efficient in execution.” In the end, the model is just an approximation of an extensive search. So I think that’s why, algorithmically, I believe the algorithms we will build are effectively those graph searches, tree searches, etc., which ultimately will feed into a simpler representation that is easier to compute in real-time.

Chris Hay: I was kind of disappointed by the paper, if I’m honest. And Bryan’s dreading what I’m about to say now. I’ll tell you why I was disappointed: the whole example was the OpenTable example. Now, unless I am wrong—and I don’t think I am—isn’t Milon the company that claimed they were the agents behind the “Strawberry man,” the “I Rule the World Mo” Twitter account? That would have been the agent example I would have wanted to see in the paper.

Bryan Casey: It is. That was actually a question I was thinking about because they talked about reinforcement learning as part of that. One interesting thing I’ve seen in the market the last few months is a light backlash happening to LLMs within the ML community, particularly from people who’ve worked a lot in reinforcement learning. You hear folks talking about LLMs being a detour on the path to AGI. As we’ve slowed down a bit in progress, I’ve seen the folks in reinforcement learning starting to pop their heads up more, saying, “Hey, it’s back! The only way we’ll make progress is with other techniques.”

So, two questions. First: if we fast-forward to a world where agents are a significant part of the software we use every day, do we think LLMs are the most important part of that? Or, Chris, to your point about this paper making extensive use of other techniques, do we think a bunch of other techniques will rise back to prominence as we try to make these things do stuff? So, yeah, anybody have a take on that?

Chris Hay: Yeah, I definitely think RL is going to come back into this. I know they were using RL in that paper, and also things like DPO. I keep thinking back to AlphaGo and the DeepMind team winning at Go—they were using similar techniques. If you take a deep learning algorithm today and get it to play a simple game like Snake or the Atari games—very simple architectures like CNN/DNN-type things—they absolutely rock that game. If you get an LLM to play, it doesn’t matter if it’s an agent or not, it’s the worst playing of Snake I’ve ever seen. GPT-4o is terrible at it, Claude is terrible at it. But really simple RL/deep learning/CNN-style architectures actually rocket those games.

Therefore, I think as we try to solve and generalize, some of those techniques that were really successful in the past have to come back. We’re going to see software engineering, improvements in architecture, improvements in algorithms—it’s going to stack, stack, stack. Hopefully, all these techniques will come together into a hybrid architecture. But when you put LLMs into an old gaming-style environment, they absolutely fail today.

Bryan Casey: Do we think there will be general-purpose agentic systems over the next couple of years, or is everything going to be task-specific? One nice thing, Chris, about the OpenTable example is it’s a very easily definable objective, which means you can pull in other techniques in ways that are harder to make fully generalizable. So, when we look at agents, do we think we’ll make progress on generalizable agents in the next year or two, or is everything going to be in this task-specific land? Skyler, maybe you have thoughts.

Skyler Speakman: No, I don’t think we’ll have general within two years. I think there will be areas—this might lead to our next topic—areas around language, creativity, that will surpass some human abilities. But the world works on much more boring, mundane business processes, and I think there’s still a lot more ground to make to get those systems to a level of trust that people will use. It’s one thing to have these methods create a funny picture or write a funny story, but to have LLMs execute financial transactions on your behalf? Different ball game. We’re not going to be there within two years. I’ll be proven wrong—you can timestamp this—but yeah.

Bryan Casey: We’re always accountable for our predictions on this show.

Volkmar Uhlig: So, Bryan, I think where we may go is we will probably go through examples—OpenTable, and try another 20. I think we will get into a tooling phase where you can explore a domain with some human intervention and guidance. You’ll have tools that can explore, say, a web page, how to interact with it, and then go through a pruning process that may be manual. But I think we will get to more automation, where it will be 10 or 100 times faster to build this. But, as Chris said, there will be a software engineering component until we are fully autonomous—where you just point at something and say “learn”—that will take a while. The question is, where does the information come from? Through trial and error? Or could we even just read the source code of the web page? We have source code in business processes; I can just give you a billion lines of code of SAP adoption...

Bryan Casey: The second story: the CEO of the company Procreate—they build illustration tools—on Sunday night, their CEO released a video saying they are “never” going to include GenAI capabilities. He actually said he “hates” GenAI. The reaction from their community and the design community was super excited and supportive. As we’re recording, that video has almost 10 million views on Twitter.

I have a bunch of reactions to that. One striking thing is how two different creator communities have reacted to the arrival of LLMs. I have friends who are software engineers, and for LLMs for code, people are generally enthusiastic, see it as a great productivity tool. I also have friends who are writers, work in Hollywood, creatives, who look at this technology like the Grim Reaper. Wildly different responses.

Chris, throw it over to you for initial thoughts. Any sense of why these communities are responding so differently?

Chris Hay: I think “never” is such a strong word. That would be one of my reactions. “Never?” So far, really? No feature at all? I’m “never” ever going to stream video content because I believe physical is more important? Well, you know what, Blockbuster? You’re out of business. I don’t know.

I think there is a general wave. I applaud them; they make tools for their particular audience, and their audience doesn’t want that. I think that’s a unique differentiator. I’m not sure how that stands the test of time. “Never” is a strong word; the industry is moving fast. Different audiences have different needs. I’m pretty sure if I use Procreate, there’s no chance I’ll produce anything of artistic quality because I have no artistic talent. I am not the target audience. But I am grateful for AI-generated art because it allows me to produce something I never could otherwise, like PowerPoint slides.

So, if they are focused on creative professionals, and those professionals don’t want AI within that, I understand. That’s great, you’ve got your audience. But I think the tide of time will push against them, and it’s a strong artisan statement.

Bryan Casey: Before we move on, Chris, what sort of PowerPoint art are you doing?

Chris Hay: I mean, generally, if I’m honest, it’s almost always unicorns with rainbow-colored hair. That’s my go-to for CEO presentations. Every CEO loves a picture of a unicorn. Sure, all the other ones do. You know, it resonates.

Bryan Casey: It resonates with me. But Skyler, Volkmar, I’m curious if you have takes on the community’s reaction to these two different sets of tools.

Volkmar Uhlig: I think we are in a world where we have artists and craftsmanship, and we are going through a phase of automation of this artistry. The bar will be really high, and there will always be unique art. Today, I can buy photography or a copy of a Monet and hang it on my wall, but there is still a demand for people to have their art. I think that will stay. We’ve seen this progression: horses used to be transportation, now they are a hobby. Old cars are going the same way. Hopefully, at some point, airplanes.

These unique pieces of art... if I can automate the creation and industrialize it, the industrialization wins. It always wins. But it doesn’t mean those tools and that craftsmanship shouldn’t be supported; it will just shrink dramatically because the capabilities become more accessible. If you used to have typists, now everybody can type; all the typists are gone. The same thing will happen.

Bryan Casey: One thing I thought was interesting is your point about craft. A lot of people choose their life’s work because they like the craft of it. They chose to be an artist or a developer because they like doing that work. Having a tool come in and do all of it can feel like it’s robbing value from what they do day-in, day-out.

Another place I see tension is in the relationship between management and practitioners. My observation is that management is often particularly enthusiastic about adopting these tools for productivity benefits—get more done, reduce cost, drive revenue. As they’ve gotten more senior, they might be one step removed from the craft, so the loss of the craft feels less consequential. But to practitioners, it’s like, “This is my thing, and this tool is doing it for me.”

I’m curious if you’ve observed any tension between management and practitioners in terms of enthusiasm for adopting this technology.

Skyler Speakman: I’m not sure about tension between management and practitioners. There might be some I’ve witnessed about which flavor or version. They’re going to say, “We’re going to use this one,” but behind the scenes, somebody’s using a different tool. There’s tension back and forth on that. So it’s not necessarily the adoption, but maybe the channel or the tool has had a bit of that. So yeah, that’s what I’ve observed.

Volkmar Uhlig: I think it’s also the question: when you look at craftsmen, there’s 20% of work you love and 80% you hate. Often, the majority... ask a data scientist: 80% is data cleaning. Do you think they like data cleaning? No. So, if the tools support the toiling, the useless work, and make people more productive, you shift more into the work you actually like and appreciate.

From an engineering perspective—I’m mostly talking software engineers—I think it’s an improvement. Nobody likes Jira ticket reviews and writing comments. If that can be automated away, that’s an improvement. Or I don’t need to go to Stack Overflow to find an algorithm; I can just ask the model. I’m more at the architectural level.

From a management perspective, they want productivity, but productivity in an engineering process often means convincing people to do work that’s necessary but everyone hates. To a certain extent, it’s an improvement on both sides.

Bryan Casey: That’s a great point. I always say—probably not safe for work—some percentage of your job is work none of us want to do, but we’ll spread it around the group. That description resonates. A lot of teams I work with operate on IBM.com, dealing with content. The property has millions of pages. We’re trying to do more with automation, connecting content. It turns out, to do that, all your tagging has to be really good across tens of thousands of pages. The amount of time spent cleaning up metadata... it would kill your calendar for three days. If we can build a good classifier to do that, it’s a huge relief and lets us focus on the work we signed up to do.

So, on my team, we’re looking at tedious but important work. Can we automate it so we focus on the core work? But when it comes to using LLMs for the core thing, everybody’s still a little skittish, honestly—not on the software engineering side, but on the more creator side. So announcements like Procreate’s resonate with me because I see it with folks I work with.

Chris Hay: I think another thing is, it’s not just tedious stuff. For prototyping and ideating, it’s really good. It doesn’t matter if you’re producing content, code, or images. Sometimes you think, “I have an idea, is this going to work? It will take me a lot of time to build it up.” Let’s get the LLM to do something, or the image generator, to get an idea, then I’ll start pruning and building the idea more.

Personally, from a software development side, that’s how I work. Right now, I’m trying to create a distributed parameter service for training LLMs. There’s no chance I could just sit and code that straight up. I need an LLM to help me figure it out, and then I’ll engineer through it. I think that’s true for image generation, too. If you’re doing a concept and need that unicorn with rainbow-colored hair, get the image model to generate it, then you can go, “Okay, that doesn’t quite work in context,” and then draw your pretty unicorn. Prototyping is a really important use case.

Volkmar Uhlig: And, Chris, when you’re doing that prototyping, you can have a dialogue with a machine and get major refactorings done in seconds. “I want this other thing, let me split this into four classes, or collapse them.” The amount of work you’d have to do... that’s all the tedious stuff, refactoring code. We have IDEs, but they kind of suck. If you can get an LLM to do that, it’s amazing. The time... you can do it in an hour on a plane. You can write massive amounts of code and experiment with it.

Bryan Casey: Before we leave this topic, Chris, I think we need to remind ourselves that you asked kind of an art question to three nerds. I’m safe in saying that, right? Just put a disclaimer here. It would be a fascinating conversation to have artist representation on this question. All of this—we’re talking about inevitability and tools—that’s where our brains go. A really fascinating conversation to have with the artists.

Bryan Casey: One reason is I have friends in both fields and have observed how different their reactions and their communities’ reactions are. There are interesting economic factors. I think there’s less concern about real industry disruption in the software engineering community than on the creative side. There’s a core underlying economic anxiety that’s not quite the same, even though you’re dealing with models helping productivity in different domains. It will land differently. It’s a great point; we didn’t fully represent the other side. It’s a super interesting topic.

To the point about “never,” I feel like many tools are used as part of a workflow, and you don’t even know the underlying technology. If you want to remove a background from an image, do I know if it’s GenAI or something else? Do I even care? In some cases, “never” seems strong. It will be interesting to see how this space evolves over the next couple of years.

Earlier this week, AMD announced the acquisition of ZT Systems. As everyone knows, the hardware space has been one of the biggest winners so far in the GenAI/LLM cycle. AMD is making a big play in this space. Their CEO, Lisa Su, was on CNBC talking about the acquisition. One thing she said is that AMD historically invested a lot in silicon and software, but this acquisition is about bringing together stronger systems capabilities.

Maybe, Volkmar, to kick things off: why is this market moving from just silicon to systems? Why are these vertically integrated systems so uniquely important?

Volkmar Uhlig: If you look at AMD’s offering... AMD acquired ATI a decade or two back; that’s the heritage of their AI accelerators. They are kind of head-to-head with NVIDIA. NVIDIA did very well over the last couple of years by looking not only at the GPU but at many GPUs in a box, and for training, multi-box—you need many machines. The integration... NVIDIA acquired a company providing the software stack for large-scale clusters (Base Command) and Mellanox, the leader in reliable network communication.

AMD is sitting there thinking, “What do we do?” They don’t have a consolidated story for putting a 10,000-GPU training system on the floor. They’re locked in the box, not yet at scale to compete on the training side. That’s why NVIDIA owns like 96% of the market. When you’re training, you pretty much only use NVIDIA. You’ve done all the coding and optimization for CUDA. Running inference is trivial; switching an ecosystem is hard.

NVIDIA went down the route of having the DGX system—full systems with network communication. AMD is just now catching up. They’re catching up on the network against Mellanox with Ultra Ethernet, and now catching up on how to get big systems at scale into the industry. They need to get into cloud providers. So, systems... being a boutique shop that makes large-scale infrastructure deployments happen is a good conclusion.

Bryan Casey: One observation I have about the GPU market is that it feels more vertically integrated than the world of CPUs, at least somewhat. Would you agree? And is building out the unique requirements around the training stack the core force behind why this market is behaving this way, differently? Or do you see it differently?

Volkmar Uhlig: The training system market is traditionally the esoteric high-performance computing (HPC) market. At IBM, we built #1 and #2 Top500 supercomputers with Blue Gene. Suddenly, we’re in a world where that is not just a domain of labs dropping $100 million on a computer. Now, every company that wants to train a network at scale needs similar technology.

After 20-40 years of HPC being an esoteric field with maybe 50 supercomputers in the world, suddenly it’s a commodity. You start up and say, “We should all have a supercomputer!” But you don’t. The joke: “I’m GPU-poor, I only have 100.” To play in that market, you need to offer a solution. AMD has been traditionally in the desktop or enterprise GPU market; they sell silicon but never built these systems.

NVIDIA, being a GPU vendor, has captured like 85% of the dollars spent in the data center. Your Intel chip, a bit of memory, and everything else—NVIDIA takes the switches, the Ethernet cards, the GPU. For AMD to get something deployed at scale, they need an offering on par. Intel, with Gaudi, is in a better shape because they have 50-year partnerships with Dell and Lenovo; it’s easier for them to get into that market. That’s not the case for AMD.

Chris Hay: This is why I don’t get the acquisition, Volkmar. If I was a company that sold Granny Smith apples, and nobody ate Granny Smith apples because everybody bought Red Delicious, why would I buy a company that makes better packing boxes for my apples? If I’m spending $5 billion, spend it on getting better GPUs and go compete with NVIDIA. That’s where I don’t understand it.

Volkmar Uhlig: I think NVIDIA figured out a way to deploy to partners, and AMD got locked out. They need a way to market. A huge percentage of training happens with the hyperscalers. Companies want to put NVIDIA cards on-prem, but often, for early beginnings, they go to the cloud. NVIDIA is delivered to hyperscalers. For AMD, this is a way to get into hyperscalers with a solution, saying, “We give you the whole thing,” reducing the risk for the hyperscalers.

Chris Hay: I’m not sure people want to use NVIDIA. I think NVIDIA’s got this market locked, and they’re awesome, they make great GPUs. But at the same time, Apple is doing well on the desktop/laptop market with their chips and MLX as a framework. Google invested in their own ASIC-based chips with TPUs. Other people are moving into ASICs. I think there is a space for a low-cost alternative to NVIDIA chips. Otherwise, other companies and hyperscalers wouldn’t be investing in that. That’s why I don’t get it. If I was a competitor, I’d try to find a niche space that isn’t the “packing boxes.”

Volkmar Uhlig: In the training market right now, NVIDIA is the only choice. I think that’s primarily where AMD is trying to break in. In the inference market, there will be Apple, Qualcomm, a ton of chip vendors, and startups making low-power chips. But in the training market, if you look at the wattages AMD is putting down—going above a thousand watts on a GPU in next generations—NVIDIA is effectively the only game in town. AMD wants to put something up against it.

For pre-training, you need huge systems; for fine-tuning, you can often do it in a box. But in the pre-training market, you buy NVIDIA or you buy NVIDIA. Gaudi isn’t there yet, AMD isn’t there yet. This is an attempt to break into that large-scale training market. Companies run 100,000-GPU training clusters; building that takes a year and is a billion-dollar investment. To capture some of that revenue, you need to have that capability. It’s not three engineers putting up a supercomputer; it’s a construction process. With this acquisition, AMD finally has a chance to bring in the guys with the hard hats because you need to put in power and cooling. They don’t have that experience right now; they’re buying the competence.

Bryan Casey: That point about competence came out a lot post-acquisition. This company has capability around building large-scale clusters, some of the biggest in the world. It’s a theme I’ve heard at every level of the GenAI stack. You’re rate-limited by the amount of expertise in the market—on the hardware side, the training side, even prompt engineering was called “magic incantations” for a while.

Over the last two years, as this has blown up, it feels like some skill shortages are getting less acute. More people know how to train models, work with models, and are attracted to hardware. I’m curious: across the board, how much do you feel our progress in AI is still rate-limited by raw expertise? How much has that improved over the last year or two? Skyler, kick it over to you.

Skyler Speakman: I have this conversation regularly with our director. I would say it’s not necessarily the overall amount of skills—that is monotonically increasing—but how it’s distributed across the globe is becoming more extreme. That’s something we are experiencing. We’re IBM Research Africa; we represent a billion people, but the talent here is likely to immigrate. What does it look like to have that talent here and bring that culture here? Yes, it is increasing, but at very different rates across the globe. That’s my short summary. It’s something we talk about regularly: what does capacity in generative AI look like on a truly global scale? That’s probably another session entirely.

Bryan Casey: That was a fascinating perspective I wasn’t expecting. Chris, Volkmar, thoughts?

Volkmar Uhlig: There is such a gold rush; it’s a new technology. It’s about trying it out; every day there’s something new. You need people who are passionate, spending their living and half-sleeping hours on it. The skill set will develop over time. I feel like we’re repeating the gold rush of the web era: “Oh my God, you can write a web service!” Now, everybody can do it. We’re in this uptick with an extreme supply shortage.

It’s deep. When you just plugged a computer into a network, it was relatively easy. Now, training is different. Do you need to understand math? Most engineers hate math; that’s why they like computers. This set of skills needs to be built up. Until it rolls into universities and we get true practitioners—first you need education, then to become a practitioner and toy with it for 5 years—I think for the next 10 years, we’ll be in this world of supply shortage everywhere, plus the speed of change.

On the flip side, coming from systems, it’s nice to see we’re building big computers again. I really like that. We’re going away from “the cloud providers do everything for us,” and we need to look at system design with a fresh angle. That’s a goodness for the industry. It was kind of locked in; only five companies in the world knew how to plug a computer into a network and power socket. It’s good we’re going through a renaissance of computer architecture and design.

Chris Hay: I’m the total opposite. I think people are learning the skills and doing a great job across the globe. But at the end of the day, to train a large language model, you need an awful lot of GPUs and access to an awful lot of data. That is outside the access of the average human being. There’s a lot of great skill talent that won’t be able to practice their craft because access to GPUs to learn the effect of data isn’t there. They can learn from fine-tuning and training very small models, but for larger models, capabilities emerge at higher scale. The scale now is tens of thousands of GPUs. That is locking out the average practitioner.

Personally, I want to see more distributed compute, more access to GPUs and skills. That will open up a talented set of people distributed across the globe to make great contributions. But at the moment, it’s concentrated in big tech companies because they have the GPUs.

Skyler Speakman: Chris, I want to fight back on your fighting back. That’s why we do this, right? If a researcher comes to me and says the only way they can make their case is they need 10,000 GPUs, that’s not a good argument. That researcher needs to make their case off of two GPUs. Show that, then we can talk about 100, 2000, 10,001. I don’t think it’s fair to say, “I can’t make progress unless I have 10,000.”

Chris Hay: I agree, Skyler. But again, we’re sitting in a company that has 10,000s of GPUs. They can go to you, make the argument with two GPUs, and then you can give them access to scale. The average person might get so far with two GPUs and then say, “Huh, I don’t have the money or access. I’m going to do something else.”

Bryan Casey: So we’re moving to a world of universal basic compute? It sounds like that’s been a little meme-y recently. We’ll call it a day there. Volkmar, Chris, Skyler, thank you all for joining. Great discussion today. For those listening, you can grab Mixture of Experts on Apple Podcasts, Spotify, and podcast platforms everywhere. Until next week, thank you all for joining. We’ll see you next time.