The text below is a transcript of the audio from Episode 59 of Onward, "Life after AI, with Kanika Bahl, founding trustee at Anthropic".

---

Ben Miller [00:00:01]

You got a good ball. Welcome to Onward.

Kanika Bahl [00:00:02]

Thank you, Ben. Good to be here.

Ben Miller [00:00:06]

So I'm excited to dive into this because you have a unique point of view. I think of you as an insider and an outsider. And so you bring that like kind of that combination, which is rare. And so I wanted to start, but maybe you could just give everybody a little bit of your bio and then we can go from there.

Kanika Bahl [00:00:26]

Yeah, that sounds great. So I'm a founding member of Anthropic's Long-Term Benefit Trust. So that's an independent governance body. It was created with the authority to select a majority of Anthropic's board. And it was really designed to ensure that the company developed AI for the long-term benefit of humanity. And so I started that role when Anthropic was valued at about $5 billion. And I was there through its growth to a $380 billion company. So a period of really rapid growth and transformation. And then simultaneously, I was the CEO of Evidence Action. And Evidence Action reached 530 million people. And it really focused on health and economic interventions based on really sound evidence. that could massively improve people's lives. And so to put that in perspective, we reached one in seven people living in poverty today and drove something like $23 billion in productivity gains. So really the focus was on scale. And so I was simultaneously on the Anthropic Trust. So I had a front row seat to what is arguably the most consequential technology and how quickly it was moving. And then wearing my Evidence Action hat, I kept returning to this really core question, and there wasn't a good answer, which is, you know, if transformative AI arrives in the next five, 10 years, who ensures that the three and a half billion people living in poverty aren't left behind? And so taken together, that turned into a new organization that I've started, the AI Access Initiative to AI. This is a nonprofit. We're focused on ensuring the benefits of AI, health, income generation, education, reach tens or hundreds of millions of people in lower middle income countries. And I know we'll chat more about that later, Ben. And then just to round out, you know, before that, I was at the Clinton Health Access Initiative. I built out operations across 17 countries and drove an 80% reduction in HIV/AIDS drug prices. And to put that in context, when I joined, this was around the mid 2000s, 2.4 million people were dying of AIDS in sub-Saharan Africa. The drugs existed, they worked, but less than 5% of people who needed treatment could access it. And the problem wasn't science. It was coordination across those countries, negotiating with the manufacturers and pricing. And that's what unlocks scale. I just want to spend a second on that because it was the most urgent work I've ever done until now. And I would also say it really highlighted a common theme of my work, which is when it comes to reaching the poor, the best breakthroughs aren't what you invent. It's not the HIV/AIDS drugs. I don't even think it's the AI, it's actually how you reach those people. And so that's the focus of the AI Access Initiative, in this case applied to, I think, the most powerful technology ever created.

Ben Miller [00:03:40]

Yeah, there's so much about your career, which is why I want to have you on the show, that addresses a lot of the fear about AI. It's about access, about distribution, who gets the gains, how is it going to change people's lives and jobs. And in a way, you've been addressing that question, not with AI, but with just wealth and access to, like you said, life-saving medical technology, there's just been a question you've been working on for your whole career. So it's a different technology, but it's still the same question about who benefits, how do they benefit, and how do you make sure they benefit?

Kanika Bahl [00:04:27]

Yeah, that's exactly right, Ben. When we look at even really simple health technologies in the developing world, can take up to 20 years to get to scale versus one or two years in the developed world. And that's really simple things. That's like a shot in arms for a child. And now when we're talking about embedding AI into national systems, it's just a different level of complexity. And I think just assuming that that is going to happen is a really, really grave mistake.

Ben Miller [00:04:58]

So the idea I'm trying to flesh out here and having you as sort of the founding guest is to try to answer this question that most people aren't asking, which is that what happens after AI? I think most people are asking the question of, will AI get to scale? Will it be transformative? Will it end up replacing human jobs, things like that? I want to accept the premise that AI is going to get to whatever we call artificial general intelligence or artificial superintelligence. It gets to a level that's comparable or better than people. I want to accept that and say, OK, now what? Because I think that now what has gotten much less thinking and And I want to do that, but before we do that, I wanted you to help me get the audience to get us there because I think accepting the premise is going to be something that's hard for people to swallow because it seems like, yeah, I think a lot of people are skeptical at the moment. So how would you help me get there? You've been in the room with a lot of AI scientists and researchers. So what makes you think that could be true?

Kanika Bahl [00:06:17]

Yeah. So I'd say when I first started on the Anthropic Trust in, I think it was mid-2023, AI was a black box to me, right? And why these assertions were being made weren't entirely clear. And I went through a lot of training, a lot of learning, and I think just to break it down, I came to actually see how this is based on pretty reasonable hypotheses and empirical data. So... In terms of why people think various labs and others think AGI might be coming or something approximating it, the core belief really, in my experience, rests on something called scaling laws. And it's basically the observation that as you add more compute, more data, and better model architecture, AI just gets reliably better at essentially every cognitive task. It's not linear, but it's pretty consistent. And that curve has held for a decade I think we're actually starting to see this. Obviously, the improvements are showing up in the real world. AI is now writing a lot of code at labs like Anthropic. But I also think there's other tests which are happening that are interesting. So there's this research organization called METR, and they track one specific measure of pace, which is the length of software engineering tasks that an AI agent can complete autonomously with 50% reliability. In 2019, it could do two seconds of work. Today, they've actually, METR has had to come out and say it's hard to actually use a benchmark because they're at 16 hours and it's getting unreliable because the tasks aren't hard enough. And so basically to do that, AI capabilities for several years were doubling every seven months. Since 2023, it's compressed to four months. And now it's been doubling roughly three months. And someone was telling me the other day, I have not confirmed this, that it was something like six weeks. And so the doubling trajectory is moving up faster and faster. And the reasons behind it are pretty, I think, quite intuitive. I talked about one of the factors being compute. We have about $450 billion that was spent by the five major tech companies on AI, and that was mostly for compute, data centers, chips, power infrastructure. So you have a lot of money going to this. You now have AI itself generating the code, and that actually acts as a force multiplier across the compute, the data, and the architecture simultaneously because you can make the compute more efficient, You can do training faster, attempt for the architecture. It's like a massive team of very, very low cost, very, very fast engineers who never need to sleep, right? And so I think that's why you're seeing this, the combination of the scaling laws with the force multiplier of the increasing ability of AI itself to do the coding is leading to this massive increase in capabilities and this exponential increase in and what AI can do.

Ben Miller [00:09:39]

Yeah. Let me try to drill down on that a little bit because I think we're trying to move past this debate. So I want to do that by talking about a few people you may know who have moved past skepticism. So you must know people probably at the AI labs who are absolute conviction that we're getting to AGI or maybe ASI. So when you talk to them, and maybe you have that conviction too, like what is different about their thinking? Like, you know, this is like a very kind of wild idea. And so, you know, one way, a lot of times people make decisions not by actually knowing themselves. They just see somebody who they know knows and then follow what they do. There's like a modeling aspect to it. So when you talk to those people, like what convinced them?

Kanika Bahl [00:10:34]

I really think it's pretty careful empiricism, Ben, right? Like it's looking at, if you believe the scaling laws, which by the way, could break, right? That's always a possibility. And even the most serious believers I have encountered recognize that could happen. But it's really, the scaling laws are if you have more compute, if you have more data, and if you have better architecture, then the models improve. And we are just pouring massive, massive investments into those factors, right? Particularly the compute, but also the training and the architecture. And then if you have, again, AI itself that is more and more capable of driving improvements across those three factors, it's, I don't want to sound simplistic, like in my head, not an engineer, but a math econ person, like it's just math, right? And so Um, it's both that factor combined with the fact that like we are genuinely seeing exponential increases in the capabilities of AI. It's not that it's gotten so good that we're now starting to see a slowing in its capabilities. We're actually seeing that like go faster and faster. I can talk about why that might not be true. I don't want to sound like I'm making the case sort of, um, from someone who, uh, as to why to me, it seems like a pretty rational assumption. The other thing I would say actually, Ben, and then I'll, I come to the counter if that's helpful. The other thing I would say is that maybe this, I have been in the room many times where I was just like, that just seems like a fantastical assertion. Like that is not going to happen by X time period. And then by X time period, give or take a few months, like X happened. And so I also think it's, increased my belief that there is just real grounding behind a lot of these forecasts. That's much more anecdotal.

Ben Miller [00:12:37]

Yeah, that's what I was trying to get at by my question that maybe wasn't as elegant as it could have been, that when you're saying we've seen this exponential growth, it's really like you have. I have. Yeah, and some people close to it, but my mom hasn't, right? It's a lot of people aren't really seeing it change their daily lives. And so it seems, as you said, fantastical that this thing would all of a sudden become – you know, artificial life, artificial intelligence of at a level that's comparable or better than people. And then why I want to move to the, except the premise is that like, that scares people. And most people actually are, think that's like a negative.

Kanika Bahl [00:13:24]

Yes. Yeah. And just one other, like, this is an anecdotal thing, but yes, I have seen it. And yeah, Also, if you actually take a step back, when was ChatGPT? I should have known. Was it 2022, 2023? In November 2022, I think. Yeah. And at the time... I remember me and all my friends thought it was so cool that it could like write a fun bedtime story that was pretty short for a kid. Right. I remember being like unicorns and whatever. And that seemed like a big deal. And now it's doing coding for like the biggest firms in the world. Right. It is. If you think about the objective capabilities of that seeming like. just blew everyone's minds when it happened to where we are today with the capabilities. I don't think it's happening, but if you don't understand the underlying factors, maybe it's not as clear to connect the dots in terms of like what that means in underlying capabilities.

Ben Miller [00:14:21]

Yeah. And so this question or this, like, explanation that you've gone into, I think that's where most people's energy is. And that's why I want to move past it and say, okay, you know what? It's happened. You know, if Robert was right, you know, OpenAI was right, we end up having, you know... They call it AGI or ASI, but for normal people, it would essentially be human-like intelligence in a machine or beyond human-like intelligence. So what now? What does that mean? Is everybody out of a job? Does it just mean – Ten people are worth $100 trillion. Everyone else is impoverished. I feel like you actually have a lot of firsthand experience dealing with, okay, there's some people who have a lot of wealth. Probably your funders have worth billions or maybe you have funders even worth more than that. And you're going to subsistence farmers who live on $1 a day and you're trying to help them. And you're genuinely trying to help, right? But I can see it be causing resentment. I can see it be very difficult from a social point of view. Is that the right way to think about this or is that just totally the wrong way to frame this? Is AI that much further ahead of most people and is really this paternalistic dynamic?

Kanika Bahl [00:15:48]

My hope then is that it's not, right? And that we have an ability to shape how it works. moves forward and how it develops. And so that's such a big question. And I think we're, maybe I'll start with just how I think about it for the Global South and for lower middle income countries, right? Africa, Asia, Latin America. There was a point in the trust where, I think this may be like a year and a half ago,

Ben Miller [00:16:21]

Anthropic Trust, you're saying?

Kanika Bahl [00:16:23]

Oh, sorry, the Anthropic Trust, yes. The Anthropic Trust, where the leadership of Anthropic and the board started to discuss, the trust started discussing, you know, how do we actually ensure AI benefits three and a half billion people living in poverty? And it was myself, it was several other international development leaders, and we didn't have a good answer. And that that seemed really deeply troubling to me. But I went home, I talked to my husband about it. I'm like, what do I do? And I felt this sense of responsibility, but also kind of surprised that nothing was happening and asked myself, what should we do? And so I basically said, okay, well, I want to look at this really rigorously and in an evidence-based way and say, what can we do? What is possible that could change the outcomes for these countries? And So I assembled a team of experts, Dario himself, Nobel. There's a Nobel laureate, Michael Kremer, who's amazing. Kent Walker, who's a president at Google, all signed on as formal advisors. I sourced 50 ideas from very talented academics, philanthropists, NGOs, others. And I said, what's AI enabled? What's scalable? What's evidence-based? What will be cost-effective? And then we put it through a process where we looked at all of those things and we got down to 10 and then we said, we really want to focus on three big bets that could improve the lives of tens, hundreds of millions of people. And I started the work being like, are there actually things we're going to want to do, right? Are there things that could transform people's lives? And the exciting news was there were actually a number of things in health, in education, in income generation that could be transformational. And we're now working to scale those directly. We are also thinking just much more broadly about how do we build an ecosystem of funders, of other NGOs? Can we incubate new ones to take this on? And so that's a very long-winded way of saying, basically, Ben, I think there's a lot we need to do. humans are very adaptable and we have agency. And so there is an ability to shape the future in the way that we want to. And so I can speak to it from the lens that I, you know, in the hat that I wear, but I think that individuals in government can be thinking about this and companies, right? We have that ability to shape the outcomes positively and ensure that actually AI is a force for benefits much more broadly. And maybe even just taking a step back on how we could see negative outcomes.

In the Global South, I think a lot about the fact that if the US, Europe actually uses AI to create hyper-personalized education, hyper-personalized healthcare, you could get to a situation where incomes go up really dramatically and just overall livelihoods and Africa doesn't uptake it, for example, or does it at very low rates. At that point, you actually have a 10xing potentially in the inequality globally. Alternatively, if you use AI effectively, we can get to a place where, you know, there's massive doctor shortages in the countries that I'm talking about, right? Where you're actually leapfrogging the lack of high quality diagnosis in healthcare and are using it, using AI to actually improve healthcare, right? Really, really significantly. And so I think there are really positive case scenarios of AI, AGI, ASI, what have you. And I think there's some really negative cases that could occur. And I think we need to be working across society to be arriving to those positive cases.

Ben Miller [00:20:23]

Yeah, let's stay on the positive for a bit because you're saying there's a lot of positives and there's potential negatives. But you didn't say sort of how, here we are, like in a way this is meant to be like a public awareness campaign. Like what do people need to be doing, you know, that drives positive outcomes? Because I think one of the ways that people, their knee jerk is like resist data centers, you know, stop the rollout of AI. And I don't think that's what you mean. So how, maybe you could do a stylized like story of like, how does AI roll out over the next three to five years or two to three years where you seeing it hit one positive milestone after another in terms of people, in terms of people, not in terms of technology.

Kanika Bahl [00:21:13]

Yeah. Great. Okay. Let me, let me start with spending a lot of my time, which is on this work in the Global South. So as an example, right now, there's, 475 million smallholder farmers. Those farmers actually across Latin America, Asia, Africa, they earn $1,500 a year. They are very dependent on forecasts. And forecasts actually are, you know, the weather is getting increasingly unreliable and unpredictable with climate change. Right now, most of those farmers are do not get high quality forecasts. They may not get any forecasts at all. They may get them, you know, right before they need to make a decision such that they can't make a decision because it takes a long time to get physics-based forecasts available for these geographies. And AI actually transforms that picture. You can get forecasts that are much, much more accurate, much faster, more local. And it costs like fractions of pennies to deliver this directly to farmers' phones or via WhatsApp. And so we are actually working over the next, I don't think it will be two to three years. It's going to take time to get it out to tens, hundreds of millions of people. But there is a very, very feasible pathway to getting this information into the hands of tens, hundreds of millions of farmers. What's exciting about it is that it's not just getting the forecast. We used actually in India this year to reach a million farmers, we use the AI to translate dry local guidance that was text-based into voice messages, which is really helpful for illiterate farmers. We translated it into in local languages, which the LLM was able to do. We translated it into image-based. We're thinking of doing some like recursive learning on what messages are most effective. And so you could get to a place where at just an low cost, you're able to deliver high quality messaging that can improve incomes and welfare in the billions annually quite easily. And so that's a positive case. I'd say another positive case that we're looking at is clinical decision support.

AI has been shown even today that to deliver very high quality healthcare with actually like pretty good patient friendliness. There was a study that was done that actually showed that many clinicians rated the advice from the AI as more empathetic and sort of patient friendly, as well as more accurate than a typical diagnosis. And so there are many countries in Asia, Africa where diagnosis rates are, or misdiagnosis rates are upwards of 50%. So less than one in two people or about one in two people get diagnosed accurately. Rural specialist shortages are something like 80%. And then you have opportunities like India where they have a national telemedicine system where people are, you have a national telemedicine system where people are calling in and getting advice from specialists, you could use AI to make that much more user-friendly, to shorten weights, to get much higher quality diagnosis, such that over time, you can be improving the healthcare that is provided nationwide in the diagnosis and have a picture where, you know, after not too long, sorry, Ben, should I pause?

Ben Miller [00:24:55]

No, it turns out that I have too many applications open. Okay, got it. That's why it's struggling. I have like 100,000 tabs open. So I'm closing like tabs on quantum physics and tabs on like – Got it. Got it. Got it. So anyways, it's always like, oh, okay, that's why it's so give me a second here. I'm just yeah, take your time. Oh, there's my Spotify. So okay, almost almost through, you know, this is way too many open tabs. Okay, that was my own fault. Okay, so I'm the problem. What else is new?

Kanika Bahl [00:25:35]

I was wondering if it was me. I was trying to change my internet along the way.

Ben Miller [00:25:39]

No, it turned out to be me. So I didn't know either until Riverside, probably AI. Okay, good. Heads up. There we go. Okay. Okay. Started recording. Beautiful. Now we're back on track. So, yeah. So you were talking about... Oh, sorry.

Kanika Bahl [00:25:55] India clinical decisions.

Ben Miller [00:25:56] Yeah, yeah. Right.

Kanika Bahl [00:25:58]

So, yeah. So in India, you have this amazing telemedicine system that's already operational. What's exciting about it is you have hundreds of millions of consults that are happening. As an example, you could actually embed AI quite centrally and use that to dramatically improve the diagnosis, to shorten wait times, and have very rapid reach across India. And that's not just India. That could then happen across multiple systems in Africa. And so those are just a couple of the options that come to mind. But if you look at it, we're talking about... about dramatically improving diagnosis and addressing massive doctor shortages. We're talking about improving income by the billions. There's also some really exciting education initiatives. AI can provide just hyper-personalized education. To paint a picture of that, there are about 617 million students in lower-income countries who fail minimum proficiency in reading and math, and 70% of 10-year-olds can't read a simple text. There's a huge teacher shortage, and it's only getting worse. And there's now randomized control trial evidence from Ghana, Nigeria, that AI tutoring helped produce one or two years of schooling gains in just a few months of use. And so, you know, that's a pretty stunning number. Obviously, you have to go from pilot to real world. But I think what we're starting to see is just this promise of AI to, if leveraged well, to drive dramatic improvements in some of the areas that the Global South has struggled with most. And if you take that and extend it, those are the similar gains that you could be seeing for lower access regions. In the US, you can envision a world where we're able to drive just improvements in morbidity and mortality. There's obviously the massive improvements that are available in terms of scientific breakthroughs for medications, including some of the hardest to treat medications. And that applies worldwide. Actually, there's richer data in the developed world. So I think we'll see more of those gains here, at least initially. And so, yeah, I don't know, Ben, there's like a lot of really wonderful things that can come about. And it's really thinking about How do you harness those and how do you drive them? Because they don't happen totally organically. I guess maybe the PSA is think about the thing you care about and then say what is it that's needed to sort of bring it into the world.

Ben Miller [00:28:45]

Yeah, I think that in a way, if AI is limited, like not super intelligence, but it helps people make better decisions, better education, better healthcare, that people actually will love it. People are very supportive. It's where it... gets better than them. It's where it takes away their jobs. It's where they are displaced, that people are no longer supportive, right? And so I think it's the, okay, AI actually is a better doctor and a better teacher and a better decision maker than 99% of people or 100% of people. You know, most people are not on board saying that's a great idea. But I think that's what people at these AI labs or some of the people at AI labs thinks is going to happen. And I'm saying, OK, well, if that's true, which I'm saying, I'm accepting the premise. Yeah. OK, like it's not going to replace a subsistence farmer. And most people in America who are voters are like, well, you know, great. I hope the subsistence farmer does better. But what about me? What's going to happen to me if I don't have a job?

Kanika Bahl [00:29:57]

Yeah. So there, I think going back to your question of what do I think needs to happen, I would say we need to get really – there are two things to separate out. There is actually will the AI technology move as quickly as people think it might, right? Whether it's over two years, five years, ten years – It is very fast in human history. And so there's that piece. The second piece actually is how do we adapt, right? We may actually find that AI is only beneficial as an augmentation of teachers, but that ultimately an AI cannot stand at the front of the class and tell the kids to behave and understand the social dynamics and manage them, right? And so then you still need the teacher for that human element of it, right? That's just one very simplistic example.

Ben Miller [00:30:46]

Yeah, again, that's the world where I think people, society would be happy. But when I met you, you were with another person. I'm not going to say her name, but she's a senior person in AI lab. And I was asking her about, like, does she really believe that AI is just going to become super intelligent? And she was like, yeah, I totally believe that. And I was like, what about all the physical world? She's like, well, AI will build robots. And then I said to her, isn't there some limit? And she said, I guess there's a limit on how many robots it could build.

Kanika Bahl [00:31:21]

You can totally see her saying that. I was like, oh, my God.

Ben Miller [00:31:31]

And so there's these people who are really, really close to the actual, like, nuclear fusion, who believe it's going to infinity, and I'm trying to accept the premise because I think it helps — in some ways the extreme view helps, um, like, better illuminate what's — what the potential could be.

And yeah, but it's — it's also — it's, but I think mostly people — it's mostly, it scares people. Yeah. Look, I think there's a world where

Kanika Bahl [00:32:01]

I want to be honest, I am also a person who both thinks that is plausible. I actually think it's less about what the technology can do and more about what society is willing to accept and how humans operate. I have less of a question that actually AI could get there and that the robotic technology could get there. I think I have more of a question of like, will children listen to a robot, right? As an example, like where will the limits of society and human interaction reach such that actually that creates its own natural ceiling. But to like, let's say that we have robot teachers and we have robot farmers and that is our future. I think then what we're starting to pivot to is how do we create meaning and purpose? And that to me probably falls along two different dimensions. One is how do we use AI to really improve upon our lives. And I think that, again, I've spent much more of my time thinking about that for the Global South, but I actually think a lot of the tenets are very similar. How do we use it to dramatically improve our quality of life, both mental health and physical health? How are we using AI to improve education? Education not as an instrumental usage, but actually as a way to enjoy and experience life and augment our knowledge base. How do we use AI to ensure that gains of AI are equitably and fairly distributed? Hopefully, knock on wood, I think that is something we really, really need to work very hard towards and it's not a foregone conclusion. And so that's one sort of element of it. And then I think the second element in terms of painting a positive picture, is saying, okay, we all now have a lot more time on our hands. What does human purpose look like in a post-AGI world? And that's something I've thought a lot about. I have a 13-year-old and I ask myself, what will his world look like if we do have the robot teachers and farmers? And I go back to thinking about someone like my mom, actually, She's 83 and she spent her whole life caring for other people, right? She cared for us. Now she actually volunteers at soup kitchens. She cares for like her elderly friends. When they're sick, she takes them food. And actually, Ben, we act as though this is going to be this giant social experiment that's never been done. But actually for like many generations of women and actually many people across the Global South, they haven't had access to paid work, right? The expectation has been that they will be caretakers. And so they're actually, I don't want to overplay that, but there's a rough precedent for this in some ways. And I think many have found meaning in it. I think at a more societal level, what I want us to do is to really think about how we institutionalize caring and actually use it for a way to both drive human purpose and meaningful improvements. And so, as an example, one of my, a good friend of mine, Vicki Hausman, is starting the American Service Project. So that's actually meant to reimagine and scale national service in the United States for the modern era. Anthropic's launched Claude Corps. Claude Corps is going out there and getting young people and placing a thousand early career fellows in nonprofits. You can imagine like similar efforts globally, revitalized Peace Corps, more equivalent that's targeted at all ages, and really thinking about what are the physical problems in the world that we could be addressing. And so, you know, I guess if I were to sort of pull back on this, what do I see as the key pieces of a positive future? One is we're really, really thoughtful about deploying AI in ways that actually meaningfully improve people's lives. Two, we're starting to think about how we use that excess energy. I am more of the care archetype. Others may have different, you may have academics, you may have you know, various others who have other ways of continuing to like advance and find that purpose, but how do we actually create like really thoughtful institutionalized scale channels to harvest and harness these energies and prevent societal disruption? And then three, how do we work towards more equitable distribution of AI benefits? Because I think it's both important for the world and it's also just really important for global stability. I think having a massive underclass is not what we want, both from an ethical perspective, but also from just like a social stability perspective.

Ben Miller [00:37:06]

Okay, there's a lot in what you just said.

Kanika Bahl [00:37:08]

I was going to say, you asked me to rip, so I'm just ripping then. I'm more focused on the next 10 years, so I haven't spent as much of my time on the full transformation, but yeah, that's where I think I'd end it out.

Ben Miller [00:37:22]

I feel like you got – I think you put your finger on the pressure points. The limit is what society will accept. Yeah. I think – let me just stay on that for a minute because I feel like most of your experience – and this is me having read about you and talked to you a little bit like – That most of your experience in change management in the Global South is actually what society – it's about the change. It's not about the technology. It's not about actually fundraising. It's actually like how do you get this change, this medicine into the people's daily lives and the change is the hardest impediment. That's exactly right.

Kanika Bahl [00:38:04]

Yes. Now, I will say there a lot of the change is government systems. And government systems are inherently built to be risk averse, to be slower moving. I do ask myself to the extent that if it is more driven by private sector uptake, then I could see a faster uptake. I'm thinking more of AI, but I think the cap will be, is there enough societal backlash? And then you could foresee regulation coming in, for example.

Ben Miller [00:38:35]

Foresee? I guarantee. I was trying to be diplomatic. But if I think about some of the projects you've done with deworming or HIV or clean water, every single time it felt like what made it work was that you made it easy for people. Not that it was like making it easy was the magic that made it like scale.

Kanika Bahl [00:39:01]

Making it easy was the magic. Yeah. It's, it's with our deworming work. So just, you know, deworming, it costs like 50 cents per kid. And it's been shown that for every dollar you spend, the kids are in 160, their incomes improved $169 over their lifetime. So just like an unbelievable bet. And when we started this work, you know, like 2012, you know, a handful of kids were getting the treatment and we kept, you know, people kept trying to deliver via clinics. It was sort of ad hoc treatment and it wasn't scaling. And so we said, what's the simplest way to do it? And we actually started doing it in schools. And so you're meeting the kids where they're at. And then we did mass deworming days. So it was like once or twice a year. So there was a big logistics push. And so, yeah, we made it easy. We made it easy because we were meeting the kids where they were at. We made it easy because it was a single day. We made it easy because it was turnkey. And yeah, I think that's the magic. And even as we're picking like a lot of the stuff for the AI Access Initiative, one of our big pieces is, is this tractable? Does anybody need to change their behavior? If not, like that's a big thumbs up, right? And yeah, so I think simplicity is really important.

Ben Miller [00:40:13]

Yeah, because I think about the two things that people dislike most about AI, I think is they're worried about it affecting them economically and they're worried about change. That's right. The analogy I was thinking about earlier today is that for most of human history, so let's say 100,000 years, caloric scarcity was what defined our entire life. You never had enough fats and sugars. And then the last 30 or 40 years, we had sort of like completely, fats and sugars are infinite. You can have as much fats and sugars as you want. You can have as much calories as you want. And our society never knew how to deal with that problem. So I think that's what's going to happen with AI. We're going to have an abundance problem, not a scarcity problem. And then in that situation – and in this case, it will be abundance of economics. And the challenge is going to be where do you find meaning and purpose if you essentially – like most of your day, most of your time was spent earning money. You no longer have to do that. But actually, you don't know what your identity is if you don't do that. Most people, maybe not to be too un-PC, but a lot of men find their identity in work especially. It's very male, in my opinion. No work, no purpose, no joy. And that's why you said this archetype. I think if you as a carer, you're trying to care about the world. I have some friends who never think about that. They think about building, building things, or exploring the universe, whatever their thing is. And if they could somehow... See AI's encouraging that passion like it lets them get to the stars or lets them whatever it is that they can't that they that basically defines their inner um self like they would all of a sudden see it as a positive and it wouldn't be thinking about it as a zero-sum game or you know give me universal basic income they'd be thinking about like what does this enable me to do but that's I think that's hard because changing people, changing behavior is the hardest thing of all the things you've tried. You say that's the thing you try to avoid most. When you're trying to pick the dimensions by which you're trying to make positive impact, just don't change people's behaviors and you're in good shape.

Kanika Bahl [00:43:02]

That's an interesting point, Ben. Yeah. And when you think about it, there's like so many dimensions of change that we're talking about, right? So first, we're asking people to envision an underlying pace of change that has never happened in an abstract way with something that seems like a black box. That gets to our earlier discussion about the scaling laws and like why it's hard to even like fathom this, right? That's one. I think two is There's a lot of pretty legitimate fears right there before they can before you could unblock using AI to get to the moon you have to get over the fact that you don't have the job you've done that you've built your whole identity around right and so there's fear of how will society manage this right will we actually get to the UBI will the UBI be enough right and then there's the like if I do have all of that taken care of like how do I shift to that state but I think Those are all the very legitimate concerns that people have. And also, unlike you, Ben, I think that if we would deploy AI well, we are talking about abundance, right? We are talking about a different frontier of what we can be accomplishing medically, scientifically, economically. If we can get those gains distributed sufficiently equitably, I think everybody has more to benefit from first. Second, I like you. I get a lot of my purpose from work. It would be very hard for me to, you know, it is scary. Like, I think about it. I think about it for my son and, you know, what will get him up in the morning, for example, you know, when he's an adult, if there's no work. But I do think this unlock of saying, can we re-envision what we're doing with AI as a lever is a really, really interesting one. And a, you know, can we use AI as a force multiplier to do even more?

Ben Miller [00:45:04]

So, if you had an AGI or even super intelligence and you're running AI Access Initiative and you're trying to affect three and a half billion people and you have the super intelligence in your area, what do you do? Like, what do you change? Yeah.

Kanika Bahl [00:45:21]

I mean, I think I use AI as a force multiplier to actually say across all of these areas, what do we need to get done, right? And maybe it just massively multiplies what we can be doing, right? Instead of me building an institution, I haven't thought about this, Ben, so I'm riffing, but like, instead of me building a hundred person organization to take forward the health work, right? Maybe I'm able to have, I still think you're going to need that relationship elements with governments. I still think there's going to be like some human interaction. We will find out. But I use that as a force multiplier to achieve the objectives. And instead of reaching tens or hundreds of millions of people, now maybe my goal is hundreds of millions or billions, right? Instead of saying, I'm going to do it in two areas, maybe we're able to do it in four areas. And so For me, it really lets you step up the audacity of what you're trying to do. Maybe instead of saying we're just going to improve diagnosis, maybe I'm using AI to transform how national health systems push out drugs and connecting it to supply chains so that when there's a diagnosis, the medicines are there. I don't know. I think there's a lot of moonshots we start being able to go for.

Ben Miller [00:46:40]

It just seems like everything you're saying, the possibilities of AI bias towards people who have certain proclivities, willing to take moonshots, willing to change, willing to do more. A lot of people like to be told, maybe not literally, but they want a place in society, not to define their own place in society. Like the 1950s view of like, man, I just want social recognition and a job that's safe. I don't want to go out and be an entrepreneur and take all that risk and every day just be completely new and different and uncertain. Uncertainty. People hate uncertainty. But you're describing a world where the people who are most comfortable with uncertainty – are the ones doing great, but what about everybody else who feel overwhelmed?

Kanika Bahl [00:47:41]

Such an interesting question. I mean, now you're taking me out of my comfort zone, Ben, into bigger questions I haven't spent as much time pondering. I would say that then for those We need to start thinking through what solutions or what options exist, right? Maybe they're more involved in things that involve the physical world and that are more routine, more building, more. I don't know. What's your take on this, Ben? It's a really, really interesting question.

Ben Miller [00:48:12]

I mean, I'm assuming that AI, this AI you're describing, that we're accepting, the super AI, ends up being the one that decides a lot of this. And so it would, I mean, this is, some people won't like this, but it will know a lot about my son and it will say, okay, this is actually what I think my son would really like. And then it would put those things in front of my son and those things would be hard, but achievable. Yeah. And it sort of becomes like you're maybe a mentor if you're on one way or it's like a game designer, another way to think of it. Because in a world where the AGI, there's still problems. There's no such thing as a world without problems. And as long as there's problems, there's purpose for people.

Kanika Bahl [00:48:59]

That's right.

Ben Miller [00:49:01]

And so AI is helping us solve problems and pick problems that fit, you know, each one of us in a more customized way rather than like, you know, back in the 1950s, everybody had to be wear a suit and go work at IBM and that a lot of people couldn't, weren't allowed to or didn't want to do that. So it's... So I see AI as the problem and the solution, but I think that there's a lot of people who won't trust that, don't trust it. Maybe they shouldn't trust it because, as you said, it's highly contingent on a good AI with good decision makers having it actually produce decisions that are good for not just a trillionaire, but also for everyone else.

Kanika Bahl [00:49:48]

Yeah, super interesting. So it almost feels like you're talking about like a hyper customization of everyone's purpose. And so some people may actually want the purpose. Others may just love spending their day playing chess and mahjong and whatever, right? And you can create that sort of world. And others may really enjoy being in the physical world and going out and building houses with habitat or what have you, right? And so But the ability for AI to really tailor and say, okay, you actually really enjoy sort of this set thing. Let's create that like environment. It's a really intriguing idea, Ben.

Ben Miller [00:50:26]

I mean, the thing I feel like is the worst idea is the only idea everybody in AI world talks about, which is universal basic income, UBI, which 20 years ago everybody called welfare. I don't think taking everybody and making them unemployed and then giving them money from the government or an AI lab is going to have good outcomes. Is it what people really want and really need? Purpose, meaning, and work. Whether that work is... being an exhibitionist and being the most popular person on Twitter, like it's an activity. It's not no activity. I think that's a recipe for a revolution. But anyways, that's why I want to have this conversation, have people start to develop other kinds of ideas that I think are more likely to be successful that moves the society's limit up to accept change.

Kanika Bahl [00:51:24]

Yeah, I love that. I love that. Yeah, so I think that... One thing I'd call out is that your premise is that being paid is actually what gives it felt like a little bit like getting paid gives purpose. I think you can divorce those two. And again, that's why I was using a generation of women who didn't have access to work, but did their purpose was caring for their kids and their families and their communities. Right. So I do think there is. some decoupling of that. I think in our society, there's a sort of combining of being paid and status and identity and what you do with your day. And so just to say, I think there could be some decoupling of that that would be useful to move forward. And then, yeah, and I think what I was saying is really, are there mechanisms to institutionalize. And I think this gets back to the point you were making of like some people just want to know what their job is day to day, whether they're paid for it, whether they're not. And so to me, that does get to how do we create institutional solutions so that this may be like more trusted, whether that's like a national service corps, which could be deployed domestically. what was that during Hoover? Didn't they have something where they were like building a bunch of things, right? There's plenty of infrastructure we need even in the U.S. that's out there working in like less developed countries. There's a ton that needs to be done there from a physical basis. You could set up those types of core. There could be other things for like academic tracks or something. But I think I really like the idea of saying, okay, across society in this AGI world, what are tracks that people could be pursuing and seeing if we can go after it? I would say the other thing that I would be really excited about is I think uncertainty increases the resistance to change. So I think the more that we can actually have sort of a pathway of saying like, okay, we actually think that here are flags that that AI is moving at a certain pace, but also here's flags that like society is actually able to like metabolize and digest it. And we're able to get signals on jobs, which jobs are more affected, which jobs are less, which regions, et cetera. I think that will immensely ground our policy solutions. And so my push, which is like less, I guess like completely blue sky, but it's more thinking about like a transition, whether that's 10 years, whether that's 50 years, would be to argue for much better data about what is actually happening in the world, combined with much more iterative policymaking and solutioning. You and I are like pie in the sky in it, but like there could be a lot more work to be like, okay, we try this, that really worked, that was a real failure. And then actually then saying, how do we cost and actually get these to scale? And so I think that work, the more chaotic, the more uncertain, change feels, the harder it is for people to accept. And right now, I think part of what's so hard is that it just all feels ungrounded. Yeah.

Ben Miller [00:54:47]

It makes me kind of go to a different question, which is that like you can, I feel like where I'm doing is the kind of classic mistake of an intellectual. So you can, you can imagine a world where AI enables, as you said, like a special tracker of a person, it's super customized. But at the end of the day, you're describing sort of a central planning kind of concept. Now it's central planning in the sense that it's centrally planned by not people. And so maybe it works better. But the Achilles heel to central planning is authoritarianism. And people don't trust authoritarian power, like, for lots of good reasons. And so it goes back to the central question of, like, well, okay, AI might be able to do that, but who decides? Like... there's this question of some group of people deciding how AI gets built, how AI gets distributed. The, you know, like it has good for it to have good outcomes. It's, it's not just the AI labs. It's not just the governments. There's a lot of, but like how did that, you know, this is as transformative as we're accepting within five years, we would better have gotten this right. So how does that happen? Yeah.

Kanika Bahl [00:56:10]

I don't think we should accept that it's the government or AI labs alone doing this. I think we need to have individual actors, leaders, testing, monitoring, ideating, scaling ideas that work. Because I think the more my... I think the planning can happen at different levels. It can happen at a national government level, at a state level. It can happen with corporations. It can happen with what I'm doing, which is an NGO trying to figure out how to do it, right? It can be happening at a, you know, philanthropic level. But to me, it's really saying, what is the future we want to build to? And then how do we at both a public and a private level gather the information, resourcing, solutioning we need? Because I think the alternative, I guess the alternative I was sort of envisioning as a like, not that you were suggesting that, but I think there is a not crazy world where just the pace of everything takes people by surprise, the combination of like change resistance, uncertainty, black boxiness of AI results in a place where we haven't thought through those questions. And five years from now, we're sort of scrambling to, or 10 years or 15, whatever, scrambling to adjust. And I think my push is more to say, okay, at every level, every node that is out there to actually say, okay, what are the potential scenarios?

Ben Miller [00:57:40]

Mm-hmm.

Kanika Bahl [00:57:41]

Let's say I lead a massive corporation, what would be signals that actually my company is likely to undertake massive layoffs in the coming years, right? How am I responding to that? How am I actually like testing, iterating, watching what others in the industry have done? I think that's influential. And I also think the government doing that is very helpful. So I don't think you can get away from the government ideally having an important role in all of this or multiple governments. But I do think private citizens have a really important role to play as well. I don't know. What do you think, Ben?

Ben Miller [00:58:22]

There's sort of two different ways to answer your question. It's like, what can I do and what should someone do?

Kanika Bahl [00:58:28]

Yep.

Ben Miller [00:58:29]

And so I think what I'm trying to do is I think that the amount of success of AI is going to be dictated by, I think, as you said, the limit of which we can accept the change that needs to happen. And that the response of resisting that change makes it less likely to have good outcomes. And embracing the change is actually will drive the success we want. And you embrace that change by having conversations like this and having more people think it through. Not that you're thinking it through because you're going to actually be able to make it. you know, a good decision ahead of time, but you're mentally prepared for that change. Because I think that in my experience as an entrepreneur, after now 15 years, 20 years doing it, nothing works in theory. Everything works in practice. And so I'm not saying that we should come up with theoretical framework, some master framework that will hold true over the next five years. I'm saying at least I, what I can do is I think if we see change as positive, not change as negative, then we're like a long way there towards success. And there's this everything, what should someone do, someone with authority, someone who's in power? I think that's a lot harder. I don't know the answer to that one. Like, you know, hopefully they are selfless, right? Hopefully they are worried. But I don't know what the answer is.

Kanika Bahl [01:00:05]

I think that's exactly right. We talk a lot at 2AI about trying to find solutions that are evergreen, where If AGI comes in two years, the solutions are good. If it comes in 15 years, it was also good, right? And that's not always the case. I think we need to have sort of, again, we call it short and long timelines, right? We need to have a variety of timelines and do some scenario planning. But the prerequisite for any of that is accepting that there are fast timeline situations.

Ben Miller [01:00:38]

Mm-hmm.

Kanika Bahl [01:00:39]

And again, I would argue whether it's two years or 15 years, that is a fast timeline in human history. Right. And then if you once you've accepted that. And I think your point is a good one, which is like, let's do sessions like this where we're like, no, let's actually accept the premise that this has happened. That seems wild. And then like from there, we can actually start doing some planning and then. As an entrepreneur, I found exactly the same thing as you, which is, I call it like learning by contact. One of my former military colleagues share that. But the more that you're actually getting in there and grappling with the problem and having other smart people grapple with it and trying one thing and failing, and you're at least in there and you're advancing the frontier of your thinking and knowledge. And I think we just need a lot more of that along a lot more dimensions of this problem societally.

Ben Miller [01:01:27]

Yeah. I think most entrepreneurs learn similar things. And so Jeff Bezos is famous for this idea of working backwards. And the way I think about it is you have to know what you want. Yep. And so we're both talking about some world where AI is enabling abundance and purpose. And you're going to work backwards from that, but you don't have any idea how. Yes. Yes. And that's the reality and that doesn't satisfy most of my team because they want to know like, okay, now tell me exactly what we're going to do. I was like, I have no idea. Let's get going and we'll figure it out as we go. And that's where I think the society is at the moment. They don't like that. People don't like that sense of change and uncertainty.

Kanika Bahl [01:02:13]

That's such a good point. Like we even started for the AI Access Initiative. I'm like, we're going to start with a few big bets. And then as I look at those big bets, To be transparent, I'm like, this is so, it's both so important and so not fitting the overall need, but I see it as an entry point to that learning by contact and saying, okay, like now we're seeing this, we've assembled a lot of really smart people running at this with the right relationships, and now we can expand out from there. But I think the thing not to do is just sort of say, okay, I am paralyzed. I feel like we need to be moving with urgency with our best possible hypotheses And then testing against those and iterating and sort of expanding our approaches as we will and won't work. So yeah, I 100% agree.

Ben Miller [01:02:58]

So as we sort of tail, I'm sorry, as we finish up this episode and I think about the next year or two, are there certain signs you're looking for? Like how do, as people try to embrace change positively, like what are the, cause I think it's going to be faster than everybody expects. It's going to happen slowly, then all at once. And I think it's within five years. Um, so it's, you know, five years is like one college graduation, right? So, what are the things you're looking for looking at as you keep track of the pace of change?

Kanika Bahl [01:03:36]

I am looking at the capabilities, the model themselves, looking at a range of tests. I mentioned METR is just one, but like there are a range of capability assessments of AI that are out there as well as just what the labs are sharing themselves. So really tracking the capabilities of the model as well as the risks of the model. And then separately looking at signs of societal adoption. The job losses is kind of a messy figure, but I'm very closely watching data on that. I'm also listening to the anecdotal reports on where various companies, I heard one company recently that had said they're hiring a lot more AI native junior people and then much more senior people. And it's actually the mid-level people who they're not hiring, but just starting to see what some of those curves look like for job loss. But again, my expectation is that we may see much more rapid gains in the AI models and slower societal adoption, but I still don't think it'll be that slow, especially where there's profit incentives to drive that change.

Ben Miller [01:04:41]

Yeah. Tyler Cowen calls that the diffusion problem.

Kanika Bahl [01:04:45]

Yes. The diffusion problem. Yes. I'm hoping on the diffusion problem. Look, I'm hoping and I move slower. I was told the Anthropic team and I'm hoping for slower diffusion. I think society needs a lot more time to prepare.

Ben Miller [01:04:59]

Yeah. Yeah. The negative of diffusion is that I think it causes more asymmetry. That's fair. That's fair. So the people who have embraced it are seeing like, you know, a thousand times gains. And then, you know, sectors that haven't are just still in the stone ages. And so it's and that and that, you know, I don't know what percentage of societal strife is from, you know, feelings of. disparity, you know, feelings of like unfair allocation of opportunity or resources, but some huge part. And so that, yeah. So if we're going to be in a world where diffusion is the problem, then distribution is the, is the main question.

Kanika Bahl [01:05:43]

Yeah, I think that's right. I think that to me, distribution is one of the central questions. Once you get past the big things like AI model alignment and, some of the major risks from biosecurity, nuclear. To me, the distribution question is really core.

Ben Miller [01:06:00]

Once I have you, I'll give you my – I think the solution to that is universal basic ownership. It's everybody gets percentage ownership, not government ownership, but literally like Trump accounts where everybody got an account. I would give shares to every – I don't know if I can give every person in the world. But it's about ownership society, and that's really what I think changes how people feel about things, the endowment effect they call it. And so I don't know how you would do that, but I think that would be the kind of path-dependent change. In retrospect, everything downstream of that becomes possible.

Kanika Bahl [01:06:42]

Super fascinating. Yeah. It's interesting, Ben.

Ben Miller [01:06:45]

Well, anyways. Okay. Thanks for letting me also sped off here.

Kanika Bahl [01:06:49]

No, this was so fun. This was great.

Ben Miller [01:06:52]

Well, you know, I really appreciate taking the time. I mean, I think what you're doing is critically important for so many reasons, so many reasons. And, yeah, and I love that you do. I love that all the people around you, like, wanted you to do it, too. Like, it's, like, such a positive sign that people in your world, like, it's, like, went off and said this is the problem they want us, one of the main problems they want to solve. Like, that's a good sign. Looking for signs of us heading in the right direction, like, you're one of them.

Kanika Bahl [01:07:19]

Thank you, Ben. I so appreciate that. And I appreciate you doing this podcast. I think it's all, the more we can all get on the same page about the potential outcomes, I think the better this is all going to go. And so thought experiments like this of saying, let's assume what seems crazy to people and start from that premise, I think are phenomenal. So thanks for having me.

Ben Miller [01:07:41]

All right. Onward. Yeah.