Tim Hwang: Exciting news, listeners: we’re recording an episode of Mixture of Experts on site at Think 2026 in Boston this year that takes place between May 4th and May 7th, and registration is now open. We’re going to put the link in the episode description, or your IBM representative can help you with the next steps. I’m Tim Hwang and welcome to the 100th episode of Mixture of Experts. Each week, MoE brings together a lovely group of thinkers and commentators working at the frontiers of artificial intelligence to make sense of the week’s news. On this week’s episode, we have Ritika Gunnar, General Manager for Data; Volkmar Uhlig, CTO and VP of Data Platforms; and Kaoutar El Maghraoui, Principal Research Scientist. Thanks to you all for joining. We’ve got a lot to talk about today. We’re going to talk a little bit about scientists using Claude Code. We’re going to talk about the future of AI and finance. But first I want to start with this really fun story about using ChatGPT to sell a house. This first appeared in the New York Post, and it was a story about a guy who was looking to sell his house and got frustrated, as we all do if you’ve been here before, with realtors and the whole world of real estate law and transactions, and decided to try to do the whole thing by himself via ChatGPT. The results, at least as they claim, are pretty good. He was able to sell his four-bedroom, three-bathroom house for around 100,000 over what the local real estate agents had been giving him as an estimate. And ChatGPT really helped him through the process — using it for listing strategies, improvements to increase the value of the house. Ritika, you’re our special guest for today. Curious about how you read this story. Are realtors doomed?

Ritika Gunnar: Well, I think it’s a great story. This isn’t a story about selling a house or realtors themselves. I think it’s really about how AI is becoming essential in almost everything that we do, and those that do work with AI are going to see those gains and those right kind of outcomes. This is about how AI has moved from being a general purpose operator to handling complex workflows, and how we’re decreasing the skills that are necessary to be able to work with AI to achieve those outcomes. That’s exactly what this guy did. He used it for every part of the workflow, and he used that to reduce the barrier. He’s not a real estate expert himself, yet he was able to show those kind of outcomes. It really shows the power of those that can use AI and the kind of outcomes they can achieve, versus the other pool of real estate agents that were giving him a lower dollar value — $100,000 less. So I think this is what we really have to see.

Tim Hwang: Kaoutar, it was recalled to me by Pedro, one of our producers, that our very first MoE episode actually covered ChatGPT when it was released. That was about 100 episodes ago. The technology really has come a very, very long way in a very short period of time. I’m curious if you want to reflect on how quickly things have been moving, because this is definitely something that GPT-2 would not be able to help out with definitively.

Kaoutar El Maghraoui: Definitely. This is showcasing this compression of expertise. These models have come a long way — the low latency, interactiveness, human in the loop, all these interactions with these models, requesting, prompting, and so on. These interactions are short, context-iterative, latency-sensitive. Since the beginning, these models have perfected a lot of these things. There are still challenges, but this is also showing how far we’ve come along in terms of AI just being a point of view to really doing real things that are useful for people. Ritika pointed that out very rightfully. What we’re seeing here is not just about the real estate market, but a compression of expertise. These LLMs are effectively acting as a distilled representation of that collective expertise. Having seen thousands or millions of similar scenarios, when you prompt them correctly, they can simulate what a top-tier agent or expert might do. This is not just automation. It’s also about changing who gets access to expertise. It’s democratizing that in a way. But it’s not just automating decisions; it’s also kind of standardizing them, and that might reshape how markets behave. So I think it has some really strong implications. This is also raising a subtle question: are we increasing intelligence in the system, or are we just synchronizing behavior? Are we homogenizing some of these markets? Because if we’re all using the same AI agents, the markets might become more efficient, but they might also become more homogenous or algorithmically shaped.

Tim Hwang: That would be a very funny outcome — every house starts to look even more alike in the listings because everybody’s following what ChatGPT is telling them. Volkmar, I’m going to prey on your biases a little bit. As a lawyer, I am very proud to say that they note in the article that they used ChatGPT for everything except for reviewing the contract. For that, they actually used a human lawyer. Do you still think there are limits to how far this technology will go in terms of automating expertise? Is our lawyer going to have a job, or is it just a matter of time before they also become subject to the same forces?

Volkmar Uhlig: I would guess that the lawyer is going to be replaced as well. I’m leaning here on Elon’s comment about AI. You cannot have — like if you imagine Excel and you had everything in Excel automated except for a few functions, then the utility of Excel goes down dramatically. His argument is that only when you’re closing the loop 100% can you actually get the economic gains. So I’m on that side. The lawyer will go as well.

Ritika Gunnar: Can I be on the other side? I’m going to lawyer that. My perspective is a little bit different. I think there’s a difference between lawyers that will use AI and lawyers that will not use AI. The human-in-the-loop aspects of taking the expertise that a lawyer has — decades of experience across different domains and functions — you want your lawyer to be able to use AI to automate and catch as much as possible. That lawyer ends up becoming a lot more efficient at their job and able to do tasks that are more about interacting with the human — getting input, really working with the human, understanding even some of the emotional aspects of things. In this particular real estate case, there may not be that, but when you’re talking about things that are more ephemeral in nature — like marital or parental — there’s always an aspect where that human part and that human in the loop should not be replaced. So you should definitely use the AI, and you have to use the AI to be able to have differentiation in how you accelerate what it means, where you need to focus, and still have that human loop. So I think there still will be lawyers, but the bar for those lawyers is now a lot higher.

Volkmar Uhlig: It is much higher. I agree with the cases you chose — anything which has an emotional component or complications that are not high-volume transactional. Anything which is high-volume transactional will be automated. If I look at the real estate market — I didn’t even have a lawyer for my real estate transaction. Anything that is high volume has a high chance of automation, and then we will get the exception cases flagged. What will also happen is because the lawyers can wave 98% through, the transaction costs will become almost zero. So I totally agree that we will have no lawyers? No, I don’t think that.

Kaoutar El Maghraoui: I also agree. Lawyers won’t be replaced by AI — all of them. But the legal profession will be transformed more than almost any other white-collar field. Legal search, document review, drafting — all of that will be completely automated. Entry-level roles will be displaced. The bar will be much higher. But what AI cannot replace is judgment and strategy, legal interpretation in ambiguous situations, courtroom advocacy, client trust — those still need a lot of human accountability.

Ritika Gunnar: My perspective on entry-level workers being displaced is a little bit different too. I do believe you’re going to need entry-level workers, because entry-level workers need a place to escape to be able to have more mature senior-level workers. You need to be able to start out and have entry-level workers, and the way in which those entry-level workers can progress to being more senior or knowledgeable probably accelerates as well. This notion of only having senior professionals that know a certain job task is incorrect when you take the long-term view, because you’re always going to have to refill the bucket of who are the entry-level people today that are going to be your experts tomorrow. So while the way in which you drive the workforce into that funnel will change, I still think you need entry-level workers — because who are the experts of tomorrow?

Tim Hwang: Ritika, a final question for you about where some of this goes in terms of what it means for the professions. If you’re a realtor, one of the reasons ChatGPT can do this is that there’s lots of data with which these models can learn about how to sell a house well and how to position a house well. I wonder whether there are going to be incentives for professions over time to become almost more guild-like — to keep more things secret and away from these models, because the minute the models get access to them, they can do the things that these people are offering. Do you think that’s going to be an increase in secrecy because of the pressure of these things?

Ritika Gunnar: Two comments. Number one: if you look at a lot of these LLMs, they are a great representation of all the public data out there in the world that is known. This gentleman took all the public data that was out there and used that to become a better real estate professional to help him in his personal endeavors. So what do organizations have to claim as their value? Their private data. In the case of companies, their enterprise data; in the case of personal, their private data becomes the way that differentiates them. This is why I fundamentally believe that the value in applications — we used to say “AI is only as good as your data” when AI was really about building machine learning models and your data was the fuel to drive those models. Now, I believe your enterprise data or your personal and private data is the application, because your LLM is a great representation of the public data. Your private data is your value, whether you be an enterprise or a personal person. That’s number one. Number two: I think many of us have children in college or early in their educational or professional career. We used to really talk about reading, writing, and arithmetic — the things we wanted to focus on to drive our kids. I used to make waves even a decade ago saying, “I think the fourth thing we want is coding.” Well, now we know that a lot of the AI tools can do the coding, but there is a foundation of understanding how to work with gen AI tools that is going to be essential for every profession. That’s one of the things this teaches us: the skills barriers have lowered, but they’ve lowered only if you know how to work with these tools and how to drive value from them. Should we be better at really enabling everybody to work with these tools? Because that’s going to be the difference in the differentiation of a lot of work that happens in the future.

Kaoutar El Maghraoui: It’s not really the entry-level jobs that are disappearing, but the tasks — the entry-level tasks that we know today are changing. We will expect these entry-level professionals to know how to use these AI tools. That is important as part of their upskilling.

Tim Hwang: Great. I’m going to move us on to our next topic. This was a fun blog post on a Substack I follow called Republic of Science by a guy named Charles Yang. This got a whole lot of play online. It’s an interesting analysis — certainly the first version of this analysis I’ve seen before. Charles, I think, is interested in the adoption of AI agents in the sciences. He does a rough-and-ready analysis, probably using Claude Code himself, that pulls what are known as ORCID profiles — identifiers for academic researchers — and then sees whether those profiles have GitHub accounts where they’re using Claude. He comes out with this analysis, and I’ll quote: out of 14 million total ORCID profiles, we identify about 15,000 profiles that fit the definition of using Claude Code. The rough estimate he ends up giving is that about 2.1% of scientists with these linked GitHub profiles are using agents. One of the most interesting factoids from the analysis is that you would assume the distribution of adoption might lean younger — people more familiar with the tech. But he finds this interesting bell curve where more established academics and younger academics are using these agents. Volkmar, I’ll start with you. Do you buy this analysis? Is it too rough-and-ready? Is it a good indicator for how adoption is happening, or are there a lot of biases?

Volkmar Uhlig: As a scientist, I’m not sure. If I take the sub-cohort of public GitHub repositories of my total cohort, that is biased towards people who are more on the coding side and therefore have a higher adoption. If I step back and ignore the numbers a little bit, I think what it shows is that there is wide adoption. I would have expected the percentage to be much higher — like 30 to 40% — because the productivity is just so much higher. If you get a 10x to 100x productivity gain, why would you not take that? So the real question is whether there is a skills bias towards using AI. Very young people use it aggressively, more established people use it, and in the middle they still try to weasel themselves out of AI. My expectation would have been a much more even distribution, so that was the interesting fact. Why is the adoption so low? Then you need to ask: how did you find out that it’s Claude Code? Do they use other tools? Do they copy-paste? There are lots of questions. The second question we should ask is: if research uses AI to that extent, what percentage of research is actually true innovation versus just repetition? Because the model needs to know the answer. Are we accelerating or decelerating science? If we decelerate because we’re moving the bulk to the middle — everybody knows what’s in the public domain — or do we suddenly have ten times the output because it’s so much faster to make the science work? Maybe you get an acceleration in the sciences. I don’t know.

Tim Hwang: That’s right. This is in some ways a variant of what you talked about in the context of real estate on this last point: if everybody starts adopting these technologies, is there a lumping towards the middle? Everybody’s doing the same thing, so all this AI adoption may actually not have that big of an effect on research progress. Is that what you’re trying to say?

Volkmar Uhlig: It’s a new middle, right? Everybody kind of converges on the middle, but the middle moves dramatically because the quality of the middle is actually much, much higher. So it may be a jump. Everybody’s doing the same thing, but it’s all better.

Kaoutar El Maghraoui: Correct. I also found the results lower than what I expected, but there are some other aspects we need to think a little bit more deeply about. For science, if you look at software engineering, if an agent writes code and it compiles and runs, we often accept that as good enough. But in science, that standard sometimes breaks down. It’s not enough for something to run. You need to know why it works, whether it’s reproducible, whether someone else can independently verify it. Today’s agents are fundamentally not designed for that. They’re probabilistic by nature, non-deterministic systems. They don’t produce a clean execution trace sometimes, and they don’t give you a guarantee that if you rerun the same workflow tomorrow, you’ll get the same result. That is a big gap we have today, especially for scientists to adopt it much more broadly. There is maybe a mismatch between deploying productivity-oriented AI systems into truth-seeking workflows. So we need to look at how we make these systems more deterministic, calibrated, traceable for science. But I also think using AI as a trusted assistant is going to be very important to accelerate science. They have to work hand in hand with scientists, and it’s a big productivity boost.

Tim Hwang: Ritika, I want to use the opportunity to talk about the data story that’s maybe implicitly hiding here. We all have beef with how Charles did the analysis, but the basic story is interesting: is the data that you would need to accelerate science using AI even there? That seems to be one of the big questions. I could assume that certain fields where the data is ready to go may see these accelerations, whereas ones that don’t are going to be a little bit slower. So AI is going to have this very lumpy effect on scientific progress. Do you buy that?

Ritika Gunnar: I definitely think so. Most researchers working on their projects have some sort of data sets that are very personal to them. The ability to have those data sets in the right format — whether structured databases of some sort, unstructured documents, images, or videos — and to have the right context being used by these models becomes really essential to get the right outcomes. For that to be useful, you need to make sure — like in this study, some of the things it assumes is that checking in code to a GitHub repo is analogous to a lot of the work done by these researchers. But a lot of research work done today uses so many other tools, especially data-driven ones — MATLAB, SPSS, Excel, whatever databases they have. That’s where a lot of their work resides. Driving AI where their data sets reside becomes really important. That may not be where a coding agent like Claude Code or any other code tool is really going to excel. That’s a different kind of AI usage on your data sets, or providing the right context to that model. We completely miss that when we take a look at this.

Tim Hwang: I love that. It’s unrepresentative. It just shows you which fields overly adopt GitHub. He draws a conclusion at the end that economists use Claude Code more, and it may just be that economists have work that takes place more through the context of GitHub. So through this lens we can see more. Most scientists working with sensitive proprietary data — patient data that might be HIPAA data, genomics data — it’s not easily accessible. So we underrepresent how critical some of that is and how we want to deal with that as we work with AI.

Kaoutar El Maghraoui: The data set he used only includes scientists with GitHub and scientists who link GitHub with ORCID. That’s a very tiny and highly technical subset — a very narrow slice of computationally active researchers. It also excludes wet lab scientists, clinicians, field researchers, social scientists — a huge number that’s not represented. If I recall, the size also only lists 331 users. I don’t think that’s enough to generalize across global science. There is definitely some measurement bias here.

Tim Hwang: Absolutely. That’s a great point. This debate has played out a little bit in healthcare and AI, where a lot of people have said it’s going to take so long for AI to shape healthcare. One counterargument is that a lot of patients are using ChatGPT already to do medical-style things. So there’s this vast shadow AI economy that exists in a lot of these spaces that is just hard to measure for one reason or another. That’s pretty interesting.

I’m going to move on to our third topic. This was a fun executive profile that popped up in Fortune magazine about Adobe’s CFO, Dan Durn. Overall, it’s a glossy feature about all the cool work he’s doing at Adobe, but it’s a nice hook for talking about how teams within enterprises are adopting AI and what that looks like. The substance of the piece is that his vision for his finance team at Adobe was to do a lot of the AI stuff in-house. The article touts his team as sort of an AI lab unto itself. They’re using AI not for anything particularly shocking, but it’s apparently saving them a lot of time — using autonomous software agents to forecast results, scan contracts, and manage inboxes. The question I want to ask — I’ll kick it to you first — is: who’s going to own AI innovation in a company? One model is that the IT department or the AI department handles this for the company. But some of these technologies are so easy to use that everybody can kind of be building their own tools all the time. Do you have a prediction for how this will eventually evolve? Will there be an AI team inside these companies, or will it look a lot more distributed with everybody launching their own local projects — which is crazy from an IT standpoint but might also drive a lot of progress?

Ritika Gunnar: We actually did a CEO and CIO study with IBM called the IBM Institute for Business Value Study, where we queried hundreds of CEOs to understand how they’re driving their AI projects, including our own within IBM. There are different representations of how everyone’s doing it. What I see as being successful projects starts with the technology aspect, but there are two other pieces that determine whether projects are going to be successful. The second is the processes. You have workflows or processes in your organization for how you build these apps or institute these apps. As you’re inserting something like AI, which can actually take action and do a lot of the automation itself, the process itself needs to change. You need to have a top-down approach — and you see that in the Adobe article, and even at IBM with Jim Cavanaugh and his office, doing the same thing, top-down — and also a bottom-up approach where we allow experimentation. In the middle, you’ll see a lot of AI centers of excellence that help bring together the technology and the processes to do these kinds of things. The third — and most fundamental — thing between projects that succeed and those that don’t is the cultural transformation. Now you have finance experts who’ve never really dealt with these tools, working with them and understanding how their work is changing. Projects that used to take years are now taking weeks and months. What happens? How do you reallocate resources back to new projects to start innovating? So when I think about this, it’s the process and the cultural aspects that become really important. I think every function is going to transform. It’s how they set up their center of excellence to really enable people, and how you have a top-down and a bottom-up culture that becomes really important.

Tim Hwang: You’re almost saying “yes, and” — it has to happen in all directions to actually work.

Kaoutar El Maghraoui: I see an inevitable shift to distributed AI. We still need an AI platform team — a central team that owns the infrastructure, the models, the APIs, the data pipelines, the guardrails, the tooling, the GPUs, the hardware infrastructure. But we’re moving towards distributed AI builders. Every team will become an AI team because AI will be embedded everywhere — in sales, legal, HR, finance, product. So every team will be partially an AI team. That’s a deeper shift from having just a central AI team to everyone doing it, but centralization will still be there for security, platform access, hardware, etc.

Tim Hwang: It’s a really interesting one about the history of how functions evolve within companies. A lot of companies started just installing PCs, and after a while they said, “We really need a team to centralize and manage all this stuff.” But maybe initially it was as chaotic as the AI transition has been. Volkmar, do you want to comment? Ritika has this really interesting point about how much getting the culture of this works. You read the article and you’re like, “What are you using AI for?” It’s extracting information from PDFs, managing inboxes, doing contract review. This was not the hyper-futuristic, energetic future I was sold, but it may just be that these very boring use cases are the major cultural lever because people are like, “Oh, it’s really helping me out with something very practical and boring.” For me, it rewrites what it means to reshape culture. It’s not about inspiring people. It’s about these very day-to-day things.

Volkmar Uhlig: If you look at an organization of that size — Adobe, IBM, Microsoft, you name it — there are a lot of mundane tasks where there’s a very large workforce just cranking the crank, moving stuff from left to right. It’s an absolutely critical business function. If you don’t do it, your business stops. But on the flip side, it’s very mundane. If I make decisions about the life and death of a patient, that’s probably not what I want to put in the hands of an AI right now. But extracting a bunch of fields from a PDF — if I have a bunch of safeguards in the middle — that’s a very easily automated task. When we started with AI adoption in the company, the first thing was always, “How do I get all my documents into a data store that I can actually do something with?” I think we all went through that phase. Now you can download stuff from the internet — we have open-source tools — you can just point at a bunch of documents and boom, you have them in a database. That’s a logical first step. Once you go through that value tier, you can start looking at much more generic and smarter implementations. But these are the low-hanging fruits of very high impact because they’re cumbersome, require labor, and when they did comparisons of the quality of labor, AI already outperforms because it’s just... So from my perspective, it’s a very logical thing. It has an instantaneous impact and very low risk to the business. That’s usually your success story. But it is the building block for higher-level thinking. Once that is solved, we free up resources and we can do something else.

Tim Hwang: You’ve hit on this theme a couple of times, which I really like: there’s a tendency to say, “AI will never do that” when it’s actually already been doing that for a few months. The rate of progress is fast enough that your skepticism is not warranted because you just haven’t been paying close enough attention. This example is a great example of adoption, because what’s happening in finance is actually one of the clearest signals of where enterprise AI is also going.

Kaoutar El Maghraoui: Finance is uniquely positioned for agent adoption because it has structured data, clearly defined workflows, strong incentives for automation, and built-in validation mechanisms. It’s almost the ideal environment for agents to thrive. What’s interesting in this example is that we’re not just seeing automation, we’re also seeing orchestration. The LLM is not just replacing models or tools; it’s also sitting on top of coordination — extraction, forecasting, reporting, generating reports and narratives. It’s a different paradigm from traditional software. It’s setting very good examples of enterprise AI.

Volkmar Uhlig: The areas where I see the fastest adoption are primarily where data sets are readily available. In the finance industry, because of regulatory and recordkeeping requirements, everything is already kind of organized — if you ever have an audit, you better pull out the document. The other area where I see a lot of adoption is coding. We put everything in GitHub, every change is tracked, everything. The tooling was built around organizing large groups of people working in collaboration on a common code base. That is also highly tooled. So these are the two areas where we see very rapid adoption.

Ritika Gunnar: I’m going to add a third. We did a study over a few thousand AI engagements and found three primary areas. The first is coding — how do you start automating not just coding but the software development lifecycle. The second is back-office tasks — finance is an example, supply chain is another, marketing tasks, human resources tasks. These are areas where you have high access to your own enterprise data, where you need to not only automate but, as Kaoutar pointed out, orchestrate across all of these. You can not only see efficiencies but also start to find new ways of doing things. The third area is customer care — how do you have better support, better knowledge of your clients, better ways to target for adjacency targeting for different products or combinations of things. That is probably the biggest area that is even further transformed from when we started in the AI era. When you think about customer care, back office, and the software development lifecycle, we think there are dramatic improvements you can get from generative technologies.

Tim Hwang: Ritika, a final thought before we close up. I’m wondering if you have any thoughts on the reverse of what you just said — what’s going to be the hardest or the last to change. That’s also a really interesting set of questions.

Ritika Gunnar: Where emotional intelligence is really needed in day-to-day jobs, that’s an area where I think it’s going to be really hard to replace what humans bring. That EQ component is so critical in everything that we do today. That’s one. Two: physical labor jobs — plumbing, electricians. There’s been a lot talked about jobs that really require physical laborers. Those jobs are going to be hard to be completely replaced by artificial intelligence. So those are just a couple of examples. I think there are more. And as I said, every profession — it’s not that there is not going to be work in those areas for humans. It is how that work actually changes and how you become an AI-first organization as you do that.

Kaoutar El Maghraoui: A different take on the plumber jobs: there is a lot of work on physical AI and humanoid robots that are trying to bridge those gaps. The next phase after agentic, I feel, is the era of robots and autonomous agents and autonomous cars. That’s the next big thing coming — having some of these things also be disrupted by AI. We’re just not there yet, scientifically and algorithmically. But I see that happening next.

Tim Hwang: I’ve retooled from being a lawyer to a plumber for a few years — at least until the robots get there. This has been a great conversation. Kaoutar, Volkmar, Ritika, thanks for joining us. That’s all the time that we have for today. Thanks so much to all our listeners for following us for 100 episodes. If you enjoyed what you heard, you can find us on Apple Podcasts, Spotify, and podcast platforms everywhere.