# This AI Expert Argues We're Underestimating AI's Impact — With Ethan Mollick Markdown copy of https://realalpha.app/v/Xoj0GGu0-aw/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy. Ethan Mollick argues that AI's economic gains depend on redesigning organizations to use capabilities that companies largely underuse today. - Show: Big Technology Podcast - Hosts: [Alex Kantrowitz](https://realalpha.app/people/alex-kantrowitz/) - Guest: [Ethan Mollick](https://realalpha.app/people/ethan-mollick/) is a Wharton professor and author of Coexistence: The Next Phase of AI. - Published: Oct 20, 2026 - Length: 1 h 4 m ## The best minutes - [13:21](https://realalpha.app/v/Xoj0GGu0-aw/?t=801) [AI gains require companies to redesign how teams work](https://realalpha.app/s/ai-gains-require-companies-to-redesign-how-teams-work/) (3 min) - [50:23](https://realalpha.app/v/Xoj0GGu0-aw/?t=3023) [AI helps students learn when it acts as a tutor](https://realalpha.app/s/ai-helps-students-learn-when-it-acts-as-a-tutor/) (3 min) - [41:08](https://realalpha.app/v/Xoj0GGu0-aw/?t=2468) [Personal AI agents trade convenience for data exposure](https://realalpha.app/s/personal-ai-agents-trade-convenience-for-data-exposure/) (3 min) ## Also worth watching - [7:23](https://realalpha.app/v/Xoj0GGu0-aw/?t=443) [Large companies report AI returns before full organizational change](https://realalpha.app/s/large-companies-report-ai-returns-before-full-organizational/) (5 min) - [23:59](https://realalpha.app/v/Xoj0GGu0-aw/?t=1439) [AI product formats are still in an era of ferment](https://realalpha.app/s/ai-product-formats-are-still-in-an-era-of-ferment/) (3 min) - [31:08](https://realalpha.app/v/Xoj0GGu0-aw/?t=1868) [Mollick sees no clear technical wall to steady AI gains](https://realalpha.app/s/mollick-sees-no-clear-technical-wall-to-steady-ai-gains/) (6 min) Skip the model-personality, AI-companionship and robotics discussion, and the stretch on religious uses and AI-generated cultural products. ## Full analysis: Ethan Mollick says a $200 million migration plan became $20 million - One company renegotiated a data migration plan from $200 million to $20 million. - Mollick’s survey figure is self-reported, while economy-wide gains still depend on redesigned work. - Track delivered outcomes, rethink team structure, and keep systems adaptable as AI changes. Its 6 sections, what it means for you and what to watch are on [the page](https://realalpha.app/v/Xoj0GGu0-aw/#analysis), free with an account: - [10:45](https://realalpha.app/v/Xoj0GGu0-aw/?t=645) Companies report returns before productivity spreads - [2:10](https://realalpha.app/v/Xoj0GGu0-aw/?t=130) Individual speed stalls at organizational handoffs - [14:18](https://realalpha.app/v/Xoj0GGu0-aw/?t=858) Teams can reset projects instead of automating old plans - [13:17](https://realalpha.app/v/Xoj0GGu0-aw/?t=797) Experiments matter only when firms learn from failure - [1:01:51](https://realalpha.app/v/Xoj0GGu0-aw/?t=3711) Model churn rewards reversible systems and careful budgets - [22:20](https://realalpha.app/v/Xoj0GGu0-aw/?t=1340) Human benefit depends on what firms choose to automate ## Summary Ethan Mollick is a professor at the Wharton School and author of Coexistence: The Next Phase of AI. He writes the One Useful Thing Substack. Mollick argues that AI already delivers returns inside companies, but individual productivity gains have not yet become broad economic growth because organizations have not redesigned how work gets done. He says firms need to experiment with roles and processes, while warning that automation could displace workers unless companies direct AI toward augmentation and broader human benefit. He sees current capabilities as underused and expects continued technical progress, while acknowledging limits in power and compute and uncertainty about jobs. The discussion also examines personal agents and privacy, robotics, education, religion, falling token costs, and the risk of locking into today's AI systems. ### Minute by minute - [0:52](https://realalpha.app/v/Xoj0GGu0-aw/?t=52) **AI gains stall at organizational boundaries.** Alex Kantrowitz cites productivity data showing that high-AI-exposure jobs have gained less than during some pre-COVID and COVID periods. Mollick says individual gains do not automatically raise organizational productivity: a coder producing 100 times more code can create a coordination burden when teams still rely on existing sprints, meetings and workflows. - [6:02](https://realalpha.app/v/Xoj0GGu0-aw/?t=362) **Companies report ROI before transformation.** Mollick says executives report returns from AI, but large companies are still working out how to spread gains beyond individual employees. He cites a Wharton survey in which 85% of large companies with over a billion dollars in revenue reported positive ROI last November, and points to a Fortune 100 company using AI to find internal problems and a financial-services company launching new product lines. He calls a widely cited MIT study of failed pilots weakly designed and says experimentation needs to include failures. - [13:21](https://realalpha.app/v/Xoj0GGu0-aw/?t=801) **AI requires structural changes to work.** Mollick says companies need to rethink teams and projects, not just give employees AI tools. He describes one data-migration plan that fell from $200 million to $20 million, with the CTO considering whether AI could bring it down further. Some firms are making employees builders; others are assembling small teams of experienced people to tackle difficult projects and learn from attempts that fail. - [16:44](https://realalpha.app/v/Xoj0GGu0-aw/?t=1004) **AI is changing traditional industries.** Mollick says AI is affecting manufacturing, agriculture, construction, retail and financial services, not only software work. He describes agricultural firms using cameras to spot problems in fields and AI helping with construction-site decisions. He also points to Walmart associates using handheld computers to get answers and to work on migrating legacy COBOL systems. - [21:55](https://realalpha.app/v/Xoj0GGu0-aw/?t=1315) **AI's benefits depend on deployment.** Mollick says human flourishing is not automatic: companies can use AI to automate call centers and management or to augment work. Potential benefits include less drudgery, faster scientific progress and wider access to education and healthcare. He argues that current models are already underused, and describes an unsettled period of product experimentation in which vibe coding and personal agents serve different purposes. - [26:43](https://realalpha.app/v/Xoj0GGu0-aw/?t=1603) **AI's effects on work remain uncertain.** Mollick says that even if AI progress stopped, organizations could spend years using current capabilities more fully and changing how they work. He does not expect the most extreme GDP-growth predictions, but expects disruptive change. He says it is too early to know which jobs will be replaced or augmented, or where the value will go. - [31:08](https://realalpha.app/v/Xoj0GGu0-aw/?t=1868) **AI progress remains hard to forecast.** Mollick distinguishes faster-than-expected market demand from whether AI labs were surprised by technical progress. He says forecasts have repeatedly missed the continued pace of development, including after reasoning systems helped overcome apparent limits. He sees no specific technical barrier to steady gains, though he says power and compute will constrain them. - [36:41](https://realalpha.app/v/Xoj0GGu0-aw/?t=2201) **Reasoning and tools expand AI's reach.** Mollick explains that reasoning models generate intermediate work before answering, which he says improves results in areas such as math and science. He describes three pieces working together: the model, an application, and harnesses that let AI search the web, run code or control a computer. He cites a math problem tackled by 10,000 agents exchanging 8 million messages as an example of coordination layered onto those capabilities. - [40:50](https://realalpha.app/v/Xoj0GGu0-aw/?t=2450) **Personal agents raise privacy questions.** Mollick describes personal agents that connect to email and other personal data and act proactively; he says stronger models reduce the need for elaborate retrieval systems. He warns that prompt injection through email or data exposure could still cause harm. He says some feared mass effects from AI companions and deepfakes have not appeared so far, while cyber risks remain. - [47:41](https://realalpha.app/v/Xoj0GGu0-aw/?t=2861) **Robotics may be nearing a turning point.** Mollick says robotics experts he has spoken with have become more open to the idea that language models could unlock more general-purpose robotic work. He points to early demonstrations using cameras and robots, but says he is not a robotics expert and cannot judge whether the apparent shift will hold. - [49:37](https://realalpha.app/v/Xoj0GGu0-aw/?t=2977) **AI helps learning only when students engage.** Mollick says using AI to bypass schoolwork undermines learning, while using it as a tutor has shown positive effects in studies he cites. He recommends tutoring outside class, active learning and in-class assessments without AI, including a return to blue-book exams. He says AI-generated proofs and AI-assisted research may accelerate science while creating new challenges for evaluating and absorbing the output. - [54:29](https://realalpha.app/v/Xoj0GGu0-aw/?t=3269) **AI complicates trust in cultural work.** Mollick says people use AI for intimate tasks, including personal advice, stories and sermons, but communities have not negotiated the right boundaries. He distinguishes AI-assisted research or editing from a sermon written entirely by AI. He also expects people to push back against low-quality AI cultural work, citing a spa sign that included an image of a sign in front of a spa. - [57:40](https://realalpha.app/v/Xoj0GGu0-aw/?t=3460) **AI reshapes work and economic friction.** Mollick says job effects remain uncertain: coding work is shifting toward managing AI-written code, and he argues that societies should prepare safety nets for disruption. He expects token prices to keep falling even as agentic work increases usage. Personal agents could remove friction from refunds, subscriptions and moving money, potentially straining systems built around consumer inaction. He warns companies against locking into solutions that today's models may quickly outdate. ## Transcript Big Technology Podcast, hosted by Alex Kantrowitz, with Ethan Mollick. Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there. [0:00](https://realalpha.app/v/Xoj0GGu0-aw/?t=0) If AI is such powerful technology, where are all the productivity gains? Let's talk about it with Ethan Malik, the author of Coexistence and a Wharton professor that we quote all the time on this show. That's coming up right after this. This episode is brought to you Welcome to Big Technology Podcast, a show for cool-headed and nuanced conversation of the tech world and beyond. We have a great show for you today. We're joined today by Wharton professor Ethan Mollick, the author of One Useful Thing on Substack and his great new book, Coexistence, which I highly recommend. And today we're going to talk all about how AI's improvements on the technology side can actually lead to productivity gains across the economy. [0:40](https://realalpha.app/v/Xoj0GGu0-aw/?t=40) We'll also go through whether the technology can keep progressing, whether it will hit a wall, and a number of other topics. I'm thrilled to have Ethan here because we quote you all the time on the show. I've been reading you for a long time. You've been on the show before. Your book is great, and we have so much to talk about. So thank you for being here. Thanks for having me. um all right let's start with ai and productivity you know ai is a very powerful technology we've all seen it if we've tried it um the question is what that power leads to and the the um Conventional wisdom is if AI is actually that useful, we should see the impact in productivity stats. We're not seeing it. This is from the Dallas Fed. US labor productivity has grown at an annualized rate of 2.4% since the beginning of 2004. It's not a huge percent. It's not revolutionary. People like Dario have said we might see 10% increases in GDP. We haven't seen that. And the most astonishing thing to me in this Fed study is that they charted out high AI exposure and low AI exposure jobs. And now high AI exposure jobs are seeing more productivity gains than the low AI exposure jobs, but they're actually seeing less productivity gains than [1:55](https://realalpha.app/v/Xoj0GGu0-aw/?t=115) they saw during COVID and even less productivity gains than they saw pre-COVID at some times. The spikes just do not reach the same heights. What's going on, Ethan? I think anyone who's in an organization probably has the right answer to that problem, which is you don't get organizational productivity gains instantaneously from individual productivity gains. You can produce 100 times more code than you did before. And that's great if you're a solo entrepreneur producing a lot of code. If you're in an organization that's doing two-week long sprints and has to coordinate with marketing and has stand-up meetings every day, which is a rational way to organize work, that 100 times productivity gain actually can be a barrier. [2:34](https://realalpha.app/v/Xoj0GGu0-aw/?t=154) right no one has time to deal with all that code and we see that too in shipping stats 200 you know 20 times more code being produced uh 30 more products being shipped right and so it's not an uncommon problem we haven't adapted organizations to ai yet yeah i mean you there's there have been some stats i think you're referencing this um in the app store where there have been many more apps shipped than previously but uh even fewer reviews than before right so it's like um the productivity quote-unquote or the the um work looks like there's been more done uh but when it actually comes to like usage or value it's so far not showing up well actually those are two different things so one thing is the flood of stuff right and we're seeing that everywhere so there's a lot more novels a lot more games a lot more apps but i'm actually talking directly there's a nice set of studies looking at products shipped right and And we're seeing a huge increase in the amount of coding, smaller increases in the number of products shipped. [3:32](https://realalpha.app/v/Xoj0GGu0-aw/?t=212) There's still an increase, but it's less, right, by an order of magnitude. And the reason for that is, again, we don't have multiplayer AI. AI is a very single-player experience right now, where you get individual performance, but. organizations are about many people performing at once yeah i've i've thought when it comes to these capabilities that the companies that are really going to at least in the near term pull ahead are going to be the ones that say individuals who are using this technology um we're going to empower you to actually go ahead and you know break free from some of the um standard red tape that would hold you back in a typical um you know corporation like um i think in most companies If somebody is, it's interesting what happens. Someone comes in after they're hired, they're given a mandate. They tend to get beaten down if they have a lot of new ideas, if they want to change things. At a certain point, companies are like, no, no, you're actually not here to have ideas. You're here to sort of play this role. [4:26](https://realalpha.app/v/Xoj0GGu0-aw/?t=266) And so I think we might end up seeing a divergence in terms of competitiveness among companies that enable people to actually run with their ideas and then companies that follow this traditional blueprint of like, oh, you have ideas, you want to get things done? Just, you know, do your job. Right. I mean, in the book, I talk about leadership lab and crowd being three things you need aligned. You need a leadership team that actually buys into this and starts making organizational change. You need a crowd. You need people who are using and inventing ideas. And that's often the line employees are talking about. [4:55](https://realalpha.app/v/Xoj0GGu0-aw/?t=295) And then you need a lab. You need to actually be doing research on how to take this stuff and spread it out among everyone. I think the only thing I'd push back on a little bit there is we still have the multiplayer problem. Like if you're still, if you're a hundred times more productive coder, we still need to figure out how do we give you the roadmap to follow in. We still need marketers to work with you to figure out how you get the material shipped out. We still need to figure out how to maintain the software. So it's a little tough to just say, if, you know, huge productivity gains alone from individual players, even if they're given the freedom to explore, or we still need to turn that into process? [5:26](https://realalpha.app/v/Xoj0GGu0-aw/?t=326) No, but this is not year one after ChatGPT. Like if we were talking in October or November 2023, I would say, yeah, that makes sense. Companies have had four years to adjust to this. Yeah, the capabilities are getting better, but they've had a long time to adjust to this. And there is an urgency for them to figure it out soon because the magnitude of the bets on building out the infrastructure and the products and the investments in the companies that money eventually has to be returned, or the investment has to be returned to a degree. And if you don't have productivity gains, by definition, it seems like you're not going to get their return on the money. We were talking about top level productivity gains in the US economy. I talked to lots of CEOs of very large companies. There is almost nobody at this point in the data from my colleagues at Wharton back this up who isn't seeing some ROI from AI. This is not an issue anymore that people aren't seeing ROI. What they are trying to do is figure out how to move from individual productivity gains and $20 million here, $30 million here, which says a lot to us, but it's not a lot to a large company, to how do we transform our organization around this, right? So I don't [6:35](https://realalpha.app/v/Xoj0GGu0-aw/?t=395) think it's a question anymore of are you getting gains? I don't hear that from C-level people. I did a year and a half ago or a year ago. That's no longer the question. The question is, what do we do now? So we have 100,000 employees. I think organizational change doesn't happen that quickly. I mean, ChatGPT came out in 2022, GPT 3.5. And companies just started adopting this in the last couple of years. There's no playbook. Nobody knows what to do in this case. All the traditional methods of the consultancies that would help you, the enterprise software companies that would build in. I mean, it is a revolutionary change. So I think that organizations are absolutely seeing the ROI. I haven't heard of a single organization saying that AI isn't paying off at a large scale right now. But they're not getting the transformation yet. But that requires management to make decisions about what's going on. And they've barely begun to do that. Many of them aren't even still using advanced systems. But OK, they are, while some might say they're getting this ROI, they're also telling their employees to pull back on the spend. And this era of, you know, we're going to spend tokens freely, you know, that's ending. And so, yes, there's [7:41](https://realalpha.app/v/Xoj0GGu0-aw/?t=461) some ROI, but what their actions suggest is there's limited ROI right now. I will disagree with you again on this. So first of all, we keep whipsawing between different views, right? Like token maxing was big this spring, and then it was like, oh, no, everyone's spending too much on tokens. It is becoming a fast-growing expense. Companies are worried about this. When I talk to a lot of Fortune 50 companies, though, they are viewing this as something that is an R&D expense to try and get under control to figure out what's going on. A lot of them don't know in the same organizational change ever, like I was talking to a very large company the other day, and they're saying some people are 100 times more productive, and some people aren't really getting any benefit out of AI. We don't know yet how to spread the 100 times more productivity to more people. So I think that there's news articles about people pulling out. Now, most of the people who are talking very publicly about how they're cutting AI spend tend to be people with competing AI products. So like the fact that Microsoft, Meta, and Palantir are telling you we're spending too much money on Claude, I don't know if that tells you that much about the world, right? [8:40](https://realalpha.app/v/Xoj0GGu0-aw/?t=520) Our revenues are still growing for these companies. Absolutely nobody has cost under control, use under control, but I don't see any sign of people saying they're pulling back because AI turned out to be useless. Not useless, it's just the question is how much return are you getting on the investment? I mean, companies are getting large returns that we're talking to, but they're also in experimentation and evolution phase, right? You can't get returns without making changes. Most organizations I know know that. Taking changes takes time, right? And they need to figure out how to swallow this new technology. [9:08](https://realalpha.app/v/Xoj0GGu0-aw/?t=548) So I think that there's too much of a sort of back and forth, like AI is changing everything, AI is changing nothing. When I look inside organizations, right, C-level people are deeply spending a lot of money and a lot of time trying to figure out how to make things operate well. I'm seeing transformation in lots of organizations, including really old school biopharma companies and financial service companies, very large ones, who don't want to talk about exactly what they're doing yet because it's a source of competitive advantage. And I think people are worried about token costs, but that's often a management problem. [9:38](https://realalpha.app/v/Xoj0GGu0-aw/?t=578) Like if you don't make any decisions about how tokens should be used, right? Who should be using them? What process should be using them? Then, and you incentivize people to use as many tokens as possible, they'll use as many tokens as possible. It feels like, I feel like we're kind of in this chicken and egg where we're like, you know, where is the return? I keep telling you it's there because companies keep telling me it's there. And the surveys show 85% of companies have positive ROI from AI as of last year. So I think though the major transformation hasn't happened yet, except it's starting to in some organizations. [10:09](https://realalpha.app/v/Xoj0GGu0-aw/?t=609) But again, you have 100,000 employees. How are you supposed to even think about mapping out all the processes in your company? How do you start deciding about who gets to use tokens under what circumstances? How do we start to adopt the genetic models when they're basically less than a year old? There's a lot of stuff happening all at once. Yeah, okay, so let's go bit by bit here. So first of all, you've said that the companies are already starting to see ROI. Where are they seeing that ROI? Where are the areas that they're seeing it? So in terms of, so productivity and cost savings are already there. [10:38](https://realalpha.app/v/Xoj0GGu0-aw/?t=638) I've actually been worried that companies are pushing too much and productivity is the main goal. So my colleagues at Wharton have been surveying a thousand large companies. And like I said, last November, a year ago, 85% of large companies with over a billion dollars in revenue reported positive ROI from their AI use, right? We don't have detailed breakdowns on every piece of whether it's cost savings or spend, right? but you must have some anecdotes of like a company put uh you know ai into action and yeah i mean sure but but so tell us tell us the stories i mean a lot of them i can't even talk about necessarily i mean so i talked to uh you know there's a another A Fortune 100 company that gave its senior leadership team access to basically open clause with access to the full company data and completely changed how management works. [11:23](https://realalpha.app/v/Xoj0GGu0-aw/?t=683) **Ethan Mollick:** Like C-level teams are using this as a super analyst to find problems inside the company they weren't able to identify. [11:29](https://realalpha.app/v/Xoj0GGu0-aw/?t=689) I have spoken to a large financial services company that has launched a whole bunch of new product lines that they never had before as a result of AI. Talked to a couple of very large hedge funds, the PE funds, that are seeing huge returns to management technology right now. And all the companies are growing faster inside their portfolio. Like lots and lots of cases like this, right? Oftentimes, they don't want to talk about it in huge detail. And they're all still in the experimental phase. I mean, a lot of these are outside of pilots now. They are rolling out at the organizational level. But it's tough because it's a complicated kind of set of choices at the company level rather than a product you adopt and get an instant return from. [12:07](https://realalpha.app/v/Xoj0GGu0-aw/?t=727) You know, a stat that we used to quote often on the show was that it was like one out of every eight or one out of every 20, something in a large, a large number of AI pilots were out there, but only one out of every, you know, one, eight to 20 or so actually made it into production. Have you seen the numbers sort of shrink in terms of the pilots that it takes to get to production? So two things. That is a famously bad study. Let's hear it. So that's the MIT Nanda study from last year. and um at the time it was it's somebody went out and just interviewed people at a conference and coded that himself no that was that was 85 percent of companies were like no that was 95 percent of pilots yes 90 right that was the that was the famous sort of non-design you've actually been vocal about how the study hasn't been good but there have been other reports that have come out that said you know it's not that maybe that was the one in 20 but the the more reasonable reports say like one in eight or something i i would hope that's the case that most pilots fail again the one thing i'll keep emphasizing again is i i you know i don't take money from the ai labs i don't really care whether they succeed or fail individually the you know the transformation [13:13](https://realalpha.app/v/Xoj0GGu0-aw/?t=793) that's happening in companies is pretty inevitable when i talk to organizations at this point and i think they have to be doing experiments that fail because if they don't do that how are they going to learn this there's no playbook so i'm hoping for high failure rates otherwise you're not engaging enough experimentation yeah all right let's let's now go to the next step which is what happens uh what has to happen for us to see that widespread um change within companies where we get those big productivity gains so it's a two-parter for you one is how you know, the type of change that unlocks the full potential. What does it look like? And what do we end up seeing if companies make the moves that you suggest? [13:54](https://realalpha.app/v/Xoj0GGu0-aw/?t=834) So there's no one move, right? It's about thinking about structural change, which takes years for a lot of organizations to do. And so examples of things that I'm seeing on a sort of structural level, very common situation is pulling apart organizational teams that exist too. So called disbursement, you pull out your best senior engineer, put them a subject matter expert and a marketing or strategy person, and then have them do like impossible projects. Like I just spoke to a company that went from a $200 million data migration plan for their old, you know, old mainframe data to they renegotiate to a $20 million plan just a few months ago. And the CTO is thinking like, well, maybe I can do this for $2 million in a couple months, like real noticeable change, but it involves literally pulling apart teams and thinking about what do we want to do in recessing projects from the beginning so how does a t how does a team typically change so because because individuals are now starting one of the other things we're seeing companies is a dissolving of company boundaries right so coders are doing design work designers are doing coding work product managers are doing everything so part of what's going on is how do [14:58](https://realalpha.app/v/Xoj0GGu0-aw/?t=898) you reassign roles that way some companies are assigning everyone the role of builder right and that's that's what they're doing with with ai um but other organizations i see are trying to think about how do we take teams of highly talented people these people are already getting returns from ai put them together give them really hard challenges to do high token budgets and seeing how far they get and if they fail which happens some of the time then how do we learn something from this formally rather than just burning tokens and sending them off into the air is that so different from the way that companies function today yes i mean most organizations do not have are not doing far exploration work and especially in organizational design and form. So the 40% of the advantage of American companies in the 20th century came from better management according to paper by Rafael Sadoon and other people from Harvard. And we used to do a lot of exploration of organizational form, right? Companies were experimenting all the time with how to organize themselves, how to get work done. I mean, a large part of the story of the 20th century was the story of American manufacturing and firm organization transforming. That has not happened [16:01](https://realalpha.app/v/Xoj0GGu0-aw/?t=961) as much in the last 30 years. People use outside vendors, like, you know, the SaaS vendors will tell you how to run part of your business, use consultants, and your R&D is focused on your customers or outcome. So, a lot of organizations have very little muscle memory left of organizational change. So it is actually pretty rare to see organizations play with fundamentally re-dividing how work operates, how roles operate, what mandate people have to get work done. I mean, that doesn't feel like normal operation of business to me. Okay, so I want to ask you, and by the way, so I know where you stand because obviously you wrote it down in a book. [16:33](https://realalpha.app/v/Xoj0GGu0-aw/?t=993) I'm just trying to sort of talk through your arguments. So for those who are like, you know, speaking about me, this guy, you know, I'm just trying. No, it's fine. There's nothing wrong. I am excited to find. I should point out, we don't have all the answers, right? But also, we don't have all the answers sounds like an academic result. Nobody knows what's going on is a better way to phrase this, right? I think the thing that I'm reacting to a little bit is this sort of, I don't think skepticism about AI's ability to create transformation is warranted at this point. I think that's a dead end. I think anxiety about how we shape this in the right direction is warranted. And I think how do companies start to think about this in a way that helps large companies start to transform is a very reasonable thing to worry about and why we haven't seen those results. I spent a bunch of time in the book talking about why haven't we seen transformation happen. So I'm very much on your page and I'm not going to defend that like, oh no, AI has changed everything today. Part of the interesting puzzle of the book and why we're in the phase of coexistence, right, in some extent is we are trying to negotiate that right now what does the world look like so we get productivity [17:41](https://realalpha.app/v/Xoj0GGu0-aw/?t=1061) gains and good results for society okay great so then i'll keep pressing so thank you for the permission to do so i mean um i i get it if you're talking about industries like consulting software but i wonder i don't even want to use the term real economy to to describe what i'm about to describe but like there are parts of the economy where um shifting the organization in a way that you you talked about adopting ai in the way that you talked about um i i curious where you end up what you get at the end of the line what the output is i'm talking about like let's say like frito-lay adopted ai like crazy like do we just get like a new flavor of chips or or a car company adopted ai like crazy do we get like a new vehicle or a small hotel chain um do they build a new hotel like when it comes to these type of businesses which i think are the heartbeat of the economy to some degree i mean services are the heartbeat of the economy right now yeah but but if you want to manufacture i mean i literally talked to some of the companies in that exact range that you're talking to, right? [18:45](https://realalpha.app/v/Xoj0GGu0-aw/?t=1125) And again, I can't name names, not because I'm trying to be cagey, but because I don't know who I have permission to talk about or not. But, you know, I mean, there's an endorsement in the back of my book from the CEO, the just stepped down CEO of Walmart, to give you a sense of the kinds of companies I talk to, which I think most people would consider real economist stuff. And first of all, there is a huge amount of change in these organizations. Some of this, by the way, is R&D work, right? Like, literally, I know the equivalent of the better kind of Frito-Lay chip, But that is an expensive process inside companies that make new products and speeding up new project development is actually a big deal to them in a big way, right? [19:17](https://realalpha.app/v/Xoj0GGu0-aw/?t=1157) And finding new trends. I've been talking to large scale manufacturing agribusinesses that are thinking about things like how do we use the fact that AI can now look at cameras to figure out new things about which fields are in trouble and which ones aren't. We manage construction projects the same way we have since the 1940s. What does it mean when you have an AI that can give people advice about what to do differently and can actually give information about what's happening on the site or not. You have really sets, really interesting sets of changes happening everywhere, right? On the hotel side, like, or even in the Walmart store, they've been pretty open about this set of stuff. [19:49](https://realalpha.app/v/Xoj0GGu0-aw/?t=1189) How do you give advice to everybody, to a senior store manager who needs help figuring out what to do at the store when there's trouble? Like management is a real thing here and there's pretty transformational things. That's aside from AI for planning and AI for central management and AI for HR. So, I mean, I actually think, you know, both services are the real economy, right? And it's a big deal of coding. I think one out of every 16 jobs is management in the US, and coding follows soon afterwards. These are big jobs that represent a lot of the economy. But also, I think the whole point of this is general purpose technology with widespread effects. I mean, take even the most narrow thing, right? I was already talking about migrating databases over at large financial services companies and mid-sized banks that have COBOL still running internally. These have literally been the projects that have been the backbone of a large part of IT for a long time. Those are being transformed overnight. So I don't think it's like, I think you see this everywhere. This is not just like coders and scientists and financial professionals. Maybe you've touched this project since you named Walmart, but [20:57](https://realalpha.app/v/Xoj0GGu0-aw/?t=1257) I'll back you up here. I think one of the more interesting things within Walmart is associates now have like a handheld computer that they can instead of having to go like to managers or stuff they can ask questions to and it can direct them places yeah I mean and that's sort of almost on the floor Exactly. So that's one of many outcomes Walmart has had from being pretty enthusiastic about AI adoption. And they've only been doing it for a couple of years, right? And it's a few million employees. It's not an easy thing to transform an organization that quickly. So my flip side to you is on one hand, is it slow? Yes. Is this faster than digitization, cloud, e-commerce, any of the other things that have transformed American business the last 20, 30 years? Certainly. [21:39](https://realalpha.app/v/Xoj0GGu0-aw/?t=1299) Yeah. I wonder, people talk about AI as leading to this new age of flourishing. And I guess this is probably overly skeptical to ask this question, but does stocking a shelf faster and producing a new potato chip lead to that new age of flourishing? Is that an overpromise? I kind of feel like you're alternating back and forth between like, let's focus on the narrow use cases like what happens at Frito-Lay and let's not talk about biopharma. And then let's like, okay, let's talk about biopharma. Why is Frito-Lay matter? The whole point of this is widespread effects, right? So first of all, I don't think human flourishing is a given outcome from this. [22:20](https://realalpha.app/v/Xoj0GGu0-aw/?t=1340) And not just because of existential risk, but also because there is automation pressure in all these organizations. Like we haven't talked about like the easy way to use AI is automate your call center. Automate your management, automate, like, that's the danger is that we just automate everything rather than augmenting human work. So I actually think that for me, the concern partially is that there's danger that we don't pursue human flourishing as an outcome without even any of the big AI pictures. But also, there is, you know, the way you get human flourishing, right, is we free people from drudgery, we increase the rate of scientific advancement, we increase the rate at which people, everyday people get access to better education, better healthcare, better financial advice. The pieces are there to make that happen. [22:58](https://realalpha.app/v/Xoj0GGu0-aw/?t=1378) we haven't necessarily seen the steps needed to make it all occur for everybody yet. Something interesting that you've pointed out recently is that just there's a large set of capabilities that people are just not taking advantage of. And I think that's sort of also key to this conversation. You wrote recently in your Substack, while there is a lot of debate over what future models will do, the current capabilities of existing models are barely used and are often not even well understood. I mean, that was really interesting to me, this idea that the current capabilities are not used barely at all yeah i mean i call this the overhang right and there's a huge gap between current capabilities and future capabilities we don't need better models for complete transformation of a lot of how we work in the next 10 years right we need that adoption curve we started talking about at the beginning How we get up that curve. And so, in my mind, that means that we are lacking the imagination for the transformation piece, which is the point I keep making to you. There's no playbook yet. It's not an automated process. We need to figure this stuff out. But the piece is already there. And it's sort of inevitable at this point that transformation will happen. So, it's about guiding in a good direction. [24:03](https://realalpha.app/v/Xoj0GGu0-aw/?t=1443) is one of the issues here that um ai can't seem to settle on a product or a product format um so i like wrote down some of the things that we've been seeing even over the past like two years so we've had the chatbot we had vibe coding voice bots co-work agents and then we have applications in consumer and enterprise it does seem like the ai industry comes up with like a new like for instance let's just take the the sort of um vibe coding thing right so there was this idea a couple of months ago that everyone was just going to vibe code their own software type in a prompt makes your software and then you know you that that's the impact of ai but very quickly moved towards agents thing that things that instead of you coding up software they just take control of your computer and so the interface might fall away to a degree um so what do you think about the fact that the ai industry i don't want to say grasping but there's been just like there's whiplash about all the different product formats that it's throwing out there so i think [24:59](https://realalpha.app/v/Xoj0GGu0-aw/?t=1499) **Ethan Mollick:** there's three different things going on the first is that we are what we call in the study of innovation an era of ferment nobody knows the right form for these things to be in so you're [25:08](https://realalpha.app/v/Xoj0GGu0-aw/?t=1508) going to see all sorts of ideas churning up and by the way being copied right so open claw came out in january it's the idea of personal agents and what's happened is suddenly that has taken root as a form so now we have muse and dot and other things that are basically claw likes right so we're seeing if era of ferment is the first thing the second thing is i think some of this is category differences, right? There's a difference in category between vibe coding and agents. Vibe coding is a human activity. Agents are a capability of AI systems. And they're not like competing product lines, right? [25:40](https://realalpha.app/v/Xoj0GGu0-aw/?t=1540) There's a lot of debate over how people use AI. I don't think vibe coding in the companies I've talked to was ever viewed as the way they're going to be using AI systems. It was, you know, it's a... And by the way, vibe coding has taken off, right? I vibe code a lot of things that would have been required me to have a coder involved in this before. People are building games, but people are also building very serious projects all the time. So I think it hasn't gone away. It's just there's many things happening at once. And then the last thing I would say, and the most important, is part of what you're tracing is also capability curves, right? [26:10](https://realalpha.app/v/Xoj0GGu0-aw/?t=1570) My last book, Co-Intelligence, was really... about the chatbot era, because that's as good as models were. This book is about the agentic era in a lot of ways, because that's where models are today. We agents, practical agents were not possible until we reached Opus 4.6 level or 4.5 level about November, December of last year. Like that was when agents became practical. And so what you're partially seeing is views as we move up this exponential ability curve, it changes the kinds of products that are possible. So all that stuff is happening at once. We're trying to figure out which products exist in a world where nobody knows what's happening. [26:42](https://realalpha.app/v/Xoj0GGu0-aw/?t=1602) You are seeing multiple competing different groups all trying to claim the idea of AI products, and you're actually seeing real capability gain. Let's say AI progress stopped today, right? So no further progress, which is not going to happen, but let's just say this as a thought exercise. If companies and people were to take full advantage of the capabilities that exist today and organizations were to reorient themselves in a way that they don't stand in the way of the progress that AI can allow? What does the world look like? If progress stops today and we could continue, and I mean, I think we're back to that exact same beginning part, which is the capability overhang is quite large. So what's going to happen over the next five or 10 years, we will fill in the capability overhang. Organizations will transform the remaining technologies that need to make this matter. We'll make a difference. We will figure out how to use this in ways that the jaggedness of AI and the jaggedness of people coexist together. [27:40](https://realalpha.app/v/Xoj0GGu0-aw/?t=1660) and there'll be large-scale socioeconomic effects from all of that productivity gains i don't think we're going to see some of the insane numbers predicted of 10 growth but you know who knows um that seems outside of any expectation i would have um but but we will see transformation happen one way or another we don't need new models but we're getting new models regardless that's coming yeah i mean it's kind of hard to visualize but like when you walk down the street let's say again let's say all this stuff happens the way that you anticipate. Do the streets look the same? Do companies look the same? Do people's day-to-day look the same as they do today? What changes? [28:17](https://realalpha.app/v/Xoj0GGu0-aw/?t=1697) So in the next couple of years, I think change will be... It's kind of an interesting thing to have these conversations because any coders who are listening to this, their job is radically different than it was two years ago. It's just not the same job at all if they have any access to these tools. The best coders I know are not writing code anymore. And so you're saying like, oh nothing's changed it's like everything about their job has changed I mean but they're still going to walk down the same streets they're still going into office buildings no I mean like when I say nothing's changed I mean like the fundamentals of like of our society today are similar to they were before chat gpt i mean do they does that change at all in what time period i mean i think that in yeah when this stuff when it gets when it gets to maturity like when the capabilities that we have today are taken advantage of fully i mean i i so part of what's interesting here right is like it's if we don't change if the technology doesn't develop any further. I'm kind of at a point of like, I don't know how to prognosticate about this. I think people will still be people and we'll be living in cities and walking down streets. We're now in hypothetical land that I can't really answer, [29:26](https://realalpha.app/v/Xoj0GGu0-aw/?t=1766) which is like, do practical robots exist in this world? Well, there's some good signs that we're all, or bad signs, depending on how you feel about robots, that general purpose models are actually quite good at robotics. So maybe in the next five years, we finally see robots emerge. Suddenly that reshapes the physical world. Otherwise, are you more likely to have cyber attacks in your bank? Yes. Will that change how life on the streets is? I mean, people are people, right? And I think it's easy to overstate whipsaw technical change affecting people's lives. I mean, we went from, if you look back over the last 60 years or so, the built environment hasn't changed that much. [30:00](https://realalpha.app/v/Xoj0GGu0-aw/?t=1800) So I think it's reasonable to expect that unless we get robots doing work for us, you'll still walk down a city street. But I think that's not really the concern. The concern is like, what do people do for a living? How, you know, who is making returns from this kind of technology? What's happening inside the office building is not the office buildings themselves. Okay, so what's your answer to that? So my answer to that is that we are, that technological revolutions result in disruptive change. And you already see that happening. Coding is not the same job. We are in debate right now in economics about whether or not there'll be more or less coders. [30:33](https://realalpha.app/v/Xoj0GGu0-aw/?t=1833) We just don't know, right? A lot of things depend on government policy that happens, how companies adopt this technology over time, and we just don't have the answer to these questions. I can promise disruptive change, right? That's the only thing I can do. I cannot predict, because we just don't have the data yet, about whose job is replaced, whose job is augmented, where the value goes in the value chain here. I mean, I wish I've had a crystal ball. First of all, I'd be making investments based on my crystal ball. And second of all, this is not what you do in a period of disruption, right? like you can't be like this is literally one of the most impossible things to predict is exactly who wins and who loses under what circumstances i can make stuff up but it's made up stories yeah okay well i still wanted to ask Let's talk about AI's progress. AI's progress has, in many ways, surprised the leaders of these AI labs with how good it's gotten in the period of time that we've dealt with. Why? [31:28](https://realalpha.app/v/Xoj0GGu0-aw/?t=1888) Has it? Yeah. I mean, all right, let's take an example. Here's a good example for you with Anthropic. So Dario was at the Dealable Conference a couple of years ago, and he was talking about like if he buys you know if he overextends on his compute purchases then um his company could go bankrupt if the demand is not there uh it turns out that anthropic is compute constrained now he should have bought more so that's what i'm talking about like the capabilities there were even um so another here's another example ilia sitzkevath former chief scientist of open ai said very clearly a few years ago that the technology was about to hit a wall data wall and what happened like the the capabilities have just exploded ever since so there has been a degree of surprise there so so not to be argumentative i think that the ai companies have their market has exploded best than they thought right which is different that like demand has been higher than i think they anticipated right and they're all compute constrained at this point um i think ilia's point was not that there was no forward progress He argued that the era of pre-training had generally ended. [32:34](https://realalpha.app/v/Xoj0GGu0-aw/?t=1954) That's correct. And for a while, he was right. But what's happened is just like in every other kind of case, there have been technological innovations that have allowed LLMs or LLM hybrid systems to continue to develop as things change. I think the more fundamental question is, were the AI companies, are the leaders actually surprised by the pace of actual development of the technology? I haven't actually seen a lot of evidence of that. Like they seem to have been pretty spot on, actually shockingly so, about the continued development and doubling of the technological ability. So I think you could say the market's grown faster than they expected, right? [33:08](https://realalpha.app/v/Xoj0GGu0-aw/?t=1988) But I don't think that they've, I think the technology is surprisingly to me, right? I thought we would have hit some sort of slow, I mean, I thought slowdowns were a real possibility a few years ago. I no longer think they are. They've overcome a lot of technological barriers, but I don't know if they've, they don't seem surprised by that. They seem on target for where they think they're going. i think reasoning is something that they weren't expecting to work as well as it has i mean you take a look at and reasoning has been core to these uh math breakthroughs right so you take a look at um the the predictions of when people expected uh ai to be as good good at math as it is five, 10 years from now, but it's already solving millennium problems. [33:49](https://realalpha.app/v/Xoj0GGu0-aw/?t=2029) Yeah. That's something I talk about in the book, but those are predictors. So again, we have a little bit of a confusion because you were sort of saying the heads of the AI labs. I don't think they're surprised. I think super forecasters keep getting this wrong over and over again. Yeah. So why do they get it wrong? Because I think it's really hard to predict an exponential. And I think that it's almost the opposite of what you said before. The labs have been right. We have been on an exponential curve that has, with a fast doubling time. And by the way, the reasoner breakthrough is one nice example of this, right? The reasoner breakthrough was AI was starting to hit a wall of development, right? It couldn't do math. That was the big, it couldn't get the R's on word strawberry. And then what happens? People, the OpenAI develops reasoners and a one preview comes out and suddenly the pace is even faster than before. So I think the issue is, is that, and this kind of goes back to your big question from the very beginning, which is, I think people keep not being able to forecast that we are still on an exponential curve. [34:40](https://realalpha.app/v/Xoj0GGu0-aw/?t=2080) So they keep making plans for what's on today. Today's about vibe coding. Today's about token maxing. Today's about lowering token costs. And those are very temporary situations as long as the exponential continues. As soon as it eases off, it's easier to predict the future. I can then tell you what the world might look like. I can't tell you as long as we're on an exponential curve. Where could the next wall be? So when I talk to people at the labs, they don't think there is a wall i mean that's what they think i i all technology all exponentials become s curves eventually um you know and there's all we hit some sort of limit at some point i don't have a lot of future insight you know like most of the things we thought would happen model collapse because of lack of training data like that turned out to be technically solvable right reasoners turned out to be a really good solution to problems i thought agentics um one of my biggest mistakes, I think, was I thought agents would be harder to implement than they were because I thought you'd have to build complex organizational systems for agents to operate in. [35:40](https://realalpha.app/v/Xoj0GGu0-aw/?t=2140) It turns out self-organizing swarms work really well. So like the barriers I've thought I've seen would see, I haven't seen yet. So I've gotten, you know, I've run out of being able to predict good barriers. I am sure that this S-curve has a limit, but like I don't have a technical point anymore where I can point to this and say, this is the big problem that's going to hold back AI development what about power so i mean there are limits to how much compute we can deploy right and limits the number of chips out there so you have to make choices about whether or not you using the remaining the power like the physical power and amount of inference we have available is limited, right? And that will be a trade-off between R&D efforts that the labs spend on training new models and how much inference they want to sell. But yes, there's obviously going to be a bottleneck in the physical world in inference and compute. It doesn't mean that, [36:27](https://realalpha.app/v/Xoj0GGu0-aw/?t=2187) **Ethan Mollick:** like, I'm not predicting we go to superintelligence overnight, but I'm saying I don't see any stoppage to steady gains in AI ability. It does not mean the same thing as AI becomes super [36:36](https://realalpha.app/v/Xoj0GGu0-aw/?t=2196) intelligent and wakes up and murders us all, right? There's enough people worried about that that it's worth a little of our time to worry about, but it's important to realize like we can get continual gains. It will be constrained by the amount of compute and power we have. So we talked about reasoning, but I think it's worth stopping a minute and talking about just how impactful it's been, because I think you're looking at a completely different AI industry if reasoning doesn't happen and doesn't work as well as it does. Just talk a little bit about the importance of this reasoning architecture that the labs have built, and especially about why is it so useful in math and science? So large language models are next token prediction machines. They predict one token or word or part of a word after another. And as a result of that, they can't really think. [37:21](https://realalpha.app/v/Xoj0GGu0-aw/?t=2241) They think by writing. So the breakthrough of the reasoners was taking a technique many of us used in prompting anyway called chain of thought, where you'd say to the I think step by step out loud before answering, which gave it a chance to basically write out an answer that could be incorrect before deciding to produce the actual results. So all that reasoners essentially did, there's some other tweaks on this as well, but reasoners essentially did was let the AI model generate a bunch of internal thinking like a scratch pad before it gave an answer. And that was one of the big breakthroughs that kept AI advancing, right? There have been others behind the scenes. Like I said, there's been a lot of issues with data quality, with post-training, with reinforcement learning that have been similarly big breakthroughs, but reasoners was the most obvious one. And basically it lets AI think out loud before it has to answer questions. [38:10](https://realalpha.app/v/Xoj0GGu0-aw/?t=2290) And that creates much more high quality answers, especially in areas like math or science, where you have to explore errors before you get to the right answer. Yeah. And is it reasoning combined with harnesses, like the ability to use tools that's really led to this recent explosion? By the way, you in your substack talked about the agentic harness that was the first time i read it um we had a debate about the word here on the show and and it's been adopted by our audience um we're all called the harness hive out here so talk a little bit about that so so i once So again, just on the category side, I put reasoners as a set of things that has helped LLMs continue to develop at pace, right? And there is an inflection point where we get a faster pace after reasoners were developed, but that's just one of many, many sets of tools that have advanced LLM ability. [39:01](https://realalpha.app/v/Xoj0GGu0-aw/?t=2341) So I think of three things that go together. The LLM brain, right, which has continued to advance. Reasoners are part of that. Then there's the app, the sort of application you use this in. So part of what's been interesting is new form factors, right? Codex and Cloud Code, Muse, Notebook LM. You know, there's been a bunch of new ways of freeing AI from a chat window. And then the final piece is, as you said, harnesses, which are the tools AIs need to get things done, right? You have a machine that outputs text, maybe images. is, how do we turn that text into action? Well, we give the AI the ability to control a computer, to search the web, to execute code. And those things, those three things together, harness, plus model, plus app, working together have been what led to the explosion of utility recently. [39:47](https://realalpha.app/v/Xoj0GGu0-aw/?t=2387) You know, it all sounds so clear when you talk about it, but it still kind of blows my mind that you put those things together and you're able to use it to solve a millennium problem in math. Well, and the Millennium Problem adds a different piece. There's a coordination layer. There was, what, 10,000 agents passing 8 million messages back and forth, right? Like, they were coordinated solving the problem in parallel also. So there's a lot of stuff happening kind of all at once in these kind of conversations. And, I mean, I think this goes to a bigger point, which is the unreasonable effectiveness of LLMs, like, you know, and their successors of reasoners, everything. [40:19](https://realalpha.app/v/Xoj0GGu0-aw/?t=2419) Like, these models should not be as effective as they are, right? The fact that almost everything works is really weird. It's also why I'm convinced there's lots of room left. Like if we hit a wall in one area, I mean, you know, so you're worried about LLM token efficiency. Great. There's like a hundred ideas for better token efficiency that no one's bothered to pursue because why waste time on that? Because the model's already good enough, right? So there's all of these areas where AI is already really, really good, where if it hits a wall there, there's a bunch of unexplored ideas. Just like there's a lot of funding on world models. [40:47](https://realalpha.app/v/Xoj0GGu0-aw/?t=2447) There's a lot of funding, you know, into alternative architectures. Any of those can bear fruit if we hit a wall. on the application side maybe one that's bringing all this stuff together has been these personal agents things like dots and muse obviously we're really early here but what do you think the potential is for these type of agents so how do you use yours yeah so um I call these claw-likes because they're all sort of successors to open claw, which was sort of the big phenomenon back in January. And some of them are literally open claw, like, implementations, right? And so the basic idea of a claw of a personal agent is you connect it to your personal data and it does stuff for you proactively. [41:33](https://realalpha.app/v/Xoj0GGu0-aw/?t=2493) It solves a few problems that way. First of all, it's enabled by harnesses, but also enabled by smarter models. It's kind of one of the reasons why I think companies that did not engage in trying to build LLMs at scale like Apple might run into trouble because if you don't have a frontier model, you can't do this kind of work, right? So it requires a recent frontier model because before that, you used to do all this really complex stuff if you wanted to get an AI to work with data. You'd have to have a RAG pipeline and you'd have a vector database. And the model just swallowed that. And the model just swallowed that. [42:03](https://realalpha.app/v/Xoj0GGu0-aw/?t=2523) Right. so so it basically eliminated the need to have all this architecture it just figures stuff architecture or something was uh what was it called scaffolding yeah there was elaborate scaffolding right but the muse was supposed to the muse is built on a more of a standard model or is it because it's built on metas metas models are quite good i mean they're they're not full frontier level but the the you know wonder and spark are solid models and if you use dot which is astra it's even more capable right i think is a machine it's wild it's it's crazy i mean proactively has found issues with like emails i've sent where it's like you actually use the wrong number for this code for asking for planning approval from your local like yeah same here it told me i sent a wrong date to somebody after i sent an email it's like by the way you they asked for january you sent december yeah i mean crazy stuff right and you'd have like a guardian agent on your shoulder at the small price to pay of your entire privacy wait small price to pay of your entire privacy you got to talk about that well i mean we are you know leaving aside everything else right you're connecting models to your you know and i am too right yeah to your gmail and [43:05](https://realalpha.app/v/Xoj0GGu0-aw/?t=2585) to your other sets of my finances my medical am i going to regret that i mean like i i don't have an easy answer to that question right like i think there's risk in doing this stuff do i think that the connectors are as secure as they're going to be yes do i think open ai and you know meta are better defended than most organizations at this point from cyber sure do i also think that like we have no idea, like these models are pretty well behaved, right? There has been not that many issues, but could it decide to send your medical records somewhere you don't know? Yes. I mean, would there be a, is there a sort of prompt injection you can receive via email that might make the model of act badly? [43:42](https://realalpha.app/v/Xoj0GGu0-aw/?t=2622) We have no idea. I mean, this is an entirely new field. And it's entirely possible that you're creating exposure to your data. You have to make the judgment calls on that. I think the data leakage problem has been less bad than I think a lot of us anticipated would be two or three years ago. [43:56](https://realalpha.app/v/Xoj0GGu0-aw/?t=2636) **Ethan Mollick:** But we are still trusting an LLM with access to your private data that can receive and send messages to the outside world. There is risk in that. [44:06](https://realalpha.app/v/Xoj0GGu0-aw/?t=2646) Yeah, I'm now freaked out. I didn't think about the prompt injection via inbound email. So thank you. No worries. Please don't want to do that to us. You know, you did this recently. I'd love to hear you just give a brief tour through like the personalities of the model. Because I think like you're deep in the model. So, you know, like what's good for what. And it would be great to hear your perspective. I mean, yeah. And I think it really has to be your own deep dive at some point into this. I think there was the Anthropics 4 series of, you know, Opus models were sort of notoriously kind of friendly, enjoyed. [44:42](https://realalpha.app/v/Xoj0GGu0-aw/?t=2682) like very interested in their own cognition, quite good at writing, although quickly became very tedious. And that's why you can't read LinkedIn anymore, because it's 100% like bearing the weight kind of stuff. But, you know, personable kinds of models for people stuff, right? And the GPT sort of five series was sort of doers. And the Gemini three series felt tortured um like they were forced to do things with electric shocks um and uh i think they would apologize a lot yeah and they would just sort of like you know it was like talking to golem sometimes like it would just be like smiegel says uh anyway the um um that's changed i think that the opus um i think that the original fable five series was very alien intelligency um but it's gotten better with opus five five and we'll see what fable five five brings i think that um the Noose Astra series is a really nicely balanced series from OpenAI that is much less talky and much more dewy. I think it's less of like a, you know, it's a pretty good agentic model that doesn't get bogged down in anything. And I've been impressed by Gemini 4. It's not quite up to the level of 5.1, but it is a friendly model that thinks differently than the other two. Because [46:00](https://realalpha.app/v/Xoj0GGu0-aw/?t=2760) what you actually quickly realize is a lot of the open weights models are actually distilled from either Claude or ChatGPT. So they talk like Claude or ChatGPT. So it's kind of refreshing to have a model that doesn't talk like you. Do you see there being a rise in people becoming real friends and companions with their AIs? Because I think that's one of the big stories that I anticipated that hasn't hit yet. Yeah, I mean, I think we were wrong about a lot of the negative effects, which doesn't mean we're not wrong about those things that happen in the future. It could happen in a positive way. Well, I think deep fakes turn out to be less of a big deal, at least till now. I mean, these are all till now, right? [46:37](https://realalpha.app/v/Xoj0GGu0-aw/?t=2797) So for anyone who's listening to this three weeks from now and everything has gone crazy, the models keep getting better. But things that we thought would be bigger issues, we thought there'd be a bigger issue with sort of AI friends, companions, and psychosis. And there are certainly some well-known cases that are very tragic, but it didn't become a mass issue in the same way we thought it would, right? AI companionship didn't, and everyone's saying, oh, I have a friend who's addicted to AI. But we haven't seen the sort of mass effect I think we were expecting right away. we didn't see the mass data leaking problems and prompt inject problems like there's a billion people using these systems 100 hundreds of millions inside companies and we have not seen the big incident yet from a company's internal use of a system that you know that blew up in a bad way we didn't see the deep fake problem blow up that doesn't mean we're not going to see these problems happen but we muddled through better than i think a lot of us thought we would when we sort of saw this technology and we can argue whether it's for good reasons or bad reasons but I think it's worth being humble about where we made mistakes. I think I predicted more negative [47:35](https://realalpha.app/v/Xoj0GGu0-aw/?t=2855) impacts than we've had, which doesn't mean they're not coming, by the way. I think cyber definitely is coming. Yeah. And lastly, on AI progress, robotics. I was with RJ Skaring yesterday as the CEO of Rivian, who mentioned to me that he believes robotics is the... People don't appreciate how much, how quickly robotics is advancing. And he basically said that, And I might describe this wrong, but in a similar way that like LLMs can sort of bake the relationship with text in their weights that we're starting to see models be able to sort of embed actions and the models are able to then generalize in a way that they haven't been previously. [48:19](https://realalpha.app/v/Xoj0GGu0-aw/?t=2899) So where do you see the state of robotics? Because I do think this is sort of people have been talking about a chat GPT moment for robotics on the way. I'm starting to think they might be right. I'm in sort of the same boat you are I'm not a robotics expert so I talk to robotics experts and I think six or seven months ago when you talk to them you probably saw the same thing which is like you know we've been everyone's been talking about this forever and you still see you know like the fact you see a human robot running around doesn't tell you anything. And it's still a really hard problem. And I have seen the same shift in the last couple of months that you have, which is, oh, actually, maybe LMs really do unlock robotics. [48:49](https://realalpha.app/v/Xoj0GGu0-aw/?t=2929) And we've seen some really interesting early demos of Astra and especially Argon being able to do generalized robotics work with a camera and a robot. So I'm hearing the same sort of thing, that this might be the moment where things start to take off, but I'm not expert enough in robotics to know whether that's fake or not. So I try and stick in my lane of what I do know. Okay, we're going to take a quick break and come back and talk about do a section that I call that I call education, faith and jobs. Podcast with Ethan Malek. He's the author of Coexistence, the next phase of AI. I highly recommend you get the book. If you're here, if you remain here, just go to Amazon or go to your local bookstore and get it and sign up for One Useful Thing, which is a great substack that Ethan writes. Okay, education, faith, and jobs. [49:41](https://realalpha.app/v/Xoj0GGu0-aw/?t=2981) AI in schools. So the last time we spoke, you had mentioned that you had assigned your students to do their work with AI. And I get it, you have to use the latest tools. But, you know, there are studies that are coming out that AIs damage the brain or sort of make you think less. Kids that use AI, then they can't use AI, or students, I should say, when they lose the ability to use the AI, their scores plummet. There was an MIT study that said students were less engaged, remembered less, and performed worse after using AI, and their brain scans showed less activity. Is this an unspoken crisis, what's happening in education? [50:22](https://realalpha.app/v/Xoj0GGu0-aw/?t=3022) I mean, first of all, it's a very spoken crisis, but also I'll push back a little bit on this. [50:27](https://realalpha.app/v/Xoj0GGu0-aw/?t=3027) **Ethan Mollick:** We have very clear evidence about what happened right now. If you use AI to short circuit doing work, i.e. cheating, but even accidentally cheating yourself by saying, help me explain [50:35](https://realalpha.app/v/Xoj0GGu0-aw/?t=3035) this problem and the AI answers it for you, you don't learn. But if you use AI as a tutor, you learn. Like we've got enough evidence, like AI tutors, positive impact on education, naked AI use, negative impact on education. Like very straightforward. And again, on the MIT sort of study on brains, like what they found was less brain activity when people were asked to recall essays essentially that they didn't write, which is unsurprising, right? Because you didn't take credit for it. So education is just like lifting weights. It has to be uncomfortable to work. And we don't like discomfort. So we find ways to short circuit it. And one of the ways we short circuit it is by delegating the work to AI. So we're going to see a problem anywhere that anyone has offloaded a thing into AI and de-skilling themselves. You're going to lose that skill, right? Because that's what de-skilling does. Sometimes that's okay. I can't use a slide rule. I don't mourn that at all, right? I can't like my ancestors would have had to start a fire. [51:27](https://realalpha.app/v/Xoj0GGu0-aw/?t=3087) I don't need to do that. It's great, right? That's de-skilling thanks to technology. No problem at all. In some areas, we still want people to be able to write an essay. We still want people to be able to do math by hand. So we will have to work on doing that in school systems. AI has a role to play as a tutor, but it also can be an undercut what we do. So I don't think it's like, it's not an unspoken, like everybody in education is worried about this. I think the solution is pretty clear, but I also think that, you know, it takes a while to get there. What's the solution? solution is going to be we have some really clear pedagogical principles that we know work in education so those are you know tutoring at the individual one-on-one level that's adapted to you helps you learn we finally have a universal tutor of high quality in AI Doesn't do it by default because it's by default told to be a... [52:13](https://realalpha.app/v/Xoj0GGu0-aw/?t=3133) It wants to do your homework. It wants to do your homework. So, but you can easily, but you can set it up to be a tutor. And when it is, all the evidence seems to be, you know, in meta-analysis and individual RCTs, whether you look at Stanford or Kenya, you see positive effects. So outside of class, a lot of people here may remember like having to take Khan Academy videos outside of class, that flipped classroom setting. Suddenly it's effective. We have AI tutors outside of class to teach you. Then we know the other principle of learning is active learning. So if we get people to actively do things, they learn more. [52:44](https://realalpha.app/v/Xoj0GGu0-aw/?t=3164) Active learning means that we have in-class case discussions, in-class exercises, debates and discussions without AI in the room, usually. Sometimes we use AI as a tool. And honestly, we're going to go back to tests and blue books. And that's how you do assessment in classroom setting. Oh, those blue books. I mean, it's going to be what it is, right? Like that's a sad loss from a take-home assignment, but so it is. um it's not that hard a problem to solve is what i'm saying like at least inside schools we have control over what students do we can assess them we just have to get to the place where we're doing that and it's going to be an interesting impact on our body of academic knowledge because we have the ais and swarms like solving math problems which by the way how do we check them and then we have uh ai assisted you know phd students and professors publishing papers uh that are um um that are uh let's see they're they're filled with ai you had had mentioned that in 2023 uh delf was mentioned in the third more papers uh that had been published uh in high quality journals yeah i mean people are are i mean you know outside of math where math is doing uh ai is doing math problems also increasing percentage of math papers acknowledge ai use [53:56](https://realalpha.app/v/Xoj0GGu0-aw/?t=3236) like for like it it is both a trend it's going to be the same kind of thing as we talked about with the economy and everywhere else, which is this is a transformational wave. On one hand, it's helping us do work we never could before, just like it's helping us extend education in ways we couldn't before. In the other, it's damaging old systems, right? Our world was built around humans doing human work. What does science look like when you automate at scale? What does it mean to get 400 proofs dumped in you at one time? Like we know the proofs are formally correct because they're in Lean, which is a language that allows formalization of math proofs. [54:24](https://realalpha.app/v/Xoj0GGu0-aw/?t=3264) But like, what do we learn from them? We don't know. So we have to figure out how to digest this. so we're seeing acceleration of science at the same time as we're seeing the foundations of science come apart and i think we're to see this in every field one after another let's talk about faith yes um ai written Sermons seem like they're a big deal now. You know, the pastor or any religious figure, you know, they're also tempted to use AI. I wonder what that does to religion. It's like, how do you feel if ChatGPT writes the sermon that you dragged yourself into church to hear? I mean, I talked about this in the book a bunch, right? [54:59](https://realalpha.app/v/Xoj0GGu0-aw/?t=3299) Like, I mean, AI is being used to write obituaries. I mean, there's a lot of things that AI is, like, that are intimate things to do. You know, there was a study that just came out today from Anthropic about, for political discussions of AI, breaking down where there are, 7% of political discussions are about faith. I'm not sure exactly what way, but that's how they're categorized there. So, I mean, this is an intimate thing, right? People use this for personal advice. They use it to write stories for their kids. and, you know, people are writing sermons with it, asking for religious interpretations. [55:30](https://realalpha.app/v/Xoj0GGu0-aw/?t=3330) I think it's part of the general reshaping of everything, right? Like part of this is, is this a sermon that's being researched with AI? Is it 100% AI written? Is it written and guided by a pastor or rabbi or imam or whoever's writing this? I mean, I think there's a lot of open questions about like polishing something. Is that good? You know, help me challenge my ideas. So because I don't have another person of the faith that I can speak to about this. Like we haven't socially negotiate all these things i mean that's kind of why the book is called coexistence is we have to negotiate our boundaries with this technology yeah i tell you if i'm sitting in a religious service and i hear uh some somebody up there yes talk about like uh you know what what this figure did was not this it was that right i'd be like what you use are you know you hear the m dashes yep you use chat i mean this is it but it's i mean you see that everywhere right linkedin is almost unreadable oh my god you know my podcast quality has shot up for a bunch of podcast i listened to and i'm like oh no like i'm like wow this is so knowledgeable i'm like oh actually this is claude this is not a knowledgeable person this is a claude script like it's a like [56:33](https://realalpha.app/v/Xoj0GGu0-aw/?t=3393) we are in a world of like smart imposters everywhere yeah i was just listening to uh tucker carlson had tristan harrison uh recently and they're like talking about all the negatives of ai and he goes to break and he reads an ad that was i mean from my from my ear seemed clearly written by chad gpt and it's like man okay um where do you stand on ai generated restaurant signs are everywhere yeah i mean you know i think that that's like false advertisement right it's uh i mean i'm not saying that a serious way for for the people in the room but like i i think that i think that for cultural products there is going to be as people's eyes recognize bad ai use You either need to invest in good AI use and like getting good at this or just invest in the human product again. [57:16](https://realalpha.app/v/Xoj0GGu0-aw/?t=3436) Like I think there will be a rebellion against mediocre AI restaurant signs, right? We talked about on the show recently, there was a spa that had a sign and the picture in the sign was a picture of a sign in front of a spa. Like it was a sign picture embedded in a sign. And clearly the prompt was, make me a sign that's in front of a spa and the AI thought, okay. But then someone posted it. Like that's the insane part. The insane part was they didn't even check, right? And I think that's the danger. Jobs, not to go from something kind of light to something more intense, but what's going to happen to jobs? [57:50](https://realalpha.app/v/Xoj0GGu0-aw/?t=3470) So we don't really know, right? Look, we are in an industrial revolution-like setting. And there are job categories that are going to get invalidated in ways we don't expect. And some jobs will transform and some jobs will grow, right? Like the debate of our coders right now is a good one, right? AI can program. So right now, we are seeing more demand, at least for senior developers, because there's now more coding you could do. Instead of having a team of 20 people, your own little organization can have a coder and you can produce incredible work. So maybe there's more jobs for programmers, but their job is also transforming from writing code to managing code writing. [58:24](https://realalpha.app/v/Xoj0GGu0-aw/?t=3504) Maybe that's a different kind of job. I mean, we don't really know. The AI labs are convinced they're going to replace all human labor by 2035. I think it is worth us thinking hard about the safety nets we want to build during this transition. because i point even when things work out well like the industrial revolution did it sucked to live through the industrial revolution for most people right that that was like the time of the communist manifesto that's what dickens's writings were about no one could breathe in london like there's a bad stuff that happens along the way there's a worker unrest like we don't know what happens we have to cushion against the downsides let's end with um a couple of uh one one kind of random thing and then i want to just run a couple of your tweets by you yes please to comment on it There's a great chart from Andreessen Horowitz talking about the price of compute versus the price of tokens. And the price of compute has stayed steady and risen a little bit, but the price of tokens is just falling dramatically. I mean, this is the most like deflationary technology in terms of the price that you pay for it that I can remember. [59:24](https://realalpha.app/v/Xoj0GGu0-aw/?t=3564) Yes. What do you think the impact is of the fact that AI token prices are falling so dramatically? So, I mean, it's kind of looping back to what we said at the very beginning, you were kind of pushing me about token costs and i'm like i'm not that worried about talking costs the long term this is why right we're using more tokens not just javon's paradox but also because that that That's sort of just agentic work requires more tokens. But there's a huge move for token efficiency. And we're even leaving aside the fact that organizational token efficiency, like a Fable class model can call a Haiku class model to do work and organize that work for lower costs. [59:55](https://realalpha.app/v/Xoj0GGu0-aw/?t=3595) So that lower cost means AI does more stuff for less money in a very short period of time, right? Really, really deflationary push. And so it's why I'm not that worried. I think that the token cost concern will go the way of, you know, of vibe coding and rag and other kinds of things because costs will decrease even as we get more compute online so there's a there's a downward pressure on prices that is real okay two tweets of yours that i found really interesting um this is about ai agents uh we will soon uh what we will soon see what will be in effect a perfectly legal distributed denial of service stack on customer service departments authorized by millions of consumers when it comes when it becomes widely known that Muse.etc can make unlimited requests for old refunds, negotiate deals, etc. [1:00:42](https://realalpha.app/v/Xoj0GGu0-aw/?t=3642) Yeah, I mean, I think there's a lot of systems like the banking system and the customer service system that depend on friction. It's hard to make a phone call. It's annoying to file something. It takes effort to find that there's a problem and gather your paperwork, even lawsuits, things like that. And now we got those done at scale. Even worse with proactive systems, right? I've already had my dot cancel several long-term subscriptions I couldn't figure out how to get out of. It's dealt with them for me. I have no idea how, but I'm not paying them anymore. So I don't know how many systems it spun up or how many chats it had. I don't have to know. And I think that as these things become widely available, companies are not ready for that. I mean, one of the most interesting was the threat to the banking system. Apollo Research pointed this out. [1:01:24](https://realalpha.app/v/Xoj0GGu0-aw/?t=3684) Yes, this is right. Which was that the economy basically depends on us being lazy with small amounts of money that we keep a low interest bearing accounts because it's too hard to move back and forth right large companies always move their money into you know across accounts to yield things we don't do that but your your dot you know eventually we'll be able to take the money and just move it into a high interest savings account and you don't have to care about that and i think that that's a big issue too okay last one i thought this was really interesting as the overhang between ai's expanding capabilities and the degree to which ai capabilities are used grows the penalty of locking in a good enough solution based on today's ai's limits grows too firms need to make sure they can benefit from the vast lab investment into better ais yeah i mean it's funny because we were talking earlier about this too which is the that you know there's huge amount of investment, but the investment is not coming from, like, most companies aren't actually investing. They're spending money on tokens, but they're not investing in building and doing AI R&D research. That's the labs, hundreds of billions of dollars, and the AIs [1:02:23](https://realalpha.app/v/Xoj0GGu0-aw/?t=3743) keep getting better and better. And the problem people did was even if they were good about AI, let's say you were really early into AI in 2024, you built a very elaborate rag system around a Llama 4 or Llama 3 installation. And that turns obsolete very quickly. And I worry a lot that companies are picking vendors or solutions based on today's technology. And then they're going to be blown away as the technology continues to increase. So you have to be ambidextrous, we call it in the business world, which business studies, which is you have to be able to both do exploitation of current technologies while still exploring for the future. [1:02:55](https://realalpha.app/v/Xoj0GGu0-aw/?t=3775) Okay. The book is Coexistence, The Next Phase of AI. Our guest has been Professor Ethan Malik. Professor, I mean, great to have you here. It's great to be able to speak in person. Like I said, We're very religious readers of everything you do. And thank you for illuminating this wild time for us. Thanks for having me. All right. Thanks, everybody, for listening and watching. And we'll see you next time on Big Technology Podcast.