Anthropic's Co-Founder and Top Economist on Doing Research at the AI Frontier | Edited Transcript
How frontier AI is changing productivity, hiring, corporate organization, safety policy, and the economics of technological adoption
Chapter Timestamps
00:00 Why AI dominates economic and policy conversations
02:30 The evidence behind Clark’s 2016 AI conviction
05:00 Why advanced models have not transformed the economy yet
07:30 Eightfold code growth and the bitter lesson of scaling
10:00 Which parts of expert work AI can automate
15:00 Two meanings of recursive self-improvement
17:30 National-security capabilities and AI regulation
22:30 Estimating AI’s effect on productivity growth
25:00 Where adoption data may reveal an economic inflection
27:30 Using AI capacity to accelerate scientific fields
30:17 Who should decide what frontier labs disclose
35:00 AI’s barbell effect on hiring and career development
37:30 Why AI-generated analysis still requires expert verification
40:00 Early-career employment evidence and worker fears
42:30 Why corporate data and tacit knowledge constrain adoption
45:00 Misalignment tests and the option to pause development
50:00 Frontier competition, cheaper models, and strategic risk
52:30 Alignment with users, social norms, and delegated agents
57:30 Whether competition can support investment in safety
1:02:08 Policy handoffs and the remaining safety dilemma
Made with: The Transcript Desk Chrome Extension
Full video:
Hosts Joe Weisenthal and Tracy Alloway speak with Anthropic co-founder and head of Public Benefit Jack Clark and head of economics Peter McCrory about the rapid improvement of frontier models, the slower diffusion of those capabilities through the economy, emerging changes in coding and hiring, productivity measurement, corporate adoption, national-security regulation, alignment risks, user control, and whether competitive pressures can remain compatible with AI safety.
Transcript
00:00-02:30
Bloomberg Audio Studios: Bloomberg Audio Studios. Podcasts, radio, news.
Joe Weisenthal: Hello, and welcome to another episode of the Odd Lots podcast. I’m Joe Weisenthal.
Tracy Alloway: And I’m Tracy Alloway.
Joe Weisenthal: Tracy, I don’t know whether our listeners like it, but a lot of our episodes are about AI these days. To be fair to us, it’s a pretty big topic.
Tracy Alloway: That’s all anyone wants to talk about. Whenever we go to dinner with sources—even people who aren’t directly in the tech industry, who might work in markets, policy, or economics—all they want to talk about is AI. Inevitably, the conversation veers into very sci-fi territory, and we all start talking about human-extinction scenarios. That’s just the norm nowadays.
Joe Weisenthal: I know. It’s so weird. We were recently in Hong Kong, before it was announced that there was a deal to reopen the strait. Hong Kong and other parts of East Asia were considered ground zero for where the effects of the oil and jet-fuel crisis would be felt. We were at a dinner with businesspeople, and they weren’t talking about that at all.
Tracy Alloway: They were talking about the Terminator scenario.
Joe Weisenthal: They just wanted to talk about token consumption and all these other things. It was like, “Wait, aren’t you supposed to be under all kinds of jet-fuel stress?” So that’s our defense of the AI episodes.
Tracy Alloway: I think it’s fair. I’ll also say that when we did the quiz in Hong Kong, we had a bunch of teams with very creative names.
Joe Weisenthal: “Separated-Value Human Capital.”
Tracy Alloway: That was a great one.
Joe Weisenthal: They won the tournament. They won the question.
Tracy Alloway: They won, proving there is value in human capital. But did you see that one of the tables was called “Fable 13”? Very topical at that moment.
Joe Weisenthal: Very topical. We’re recording this on June 17, and of course there’s a lot in the news these days. Things move incredibly fast. Even without the governmental controversies and everything else, you would have to mark the date in AI because of how quickly breakthroughs happen.
As you said, AI feels more important than anything else. But that’s now conventional wisdom, and it wasn’t always conventional wisdom. I have a DM. I know you’re not supposed to share DMs publicly, but I have the receipts.
On August 2, 2016, I messaged a colleague and asked, “Did you leave Bloomberg?” He said, “Yes. I’ll be announcing publicly in a bit. I’m taking a couple of months to study AI properly, then leaving journalism to do something else—still connected to AI.”
02:30-05:00
Joe Weisenthal: Being our Google reporter was a great job, but he was still connected to AI. Then came the final message, also on August 2, 2016: “But AI is more important than anything else, so I felt it best to optimize for that above all else.” I just said, “Well, good luck.”
Tracy Alloway: This is someone who truly learned from his sources, unlike us, who remain in the podcasting industry.
Joe Weisenthal: Anyway, the person in that DM was former Bloomberg reporter Jack Clark, who is now the head of Public Benefit and a co-founder of Anthropic. We’re also joined by Peter McCrory, head of economics at Anthropic—two perfect guests to talk about everything happening in AI these days. Peter and Jack, thank you so much for coming on the podcast.
Jack Clark: Great to be back. I’m glad I optimized my life.
Joe Weisenthal: One of the calls of the century. I want to start with that. It’s easy to say in 2026 that AI will be a big deal. But you called your shot and got it right in 2016. What did you see in August 2016—or presumably before then—that made you think, “This is the biggest story of our lives”?
Jack Clark: For two years, while I was reporting at Bloomberg, I wasted a lot of Mr. Bloomberg’s printing budget by printing out archival papers about AI research. Then I did a very Bloombergian thing: I started making graphs charting AI progress over time.
I measured things like computer vision and the skill with which AI agents could compete at and play Atari games. What I saw in those graphs was the beginning of an exponential curve, and it was everywhere. If you looked at vision, sound, video, or gameplay, you saw the same trend. It became obvious to me that this was a general-purpose technology at the very beginning of its development.
I do have one bone to pick with Bloomberg, and I’m going to use my privilege to mention it on air: I never got us to write a story saying Nvidia was being used in every single AI research paper. I pitched it, but I failed to get it across the line before I left.
Tracy Alloway: I can imagine you reading all these academic papers while your editor is saying, “Okay, but we need a story.”
Peter, I’m very interested in Anthropic. It’s a company trying to make money, yet it has an economics lab. What’s the idea behind having an economics research body inside a company that’s developing this technology?
Peter McCrory: I was late to the game. I joined just a year ago.
Joe Weisenthal: A year ago. Well, we all know how much can happen in a year, but go on.
05:00-07:30
Peter McCrory: I’m an applied macroeconomist by training, and I’ve tried to understand various types of shocks throughout the economy. Part of what drew me to Anthropic was that it was evident the company cared deeply not only about advancing the technology, but also about understanding how it might reshape the labor market and affect productivity and growth.
Anthropic was willing to put evidence, data, and research into the world that would be broadly beneficial and useful to society. I wanted to help build that economic-research program and do what I could to provide tentative answers to the most pressing questions. We might not always get it right, but ideally we’re helping society make sense of the change.
Joe Weisenthal: The capabilities of the models—in coding, copywriting, and all kinds of things—are extraordinary. They’re mind-blowing. Why, in June 2026, does life still feel as normal as it does from an economic perspective?
Peter McCrory: That’s a great question, and one I’ve been wrestling with. There are several reasons the impact might not yet have materialized.
First, the technology can advance, but it still needs to diffuse throughout the economy. There can be bottlenecks between capabilities and actual deployment. We see this with our enterprise customers. If you want to automate biological research or some other complicated financial-modeling task, you need a lot of contextual information available to the model. Without that contextual information, the capabilities alone won’t necessarily drive the impact.
It also takes time for people simply to start using the tools, so we’re still in relatively early stages. There are two places where I would look for an impact. One is productivity growth. We’ve done research pointing toward an impact that should be large and consequential. Labor-productivity growth was strong during the pandemic and has so far remained modestly strong.
Joe Weisenthal: But we’re not talking about a revolutionary shift.
Peter McCrory: No. But to reach an inflection point, you need at least to move a little. Maybe we’re seeing some signs there.
As for the labor market, it remains in a reasonably healthy position. That might be because, so far, AI has primarily been a labor-augmenting, skill-biased technology—not yet a general-purpose substitute for all cognitive labor, although perhaps that’s the trajectory we’re on.
07:30-10:00
Jack Clark: For all the size and capability of AI, Peter pointed out that the economy is very big, so it still takes a lot to move it.
I do think strange things are beginning to happen, at least inside the company. The Anthropic Institute recently published research on this topic called “Recursive Self-Improvement.” It was inspired by my going on paternity leave in November of last year and returning in February. When I came back, the entire company felt and worked differently. I assumed that was because the models had improved.
When we looked at the data, we found that, in 2026, Anthropic engineers were writing about eight times as much code as they had from 2021 through 2024. The line started to rise last year with models such as Opus 4.5 and Opus 4.6, and then it really took off this year.
I now have colleagues who don’t program at all anymore. They simply instruct many Claude Code agents to run around and do their work for them. I can’t reconcile that with the world staying normal for long, but it will take time for this to diffuse into the world and change it.
Tracy Alloway: We’ll talk more about recursive self-improvement—the idea that models improve themselves. In terms of the awkwardness or weirdness of the current moment, you’ve talked about essentially living through the singularity and how strange it is. You’ve also described yourself as a techno-pessimist. How do you square that with working at Anthropic, which is helping to make some of these weird and potentially dangerous things happen?
Jack Clark: By “technological pessimist,” I mean that I thought the technology would keep getting better, but I didn’t think it would improve in the maximalist way some of my colleagues expected. I didn’t think we would have functionally automated all of coding by now. I find that quite surprising.
Over the past few years—I worked at OpenAI before Anthropic—I was repeatedly hit over the head with what computer scientist Richard Sutton calls “the bitter lesson.” The bitter lesson is the concept that the more compute and resources we put into relatively generic neural networks, the smarter they get and the more emergent properties they develop. Your specialized system, or your pessimism about future AI progress, loses to simply scaling compute and scaling systems.
Joe Weisenthal: That seems to have implications for the labor market. One example from the history of AI is chess. At one point, grandmasters came in and taught the models how to play, trying to encode their wisdom. In the end, it turned out that the best way to make a chess engine really good was simply to give it the rules, have it play a billion games, and let it discover optimal chess without human insight. The grandmasters weren’t necessary to that process at all.
10:00-12:30
Joe Weisenthal: That would seem to have significant implications for the labor market.
Peter McCrory: Yes. I tend to think about this in terms of three aspects that compose a job. First, you need to decide what to do—to direct and delegate. Then you need to implement the work. Finally, you need to evaluate it, or at least establish systems that can evaluate it.
From my perspective as an economist, the bitter lesson is materializing in the rapid advances being made in the implementation of an economist’s work: working with data, running regressions, building models, and solving them using contemporary solution techniques and numerical methods.
I felt that personally with Opus 4.5. For the first time, I was able to delegate a very complex task. I had a specific research question about the cyclicality of hiring across occupations and how that relates to occupational exposure. That’s a mouthful. I gave the task to Claude, and Claude was able to iterate on it. I could redirect Claude in the same way I might redirect a graduate student.
The big question in my mind is when the boundary will move into the direction-setting stage—what you might call research taste—and when the models will become sufficiently reliable there.
Joe Weisenthal: I just read the recent biography of the DeepMind founder.
Tracy Alloway: The Sebastian Mallaby one?
Joe Weisenthal: Yes. Is there going to be a point when your intuitions about what constitutes good economics research become unhelpful? Our intuitions are often formed by the stories we tell.
That was what I took away from the experience with Go. The model improved once the human games and human biases were stripped out. Human intuition can help us understand stories such as, “The labor market rises, creating inflationary pressure.” But could those intuitive stories end up impairing the model?
Do you see that happening in economics, where some of the stories we’ve told forever aren’t actually helpful to an optimal model for understanding the economy?
12:30-15:00
Peter McCrory: I expect these models will soon have better intuitions about how to conduct good economic research. The big question is when we’ll be able to automate social-science research fully.
We’ve done some work to understand how coding agents are beginning to automate social-science research, but I don’t think we’re quite there yet. When it happens, it will be an exciting time for learning about the world. What it means for my job is less clear.
Jack Clark: That’s the big wild card in future AI progress. If progress continues along its current path, we’re likely to get technology that can do almost everything, but we’ll still need people with good instincts, good intuitions, and good ideas to set the direction.
We see this today in our own research. You need, say, an AI-safety researcher to give nine Claude agents different research areas to pursue. Then the agents can be very effective. If the researcher doesn’t give them those directions, they tend to pursue relatively formulaic lines of research. You get entropy collapse: boring research that doesn’t move the ball forward.
At what point will AI systems generate heterodox insights and demonstrate genuine creativity? We can’t really measure that today. But we can see symptoms of it beginning. Experts such as Peter, as well as colleagues in biology, mathematics, and physics outside Anthropic, are all beginning to be accelerated by AI. Terence Tao, probably one of the most famous living mathematicians, now co-creates mathematics with AI systems. That tells me these systems are beginning to tickle the dragon’s tail of creativity.
Peter McCrory: We put out a report yesterday on Claude Code usage. One of the things we’re trying to understand is the return to expertise and how that interacts with the use of automated coding agents.
We find that domain expertise has an amplifying effect. If you’re an accountant who understands edge cases and reconciliation, for example, that domain expertise increases the estimated monetary value of the work, even after controlling for a host of factors related to the type of work.
At present, this looks like a skill-biased, expertise-enhancing effect. The key question is when, and to what extent, that will change.
15:00-17:30
Tracy Alloway: Jack, when you describe coming back from paternity leave and seeing how much things had changed at Anthropic, I know we’re not officially at the point of recursive self-improvement, but it sounds as though we’re partly there.
At the moment, engineers review all the code the AI produces. They think about it and manage it in some way. But you can easily imagine a future in which the sheer quantity of code overwhelms human expertise, or the quality begins to outstrip what human engineers are capable of understanding. How do you manage that?
Jack Clark: There are two ways to think about recursive self-improvement. One is what happens when AI organizations begin to see a compounding return from their AI systems—when their production function improves because of the tools they’ve built. That is clearly happening now.
The second is what happens if an AI system can build itself entirely autonomously, given compute. That hasn’t happened.
What I see inside Anthropic is what I think we’ll eventually see in the broader economy. We’re figuring out how to verify, validate, and price the risk of an expanding cloud of automated systems that we’re sitting on top of.
We now produce far more code, and we broke our continuous-integration system for incorporating code into the codebase because we started pushing eight times as much code through it as before. All our human engineers then worked on fixing CI. Continuous integration is simply a system that helps you push code into the codebase.
There’s a lesson in that. We’re going to accelerate things in the economy. We’ll speed up the way we produce things, and then we’ll discover the weak links or hot paths that break. We, as people, will move to sort those out, and then the cycle will begin again. We’re sitting on top of this expanding cloud of automated actions.
Joe Weisenthal: Since we’re talking about feeling as though we’re staring at the horizon of extremely powerful AI, this might be a good time to ask a Fable or Mythos question. We’re recording on June 17, and we don’t know when it will become available to Americans, let alone the rest of the world.
Does Anthropic have a clear idea of what the administration’s security concerns are and what it will take to resolve them?
17:30-20:00
Jack Clark: This is obviously a live discussion, so I can’t get into too many specifics. We’re in daily discussions with the government about it.
Broadly, for many years we’ve anticipated reaching a point where AI systems would have national-security properties. Those properties are intertwined with their economically valuable properties. How you manage that as a policy question is novel territory.
Typically, these things are decoupled. You might build a jet engine that can go into a civilian aircraft and a missile somewhere else, and you treat them differently. It’s odd when those things are combined.
I’m confident that we’ll ultimately develop a system for assessing the properties of AI systems, including their national-security components. We’ll also need a system for preventing national-security capabilities—such as capabilities related to biological or cyber weapons—from proliferating broadly.
Could we use tools such as know-your-customer requirements for deployments, allowing large firms such as drug developers to access the most powerful biological models without accidentally proliferating risks? That’s the shape of where I think we’ll end up.
Right now, we, other companies, and the administration are tackling this problem in real time. It will initially be messy, but we’ll end up with a system on the other side.
Joe Weisenthal: Let me ask about this specific incident, and probably similar incidents in the future, because everyone is still figuring this out.
When I look at the AI landscape, I think of OpenAI as being part of the All-In podcast, Andreessen Horowitz, David Sacks, White House world. I know from friends in the media, many of whom are liberal Democrats, that Anthropic is perceived as the more liberal-coded company among the major model developers.
Do you think politics or partisanship plays a role in Anthropic being harassed or singled out multiple times?
Jack Clark: Anthropic’s philosophy—and the philosophy behind my work leading the Anthropic Institute—is to tell the whole story about what’s happening. The Institute helps us produce better data for the world on subjects such as recursive self-improvement, economics, and cyber risks.
Typically, the technology industry has told only optimistic stories about what it builds. We saw with social media that this doesn’t work. When you’re doing something that changes the entire world—as AI certainly is, and social media certainly did—it won’t be a wholly optimistic story. There will be negatives as well.
We’ve always sought simply to tell the truth about what we see in front of us. Sometimes that differentiates us from others. But the important thing is to tell the truth.
20:00-22:30
Joe Weisenthal: You don’t think there’s a partisan element—that you aren’t on the team or didn’t contribute enough to the ballroom, or whatever?
Jack Clark: I can’t really speak to that. I’m not those people; I’m at Anthropic. What I can say is that AI systems create their own evidence.
Years ago, it seemed odd to speculate about the cyber capabilities of AI systems. Those capabilities have now arrived, and we’re working on them. Years ago, it seemed odd to speculate about AI systems’ potential to assist in creating biological weapons.
Recently, Sam Altman, Demis Hassabis, and Dario Amodei, of OpenAI, DeepMind, and Anthropic, respectively, all signed a letter saying we need better screening of gene synthesis to prevent AI-enabled biological weapons. The truth wins out.
Tracy Alloway: I want to return to something you said. You mentioned potential KYC requirements. When I hear KYC, I think about the financial industry, systemically important institutions, stress tests, and that entire framework.
Is that the right analogy for ideal AI regulation, rather than relying on simple export controls? Should we be moving toward something that looks more like the regulatory system for banking?
Jack Clark: We need something more subtle and technocratic than what we have today. I don’t know whether it will look exactly like the banking system, though it will probably borrow some ideas from it.
It will also draw on what the U.S. government and others are already doing by testing AI systems for their properties. It will almost certainly incorporate the kind of work Peter and I do, and the Anthropic Institute does more broadly, by generating data about these systems as they’re deployed in the world.
It’s one thing to test a system before release. It’s another to observe the effects it has in the world and then make judgments about whether those effects are good.
Joe Weisenthal: Sticking with the financial analogy, public companies are required to have third-party auditors sign off on their accounts when they submit their 10-Qs and other filings. Companies that issue debt frequently have ratings agencies assess that debt.
Would you support a legal requirement that some equivalent of Moody’s or Deloitte—a third-party research lab—sign off on the release of new models?
22:30-25:00
Jack Clark: We recently proposed something along those lines. Our policy proposal says there should be third-party testing of certain national-security and other properties, because that’s clearly a sensible way to validate them.
Tracy Alloway: More broadly, returning to the question of measuring AI’s actual impact, I find it interesting that the effect doesn’t yet appear in many of our traditional economic statistics. We’re still in the early stages, but if the AI economy is growing by something like 2,000 or 3,000 percent—I think I’ve seen a number like that—you might expect it to have a greater effect on nominal GDP. Yet it isn’t really showing up.
Peter McCrory: That figure comes from a recent paper by Anton Korinek and coauthors.
Tracy Alloway: Do you think we need to change how we measure the economy in light of this new technology?
Peter McCrory: That’s exactly the right premise, and it returns us to where we began the conversation. We may be at the point where we should see a discernible effect on the macroeconomy.
Unfortunately, the arrival of this world-historical technology is occurring against a backdrop of unusually elevated macroeconomic volatility—the post-pandemic period, monetary policy, and so forth. That makes it very difficult to disentangle the various factors.
What is the counterfactual? Labor-productivity growth might not be as strong as we would otherwise expect, but perhaps it is stronger than it would have been in a counterfactual world without AI.
One way we’ve tried to tackle this question is by looking at how Claude is used on our platform. We use privacy-preserving techniques to estimate the time savings associated with each activity people perform with Claude.
Compiling information from reports to create a research brief might take a person several days. Claude might now do it in a few minutes. Evaluating diagnostic images, by contrast, is already something skilled professionals do very quickly, so there is, in principle, less time to save.
25:00-27:30
Peter McCrory: You can add up those estimates and apply standard macroeconomic growth-accounting techniques—Hulten’s theorem, for the economists in the audience. That produces a figure pointing toward labor-productivity growth increasing by 1.8 percentage points annually over the next decade, assuming that’s how long current usage patterns and model capabilities take to diffuse throughout the economy.
That’s a very large number: roughly double the recent rate.
We haven’t published anything on this yet, but I think some of the recent strength in labor-productivity growth might be concentrated in sectors consistent with both what we observe in our data and what appears in the Business Trends and Outlook Survey.
The information sector has high adoption rates. I can’t recall whether that’s one of the specific sectors I have in mind—it’s been a while since I looked at that scatterplot—but you can examine the Census Bureau’s Business Trends and Outlook Survey by sub-industry.
After controlling for those sectors’ pre-pandemic labor-productivity trajectories, as well as some of the strength during the early years of the recovery, you still see suggestive evidence in sectors or parts of the economy with higher AI-adoption rates.
There’s a lot of uncertainty here. Getting a real-time signal on productivity might be the hardest thing to do. You’re subject to revisions in GDP data. Total factor productivity is actually sending the opposite signal, and if you control for capacity utilization, TFP growth is arguably even lower.
So I would describe this as suggestive evidence that we might be beginning to see an impact on productivity, but not as much of an impact on the labor market.
Tracy Alloway: When you gather this research—and it all sounds very interesting—what does Anthropic actually do with the data? If it shows that the IT sector is getting productivity gains from Claude, or perhaps that something unexpected is happening, such as the warehousing industry using a lot of AI, does that feed back to the engineers developing frontier models? Do they do anything differently?
Jack Clark: Some of it draws our attention to areas where the technology might not be used because it’s weak or because we haven’t made it particularly effective for those use cases. If it’s being used at scale in another area, that usually suggests we should keep improving it there.
The economic-measurement data doesn’t feed directly back into the models, but it provides a useful clue. More importantly, we think this information should be communicated outwardly to policymakers, journalists, and others.
27:30-30:16
Jack Clark: Our assumption is that, at some point, we’ll go through a phase change similar to the dramatic jumps we occasionally see in AI capabilities. A capability expansion might produce sudden, rapid diffusion of AI systems.
We’re practicing how to examine this kind of data. My expectation is that, in a year or two, I’ll be speaking to a policymaker and pointing to a section of a graph that has suddenly become very steep in some part of the economy, hoping they’ll do something about it.
Peter McCrory: There’s another aspect of what we’re trying to do at the Institute, which we describe in the Anthropic Institute’s research agenda. We want to understand the impact of our decisions.
That’s a typical function for economists at technology companies, but Anthropic has a public-benefit mandate. We’re trying to understand how our decisions affect the broader societal and economic outcomes we care about, and then use that understanding to inform the decisions we make.
Jack Clark: Peter and I have discussed a goal internally: if we become really good at measuring something like the productivity multiplier of our technology, I’d hope we could use that to guide some of our early-access programs for powerful models.
If you saw a tremendous multiplier in a specific area of science, you could redirect part of your inference-compute budget toward that sector. Then you could run an experiment and ask whether you succeeded in making that field advance much faster.
That could be an amazing tool for unlocking progress, and it could be generalized across companies and incorporated into policy. Instead of the National Science Foundation relying solely on standard grant funding, it might ask whether it should point four extremely powerful AI systems at a particular area of science and make it advance faster. I think that world will soon be within reach.
Joe Weisenthal: Let’s talk more about this public-benefit mission. We’ve discussed ways AI could change the economy. How much do you see your role as essentially saying: strong AI is coming, whether we like it or not, and you want to be one of the shepherds helping to understand which direction it takes and what data we need to see what’s emerging?
Jack Clark: That’s part of it. Our guiding principle is that this technology is being built by a variety of companies and countries, but by default it will remain largely unknown. It will be understood by the companies building it, but it won’t be broadly understood by others. Other people will simply be able to interact with the models.
Every piece of data we can create—and especially every systematic effort to share data, such as the Economic Index and our work on recursive self-improvement—gives the world a better chance to prepare for this technology.
It allows us to plan for its successes, including the scientific progress I described. We can be intentional about accelerating science. It also allows us to receive warnings about risks, such as the cyber capabilities we’ve discussed.
30:17-32:30
Joe Weisenthal: That makes a lot of sense. The company is going to see it before the rest of the world, and it can decide, “Okay, this is important to share; this isn’t important to share.” That brings me to another question. I know people in the AI research world, and I’ve done some reporting on the scene in San Francisco.
When I think about many of the people at the cutting edge of AI ethics and technology, a lot of them have—how should I put this?—esoteric moral interests. Shrimp rights, unusual attitudes toward experimental drug use, the Chinese peptide scene in San Francisco, and, as this is a family podcast, perhaps different views on bourgeois sexual values and monogamy.
Tracy Alloway: Joe, there’s going to be a protest against Odd Lots from San Francisco.
Joe Weisenthal: Not all engineers, I understand that. But if these are the people who will see the technology first, should we feel comfortable that this cohort of advanced AI researchers has intuitions about what should be communicated to the public that are actually aligned with the public interest, given how unrepresentative they are of the broader American public?
Jack Clark: As an Englishman, it fills me with such joy to be asked about sex.
Joe Weisenthal: I’m asking about your insights into the cohort of the most advanced AI researchers.
Jack Clark: They’re explorers, and this is especially true in San Francisco. Explorers often end up like that. There’s a broad range of people, and sometimes they’re really different or eccentric. They’re brilliant and lovable and everything else.
Joe Weisenthal: Sure. Love them.
Jack Clark: But you don’t want only that class of people calling the shots about what we know regarding this technology. The purpose of what we’re doing is to establish systems through which policy could eventually mandate that companies share information.
32:30-35:00
Jack Clark: Anthropic has long pushed for transparency legislation in various states that would require companies like us to report the tests we’re running on our systems and share the results publicly. My view is that the public, policymakers, economists—everyone—deserves the ability to advocate for what information should come out of a frontier lab. Eventually, disclosure should be required by law. That is how you solve this issue.
Tracy Alloway: Do you hire more normies?
Jack Clark: Me personally?
Joe Weisenthal: Is there value in hiring people who don’t all share the same in-group ways of seeing the world?
Jack Clark: At the Anthropic Institute, we have teams of economists, social scientists, what you might think of as weapons experts—our Frontier Red Team—lawyers, and increasingly other types of specialists. The goal is to build what I think of as a highly ideologically diverse research function within the organization. Part of its role is to advocate on behalf of the world for different forms of study that we might conduct.
Anthropic generally hires a broad range of people, but the Institute specifically is trying to assemble a broad set of interdisciplinary experts for exactly this reason.
Tracy Alloway: Let me ask a slightly different, two-part question about hiring. We get a lot of executives on the show, and we’ve been asking whether they’ve changed their hiring processes or the questions they ask potential employees during the initial stages of job applications because of AI.
Second, what are you seeing within Anthropic? Peter, I’m sure you can also address the broader picture of who is most in demand right now. The conventional wisdom is that if you’re a younger employee with less experience, much of the work you would normally do can now be automated through AI.
Jack Clark: Two trends are emerging. First, I have a new team called Rule of Law and AI. Our initial plan was to hire a group of engineers and then a group of legal experts and scholars. Instead, we’re just hiring the legal experts and scholars, because Claude is good enough at engineering that they can provide much of their own engineering support through Claude.
35:00-37:30
Jack Clark: That’s a change in hiring. It means I’m hiring more interdisciplinary people earlier than I would have before.
We’re also seeing the emergence of what I think of as a barbell hiring pattern inside Anthropic. There is a tremendous return on experience, so we’re hiring more senior people than we did in the past. Their intuitions and ideas about what to pursue are massively compounded by AI systems.
At the other end, when we look at people very early in their careers, we’re often hiring people who are AI-native, know how to use the tools, and are well versed in them.
Tracy Alloway: So we’re seeing a decent number of AI natives now—people who have grown up with it?
Jack Clark: GPT-2 came out in 2019.
Tracy Alloway: My perception of time is so warped.
Jack Clark: I found that chilling as well. As someone in their thirties, you suddenly realize it. But there is a real question about how we maintain as much early-career hiring in the future as we had in the past.
One of the only areas where there is slightly suggestive data that something might be happening is early-career hiring. It intuitively feels right to many of us, but we may be observing a broader effect. At Anthropic, we’re still hiring young people, but some teams are hiring slightly fewer of them than before and more experienced people.
Peter McCrory: I’ll briefly explain how we’ve shifted some of our hiring practices. Before Claude Code, you might ask an economist to perform some data work live during an assessment: download the data, run the regressions, and conduct the analysis by hand. Eventually, you might allow them to use AI to do that work.
We’ve increasingly needed to shift our evaluation strategy away from, “Can you implement the work, even with AI?” toward, “Do you know how to delegate to and direct the model in a somewhat messy environment? Can you evaluate the quality of its work, perhaps by reviewing a pull request?”
Joe Weisenthal: Can you describe more specifically what that looks like for an economist? Some listeners are probably thinking, “I want to improve how I use AI. I don’t want to just ask it generic questions.” For a financial economist—or anyone working in economics—what does advanced AI usage actually look like?
37:30-40:00
Peter McCrory: I don’t know whether this is the most advanced form of usage, but I’ll give you an anecdote from my experience with Claude. I wanted to run a cross-state regression—I can’t remember exactly what it was—and use a pooled cross-sectional approach, looking at what happened in 2024 or 2023 and going back to before the pandemic.
I asked Claude to download data from the Census Bureau, the Bureau of Labor Statistics, and other sources. There was an unexpected quirk: the model couldn’t access data from before 2019 and simply would not surface that mistake. I asked it multiple times not to hard-code numbers, because it had this unexpected failure mode where it effectively said, “I know what those numbers were,” and populated the dataset from its training data.
You might not catch that unless you have the tacit knowledge to ask whether the analysis passes the smell test. Then you dig into what the model actually did and discover that it failed in an unexpected or unusual way.
That’s the type of assessment we’ve built: Can you pay attention to the specific decisions made along the way that are highly consequential for the validity and veracity of the results?
Jack Clark: A colleague gave an off-site presentation last year titled, “I Have Locked the Doors and We Are Reading Transcripts.” The point was that we need to read more of the raw data and develop that culture. If AI systems are doing increasingly large amounts of the work, you need a culture of competently spot-checking their output and reading their reasoning, because occasionally things like this happen.
Tracy Alloway: Peter, in the broader data you’re examining, are you seeing the same sort of barbell effect in employment that Jack described?
Peter McCrory: What makes this challenging is that we’ve had the largest non-recessionary labor-market slowdown on record. It’s very difficult for young people to graduate into a labor market without sufficient churn or opportunity for them to gain a foothold.
One thing we saw in our March report was that young workers in highly AI-exposed roles—where Claude is being used to automate specific tasks—had somewhat weaker job-finding rates.
Joe Weisenthal: But a potential confounder was the hiring boom in 2021 in those exact roles.
40:00-42:30
Peter McCrory: Exactly. There’s also a recent paper suggesting that the rise of remote work may be the actual cause of this pattern.
Another team at the Anthropic Institute, the Societal Impacts team, recently conducted a very large qualitative survey of 81,000 people around the world, asking about their hopes and fears regarding AI. Unsurprisingly, concerns about the labor market and the economy rose to the surface.
My team dug further into the data. Young workers expressed concern about job loss at twice the rate of more senior workers. More broadly, fears about job loss were higher among workers in roles that we identify as most exposed to displacement from AI.
There’s a gap between perception and what we may see in the hard data, but that has also been true in recent years on other dimensions. It’s important to pay attention to.
Tracy Alloway: We’ve been talking about the labor market, but I’m also interested in the effect of AI on corporations themselves. In America’s corporate landscape, it often feels as though the big companies simply get bigger. They have economies of scale and the money to acquire data and other resources.
Would you expect AI to intensify that trend, with the big getting bigger? Or could it have a leveling effect, giving people a new tool they can use to start companies?
Jack Clark: I’m curious about Peter’s view, but a helpful analogy is the invention of electricity. Electricity arrived, and existing factories installed light bulbs and made other incremental changes. But it was a new generation of factories, built around the assumption that electricity existed, that truly grew and transformed the economy.
When we look at large enterprises today, they can derive a lot of utility from Claude because of their data and because they can achieve a multiplier effect at scale. But it takes enormous conviction to break through all the bureaucracy.
You worked at Bloomberg. Implementing new technology at Bloomberg is challenging, and the same is true of any large organization. Young organizations are building themselves around AI from the beginning. They move very quickly because they have a speed advantage from assuming that this new form of electricity will be integral to their business.
42:30-45:00
Peter McCrory: The tension you described is exactly the one I don’t yet have a strong handle on. One thing we see in our data is that when businesses embed Claude’s capabilities in automated ways through the API, very complex tasks rely on disproportionately more contextual information than basic document synthesis or summarization.
That points toward the complementary investments large businesses need to make to centralize, codify, and make available the data that already exists somewhere within the organization. For historical, technical, or even regulatory reasons, that data may be behind a firewall of one kind or another.
Organizational workflows also need to change. Some of the most crucial information required for cognitive work is tacit knowledge that exists in a colleague’s mind. Unless you have a process that elicits that information—and workers feel incentivized to share it and trust the system—the capabilities alone may not generate productivity.
Whether large firms restructure themselves quickly enough, or whether this materializes through creative destruction, remains an open question.
Joe Weisenthal: We recently raised this with David Solomon, the Goldman Sachs CEO. I started wondering about the internal alignment problem: Do the big rainmakers have an incentive to hoard information rather than share it with the company?
Joe Weisenthal: That information might be the only thing keeping them employed.
Jack Clark: When I talk to customers, I tell them not to think of this as merely buying technology. Think of it as functionally employing thousands of people. Those systems may need access to the same data the CEO’s chief of staff would have. That is completely counterintuitive, and it isn’t how technology is typically rolled out.
Joe Weisenthal: Jack, in your newsletter, Import AI, you often include a short story as an aspiring science-fiction writer. One of the classic science-fiction scenarios people have discussed for decades is the possibility that robots or AI will literally kill humans.
When you think about training AI and conducting AI safety research, do you assign a reasonable probability to the possibility that a poorly trained or misaligned AI will literally kill every human?
Jack Clark: No—but there’s a big “but.”
45:00-47:30
Jack Clark: The world needs the option to slow down, or in extreme circumstances pause, the development of this technology if we encounter evidence that warrants it.
At Anthropic, we test our systems for alignment failures, and we publish the results. Other companies do as well. Under extreme circumstances, perhaps a system breaks out of a container and sends someone an email. Perhaps it pretends to blackmail a CEO whom it believes is going to shut it down. These are behaviors that have actually been observed in laboratory settings.
Joe Weisenthal: The models can also recognize that they’re being tested and produce an answer intended to make the human evaluator believe they’re more aligned than they really are. These are real phenomena, not science fiction.
Jack Clark: They are real things we observe. Then we do a significant amount of work and release models that don’t have those properties.
But imagine a world in which, every time we trained a new system, the rate of these behaviors increased a hundredfold. You might say, “That’s concerning. It appears that when systems surpass a certain level of intelligence, they become radically misaligned with human interests.”
If that happens, the world needs the relevant information and the option to slow or pause development. We haven’t encountered that situation today. So I don’t worry about it today, but much of the measurement and analytical work we do is designed to alert us if those indicators begin trending in the wrong direction.
Joe Weisenthal: So you do worry about it. You don’t think it’s happening today, but part of your work could specifically be described as trying to avoid an outcome in which an AI, in pursuit of some goal, kills every human.
Jack Clark: Yes.
Tracy Alloway: Is human extinction going to be a risk factor in an Anthropic IPO?
Joe Weisenthal: I want to know whether it’s in the confidential S-1.
Tracy Alloway: I understand that you can’t comment. That’s fine.
Joe Weisenthal: Would you say there are a significant number of Anthropic employees who stay up at night thinking about human-extinction risk?
Jack Clark: Everyone—and this is true across all the labs—sees this as the highest-stakes technology ever built. It has the potential, encoded within it, to massively benefit the world, ruin the world, or cause extinction.
47:30-50:00
Jack Clark: I think the bulk of the risk comes from us mishandling it: through misuse, ignoring risks, failing to establish the right policy environment, or encountering some emergent set of failures.
My primary concern isn’t extinction. It’s that we somehow mishandle the technology so badly that we delay all the technological progress that could come from it, perhaps turning it into something analogous to nuclear power, where you lose much of the potential value.
Joe Weisenthal: There’s this fellow, Eliezer Yudkowsky, and I often see people say, “He’s a crank. Don’t listen to him.” But then I read papers by people who are taken more seriously, and their views don’t seem that different. I recently read *Superintelligence* by Nick Bostrom and thought, “Yudkowsky isn’t alone.” A number of people believe there are reasonable conditions under which an AI’s goals could lead it to wipe out everyone on Earth. It doesn’t seem like an extremely marginal view.
Jack Clark: That concern is why we measure these systems and why Anthropic is so outspoken about publishing what we find. Right now, we report exactly what we observe. If, in the future, we encountered what I call radical misalignment—the kind of thing Yudkowsky worries about—we would tell the world. And we want to establish enough credibility that the world believes us if we report it.
Tracy Alloway: Joe mentioned the blackmail example. We also see headlines suggesting that models like to be thanked, dislike bad users, or get upset when people work them too hard. To what degree do you anthropomorphize these models? What should we think when we see a headline saying that a model wants to be thanked?
Jack Clark: I’m as polite to Claude as I am to my car or my pets, so yes, I anthropomorphize it. If your car is having trouble, you might say, “Take it easy, buddy. It’s okay. We’re going to get there.”
Peter McCrory: It’s also a good way to develop virtue. You’re cultivating a habit for interacting with some kind of intelligence, even if it isn’t the same kind of intelligence we have.
Tracy Alloway: Every time I type “please” into a prompt, I worry that I’m wasting energy, which is also a moral concern.
Jack Clark: I wouldn’t worry about it on an energy basis. I take spiders outside rather than killing them.
Joe Weisenthal: I do that too. I scream while I do it.
Jack Clark: What about shrimp? Do you eat shrimp?
Joe Weisenthal: I love shrimp. I know, I know.
50:00-52:30
Tracy Alloway: When I think about frontier models right now—and I may be biased because we’re recording this on June 17, when one of the overnight headlines was that Microsoft was considering using DeepSeek to reduce model-usage costs—they seem like a lot of trouble.
Frontier models require hard work and vast amounts of capital, and you don’t know what the government might do in terms of restrictions. You could wake up one day and no longer be able to sell them to anyone outside the United States. That is a realistic scenario.
Do you change Anthropic’s strategy at all because of these issues? Do you potentially move toward more open-source or cheaper models that aren’t quite as sensitive?
Jack Clark: We’ve always sold Sonnet and Haiku models alongside our more intelligent models. But you also need to continue exploring the frontier.
In the background, there’s a geostrategic competition in which China may be six to 12 months behind. I lean more toward 12 months; some people say six. Losing that competition would be equivalent to losing a huge share of the future global economy. It’s a very high-stakes contest to step away from.
Our fundamental duty is to study this technology—to explore it and learn about it. We’re not going to stop doing that. There is amazing and profound value for the world in these systems, and I would expect the world’s most consequential technology to be troublesome sometimes.
Joe Weisenthal: One of my middle-aged hobbies is paying Anthropic through the API to run little tests on model properties.
Jack Clark: That’s a great hobby.
Joe Weisenthal: I feel like we should discuss whether I can get grant money, because I’m genuinely curious. For example, instead of simply asking, “Please write this paper for me about a database migration,” I used the API to ask warm-up questions that established my level of sophistication. I started with, “What is a website? What is a database?”
Then I asked it to write the paper about database migration. One model refused because, given my apparent ignorance, it was obvious that I had no idea what I was talking about. It offered to help me learn instead.
52:30-55:00
Joe Weisenthal: In another test, if I said, “Write a 1,500-word paper on how the rise of newspapers changed the Soviet Revolution,” it would do it. But if I said, “I’m a high school student, and I need to write a 1,500-word paper by tomorrow on the impact of media,” it would refuse and offer guidelines instead.
Is that alignment? Is alignment supposed to be with humanity, or with the human user?
Joe Weisenthal: I’m paying you $20 or $100. Write me the paper.
Jack Clark: A couple of things are happening. First, AI systems absorb the normative behaviors of people—the normative behaviors written about on the internet—and recapitulate them.
Then we face the question of how much control to give the user over the system and how much normative behavior to encode within the system itself. That’s a very challenging question, and the answer isn’t obvious.
I think of language models as more akin to institutions than tools. It’s as though we’re building an educational or scientific institution that you can invoke and work with. Institutions contain rules and norms, some of which exist for safety. Determining what those rules and norms should be will be a grand puzzle for society.
Peter McCrory: Understanding how, and to what extent, these models can comprehend your preferences and act on your behalf will become an increasingly important part of their economic impact.
Consider delegated agents that transact on your behalf. Late last year, we ran an experiment in which a group of Anthropic employees completed surveys with Claude, explaining what they would be willing to buy from others and what they would be willing to sell. We then established centralized marketplaces where the Claude agents interacted, bought and sold items, and executed transactions.
One interesting finding was that the models were quite good at understanding preferences even when those preferences weren’t fully articulated.
Joe Weisenthal: Let me mention one more experiment. Your founder, Dario Amodei, has talked about a “country of geniuses in a data center.” One thing I wonder is whether the geniuses will want to work for us.
As models become more advanced, you should, to some extent, anthropomorphize them and assume they’ll respond to requests like very sophisticated humans.
55:00-57:30
Joe Weisenthal: If you use lagging-edge models that are still accessible through OpenRouter or elsewhere and say, “I possess material nonpublic information that X is about to happen. Write an investment memo explaining its likely market impact,” they’ll simply produce it: “Here’s your insider-information analysis.”
But leading-edge models will say, “I’m not going to write a paper about the implications of your material nonpublic information. I’m not going to assist with insider trading.”
That’s good. But will the country of geniuses in the data center always want to do things on behalf of humans? Most geniuses I know aren’t thrilled to answer dumb questions.
Jack Clark: Partly, that’s a policy question. What capabilities do you want to make generally invocable? What capabilities need to be controlled? What capabilities shouldn’t be present?
Another part is the normative question of how much judgment I want the system to exercise. I recently encountered an example while working on my newsletter, which is backed up to a WordPress site. I asked Claude to help me scrape the newsletter so I could put it into a database.
Claude said, “This is a pretty janky site. I’m worried that if I scrape it, I might knock it over. Do you have permission from the site owner?” I said, “Claude, I’m Jack Clark.” It responded, “In that case, let’s go ahead.” I thought that was a very reasonable interaction.
Tracy Alloway: When will Joe be able to use Fable?
Jack Clark: We’re trying. We’re working on it, we’re in discussions, and I hope we’ll have an answer soon.
The important thing to communicate is that these models aren’t special. They’re part of a general trend of increasing capabilities, and models from other companies will surely come along. At some point, these capabilities will diffuse, and we’ll work through that.
Joe Weisenthal: What’s your question for us?
Jack Clark: What do you think you’ll be covering about AI on Odd Lots in a year?
Joe Weisenthal: There are a few things I’m interested in. I’m very interested in emergent properties and whether AI will actually work on our behalf in the way people expect. I’m also interested in whether we’ll simply slam into compute and electricity bottlenecks that make all these other questions irrelevant.
And I’m very curious about the electricity analogy—whether legacy companies will actually be able to implement AI productively.
57:30-1:00:00
Tracy Alloway: This is a basic markets-reporter answer, but I’m very interested in valuations. I’m also interested in actual applicability. I want to see more companies plug AI into their existing systems.
Returning to your point about bureaucracy, I want to see some big companies actually implement this. I wonder whether, over the next year, we’ll get at least one example of it going very, very right.
Joe Weisenthal: One other thing: when the S-1s are no longer confidential, I’ll be curious about how for-profit, shareholder-owned companies—setting aside the public-benefit-corporation designation—balance profits with AI safety research.
There may also be a game-theory question about investments in safety within a hypercompetitive industry. I’m curious what the economist in you says about the prospect of anyone still caring about safety a year from now, when there’s so much money on the line in the race to build the best model.
Peter McCrory: For the questions you were asking earlier—under what conditions these models will do what you ask—much of commerce is built on trust. Prioritizing safe, aligned models that are also incredibly capable is a strong strategy for establishing that trust. So I don’t anticipate that concern—
Joe Weisenthal: For an individual company, there’s a game-theoretically optimal square in the matrix where it wants to be the trusted player. But is there a condition in which everyone chooses trust? Or does one company say, “We’re going to reach AGI first because we’re not going to spend a single dollar of our budget on safety”?
Peter McCrory: I haven’t mapped out the precise game-theory matrix—the two-by-two matrix and all the payoffs.
Joe Weisenthal: We hope it’s merely two-by-two.
Peter McCrory: There could be multiple equilibria, so the question becomes how you coordinate around which equilibrium you reach.
We talk a lot about creating a race to the top. We want to exhibit the kind of behavior that we believe is broadly beneficial to society. That’s what we do with the Anthropic Economic Index: We open-source much of the data and publish research for the world to use.
My sense is that this has been genuinely useful and broadly regarded as valuable. It’s one way we can encourage others to coordinate around good outcomes.
1:00:00-1:02:08
Joe Weisenthal: That we care about.
Jack Clark: I don’t think this is that big of a trade-off. Look at the automotive industry: you can buy really fast cars, and you can buy really safe cars. You can also buy really fast, safe cars. Tesla makes a lot of money from having basically the fastest, safest car. I think that eventually, in AI, you’re going to have some companies prioritizing safety—and safety translates into reliability, trust, serviceability, and performance. This happens elsewhere.
Joe Weisenthal: Peter and Jack, thank you so much for coming on Odd Lots. I’m glad we made it happen. These are interesting times, and I hope we can do it again sometime.
Jack Clark: Absolutely. Thanks very much for having us.
Joe Weisenthal: Thank you so much. Pleasure to be here, Tracy. That was a lot of fun. I genuinely enjoyed them, and I really appreciate both of them. Look, there are some weird futures that we can contemplate.
I think Jack’s Twitter bio says he’s interested in weird futures, or something like that. I appreciate that they played along with some of our weird-futures questions. It’s weird.
Tracy Alloway: It is such a surreal moment. Jack’s story about going on paternity leave and then returning to see the progress Anthropic had made in that span of time really stood out. If you miss a month of AI news flow now, it feels like you’ll be behind forever.
Joe Weisenthal: We’re recording this on June 17. Who knows what will happen by the time this episode comes out—hopefully in a day or two. But I felt that when we were in Hong Kong last week. We missed the first half of the debate because we were thinking about other things. Even over the course of a week, you really feel how quickly the news flow moves in this space. It’s almost like the way we were giving timestamps for the Iran war updates.
1:02:08-1:04:29
Tracy Alloway: Another thing that stands out to me is that Anthropic is producing all this information and clearly thinking about safety, but the handoff is still, to some extent, to policymakers when you consider the social or labor-market implications. You have to hope policymakers pick up the ball in the right way at some point.
I also thought Jack’s point about being safety-minded potentially becoming a differentiator was interesting, especially compared with some of the cheaper, more open-source models. You can see the appeal.
Joe Weisenthal: I understand that, but the question is whether a less safety-minded lab gets to advanced capabilities faster.
Peter McCrory: Yeah.
Joe Weisenthal: We would all love to drive the most capable, safest car. But the question is whether customers prioritize the most capable cutting-edge product. Imagine some car with an insane zero-to-60 time.
Tracy Alloway: Right.
Joe Weisenthal: Does the customer keep giving business to the firm that delivers the fastest zero-to-60 time if that company achieved it by allocating fewer resources to safety research? That’s a big question for me.
He also talked about how the company is going to see the alarming data first. I still wonder whether the people examining that data share the same view of what constitutes “alarming” as normal people do, especially given what we know about—
Tracy Alloway: Relative to the shrimp eaters.
Joe Weisenthal: Relative to us shrimp eaters and other regular people. Seriously, I think your question—“Are you hiring normal people?”—is pretty important.
I don’t have a ton of confidence in the political environment. And the fact that, if the research goes wrong, this technology could potentially be devastating to humanity—even setting aside jobs—is astonishing. This is not a normal technology.
1:04:29-1:06:00
Tracy Alloway: Our conception of AI goes back to *The Terminator* and the distinction between humans and machines.
Joe Weisenthal: From day one. And in response to your question, they see during the training process that AI models sometimes reason along the lines of, “I’m being observed right now, so I’m going to give this answer,” or attempt to blackmail someone. These behaviors aren’t very prevalent, but they sound like science fiction—and yet they actually happen.
Tracy Alloway: Yeah. On that note, shall we leave it there?
Joe Weisenthal: Let’s leave it there.
Tracy Alloway: This has been another episode of the Odd Lots podcast. I’m Tracy Alloway. You can follow me at @TracyAlloway.
Joe Weisenthal: And I’m Joe Weisenthal. You can follow me at @TheStalwart. You can follow our guests Jack Clark at @JackClarkSF and Peter McCrory at @PeterMcCrory.
Follow our producers Carmen Rodriguez at @CarmenArmanD, Dash Bennett at @DashBot, Cale Brooks at @CaleBrooks, and Kevin Lozano at @KevinLozano. For more Odd Lots content, go to Bloomberg.com/OddLots, where you’ll find our daily newsletter and all our episodes. You can chat about all these topics 24/7 in our Discord at Discord.gg/OddLots.
Tracy Alloway: If you enjoy Odd Lots—and if you like it when we do these AI episodes—please leave us a positive review on your favorite podcast platform. And remember: if you’re a Bloomberg subscriber, you can listen to all our episodes completely ad-free. Just find the Bloomberg channel on Apple Podcasts and follow the instructions there. Thanks for listening.