Show Open and Guest Intro
Adi Ben-Ari (00:00):
Welcome to the Applied Edge podcast. I'm Adi Ben-Ari, and this is the show where enterprise leaders talk candidly about technology decisions reshaping their industries, whether it's AI, blockchain, or privacy, and what it actually takes to deploy them at scale. Today's guest is Professor Michael Muthukrishna. Michael is Professor of Economic Psychology at LSE and NYU, and the author of a book called A Theory of Everyone, which is an amazing book that we touched on during the conversation.
Adi Ben-Ari (00:26):
I first met Michael when I was taking his LSE executive AI course, about a month or two ago. He's an amazing lecturer, really broad background and knowledge, and he covers a lot of bases. I thought it would be amazing to have him on the show, so enjoy the episode. Michael, welcome to the Applied Edge podcast.
Michael Muthukrishna (00:44):
So great to be here, Adi.
Adi Ben-Ari (00:47):
Michael, for our audience, you've got quite a varied and interesting background. You and I met at the LSE, when you were giving the course there on AI for leadership and management. Give us a little bit of background, help us understand where you're coming from.
Michael Muthukrishna (01:00):
Yeah, sure. So, you were in one of our exec courses on AI and the Future of Organizations, which we run at LSE because we really think the technology is kind of there, but it's the human element that's really lagging behind.
His Path: Engineering, LSE, and the AI Course
Michael Muthukrishna (01:16):
How do we integrate this into actual human systems? How do we integrate it in ways that people and organizations can actually work with effectively? I'm probably unusually well prepared to teach the course, for a few different reasons. The first is that my background is in engineering. I'm probably one of the few people, if not the only person at the LSE, who's trained neural networks from scratch.
Michael Muthukrishna (01:42):
My first paper was actually on language models too, at the time, I guess you'd call it a small language model. So I'm able to speak to the technology, but a lot of my work is also on behavioral science, innovation, cooperation, cultural evolution, and organizational change. I try to bring that human element together with the technology to give some insights to business leaders like yourself.
Adi Ben-Ari (02:07):
Amazing, it was a really interesting course. I think we had a great set of people there as well.
Michael Muthukrishna (02:12):
Oh, for sure, a real variety.
Adi Ben-Ari (02:14):
But I found it super interesting. I'm more of a pure techie, so it was great to get much more of the people angle, the organization angle, and all the challenges there.
Adi Ben-Ari (02:21):
Can you give us a bit of a summary, at a high level, what are the main points you'd like to bring to the course, and to that audience?
Michael Muthukrishna (02:27):
First, I have to say it was really great to have you on the course. The team curates this group of individuals very carefully, it was great to have a mix of practitioners and people who had very little AI background,
Michael Muthukrishna (02:36):
and of course, people like yourself who already understand the technology and are just looking for the human element. The idea behind the course is: we're not trying to teach you how to do backprop, or any of the basic algorithms, but we are trying to leave you with enough of a mental model that you understand, let's say, a neural network, or a large language model in particular.
Michael Muthukrishna (02:57):
These are not lookup systems. They're really storing an abstract world model, if you like: how these get trained, why they're like autocomplete, and why they're not. So it's a sufficient mental model for a non-technical user to understand how the technology works, without getting into the math or the algorithms. From there, we look at: given that this technology exists, and given what's on the horizon, with these models doubling in capability every seven months, where are we now, and where are we likely to be?
Michael Muthukrishna (03:29):
And where does the human fit into this? For example, how does a leader decide when AI should be integrated in a way that augments a human decision-maker, where it's essential that a human stays in the loop, and when is it better for it to be completely automated, where the AI is really taking care of things and the human is overseeing it, the way you might oversee a team of employees, where you don't need to know the details but you do need to check the output.
Michael Muthukrishna (03:56):
Then there's simple things, like, what are the concerns people might have when a leader deploys AI in a company, what are the common mistakes people make. We describe dealing with AI almost like AI communication, what are the places where humans are overconfident about what the machines are capable of, and where they're not letting the machine do enough?
Michael Muthukrishna (04:23):
I could go on about the whole course, but a lot of it, I'll say this too, there are certain subjects, like math, or programming, or even the use of AI, where you can't really learn it by reading. You can't read about math and then know math, you have to do math. You can't read about programming, you have to actually program.
Michael Muthukrishna (04:41):
Same with AI, so we try to make it very applied. We go through clear case studies. Here's a failure case: Air Canada launches an AI chatbot for customer service, it hallucinates. Who's responsible for that? Who carries the risk when organizations deploy these technologies? All of that kind of thing.
Adi Ben-Ari (05:00):
Yeah, so the business implications, and I guess the people implications too. How does this land with the team? How do people see it relative to their roles?
How Enterprises Actually Adopt AI
Adi Ben-Ari (05:07):
How do you organize this around the firm, and so on, how do you do it in as smooth a way as possible?
Michael Muthukrishna (05:14):
Exactly.
Adi Ben-Ari (05:16):
Michael, you're quite close to a lot of firms working in the space, what are you seeing? We get these reports all the time, headlines and all the rest of it.
Adi Ben-Ari (05:24):
What are you seeing, what are you feeling on the ground, in terms of where interesting things are happening?
Michael Muthukrishna (05:30):
Yeah, I mean, there's massive variability. You saw that in the course too, some companies have been testing, experimenting, and trying things for long enough that they're well ahead of most others. Then there are others where there's just a sense that AI is important.
Michael Muthukrishna (05:51):
Everybody knows it's important, so we're going to assign somebody the role of thinking about how AI is integrated into the organization, and maybe we'll buy a Copilot license if you're lucky, maybe you'll get access to Claude or something. But there's not much more than that. There isn't a cohesive strategy, and one of the challenges is that the landscape is moving so quickly that by the time somebody writes a playbook, unless that playbook is in very general terms about how to think about this in the first place,
Michael Muthukrishna (06:20):
say, how do we deal with the regulatory environment, it's out of date. So it's very much a case of: how does your organization either learn from other organizations further along in the process, or bring in the kind of training that you received, for example, or conduct its own experiments so it understands enough about what the potential use cases are?
Michael Muthukrishna (06:44):
But then, because you know your industry better than anyone else does, you know the kind of work you're doing, you're looking to identify where in your processes AI might actually be able to help with some bottlenecks. One mistake I see a lot of companies make is that AI comes first. Whereas in reality, the best companies realize that AI is in service of the business, AI is in service of the bottom line.
Michael Muthukrishna (07:10):
What you're looking for is: given that this is what our business does, where along the pathway can it help, is it triaging potential customers? Is it, once customers come in, making sure they have an excellent experience? Is it aftercare? Is it your actual product that needs an upgrade?
Michael Muthukrishna (07:31):
And when we run these courses, it's wonderful to hear where companies are. I loved your example, actually, we just built a layer so we could understand, across the organization, all of these things that were previously diffuse. I was listening to managers personally, now I've got a live report, and I know what's going on.
Michael Muthukrishna (07:51):
That's a good example of a company that's experimenting and is at the forefront.
Adi Ben-Ari (07:56):
Yeah, we had a few other examples in the course as well, which is great. Thinking about how this plays out and how companies adopt it, one of the things I've found is I've seen some companies set goals around AI, bring the tools in, and try to force it in and propagate it to get the efficiencies.
Adi Ben-Ari (08:12):
What I found in our own company, and I think I took some of this from the course, is that our company goals didn't change. Our company goals stayed the same, and it's really about how we can use AI to better achieve those goals. So the goals, even the deployment goals, don't move. It's really just rolled up into that.
Adi Ben-Ari (08:27):
So we've got sales targets, revenue targets, efficiency targets, and so on.
Michael Muthukrishna (08:34):
Yeah, you're saying the same thing I said, in a slightly different way, which is: AI is in service of the business, not the business in service of AI. So unless you're Allbirds and you decide to flip from making shoes to training models, your pivot is really about the ways you've traditionally done things,
Michael Muthukrishna (08:52):
There are new capabilities available on the market. Some of them come with some level of risk, some with some level of a learning curve. Is that worth it for us? What's the potential payoff? Is the risk actually worthwhile? And if it is, how do we learn enough and adapt the technologies well enough?
Michael Muthukrishna (09:11):
How do we differentiate in a marketplace where there are AI cowboys claiming all kinds of things? How do we evaluate whether somebody is actually accurate, or whether the claims they're making are actually true? How do we protect ourselves if the company isn't here a year from now? To what degree do we integrate the technology versus trial it in some manner?
Michael Muthukrishna (09:33):
Should we build it? Should we buy it? Should we co-create it? These are the kinds of decisions companies are making, but they're making them, and I can't emphasize this enough, in service of whatever it is that actually makes them money. At the moment, what we're seeing is a lot of money being spent and not seeing the returns, because it hasn't been spent in effective ways.
Adi Ben-Ari (09:56):
Okay, that makes sense. That's a lot of what we see too, a lot of bottom-up: here are the tools, try and do something, make your work more efficient. And that's quite difficult to then roll up, account for, measure, and actually see the efficiencies come through.
Michael Muthukrishna (10:11):
It doesn't have to be bottom-up.
Michael Muthukrishna (10:14):
There is a bottom-up kind of learning process, where there's everyday innovation and you're talking to the people at the coalface, so to speak. But there's also an important role for leadership, placing an appropriate person with an appropriate amount of power to look across the organization, especially if you're a large organization, and look at all of those little experiments, to avoid "pilotitis," where you get all of these little pilots that never bubble up into anything that rolls across the organization, or you're repeating the same experiment multiple times, or buying in a service you've already bought.
Michael Muthukrishna (10:43):
So, cataloging and identifying everything that's going on, having a very clear AI strategy, as you guys did, putting together your own AI playbook, and then having sufficient power in a leadership team, or a particular leader, so that all of that bottom-up learning can be delivered across the organization in a top-down way.
Adi Ben-Ari (11:02):
Yeah, this is part of what we had to deal with. We let everybody build things with the tools, and then I had three project managers create three versions of the same thing. Then you have the people problem: everyone's emotionally invested in their own version. Now I've got to walk that back, best to avoid it in the first place.
Michael Muthukrishna (11:21):
Yeah, yeah, yeah.
Inside A Theory of Everyone
Adi Ben-Ari (11:24):
One of the things I learned through the course review is that you have this book you've written, A Theory of Everyone. I read it, and I was actually quite surprised, I was expecting a book on AI.
Michael Muthukrishna (11:33):
Right, right.
Adi Ben-Ari (11:34):
And it's not. It's much more than that. It's a book about people, about humans, about human society, about the groups we form in society and how we treat each other and how we perform, and so much more. Can you give us a bit of a breakdown of how you got there, and what you were trying to achieve through it?
Michael Muthukrishna (11:58):
Yeah, sure. I was lucky, I knew some OpenAI board members before ChatGPT dropped. I happened to be in San Francisco, having coffee, and they said, "We're releasing this new tool, we've made these breakthroughs, we think this could be very important." As I learned more, I worked it into the book, in 2023.
Michael Muthukrishna (12:19):
But it was already in the manuscript, luckily, because I knew what was coming. So there's certainly an element of what AI represents in the human story, but there's blockchain in there too.
Adi Ben-Ari (12:30):
Which is also good to see.
Michael Muthukrishna (12:32):
Yeah, yeah, you might remember my tragic story around that as well.
Michael Muthukrishna (12:36):
So, what I'm interested in, more broadly than just AI, are the challenges humanity faces across the coming century, one of which is now AI. I don't think I would've said that ten years ago. To tackle that problem, it's a bit like pulling together learnings across fields,
Michael Muthukrishna (12:54):
everybody's got a piece of the puzzle. Economists recognize that growth and the economy are important. Political scientists recognize that politics and governance are important. Psychologists recognize there's a behavioral element that everybody seems to miss. And you can't do anything in public health, governance, or economics without understanding that human element.
Michael Muthukrishna (13:15):
And you've got lawyers who are constrained in all these ways. Everybody's got a piece of this puzzle, and what I wanted to do was try to figure out how one would put that together. So the starting point, actually, in my research, when I pivoted from engineering to a new line of research, was to understand how humans started in the first place.
Michael Muthukrishna (13:33):
As an engineer, one of the things you're trained to do is to be able to zoom in and out of a system. Especially if you're doing the full stack, you have to know what's going on in an individual component, and why, when it bubbles up to a class or something, it's not behaving the way you expect, but then also how that system fits together with every other piece.
Michael Muthukrishna (13:54):
And if you don't have that kind of model, the ability to zoom in and out, it doesn't really work. So I wanted to start from first principles. My initial work was actually on human evolution, the evolution of the human brain. What were the selection pressures that led our species, our lineage, to diverge from all the other animals on Earth?
Michael Muthukrishna (14:12):
About four to seven million years ago, we diverged from chimpanzees. We had a last common ancestor with them. What led to that? The answer was that we switched to a kind of distributed cognition. Most animals, when they encounter an environment, are forced to genetically adapt, they get more fat and fur if it's cold, for example, or more muscle if they need to outrun predators, and then they learn things along the way too.
Michael Muthukrishna (14:35):
We do that as well, we have genetic adaptations, and we also have individual learning. But what we really started to do was share knowledge with one another, and selectively learn from one another, like what's happening in this podcast, where somebody has knowledge worth learning, and I want to store a piece of that and apply it to my thing.
Michael Muthukrishna (14:53):
And even though you don't realize it, there's a kind of collective brain of humanity solving puzzles, letting ideas flow through the social network. The network itself is like one big neural network, and that neural network is actually what's solving a lot of things. So in that initial work, we identified what those selection pressures were.
Michael Muthukrishna (15:10):
We published papers on how it explains primate brains, human brains, whale and dolphin brains, and, most recently, octopus brains, a completely different lineage. From there, you start to think: how does innovation happen in this kind of collective brain? How does cooperation work, when, even inside an organization, are people likely to be at each other's throats and not work together,
Michael Muthukrishna (15:35):
and when might they actually work together toward the common good? In the book, I write about Jeffrey Skilling and Enron, and what I call the "Enron effect," where Skilling sets up a situation in the company where he's very... resilient, and very...
Adi Ben-Ari (15:52):
Yeah, exactly.
Michael Muthukrishna (15:54):
So, he's a big fan. He loves Dawkins, and he thinks he understands evolution, he thinks it's all about competition. But it's actually cooperation and competition as two sides of the same coin. As a result, he goes for "rank and yank," which was very popular at the time, but he's very vicious about it. He ranks the company and chops off
Michael Muthukrishna (16:09):
the bottom 15%, or 5 to 15%, I think it was, and does this repeatedly. What that actually does is create a zero-sum situation. And in a zero-sum situation, your win is my loss, my win is your loss, so there's no incentive to work together under those circumstances.
Michael Muthukrishna (16:24):
In contrast, if you read Satya Nadella's book, Hit Refresh, you see the opposite. You see what the models show: if you can create a scenario where you tolerate risk, allow people to try things and fail, but you're all working together toward a common goal,
Michael Muthukrishna (16:41):
it's like: if it works, we share it across the organization. It's a series of startups, if you like. Then you've suddenly created a positive-sum situation, where it pays to work together because we all get a reward, we all get a piece of this, regardless. The model makes all of these very broad predictions that apply, as I like to say, all the way from cells to societies, right?
Michael Muthukrishna (17:01):
Bacteria to businesses. These are common laws that apply across the board. And most recently, a lot of my work is on AI. The reason for that is we're in new territory, you cannot look to the past and say, "well, we know what happened with computers, or the internet, or social media, so this is what's going to happen with AI."
AI as a General-Purpose Technology
Michael Muthukrishna (17:25):
You really have to think it through from first principles, you have to understand why things happened the way they did with agriculture, industrialization, computers, the internet, and so on. Because, as everyone knows, AI is a general-purpose technology. It's like electricity. It's hard to point to. It's not a particular app,
Michael Muthukrishna (17:45):
it's not a particular feature. It's something that permeates the whole world, the way you and I are talking right now, thanks to microchips, a bit like the internet itself, I guess.
Adi Ben-Ari (17:52):
Yeah, exactly, exactly. That's a lot, let me just unpack that a little bit. First of all, I listened to a conversation this week, another podcast, where there was a self-proclaimed futurist.
Adi Ben-Ari (18:10):
It was interesting, one of the questions he was asked was what career he'd recommend for his children, given AI and how things are evolving. He said engineering, he has an engineering background himself, and I was quite surprised by that. But he gave a similar explanation to the one you gave at the beginning, about problem-solving, understanding the world, mathematical models, and so on.
Adi Ben-Ari (18:36):
So that's an interesting thread, thinking about your book. First of all, I found it amazing on many, many levels. The thing that was interesting for me, and maybe this is because I also have an engineering background, I was a software engineer, is that I feel like the world we live in, the world I've grown up into, is incredibly complex now. Especially in recent times, with social media, politics, all the bubbles we have, false information, fake news, all of these things going on, more
Adi Ben-Ari (19:12):
tribalism, so much complexity, so many effects. It's a real struggle to understand why things are happening and where you find yourself in that. And I found myself reading your book, and piece by piece, layer by layer, it all made complete sense to me. There wasn't one thing in there I could take apart and say, "no, that's not how it works," or
Adi Ben-Ari (19:39):
It actually helped me explain a lot of things I see today. And I think part of what you did there is tackle head-on some very, very difficult and contentious areas of life, how we split into groups, how these groups view each other, why certain groups perform in certain ways in different areas, and how they interact.
Adi Ben-Ari (20:01):
And today, there are so many things you just can't talk about, because they're perceived as so sensitive. But if you unpack them with an engineering mindset, they actually work mechanically, almost, and this is why. And you bring in historical context, this is why things evolved this way, and this is why that group evolved that way,
Adi Ben-Ari (20:21):
and you bring it all together, and it sort of adds up in a way that's explained like I haven't really seen before.
Michael Muthukrishna (20:30):
I really appreciate that, Adi. If I had to say what my ideal reader would say about the book, I don't think it could've been articulated any better.
Adi Ben-Ari (20:40):
I also have this,
Michael Muthukrishna (20:42):
No, go ahead.
Adi Ben-Ari (20:43):
I was going to say, for me, it's almost a must-read for society. I think if every human being who reads books read this book, the world would be a better place, genuinely, because I think we'd look at things a little differently.
Michael Muthukrishna (20:56):
Yeah, yeah. I have an annoying tendency to just be honest in everything I do, and that includes my writing.
Michael Muthukrishna (21:04):
I took the gloves off, and I said the things people don't often say.
Adi Ben-Ari (21:10):
Yeah, but you've done it in a way that's, as I said, based on data. It's not... it doesn't offend anyone.
Michael Muthukrishna (21:21):
It doesn't. Yeah, yeah, it's more like, "for these reasons, that group is like that." Like I said, it's become more and more contentious over time to talk about groups. But if you don't understand it, it's difficult to understand why things move the way they do today.
Michael Muthukrishna (21:39):
And if you don't understand it, you can't solve it, you can't change what you don't understand. You have to really grapple with the machinery.
Adi Ben-Ari (21:50):
Yeah, exactly.
Michael Muthukrishna (21:52):
So this might still sound a bit abstract to people, and if it does, well, you have to read the book.
Adi Ben-Ari (21:58):
I feel like I should just grab a copy, since we're advertising it so much anyway.
Adi Ben-Ari (22:01):
These are genuine feelings, like I said, I feel it as an individual. There's also something I grapple with because of my own background, and I felt this really made a lot of sense to me. The other book I'd probably compare it to is Yuval Noah Harari's Sapiens, reading that blew my mind.
Adi Ben-Ari (22:21):
There are obviously a lot of parallels, you go through this whole human journey and unpack it. But I think you look at it from quite a different angle, a different perspective, and bring something different to it as well.
Michael Muthukrishna (22:31):
I like to think of it as: I'm the science to his history. He's the historian,
The Deeper Argument
Michael Muthukrishna (22:37):
I'm a scientist. I don't know if he does the history well, but he does, the science is missing there, and I bring that science element to it. So, going back to the topic of AI: we laid the internet over human society, and it changed a lot of things.
Michael Muthukrishna (22:52):
We then had social media on top of that, and that's changed even more, for better and worse. You go through some of that in the book,
Adi Ben-Ari (23:05):
What does that do to society?
Michael Muthukrishna (23:08):
Yeah, so there are multiple things to say. The first is: we shouldn't necessarily expect to see the impact of AI immediately, because there's something called the Solow paradox in economics, people said the same about computers, and about the internet: you see the effect of computers everywhere except in the productivity statistics.
Michael Muthukrishna (23:23):
You see the effect of the internet everywhere except in the productivity stats too. When that was said, in the late '80s or so, it was true. But over time, you did see it. That's because, initially, when technologies come in, we don't know how to use them effectively.
Michael Muthukrishna (23:40):
And sometimes we do silly things, there's a level of experimentation required to figure out how to use things most effectively. This is true of every technology. Even going back to when we created our factories, the first factories ran on steam engines. You'd have one steam engine and one shaft running through the factory, just turning,
Michael Muthukrishna (24:02):
and every other shaft, through gears, operated off that one shaft. When electricity came along, at first all that happened was: instead of steam turning that one shaft, electricity turned it. You didn't see much of a productivity gain, the only thing that had happened was you'd shift from steam to electricity.
Michael Muthukrishna (24:22):
But over time, people realized: wait, if we're using electricity, why do we need a single shaft? You can have separate machines across the factory floor. Once that diffusion happens, you can optimize the layout of the factory, and that's when you start to see the productivity gains.
Adi Ben-Ari (24:41):
Why don't we, like, all of them, the fact we can run at night, and so on,
Michael Muthukrishna (24:45):
Exactly, exactly. All of these things start to happen. It's the same thing that happened with computers, at first, computers were like fancier typewriters. Now I can go back and edit my documents in WordPerfect or something,
Michael Muthukrishna (24:55):
if people are old enough to remember. Then, over time, we figured out computers are capable of far more than this. Excel is this massive technology that at first is just a calculator, but now it's an operating system, an entire programming language. With AI, we're seeing similar things, where people are experimenting in all kinds of ways.
Michael Muthukrishna (25:18):
Some of it isn't sensible, some of it's actually worse than the human equivalent. But occasionally we chance upon something that's actually correct, effective, and incredibly time-saving or productivity-enhancing. That's the first thing to say. The second thing is: if you think about what humans have access to for decision-making, from an evolutionary perspective, we have our genes, right?
Michael Muthukrishna (25:44):
We're human beings, we have genetic capabilities, and we have our lifetime of learning things by ourselves, running our own little experiments. But the big thing is that we have this cultural corpus of beliefs, values, and ways of doing things, where we learn information from one another. And that's been true even in terms of the internet.
Michael Muthukrishna (26:02):
The internet isn't fundamentally changing what's happening, we're just learning from one another in a more distributed, broadcasting way. Anyone can share their skills on YouTube, and you can learn anything you want.
Adi Ben-Ari (26:15):
Yeah, the access has completely changed, right? It's really changed.
Michael Muthukrishna (26:18):
But we haven't fundamentally changed the basic rule, the basic pattern.
Michael Muthukrishna (26:25):
We have always used machines to enhance ourselves physically, and then to automate ourselves physically, like with factories. Then we used machines to mentally augment ourselves, the internet, learning from one another, Excel. We have never, in the history of the world, outsourced the thinking. We've never in the history of the world said, "hey, you, you do the thinking."
Michael Muthukrishna (26:45):
For me, that's never happened. So we're really in new territory. We already know it changes what we fundamentally do, in terms of writing, our software is the same, but suddenly we can access the entire cultural corpus of human knowledge in a way that's highly personalized. If you want to learn something on YouTube, you have to scan through the video, past everything you already know, to the thing that actually matters,
Michael Muthukrishna (27:09):
whereas if you go through a textbook, it's just stuff, stuff that's not directly applicable. But with AI, you can literally get it to solve your specific problem, highly personalized to you. That's completely world-changing. That's not something we've ever had before. Even if you're a very data-oriented company, or a data-oriented person, who only makes data-driven decisions,
Michael Muthukrishna (27:36):
often, what you've had to do is rely on the average. What makes most companies succeed? What makes companies kind of like mine succeed? That's what research, papers, everything presents.
Adi Ben-Ari (27:47):
Yeah, exactly, you get the mean.
Michael Muthukrishna (27:52):
Right. But I like to joke that the average person is, say, a Chinese man named Mohammed, making $15,000 a year, but that's not a real person, it's not a real thing. The average doesn't represent each of our individual characteristics, but AI can. As it gets to know you, gets to know your company in context, it's making better decisions than you'd have had access to before.
Michael Muthukrishna (28:19):
As I described it in the book, we have a fourth line of information: genes, culture, individual learning, and now access to this cultural corpus that's highly personalized, highly individualized. That's a total game changer. The third piece is that we're in a kind of "Covid" moment, in the sense that we're seeing exponential growth in these models. The people who've been best able to predict the present moment are the ones who believed, since 2020, in what I call "the line," the fact that these capabilities have been doubling every seven months.
Doubling Every Seven Months
Michael Muthukrishna (28:48):
However you want to measure it, roughly every seven months, the models get twice as good as they were before. That's the same thing, remember when Covid hit, people were like, "okay, this is a weird thing in Wuhan." Then, "oh, now it's in Italy, I don't know what's going on." And then you see hundreds turn to thousands, tens of thousands, hundreds of thousands, millions.
Michael Muthukrishna (29:03):
And suddenly the whole world is locked down, and it happens in months. We're seeing a very similar pattern here: what models were capable of last year was okay, what they were capable of earlier this year is already different from what they're capable of today. And according to that line, we should, right about now, have a model that's too powerful to give the public access to.
Michael Muthukrishna (29:25):
And right on time, Mythos arrives, and export bans follow, right on cue. What we should be seeing, behind the scenes, is that this further accelerates the process, because frontier companies, OpenAI, say, are able to use those models themselves to improve the next ones. So we're in a scenario where things are rapidly happening, but the public isn't aware of it.
Michael Muthukrishna (29:47):
From their perspective, they don't realize that, at the moment, you can get maybe half a day's reliable work out of one of these models, but by the end of the year, it'll be a little over a day's work, and by next year, maybe a week's worth. That is a game changer,
Michael Muthukrishna (30:04):
the fact that all of this cognitive work is now potentially doable by a machine.
Michael Muthukrishna (30:08):
Yeah, I don't want to just keep talking, you should say something too, Adi.
Electricity, Steam Engines, and Jevons Paradox
Adi Ben-Ari (30:11):
Yeah, a lot of thoughts about this, one of them is just going back to what you said at the beginning of that section, around the factory,
Adi Ben-Ari (30:21):
the steam engine, and the electricity, and what that did. Having started a company in the blockchain space, that really resonates, because we had a lot of companies using it, they plugged it in, in some way, so, in inverted commas, they were "using" it. But they weren't really making the most of it, right?
Adi Ben-Ari (30:40):
They hadn't adapted or changed the model, or the way they were using the technology, or the way they actually plug into the benefits of it. I think that's part of the real struggle with AI now too, with organizations, the way they're structured, the people, the resistance, how you propagate this, how you manage the initiatives, top-down, all of that,
Adi Ben-Ari (31:00):
there's all of that going on. The second thought is what I said earlier, about what companies are trying to achieve, and their business objectives. You touched on this, I think there's an exception, which is where your industry changes: where your offering now has to change, or part of what you offer now has to change, because of what AI offers.
Michael Muthukrishna (31:21):
That's correct.
Adi Ben-Ari (31:23):
Well, your customers can maybe do it on their own now.
Michael Muthukrishna (31:28):
Yes, exactly right. Look, the marketplace is an evolutionary system. Companies either evolve or they go extinct. And now there's a new apex predator on the market, and that apex predator is capable of eating your business, and eating your consumers.
Michael Muthukrishna (31:47):
You first saw it with Stack Overflow, that's where all the engineers went, and suddenly nobody goes there. Then you started to see it with outsourced developers, suddenly people are building software on demand. Is it enterprise-grade? Does it meet all the quality assurance? Does it meet all the regulatory, ISO standards?
Michael Muthukrishna (32:06):
No, it doesn't, but for a quick app to manage your life, your to-dos, or your project management, it's fantastic. So we're seeing entire industries, illustrators, for example, suddenly disappear. And we're also starting to see Jevons Paradox. In economics, Jevons Paradox is that, because there's latent demand for something, when the price falls, usage doesn't just meet demand, the overall amount of usage goes up.
Michael Muthukrishna (32:31):
In other words: there's latent demand for electricity, for example, but the only thing preventing anyone from using as much electricity as they want is that it's expensive. If the price of electricity falls, it's not just that demand goes up to meet what was previously there, no, usage actually goes up. We use more electricity than before.
Michael Muthukrishna (32:47):
We're seeing a similar thing with Jevons Paradox when it comes to AI. There are things you couldn't do before, you couldn't have software on demand, you couldn't have analytics whenever you wished. Now you're able to ask for these things, or even just marketing, or illustrations, or any of these kinds of things.
Electric Twin and the Jobs Debate
Michael Muthukrishna (33:07):
I'm the Chief Science Advisor at an AI company called Electric Twin. What Electric Twin does is create synthetic audiences, and the nice thing about the company, partly because of my involvement, is that it's actually science-based. It's not three guys out of college trying to use an LLM to simulate a human being.
Michael Muthukrishna (33:23):
It's actually all built on the science behind it. What we're seeing is that, suddenly, in order to make good decisions, companies need to know what their customers want, what their potential customers want. It used to be a very expensive process to get that information, so the amount of research a company would do was minimal, because the return on investment was always there, but the cost of the investment was too high to justify it.
Michael Muthukrishna (33:46):
Now, because you can ask a highly reliable synthetic audience something in a matter of minutes, we're seeing companies do so much more research, they can make fitter, faster decisions off the back of it, and if they really need to check, that's when they go back out into the real market. You have a kind of human simulator, in this case.
Michael Muthukrishna (34:04):
Ben Warner, one of the co-founders of the company, calls it "Jensen's paradox," after Jensen Huang.
Adi Ben-Ari (34:13):
I think this also plays into one of the big debates of our time, around jobs, because something can now be done using AI, will you need fewer people for those activities? Or, as a company, might your customers want a lot more of everything, in which case maybe it's the same number of people doing a lot more, rather than fewer people doing the same?
Michael Muthukrishna (34:33):
Exactly. We're seeing that first in software, the fact that we now have software agents has changed what it means to be a software developer. People aren't writing a lot of code themselves; they're managing fleets of agents, and managing them cleverly, so you're doing quality assurance and reviews as part of the AI-driven process, an "AI factory," if you like.
Michael Muthukrishna (34:55):
But actually, if you look at Anthropic, they're paying more than they ever did, and hiring more engineers than ever before. We're seeing that across the board. It could be that that's stage one, there was always latent demand for software, and software engineers were paid a lot of money because they couldn't write as much software as the world actually wanted or needed.
Michael Muthukrishna (35:17):
Now that AI is there, you still need software engineers, but they're able to write so much more software, so much faster. So you actually need more engineers to take advantage of these new possibilities. But at some point, it could be that the AI itself is managing those fleets, and then suddenly you're hiring more and more engineers,
Michael Muthukrishna (35:36):
and then, at some later point, we're not hiring any engineers at all. It's very difficult to predict these things a priori. But I don't think it's straightforward that just because AI is entering your industry, it means fewer people are needed. It could actually be that more people are needed, but they're way more productive than they once were, because they're either augmented, or they're leading automated teams of AI agents that they work alongside.
Michael Muthukrishna (36:01):
I think all we can do right now is lean in.
Adi Ben-Ari (36:04):
Yeah, yeah, yeah, it's one of those things where you want to keep up. I think a lot of people from our generation remember when computers first came out, and people were looking for typing skills, word-processing skills, Excel skills, and those people had an immediate advantage in the market.
Michael Muthukrishna (36:20):
There's a similar thing here: people who have some knowledge of their own domain and industry, and are able to amplify that using these new technologies, and stay on top of it, have an immediate advantage over those who bury their heads and pretend nothing's going on, thinking everything will return to normal. It's not, because we're really just at the beginning of Covid.
Close
Adi Ben-Ari (36:47):
Yeah, agreed. I think, on that note, Michael, I'd like to invite you back and have another conversation in seven months.
Michael Muthukrishna (36:53):
Yeah, let's do it, once capabilities have doubled from what they are now.
Adi Ben-Ari (36:58):
Let's do it. Well, thank you very much, Michael, it's been an absolute pleasure.
Michael Muthukrishna (37:00):
Likewise.
Adi Ben-Ari (37:01):
So, let's see where we are in a few months.
