Tim Hwang:  I’m Tim Hwang and welcome to another thrilling episode of Mixture of Experts. Each week, MoE brings together a group of the smartest, sharpest, cleverest thinkers we can find who are working at the frontiers of artificial intelligence to make sense of the week’s news. On this week’s episode, we’ve got Merve Unuvar, Director, Agentic Middleware and Applications Research; Martin Keen, Master Inventor; and Olivia Buzek, Staff AI Engineer. We’ve got a packed episode today. As always, we’re going to talk about the new Anthropic Institute, Shopify getting into agentic shopping, and news that OpenAI might be in the middle of a pivot. But first I want to start today by talking about the big conference of the week, which is NVIDIA GTC. This is the big NVIDIA conference that happens once a year where NVIDIA and Jensen Huang get to come out on stage and announce all the things that are happening with NVIDIA. There’s a little bit of a ridiculous set of headlines coming out of this conference. At the conference, Jensen mentioned that he sees $1 trillion in orders for Blackwell and Vera Rubin systems through 2027. So I think we’re past the billion-dollar mark with NVIDIA — now they start talking with trillions and a huge number of announcements.

Tim Hwang: The really key theme that I want to maybe start with — and Olivia, I’m curious if you have any views on it — is that they seem to have shifted. I think that was one of the most interesting aspects of all this. We’ve come to know them very much as “you do the big GPU build-outs in data centers,” but it seems like even now the software is almost kind of the upstream thing. It’s forcing Jensen to change what he needs to do. So curious about what you think about all that.

Olivia Buzek: Yeah, I do think that’s really interesting. Just so folks know, I’ve been here at NVIDIA all week and hearing how people talk about it. Mostly we brought a bunch of agentic AI stuff to the show floor to help people get a better understanding of it. I think the big thing with agents — as of a few months ago, saying that agentic AI was going to take everything was the buzzword last year, but it sort of didn’t mean anything yet. The models weren’t quite there in terms of being able to automate things in this unsupervised way. Now, there are at least a few systems that you can log into where not only can you ask the main agent to do things, but it can launch sub-agents and you don’t even necessarily — as of six months ago, if I wanted to build an agent, it was a whole thing. It was a complicated layered system. But now at this point I can say, “Hey, spin off 50 agents. I want half of them to do web research, some of them should be testing my code, and by the way, a couple of them should be synthesizing results on this monitoring job that I want you to do — just go do that for me.” The amount of sub-work that you then create — each one of those is an entire LLM that’s being run behind the scenes. So that drives up the compute exponentially relative to the amount that you would use normally. I think that is an interesting development. I don’t know if it’s good or bad yet. Obviously there are a lot of implications depending on how you feel about it. But what that means is that people are starting to do that. Right now I think mostly specialists know about those types of behaviors, but I think it’s completely reasonable for us to imagine a world where people actually are doing those sorts of things on a regular basis — spawning several different agents to do their bidding. If that’s the case, then they need the compute to back it. So specializing the compute, making it so that a lot of that stuff can happen on a single chip or a single stack or whatever, makes a lot more sense than trying to spread all of that all over the world.

Tim Hwang: Definitely. Marve, I know you spend a lot of time thinking about agents and agentic middleware. Curious what you saw in the GTC announcements. Anything intriguing or surprising to you?

Merve Unuvar: I think to Olivia’s point — I was joking right before we started the recording that I have so many tabs open on my browser, it feels like we’re going to get there with agents as well, where it’s like “Where did I put that agent that’s supposed to be doing that for me?” I think it’s closer than you think. What caught my attention is the Nemo Claw announcement. OpenClaw is something that I work very closely with, and I think OpenClaw could have done those things but could have done other things that you wouldn’t want. So for folks who don’t know, Nemo Claw is NVIDIA’s enterprise-grade AI agent platform built on top of OpenClaw. OpenClaw — since February, I think whoever is in the agentic world is probably paying attention because it’s the fastest-growing open-source project in history. But it had serious security problems at a level that I wouldn’t feel comfortable putting it on my local machine and running it on my private personal data. What’s interesting is NVIDIA took this and made it like, “Okay, I put some security constraints and controls,” but the things that make OpenClaw dangerous are at the same time the things that really make it useful. Access to all your apps or the filesystem, persistent memory, integrating with every single app you use — that’s basically what makes OpenClaw interesting and a very capable assistant. Nemo Claw is basically sandboxing and giving permission controls — genuinely helpful especially for enterprise applications, but it’s like a leash. The whole point OpenClaw went viral was its independence. So I don’t know where Nemo Claw is going to take this OpenClaw game. Jensen kept making the Linux analogy — Linux became enterprise-safe through, but OpenClaw is just a few months old. Linux went through years and years of hardening, decades almost. Commercial distributions like Red Hat had mature governance, and Nemo Claw is alpha software. I wanted to try it, and I’m like, “Okay, I have to set up a virtual machine before I try it because it’s alpha.” So it’s interesting to see where this is going to go. NVIDIA moving up the stack from the infra layer to basically getting to the agent control layer is another interesting point. Now this is like an agentic operating system war. Who is going to play there is going to be an interesting thing to watch.

Tim Hwang: Martin, that was actually the question I was going to bring up to you, and Marve set me up really nicely. We think of NVIDIA as a hardware company, a GPU company, and of course they made all the announcements. There’s this new Kyber platform coming out that’s going to be even bigger and more powerful than the last platform they had. But it does seem like the focus of this year’s GTC is very much their movement on the software side. Should we be surprised that they’re moving in that direction?

Martin Keen: Well, NVIDIA just seems to have crazy growth day on day, isn’t it? I feel like every time I come on this podcast, we’re talking about some astronomical number that they have just released. The big business thing was their year-on-year quarterly revenue was up 77% — on a base that was already enormous, to be clear.

Tim Hwang: Yes, because the previous 11 quarters, the revenue growth has been over 50% every single time. These are just crazy numbers. I actually laughed when I saw the trillion-dollar number. I was like, “I don’t even know what numbers are anymore.”

Martin Keen: Right. And then you’ve got Jensen up on stage going through some of the hardware stuff, but then two hours later he’s still up on stage talking about, “Oh, by the way, there’s this Nemo Claw thing. Oh, and remember when we bought Grok for $20 billion in December? Here’s what we’ve got going on with that. Oh, and we’re doing some level 4 autonomous vehicle stuff with a bunch of car companies with the NVIDIA Drive Hyperion program.” The number of things going on is just mind-blowing — just to keep track of it all, let alone actually run this as a business.

Tim Hwang: Olivia, it sounds like you’re at GTC right now. I’m curious what your experience has been like. Is it just super scattershot with so many things going on? Are people able to keep up?

Olivia Buzek: Absolutely. It’s huge right now. Everybody has brought the same things, but a little bit of different things. My big thing — and the thing I haven’t seen talked about in a lot of the announcements — is that NVIDIA has a lot of stuff with robots going on. I don’t know if y’all have seen the kung fu robots. Anyway, anybody who’s listening, go look up the kung fu robots. It’s really crazy. Look up the difference between the 2025 ones and the 2026 ones. I was at PyTorch Con last August when NVIDIA brought an explanation of what they were doing with robots. What they explained is basically the reason that the robots are getting better is because they put an LLM in the brain of the robot. It used to be just visual language models, and then it started being large language models that are making a big difference. It means now that the robot is able to get better by analyzing the scene with words and telling its hand to go a little bit to the left, which is a very, very different situation than we were in before. So not only are there layers and layers of what NVIDIA is doing that we’ve already talked about, there are also all these other layers that they’re pushing into as well. They’re a little bit everywhere. And yeah, very much the conference — you can see that. I did see a robot on the floor the other day. Was not that fancy a robot, but it was a really cool robot.

Tim Hwang: There are two views. One view is, “Are they spreading themselves too thin?” On the other hand, if you have a trillion dollars, you can afford to be spread everywhere. I’m going to move us on to our next topic. This is a story that came out through Anthropic. Anthropic will be announcing and launching something they’re calling the Anthropic Institute. It’s going to be led by co-founder Jack Clark, who’s going to be assuming a role as Anthropic’s head of Public Benefit. The way they frame up what this new institute is going to be for is the idea that AI is going to change society from their point of view, and they are going to increasingly need to tell the world about what Anthropic is learning as AI transforms the world. It looks to be a super interesting collection of researchers from different disciplines. In effect, they’re creating a unit to figure out what their technology is really doing to society. Marve, I’m curious what you think about this. The skeptic might say, “Ah, this is just marketing.” Do you see anything further than this? Why would they even need something like this?

Merve Unuvar: I have mixed feelings. I like the fact that this has co-founder involvement and they bring all these diverse backgrounds — independent, non-Anthropic leaders to lead this effort. But the harder question is: can you build and study simultaneously? Can you be the honest auditor of your own work to see what’s impacting? In history, we had the tobacco industry funding health research or the fossil fuel industry funding climate science. In those cases, research was real — people were funded to do research — but institutional incentives shaped which questions got asked. So this is a great initiative from Anthropic, but there’s still some skepticism. Is this going to ask the right questions? They’re going to influence the questions being asked. As you said, Anthropic is essentially saying, “We don’t really fully know what we’re building and how it’s going to impact.” That’s the honest answer. I think no one — not regulators, not economists — knows an AI shock hitting many labor markets at the same time, legal systems, democratic institutions, countries simultaneously. So it’s an attempt to build that understanding in real time from the inside — what’s happening. This is reassuring to me; otherwise the terrifying aspect would be if they pretended that they knew what the implications would be and there was no research behind it. So I lean toward wanting to see what this research institute is going to do and how they’re going to study. But it’s pretty new, and there have been efforts in the past from other companies too. We’ll see.

Tim Hwang: Martin, it’s a really hard problem. I had some of the same feelings as Marve — it’s good that we’re trying to understand what’s happening, but can you do that research and execute at the same time? From the blog post, Anthropic would tell you, “We’re really well positioned to do this because we see everything. We build the systems, we have all the usage data that allows us to say things about what’s happening in AI that no one else might be well positioned to talk about.” Do you buy that argument?

Martin Keen: The more cynical among us might roll their eyes when they hear that an AI lab is going to create an institute to make sure the AI lab is being safe. There’s a bit of a vested interest there. But the interesting thing from their announcement was that they said the institute has a unique vantage point — access to information that only builders of frontier AI systems possess — and it intends to take full advantage of that. So they obviously know more about their models than we do. Anthropic has been a leader in interpretability for years. They’ve been very open about trying to understand what’s going on in their models. My favorite example is an earlier version of Sonnet where they just tweaked a few weights and created what they called Golden Gate Claude, where this thing was obsessed with the Golden Gate Bridge just by tweaking some weights. The result was a large language model that no matter what you asked — if you asked it for a recipe for cookies — it would somehow work its way back to the Golden Gate Bridge. They set up this institute with three pillars: the red-teaming pillar (stress testing, which they’re probably already doing a lot of), a pillar looking at societal impacts, and economic research specifically tracking jobs. The CEO has been very open about saying where they think this is going and the impact it’s going to have on jobs in the coming years. As eye-rolly as I was when I initially saw this announcement, because it’s Anthropic and they have this history of being quite open about this, this could be something to keep an eye on.

Olivia Buzek: There’s a role for research institutes controlled by large AI labs. I think there’s a responsibility we all have as AI engineers to make sure we are having those conversations and thinking about those things. I also think those research labs need to collaborate with more independent ones in order for us to have a complete conversation in this field around what responsibility looks like. It remains to be seen which way they actually go, but I’m really hopeful that they decide to be good players in the research sphere and take into account what’s coming from a lot of the independent places as well.

Tim Hwang: One thing I did want to touch on before we move to the next topic: is it kind of funny that they’re only doing this now? There’s a part of me that’s like, you guys are building this, but you have no idea where it’s taking us. Maybe this announcement is coming a little bit later than we would have wanted, because understanding these effects would be something you would just hope the company would be working on. It seems like even the leading people in the field are like, “I don’t know where this is all taking us.” Olivia, maybe since you had the last comment, should we be concerned that no one seems to know where we’re going?

Olivia Buzek: You get on the bus for a cross-country trip and you say, “Hey, is this our destination?” And they go, “I don’t know, I have no idea. See how things go — we’re just driving.” So it is and it isn’t concerning. I appreciate people’s reluctance to state a strong truth. I was just reading Dario Amodei’s essay from back in October 2024 about how he thinks about the future. He talks a little bit about how AGI in particular is not necessarily a useful term, but what happens with the emergence of powerful AI and things like that. I appreciated his restraint in saying, “This is exactly what will happen,” because there’s another vein of thought that’s very common in the AI space: the singularity is coming, these things are definitely going to replace you, this is how the economy will be, we will have no need for the legal profession anymore because it’s all going to be done by AI. It reminds me of back when I was in grad school, there was a joke about how computer scientists get really clever in one discipline and really like to think that they are also really good at other disciplines, submitting to physics journals saying, “I applied a computer to this and I’ve made this grand new physics result” — and it’s not the case that that’s never true, but there’s a hubris to the field. I want to see balance in that. There’s a tendency towards hubris, so I do want people to restrain their “I’m going to tell you how society works because I built this cool language model thing.” But at the same time, because we have this role of guiding what that future looks like, telling people where the bus is going would really help.

Tim Hwang: At least where we’re going — GTC seems like NVIDIA is going everywhere. One point of view is that they have some N-dimensional chess master plan they’re pursuing. It may just be that even they don’t really know where this is all going.

Martin Keen: I think it’s worse even than not knowing the roadmap. The frontier AI models today do not — funnily — understand how their AI models work. You cannot trace from an input token through to the output token and know how that model arrived there. So when you don’t even know how the stuff works today, forecasting what’s going to work in a year’s time is a bit difficult.

Tim Hwang: It’s a little bit of a weird feeling, but to the points that have been made so far, it is exciting to see them not just using computer scientists. I guess they could have done a version of the institute which is like, “We took a bunch of engineers and they’re going to tell us about this,” but they seem to have an openness to all these other disciplines that might have relevance to navigating this — or at least letting us know which bus we’re on and then maybe get to the question of where the bus is going.

I’m going to move us on to our next topic. This was a sort of fun announcement. We’ve talked a lot in past episodes of Mixture of Experts about agents and shopping agents in particular — a little bit about what Amazon has in the space. This headline caught our eye: Shopify, the distributed e-commerce application and platform, announced through its president Harley Finkelstein that they’re about to make a big push into shopping agents. The vision is that the agent will go out there and do your discovery, buying, and comparing of products for you. Marve, I’ll kick it over to you. The first thing I immediately thought is that Shopify is a very different kind of business from a platform like Amazon, and that might mean that the agents look really, really different. Does that hypothesis ring true to you, or do you feel like shopping agents will be sort of uniform across the whole e-commerce world?

Merve Unuvar: I believe it’s just like how when GPT was launched, it changed our expectations of search. Agentic shopping will genuinely change the consumer experience, and it’s going to somewhat converge because the consumer — the person who wants to buy something — is the common denominator here. But this feels to me like an extension of existing gen AI use cases, not something fundamentally new. The core pitch of Shopify was genuinely compelling, but it’s really just a natural next step of personalization that’s been progressing for 20 years — Amazon’s recommendation engine or Google Shopping. What is going to be interesting for Shopify is, as you said, they’re distributed, so they don’t really own a single platform or a single merchant. This is going to democratize things and make it a little bit more interesting if you enable agent-to-agent commerce, and hopefully normalize not promoting sponsored products or Amazon’s Prime. But what I found interesting the more I thought about it is: if agents are going to act on behalf of users to do purchasing decisions, what’s going to happen to a brand? The role of a brand is storytelling, emotional purchases. Brands have always been partly about bypassing rational decision-making and creating desires. Agents will basically optimize on whatever you want to optimize — cost, performance, past history — so you will not be discovering things you didn’t know you wanted to buy. That’s going to be interesting. The consumer experience will be different. But other than that, this is a natural extension of the personalization use case that is now going to be powered by agents.

Tim Hwang: There’s a lot there to unpack. Let’s pick up on ads. That’s one of the most intriguing things you’re putting out. A little bit like how Model Context Protocol (MCP) is a system where you access a website, but there’s also a version of the website that’s for the agent. I wonder in the future whether you’ll have advertising for humans and then advertising to agents will become its own market. Because now I don’t have to persuade the consumer; I just need to persuade the agent that’s shopping on behalf of the consumer. Olivia, I see you nodding — that feels very sci-fi to me, but it feels like a very obvious next step of where this is all going.

Olivia Buzek: Oh, it’s a very obvious next step. The way the Shopify CEO describes it, he’s basically like, “Oh yeah, this is going to be an authentic shopping experience. It’s going to be so great.” So cynical. Maybe it’s like that today. I have used some of the consumer-grade chat models to do product comparison. The reason that works right now is because there’s a trove of Reddit data and forum data that is accurate for the products today. But the problem with all of these — and the problem we collectively have to solve as a field — is the hollowing out of those knowledge resources. If we are not careful, they will be replaced by ads, exclusively ad-based content. Then your agent can’t possibly give you good recommendations when the only content out there is marketing-based. It degrades the quality of the agent simply because the information isn’t there. To a degree, that has happened with the shopping experience already, even pre-generative AI, but not to the degree that you will see. We can see an example of this happening already with Stack Overflow data. Stack Overflow, for anyone who’s listening who isn’t a computer engineer, is where we all share information about how to code. Stack Overflow was the place to be for most of the 2010s if you were a coder. It was pretty steady in popularity — never all of the developer population, but a good 5-10% of developers were on there, helping glue the community together, making sure we had the information we needed so we could all do our jobs. Nobody’s on it anymore. I’m sure there are a couple of diehard enthusiasts, but if you look at the usage graph, it took off, went to 10x, went along for several years, and then just dropped right back to where it was before the jump. That means that as coders we no longer have that source. Let’s say I’m using an open-source library and I want to know if it’s getting better and what the latest versions are. Increasingly I can only rely on the documentation and what gets posted on issues by agents, rather than things coming from other developers who are dealing with the same problems and have solved those problems. Translate that over to our agentic shopping world: if you hollow out the pieces in the first place — news organizations not being able to post reviews because they’re no longer able to monetize the content because the ones reading it are exclusively agents — the quality of the information that’s available degrades. So we have to solve that problem. This is why I think a lot about the fact that as AI engineers, we actually have a role to play in that. We could build systems — I haven’t seen anyone yet say, “Hey, my next startup is going to be the treasure trove of information. I’m going to make the agents share information publicly with each other. I’m going to keep the old internet alive, the concept of the web alive, so that things can keep working in the future.” We have to make that investment. I certainly don’t see any of that coming from any of the frontier AI labs. People have all these other risks about very serious concepts, and that’s fine, but there’s this real basic concept that I need to see someone address.

Tim Hwang: Marve, that’s a doomsday scenario. Do you buy the doomsday scenario?

Martin Keen: Yeah, that totally makes sense. I’m really interested to see what this Shopify storefront experience would look like when it’s optimized for an AI agent. If you think about what a storefront looks like today, it’s optimized for different things: search engine optimization so that your storefront shows up on Google, and user experience so that you’ve got human eyeballs on the storefront. So you need things like an attractive banner image. Does that matter to an agentic shopping agent? Or it’s going to use some sort of psychological nudges — if you’re selling a subscription, you might have three tiers and want to nudge people to take at least the middle tier by making the lower tier look a little bit insufficient. Are those psychological nudges going to work with an agentic system? A storefront that works well with an agentic system today would be optimized for things like computer use. Agents are pretty clumsy when navigating. If I give Claude access to my Chrome browser and it can load websites and scroll around, there are certain UI elements that trip it up — information that is collapsed and needs to be expanded, stuff like that. So does a Shopify storefront that is optimized for an agent get rid of the banner image, get rid of all the psychological nudges, get rid of the UI elements that are tripping it up so that it’s more easily usable? Then you want to do it the old-fashioned way and you actually go to the website and you’re like, “This is terrible, I can’t use this at all” because it’s been fully optimized for agents. That’ll be a very strange feeling.

Tim Hwang: Marve, a last question here. Do you think that Shopify or even a company like Amazon has a fighting chance here? One obvious thing to say to this kind of announcement is, “Well, we’re already going to have shopping agents. They’re called ChatGPT or they’re called Claude.” So this Shopify agent competes for user attention and adoption against our general-purpose agents that are increasingly going to be working with us. Do you think Shopify has leverage to carve out a path, or is it really hard to see them competing against the big tools?

Merve Unuvar: As I mentioned at the beginning, I don’t think this is going to be a big fundamental thing. They are the second-largest e-commerce platform, so they have that, but they don’t have, for example, Amazon’s own data over the years of users coming because it’s proprietary and from their platform. If anything, I see Amazon or Google — Google has a lot of ad and sponsored data — that may put them on edge. Shopify is distributed, but distributed to merchants. I don’t know if they can leverage that to their advantage. Otherwise, as you said, I think this is going to be yet another agentic experience for the consumers. We’ve all gotten used to looking at our flights or the weather from either Claude or ChatGPT. That has changed, and that’s still going to be the entry point for consumers. If they can integrate their shopping agents well with these entry points and leverage the data they have, maybe they may differentiate. But I don’t think this is going to be a very big, fundamental shift for the shopping experience or the industry.

Tim Hwang: We’re going to move to our last story of the day. This show has ended up being the “are they spreading themselves too thin?” episode. We’ve hit on that theme over and over again, and this final story is no exception. The Wall Street Journal exclusively reported this week that OpenAI is readying a big strategic shift. What is rumored to be on its way is the recognition among OpenAI’s leadership that the “do everything all at once” strategy is not working for them. What they’re going to do — we’ll see how this plays out — is to refocus the company specifically around coding and business users. This is a pretty big shift. As someone who loved their short-form video products like Sora and a whole host of other things they’ve done over the last few years, this is a pretty big shift. This is almost a recognition that enterprise is going to be the future of OpenAI. Olivia, thoughts? Discuss.

Olivia Buzek: This may have changed this year, but in the last couple of years, every summer when ChatGPT usage gets posted, it dips — meaning all the students go away and they stop using ChatGPT for things. I’ve read that there’s a lot of market penetration of things like ChatGPT, Claude, et cetera, but most people are doing a query every couple of days or something like that. It has never happened that AI advancements suddenly manage to penetrate all corners of society immediately. In the history of technology, that has never happened. So I don’t think it’s viable for most companies to aim for the consumer market as their first market for an AI product. Yes, ChatGPT has been a big darn deal. Everybody’s paying attention to it. All of this matters. But there are diminishing returns until and unless everybody can really understand how to use these tools well, because they are fundamentally tools that you need skills in order to make good use of. Specializing in the places where models do well, where there are populations of people who are predisposed to be able to think about these complex problems — it’s going to work a lot better, and where there’s an incentive to use them. Coding and enterprise — there are obvious incentives. There’s a reason that everybody is talking about this for code and less so for other things. As a consumer user of these products, I am a little bummed to hear that they are no longer servicing me.

Tim Hwang: We’re also talking about when the VC subsidization of the product is going to end. One of the reasons I love using it as a consumer is that even though I have a pro subscription and pay a lot of money, it is still an incredible deal relative to the amount of money that OpenAI needs to spend to service my user account. Martin, do you think the music is stopping on the consumer experience of this tech? Will they say, “Hey, if you really need this, you’re going to have to pay us hundreds of thousands of dollars to get access to it”?

Martin Keen: How do most of us know OpenAI? From the ChatGPT launch. I was using their models prior to ChatGPT and subscribed to their Discord server. Every time they had a new announcement — “We’ve updated the training data” or “We’ve added a few more billion parameters” — they would send a general announcement to everybody. One afternoon they sent this general announcement that, “Oh, we’ve added this chat interface to our model.” It was such a “meh” moment. To think that one year later that nothing announcement would mean that my mother not only knew what ChatGPT was but was using it daily to help with her recipes — it never felt like that would be the moment. But this became their big thing, and it started them on this journey of being very much consumer-focused. The generation that first started to bring out — that was a huge deal. Then, as you mentioned, Sora and Sora 2. When Sora 2 came out, it racked up a million downloads in the first five days, which was more than ChatGPT. But it was so transient. I don’t know anyone who’s doing Sora stuff anymore and watching these AI-generated videos. It was all you could see for a little while. Same with image generation — constant image generation on social media, and now it’s nothing. With the consumer stuff, it moves so fast. It’s super transient. You get a big bump and then it’s onto the next thing. Whereas with corporate business contracts, agentic coding — that’s where Claude has done so well. That’s a market where once you tie people in and they’re hooked, they’re going to keep coming back with these big contracts. It’s not going to be the $100 plan. If you’re doing this at a corporate scale, it’s going to be a lot of money. It is a shame, though. I love things like the advanced voice mode and how fun that was. It will be kind of a shame if we’re just focusing on models that are improving their benchmark scores on coding benchmarks from now on, but I sort of get it.

Tim Hwang: Marve, the cheeky response to this headline is: they’re going to focus on coding and enterprise — is this just the Anthropic strategy? This is kind of funny — another player has already been focusing on this strategy: Anthropic. Is this almost OpenAI conceding the battle to say, “Actually, where the money is going to be made is what Anthropic is doing, and we’ve got to get in there”? Is that how you read the news?

Merve Unuvar: Yes. If they’re making this decision, I think they’re making the right strategic call. But my question is: did they wait too long to switch to code over consumer? In enterprise AI, even within my teams, the coding agent isn’t just one product category. It’s the entire foundation for the relationship that you have with the enterprise. If you lose the developer, you lose the platform. If you lose the platform, you lose every single upsell opportunity to any other AI application or workflow the enterprise will eventually buy from you. OpenAI built ChatGPT with a consumer-first approach and assumed enterprises would follow with all the nice features they kept introducing. But Anthropic went the other direction. As Martin was saying, developer preferences are much stickier than consumer ones. Once an engineering organization has built internal workflows around a coding agent and trained their team to use it, changing that is going to be a big, big sales cycle from OpenAI’s perspective. So Anthropic, as you said, read the market better than OpenAI and started from the right place. I think OpenAI is following again. Do you guys think OpenAI has a potential to win? Did they wait too long? I think Codex is great, but the challenge is the developer community has already formed their loyalties around these coding platforms. I don’t know what you guys think.

Olivia Buzek: I simultaneously think that’s true — I do see people starting to develop their loyalties to a particular platform — and I also think that none of these coding platforms have much of a moat yet. At best, they have a set of skills that they do particularly well. For example, Claude Code has superpowers around certain types of things, and you can now mostly trust Claude Code to do a good job at reviewing things. But that’s going to be table stakes within months. It won’t take long to replicate those features. Nothing that any of them have put out so far is completely non-replicable by the other labs. There’s no stickiness, no network effects. If you’re using whatever coding agent on your team and the next team over — or even the next person over — is using a different one, it’s not going to be a big difference. You can pick up one one day and the next day another. If your company says, “Oh, we’re done with X and you all have to use Y,” everybody’s just going to shrug. I sort of think of it like video platforms. At the beginning of the pandemic, there were a whole bunch of people who tried to put out video platforms and say, “This is going to be the video platform.” Zoom was suddenly tip of everybody’s tongue. Then that market lead evaporated a little bit. I mean, Zoom is big, lots of people still use it, it may even be the market leader — I’m not up on that market — but I do know that there are a whole bunch of other companies that have been able to say, “We can get video platforms from anybody.” It’s more of a commodity than that. I think the struggle with coding agents is that they are in that category. Right now there’s no real incentive to use one versus the other. If they were able to actually introduce those network effects and make it so that if you’re using X platform and your team member is using X platform, you are able to work together better, then this would be a different story. But I haven’t actually seen them solve that problem yet.

Tim Hwang: Martin, do you want to have the final word here? The only other thing I’ll throw in is that it’s possible that OpenAI may be late but still good. We sometimes forget it’s still the much bigger company relative to Anthropic. Can they marshal their resources? That’s a big question. But curious about what you think.

Martin Keen: From the actual practitioner perspective, the cost of switching — as Olivia says — is so low. You build something with Claude Code and it’s making MCP calls and using some skills you’ve built (just markdown files). All of this stuff transfers over to Codex, and you can use OpenAI’s model to do the same thing. In fact, I’ve had several situations where I was building an application with Claude Code, ran out of tokens, and I still had some tokens left with OpenAI and Codex. I just pointed the Codex agent at the GitHub repo, it read the code, picked up where I left off, continued, and finished the work. These things are completely interchangeable, kind of by design at the minute. So I would say it’s absolutely not too late.

Tim Hwang: Well, that’s another one where we’re going to have to keep an eye on it and see what happens, as is with all these stories on MoE. Marve, Martin, Olivia, thank you for joining us. This is a great panel. I hope you’re all back on soon. That’s all the time that we have for today. Thanks for joining, listeners. If you enjoyed what you heard, you can find us on Apple Podcasts, Spotify, and podcast platforms everywhere. And we’ll see you next week on Mixture of Experts.