Nikhil KamathIs AI a Bubble? I Asked the People Building It | WTF is AI?
EVERY SPOKEN WORD
135 min read · 27,056 words- 0:00 – 2:46
Introduction
- AMAzalia Mirhoseini
[upbeat music] [whooshing] [electronic buzzing] [clears throat]
- AKAkshay Kothari
AI may exceed the sum of human intelligence in around five years.
- SPSpeaker
We all have a tremendous amount of responsibility, and cannot let egos or incentives for profit or anything else get in the way.
- VKVijay Krishnan
Artificial intelligence agent has infiltrated an Australian Government website.
- SPSpeaker
Whoever wins AI, and we're leading by a lot, whoever wins AI, wins.
- NKNikhil Kamath
A lot was in the news about 10% of humanity dying. As somebody who's not in this world, it scares me. As people who are in this world, how do you think about it? [upbeat music] Whenever I come to San Francisco, I come with the intent of learning what has changed, what is happening, what comes next.
- AMAzalia Mirhoseini
In five years, we need 10 to the power of f- five more compute than we have today.
- VKVijay Krishnan
The one interesting thing about this, uh, space is whatever it is we are doing today will be obsolete, uh, nine months from today because all models have learned it.
- NKNikhil Kamath
The amount of money spent on AI this year is 700 billion. AI enterprise revenue is about 100 billion. Inherently, that math sounds off, right?
- AKAkshay Kothari
We're definitely in a bubble here [laughs] in terms of how we use AI.
- NKNikhil Kamath
Anna, if the world were going to die, what would you spend your last one month doing?
- AMAzalia Mirhoseini
I think I would keep doing what, exactly what I'm doing right now.
- NKNikhil Kamath
Really?
- AMAzalia Mirhoseini
Yeah.
- NKNikhil Kamath
What would you do? [laughs]
- AKAkshay Kothari
Uh, I have two young kids, so probably spend more time with them.
- VKVijay Krishnan
I have three, yeah, three young kids. I have a fourth one on the way, yes.
- AMAzalia Mirhoseini
H- how about you?
- NKNikhil Kamath
No, I think I would, like, go, like, full-on hedonistic. Self-abuse, get intoxicated, buy a car and then break it down. [laughs] [upbeat music] Thank you guys for doing this. I know it's, uh, really last minute.
- AKAkshay Kothari
Mm-hmm.
- NKNikhil Kamath
And very kind of you guys to oblige and take the time and... Did you drive from Palo Alto as well?
- AMAzalia Mirhoseini
Uber-
- AKAkshay Kothari
Yeah
- AMAzalia Mirhoseini
... but yes.
- NKNikhil Kamath
Sure. Lots of traffic today, right?
- AMAzalia Mirhoseini
Yes.
- NKNikhil Kamath
How long did it take you?
- AMAzalia Mirhoseini
Uh, about an hour and, like, 10 minutes.
- NKNikhil Kamath
Yeah, same. At least I had Aryan for company.
- AKAkshay Kothari
Mm-hmm.
- 2:46 – 5:21
Meet Akshay Kothari and Azalia Mirhoseini
- NKNikhil Kamath
Maybe we can begin by doing some introductions. Uh, Akshay, whose brother Nikunj has been on. I did not know they were brothers up until today. But also the founder of Notion. You wanna start by telling us a bit about your life?
- AKAkshay Kothari
About my life?
- NKNikhil Kamath
Yeah, why not?
- AKAkshay Kothari
All right. I grew up in, uh, Ahmedabad.
- NKNikhil Kamath
Mm-hmm.
- AKAkshay Kothari
And, um, I've at this point spent half my life in India and half my life here, so both are home. And I came to the US for college. I went to Purdue for my undergrad, and then Stanford for my master's. And I started a company called Pulse, which was a news app back in the day, maybe 15 years ago-
- NKNikhil Kamath
Mm-hmm
- AKAkshay Kothari
... that was acquired by LinkedIn. And then I was at LinkedIn for many years, including a few years in India, running LinkedIn in India for three years. Um, and I came back here to, uh, start Notion.
- NKNikhil Kamath
And Notion began when?
- AKAkshay Kothari
Notion originally started in 2013. Um, and then I joined in 2018 when we were about eight people.
- NKNikhil Kamath
Right.
- AKAkshay Kothari
I was an early investor, and then I joined full time.
- NKNikhil Kamath
Okay. Azalia, would you like to... Is that the right way to say it?
- AMAzalia Mirhoseini
Yes.
- NKNikhil Kamath
Okay. Would you like to go next?
- AMAzalia Mirhoseini
Yes. Uh, sure. I, I, I was born and raised in Iran. Um, I came here to the US, uh, for grad school. I went to Rice University in Houston, uh, to get my PhD, and then I came to the Bay Area and, uh, I joined the Google Brain team in 2016. Um, and Anna and I, who is also here, we started the same day at Google Brain in 2016. And, um, we have been working in the same companies across so many years now, about 10 years. Uh, so I was there-
- NKNikhil Kamath
Do you always quit on the same day together?
- AMAzalia Mirhoseini
Yeah.
- NKNikhil Kamath
Why is it-
- AMAzalia Mirhoseini
And it wasn't, like, accidental.
- NKNikhil Kamath
No.
- AMAzalia Mirhoseini
We were just planning our next thing together. [laughs] Um, so, uh, we were doing a bunch of research. Uh, we had a team on machine learning for systems. Um, so we, we did a bunch of projects on applying AI to do- uh, to optimize the computer system stack, the compilers, distributed systems and chip design. And our work AlphaChip was the one that got, uh, a lot, a lot of attention, uh, externally. Uh, and that was a work on, uh, training an RL agent to do chip layout optimization, and it was a- adopted, um-
- 5:21 – 12:11
AlphaChip: AI That Designs Computer Chips
- NKNikhil Kamath
Can you tell us a bit about what AlphaChip was? I read about it.
- AMAzalia Mirhoseini
Yeah. Um, so-
- NKNikhil Kamath
And explain everything to me like I'm a intelligent 12-year-old child.
- AMAzalia Mirhoseini
[laughs] Okay. Um, so AlphaChip is an RL, uh, agent that can do layout optimization for chips. Um, and that i- that has to do with one stage of chip design when you have the chip already designed, and then there's this graph of, like, many, many nodes, like millions of nodes of memory and logic. Logic is, like, and, or gates. And then you want to place, uh, these nodes, uh, onto a canvas, which is gonna be the silicon And you want to do so such that it meets certain constraints in terms of timing, power, area, you want to optimize these things. And AlphaChip, uh, was a project that was, like, one of the first times that reinforcement learning, and to our knowledge, the first time it was used in an engineering project, used in product. And we basically, what we showed is that we can de- design this AI that just like a human expert that previously did, did this job, uh, it can do so and it can get better as it solves more instances of the problem. Basically, we brought pre-- the concept of pre-training to this AI for layout optimization expert. Um, and not only it becomes better, it becomes also faster, and it can do more and more complex tasks. Uh-
- NKNikhil Kamath
And this was in 2016?
- AMAzalia Mirhoseini
We did this project in 2018, and we were able to publish it in 2020. So after AlphaChip, we, uh, I joined Anthropic, uh, again, with Anna, same day. [chuckles] And we were early employees there. We were employees twenty-five and twenty-six. Um, so it was a very s- uh, interesting time to be, um, at a rocket ship AI startup. And this was before ChatGPT, before Claude. So, uh, there was a lot of, um, kind of learning, not on the technical side, but also on the execution side. Um, and then I always wanted to be a faculty, and I had the offer from Stanford, uh, before I went to Anthropic. So after a year, I went back to Stanford and started, uh, as a CS faculty, um, and created this lab. The name, uh, is called the Scaling Intelligence. Um, and yeah. And I also started part-time at the Gemini team, went back to my old team and worked on Gemini this time. I guess last year in 2025, we started, Anna and I co-founded, uh, Recursive Intelligence. It's an, it's a frontier lab, and our thesis is that in order for us, uh, to, uh, push the frontier of AI capability, uh, we should really focus on, like, how we can bring this cost capability, uh, trade-off and, like, optimize o- on the cost capability trade-off. And one-- And our thesis on how, uh, to approach this is through chip design. Basically, we are developing this platform that enables the AI to design chips, uh, that quickly adapt to the AI workloads, and as a result, we can unlock massive performance improvements. And what that means is that we can quickly enable the compute that's needed to train the next generation of your AI or fine-tune it or serve it, um, at a multiplier faster than existing, uh, general chips. And as a result, we can push the frontier faster, and we can make things more effic- more efficient, more energy efficient, latency efficient, throughput efficient, and so on.
- NKNikhil Kamath
And when you speak about chips, what kind of chips are you speaking about? And when you say faster, faster than?
- AMAzalia Mirhoseini
Our platform is general. So we can take one workload or a set of workloads, um, and we can design a chip that optimizes those workloads. Now, for a lot of today's applic- application, those workloads, those workloads are AI models. So then the chip would be an AI accelerator that's custom for those one or a few workloads. Um, and what we enable is everything from hardware, software, co-design, all the way to silicon. And there are a lot of opportunities here to add value, whether we do things end-to-end or help a customer that has a sophisticated chip team already. We enter at a certain stage, for example, the back end, and optimize everything on the back end for them. Or we help a customer, like a new lab that has a model that has very low performance on current GPUs. We can help-- On current accelerators, we can help them, uh, kind of, um, by designing custom compute for them so they can, they can train their models better, they can serve their models more efficiently.
- NKNikhil Kamath
So in a way, if I were to summarize, you're building AI that will design chips faster and more capable than the existing people.
- AMAzalia Mirhoseini
So I would, I would say, um, we are turning chip design into a, into a... Like, basically bringing scaling laws to chip design such that we have these algorithms that scale with compute and data and they're intelligent and they can design chips for us in a way that every iteration of every stage of chip design is a lot faster, 'cause chip design has many, many stages, and it's automated. And as a result, we are-- we can co-optimize across these different stages, make chips faster and also make better chips, uh, for our target workloads. Um, and there are a lot of complexities when it comes to chip design that are, um, that we need to handle. There's a physical reality there. There are massive combinatorial optimizations that we need to solve. Then there, there, there's a lot of interesting problems that in the last, uh, about, uh, one year or so that we had to overcome and, um, we are, we are seeing like... What I can say, we are seeing breakthroughs that our team has ach- achieved both on the optimization side and also the velocity of like, um, oper- operation.
- 12:11 – 17:28
Cheaper Models, More Compute and Custom Chips
- AKAkshay Kothari
So I think the, the cost itself for the current models are going down so fast, right? So in this case, this is probably thinking like two orders of magnitude it could be further going down. Is that the bet? 'Cause I feel like, I guess if you take the current models, they're getting optimized and, I don't know, every six, 12 months it drops by a, a factor of 10.
- AMAzalia Mirhoseini
Yeah.
- AKAkshay Kothari
And so, like-
- AMAzalia Mirhoseini
Um-
- AKAkshay Kothari
... I guess this new method could be... I guess you see a path where this could go even faster, way faster.
- AMAzalia Mirhoseini
Yeah. I think, I think it's true that the current models are, um, getting more efficient. I can get to that, uh, about some of our research at Stanford on that. But at the same time, what we are seeing is that in the past few years, and kind of the projection in the next come... in the upcoming years, we are seeing something about 10 to 100x increase in AI inference demand. And even on a conservative, if you're just conservative about it, that means that in five years we need 10 to the power four- five more compu- compute than we have today. And we are not gonna get that simply by just, like, scaling our data centers and just having that much, that many more chips, right? So every chip, like, the way we are approaching is that every chip, every wafer that is used for that chip should be mu- much more efficient. So we can bring, um, order or orders of magnitude efficiency by adapting and evolving these chips for the models, so it can help meeting this growing demand of compute.
- NKNikhil Kamath
End state, would you be competing with a certain kind of company?
- AMAzalia Mirhoseini
We don't, we don't think... I mean, of course, like, uh, we are in the chip design space, like we... at, at this stage of the company. But the way we think about our, uh, our, the, our business is that we are creating new markets. Uh, we are going to enable, again, custom chips for companies that have a workload, have a I... have an AI workload that they serve at scale, and they can get a lot of efficiency, and this could be new frontier labs, or it could be robotics, physical AI, and such, that they, because of the scale at which they serve their models, they would benefit from custom chip. Um, we can also help, and this is like we are seeing just today, uh, we can help existing chip makers, whether they're hyperscalers, they're big chip design houses, new, new AI accelerator companies, all of those would benefit from a platform that help them design their chips faster. Because not only they can make their existing chips faster, they can also have new bets on the architecture and the type of chips that they can make.
- NKNikhil Kamath
If you had to wager a guess, in five years from now, who would be your biggest client?
- AMAzalia Mirhoseini
So i- it's a little hard to say what happens, uh, think about what happens in five years. I can say what we are seeing right now and what, what I believe is going to continue. So we are gonna have a wide variety of customers, again, hyperscalers, chip designers, fabs, um, memory companies, uh, f- frontier labs, AI labs, physical AIs, all of those. Again, what we are seeing that although they, even if they have existing chip design sizes, they want to bet on new areas. So all of these we can, um, we can help all of those. And that's the type of demand that we are seeing from our customers and partners right now.
- VKVijay Krishnan
Now, much like, uh, in the model space, there's, uh... it's becoming increasingly clear that, uh, you know, we need a- all kinds of different specialized, uh, models to achieve the, uh, Pareto frontier, uh, outcomes. Do you feel like w- with, uh, lowered, uh, mm, let's just say costs of, uh, and effort of, uh, chip design, do you feel like, uh, different, uh, workloads will have their own, uh, drastically different chips?
- AMAzalia Mirhoseini
We think this can, oop, enable new kind of neural architectures that right now we don't have, and we don't scale them because the chips aren't, don't exist to enable that. We think that, um, there are model architectures that could be, and training paradigms that could be drastically different from what we see today, um, and we can enable those. Basically, uh, getting out of this hardware lottery that we have. Chips are designed for models of today, but chips, as we are planning them, the chips are gonna be taped out two years or so from now. Um, and models are designed for the chips of today. We want to get out of that deadlock.
- NKNikhil Kamath
That's right.
- VKVijay Krishnan
Yeah, when we met, uh, Jeff Dean and did an event with him at NeurIPS, uh, two years ago, uh, he was quite interested in this idea of, uh, uh, being able to put together relatively irregular, uh, neural networks like the brain and so on, with a- a- all sorts of arbitrary connections. But probably they were not quite, uh, supported by the chips, and it made many of those ideas impractical. I imagine some of those ideas will be back on the table again.
- 17:28 – 20:12
What Notion Does and How AI Changes Work
- AMAzalia Mirhoseini
Yeah, absolutely.
- NKNikhil Kamath
And before we move on, Akshay, can you tell people what Notion does and what it's going to do?
- AKAkshay Kothari
So Notion is a collaborative platform for work. Um-
- NKNikhil Kamath
I use it. My team uses it. It's very prevalent in India.
- AKAkshay Kothari
And it was, uh, built for human collaboration. Uh, and over the last couple of years, it's been extended to agents. And so we really think of it as, like, the best collaboration platform where you can, um, work with humans and agents together. And increasingly, we find ourselves in an interesting place where there's just so many models, uh, out there, so many agents, where they're all looking for a central context infrastructure where you can kind of write back to and read from. So I don't know, there's like a new consumer agent every week- And if you show up to the agent, uh, you have to retrain the agent, 'cause I have to tell it again all the things I'm doing and all the things it's learning. And so I think Notion sort of fits that gap, which is you essentially the first thing you can do is connect it to Notion, and now it will know you as if you just sort of have been working with that agent for a while. And I think the same thing is being applied to the company as well. It-- I think a lot of people refer to Notion as their company brain, which is where all the context is, where all the, um, knowledge is.
- NKNikhil Kamath
What kind of tasks are most people using it for now?
- AKAkshay Kothari
Oh, it's very broad.
- NKNikhil Kamath
Mm.
- AKAkshay Kothari
Um, yeah, if we take a cross-section of our users, it looks like a company. So I think people use it for all different functions.
- NKNikhil Kamath
Like, I'm trying to explain Notion to somebody who has no context. Not context like your context-
- AKAkshay Kothari
[chuckles]
- NKNikhil Kamath
... but no idea.
- AKAkshay Kothari
Yeah, I mean, I think a lot of-- I mean, maybe the simplest way to start using Notion would be to, um, use it as a place where you take notes. Uh, used to be people had to w- type notes, now you can essentially do voice notes. Um, and so voice notes can be converted to a summary, can convert to a task, and you can then tr- follow your tasks inside the same platform.
- NKNikhil Kamath
Right.
- AKAkshay Kothari
And I think-
- NKNikhil Kamath
I think I use it for that most. You can collaborate on a task, you can assign it to someone.
- AKAkshay Kothari
And it's built for teams.
- NKNikhil Kamath
Mm.
- AKAkshay Kothari
I think that's always been true from the beginning. And, um, yeah, I think it's, it's, uh, it starts simple as someone who could just use it as a notes, and then there's also ten thousand plus employee companies who are using it as their core context infrastructure, and it's the same product. It sort of starts simple. It's almost like a video game. You start in level one, and then there are all these different ways you can go up the levels.
- 20:12 – 29:18
Vijay Krishnan and the People Teaching AI
- NKNikhil Kamath
Right. Vijay, would you like to go next?
- VKVijay Krishnan
Sure. Um, um, so let's see. Uh, I, uh, got my, uh, undergraduate degree at IIT Bombay, and I, uh, got started in, uh, machine learning around back then, like, um, sometime during my bachelor's degree back in, uh, 2003. Uh, I mostly started off working on, uh, problems in, uh, search, uh, NLP, uh, privacy, data privacy and such. Um, uh, came to Stanford for grad school, um, uh, specialized in AI, which was, uh, believe it or not, not even the most, uh, popular, uh, specialization then. I think it was number three or something behind systems databases. It might have even been number four, uh, at that time. And, uh, uh, worked a little with, uh, Chris Manning and, uh, Andrew Ng, uh, back there. And, uh, soon after that, uh, s- uh, spent a little bit of time at, uh, Yahoo and, uh, started, uh, my first, uh, company in the... We started off doing personalized search, uh, eventually pivoted onto, uh, content recommendations, and we got acquired, uh, around a decade ago. Um, a-after the acquisition, uh, spent a little bit of time with the company that acquired us, uh, decided to take a b-bit of a break, do a hard reset, uh, lo- uh, not just, uh, sort of locally optimized areas I'm, uh, already familiar with, but l-learn, uh, about some new areas and such. So both my co-founder and I, we, uh, hung out at, uh, Foundation Capital as EIRs, and we started, uh, Turing about eight years ago. So when we started Turing, a lot of the intuition was, uh, driven by the fact that we had worked with some very talented remote engineers all over the world. And, um, uh, one of the, uh, sort of, um, common pieces, uh, of, uh, advice, uh, when starting a company is that you-- uh, it's, uh, helpful to do it if, if you have some profound insights that, uh, are not as widely understood. We felt like we understood this opportunity better than, uh, most, and, uh, that was our reason for starting, uh, Turing, uh, back then. Of course, uh, you know, with, uh, tw- 2019 and later with the COVID and everything, uh, this thing became, uh, mainstream. So that was the sort of, uh, first stage of our, uh, journey, um, putting together good quality remote teams. In, uh, 2022, uh, about a year or so before ChatGPT launched, uh, OpenAI came to us, um, primarily to, um, uh, get help with, uh, evaluating and improving, uh, those earlier models on, uh, coding capabilities. At that time, uh, the model capabilities were a lot, lot more primitive compared to what-
- NKNikhil Kamath
Can you explain in very simple words how you do that, help them?
- VKVijay Krishnan
Yeah, yeah. So it turns out that, uh, to improve these models, um, you know, people, uh, at the highest levels, uh, understand that advances in algorithms are important. We understand that a lot of, uh, compute is important, uh, again, uh, for these models. But, uh, and people at a high level understand that data is important. But it turns out that the data, again, uh, falls in, uh, two buckets. One is, uh, what is widely called, uh, pre-training data. It is, uh, like the entire internet crawl, all the GitHub, uh, code repositories, Wikipedia, all sorts of, uh, things. Um, that is one component of it. Um, OpenAI, uh, wrote this, uh, uh, paper in early 2022 called, uh, it's popularly called the, uh, InstructGPT paper, which showed that a relatively small amount of high-quality targeted, um, data, uh, can substantially improve model performance. It was this insight that made it practically viable to actually hire expert humans and, uh, um, you know, uh, target their efforts, uh, toward the specific model weaknesses and then use tho-those in turn to improve the model.
- NKNikhil Kamath
You're saying better data is more useful than larger set of data.
- VKVijay Krishnan
That's exactly right. That's exactly right. Because be- be-before that, right, like before the InstructG- GPT paper, people would say, "Well, I'm training on the entire internet. What can a team of f- 40 people possibly produce over three months, uh, which will, uh, move the needle?" But yeah, it turns out that, yes, better, uh, high-quality data from the right experts, uh, um, can be disproportionately valuable, and that is what made this, uh, line of attack, uh, valuable.
- NKNikhil Kamath
And how do you get that?
- VKVijay Krishnan
Yeah. So we, uh, fortunately have one of the largest talent platforms in the world, more than five million knowledge workers from, uh, seventy-plus countries around the world, uh, software developers, uh, um, ML engineers, um, PhD level, uh, um, people with PhDs in mathematics, physics and chemistry, lawyers, doctors, all sorts of, uh, different domain experts. Ultimately, the simplest way to think of it is that if you put together a team of human experts that know meaningfully more than what the frontier models know in some area that is important, uh, there, there is a way to design tasks and to design, uh, evaluation data sets and everything to essentially transfer that human knowledge and expertise o-onto the models.
- NKNikhil Kamath
I'm still not understanding. How does it happen practically? Say you have all these software engineers and PhDs and all of that.
- VKVijay Krishnan
Correct.
- NKNikhil Kamath
What data are you taking from them and feeding into a model?
- VKVijay Krishnan
Yeah, yeah, great question. So it, uh, so, um, uh, the fir- very first thing you do is you try to c-create different kind of problems, um, that, uh, satisfy a few different conditions. These problems should be so-- beyond the reach of the current m-models. Otherwise, you're teaching it nothing new, needless to say. They have to be grounded in realism. Like, there's no point putting some artificially, uh, you know, hard, uh, problems on the table that no real software engineer would ever do.
- NKNikhil Kamath
Can you give me an example?
- VKVijay Krishnan
Sure, sure. So for example, uh, perhaps if we wish to, uh, teach the model how to debug a code in a complex, uh, code base, we put together realistic co-code bases, put together the, the right kind of bugs. Maybe the model has to, uh, uh, navigate a lot of different things, test a lot of different things along the way in order to, uh, actually fix this issue. So this could be one problem.
- NKNikhil Kamath
So these knowledge workers you have, are they sitting and creating this data all d- all day, or are they giving you the historical code base for the model to learn?
- VKVijay Krishnan
So both kinds of things happen. Uh, very often, um, yes. So there is a place for acquiring, um, specific historical code bases, et cetera. Sometimes, uh, they come with enough of the, uh, sort of, uh, um, uh, in a rich enough form, uh, that is usable by the models. Sometimes, um, uh, some, uh, element of, uh, human, uh, work may be needed on top of, uh, this, uh, data to mine, uh, specific tasks out of them, et cetera. But yes, this, uh, uh, idea of using existing data, uh, creating some kind of synthetic data with expert humans at the right place in the loop, et cetera, are, are all, uh, evolving, uh, areas.
- NKNikhil Kamath
And how does one differ from another? Say in your, in your business per se, how do you distinguish between a Mercor or a Surge or a Scale and a Turing?
- VKVijay Krishnan
Well, the, the reality is, I mean, all the, you know, all these, uh, all the four of us, uh, that you mentioned are, uh, effective, uh, players in the space. I think the, uh, the... I mean, the market is quite hungry. The demand is large. The market will sort of, uh, eat up, uh, um, high quality data in valuable areas from, um, from, uh, multiple vendors. And honestly, the, you know, all of us are rapidly, uh, evolving. The one interesting thing about this, uh, space is whatever it is we are doing today will be obsolete, uh, nine months from today because all models have learned it. So all of us are constantly reinventing ourselves every, uh, few months, substantially as, uh, companies itself. But yeah, I would say, um, uh, fundamentally, uh, uh, I'm sure at different points of time, uh, you know, different, uh, um, um, companies among these, uh, four are stronger or, uh, um, slightly weaker in different areas. But it's a very competitive space. Um, no company looks the same, uh, one or qu- two quarters later even.
- NKNikhil Kamath
What comes next in nine months?
- VKVijay Krishnan
Well, when it comes to, uh, coding itself, the more, uh, um, the ha-harder cases, the kinds, uh, that, uh, yeah, um, that can get us, uh, uh, more and more towards, uh, automating a software engineer's life's, uh, um, you know, uh, software development life cycle completely is valuable. Uh, this getting the models to be genuinely a lot better at vibe coding, where you just, uh, uh, you know, say what you want, it gets it right at the ve-very first attempt. These sorts of things are, uh, valuable. Certainly, uh, things to do with, uh, even the whole product planning trade-offs, all that stuff, getting the model to be better at all that also, and im- optimizing the rest of the software development, uh, life cycle is valuable. I expect a lot of, uh, continuing, uh, advances when it comes to getting the models to be better at, uh, enterprise knowledge work. Like today, uh, I think we have got a lot of nice wins when it comes to the software development function. But if you look at how much have these things helped to, uh, automate an investment banker's, uh, responsibilities or a lawyer's responsibilities and other things, it is relatively a lot smaller compared to software engineering. So there's a lot of, uh, things to do in the, uh, in the, even in the white collar economy. And of course, when it comes to physical AI and everything, we are a lot, lot more behind. I don't even think we have as firm an understanding of the recipes. Like in a lot of the other areas, at least we understand the rough recipe. We understand that if you produce a certain kind of data at this cost and this volume, we can expect this sort of improvement. That's not quite the case with robotics. So I expect nice advancements in, uh, all these, uh, different areas.
- 29:18 – 32:43
Understanding the AI Industry From Scratch
- NKNikhil Kamath
Right. So I'll, I'll tell you for today, I'm not gonna enter too much about your companies and ask you questions around that. This is very academic in nature. Uh, I don't know enough. I don't know much about this industry at all. Whenever I come to San Francisco, I come with the intent of learning what has changed, what is happening, what comes next. The last time I came here, it was completely different. I've been here a week now, and it seems like completely transformed in, in not more than a couple of quarters. After Anthropic's IPO, when I come back next, I don't know what would have changed. Uh, but I'm gonna approach it from that lens that you guys help me learn about what's happening in the world of AI to somebody who doesn't know enough about it. Layman understanding of what is happening. Computers followed rules. Line by line, you would tell them what to do for the last, I don't know, seventy years. Google in-invented transformers where you could guess the next word by virtue of historical data. Then came ChatGPT and things like it blew up. The world woke up. Today, the world seems to be in a race for chips and electricity. Is that broadly where you would say we are, broad picture, history trajectory-wise? And now I'm splitting it up into some sectors that I see are really hot right now, and sectors that play a significant part in the AI race. At the very bottom, I have chips and data centers. I think Azalia to a certain extent is in there. Then come the AI frontier labs, the OpenAIs and Anthropics and Googles and Metas. Then are the new labs, the neo labs like Ilya's SSI, Recursive, Thinking Machines. Are you a new neo lab in a way? Neo lab is responsible for catching what a frontier lab is missing.
- AMAzalia Mirhoseini
Um, depending on, depends on the neo lab, yeah. I think some neo labs are trying to catch the frontier on, like, following the exact same pattern, like, "But let's change the models or change our recipes." Um, I would say Recursive is we are, uh, aiming for the frontier, but in a different way with chips at the center.
- NKNikhil Kamath
Right. And then come the trainers like Turing, Merco, Surge, Scaler. And then there are the inference guys like Fireworks, Together, Base10. And then there's the application layer, the agents and the apps on top.
- VKVijay Krishnan
We are there as well.
- NKNikhil Kamath
Who?
- VKVijay Krishnan
We are there as well.
- NKNikhil Kamath
Right.
- VKVijay Krishnan
Yeah, yeah, in our enterprise side of the business.
- NKNikhil Kamath
Right. Right. So have I covered everything that's hot?
- AKAkshay Kothari
So Notion is a collaborative platform for work. Um-
- NKNikhil Kamath
Right. Okay. Explain what services would be. So we went chips and data centers, frontier labs, new labs, the data trainers, the inference guys, apps and agents,
- 32:43 – 37:09
Why Inference Companies Get High Valuations
- NKNikhil Kamath
and then service. The inference guys are just repackaging compute open source models and selling it, right? Can I ask a question? Why do they get such high multiples?
- AMAzalia Mirhoseini
Which ones? The-
- NKNikhil Kamath
The inference guys.
- AMAzalia Mirhoseini
Um-
- NKNikhil Kamath
When you value them, because it doesn't seem like they're doing anything innovative per se.
- AMAzalia Mirhoseini
Are you, are you talking about I-ISPs, like selling model tokens or selling compute? I think the, the bot-bottom line is that compute has become much more expensive in a way. Data center, like, capacity is limited, and the demand is increasing. Um, so just the price of, like, the amount of, like, GPU hour, like how much you pay, uh, for an... Even an older chip a year ago was a lot cheaper than what you pay today. Um, so if you have access to a lot of GPUs and a lot of accelerators, then the, you have this very valuable asset that you can choose how you serve it to, and how you, uh, price it to your customers, because everyone wants to use that and there's more demand then.
- AKAkshay Kothari
I think it's useful infrastructure, especially in the, in the US, you know, for US companies to be able to leverage open weight models. And I guess the multiple is probably, like, correlated to the slope. I think these things are growing crazy fast 'cause, uh, I mean, overall demand is growing. But at least this year we've definitely gone from... I don't know. I think this is true for Notion, but also true for other companies. We've sort of gone from token maxing in the beginning of the year, everybody use AI as much as you can, to like within, like, I don't know, six months, I think we've sort of moved to token economics. It's like, wait, where, [chuckles] where is all the money going? We need to have a view on that. And for a specific task, you don't need to use a frontier model. I think you can get something that's a lot cheaper, and so that's actually really driven a lot of demand, um, especially in the application layer for supporting open weight models. And then every company I talk to here is trying to build their own system as well where they can reduce their dependence on frontier models and maybe make it more task, uh, oriented, right? Depending on a specific task, I can use a cheaper model, and I think Base10, Fireworks, we're customers of theirs, and I think they benefit from that shift.
- VKVijay Krishnan
Yeah. Yeah. Yeah, just to add to that, um, yeah, I would say, uh, like a lot of the valuations are, uh, you know, are predicated on the, uh, assumption that something like global AI revenues will, say, exceed a trillion dollars by twenty thirty, basically. And, uh, the expectation is absolutely that many hundreds of, uh, billions, uh, will come, uh, on, uh, uh, with, uh, something, uh, to do with, uh, open source, uh, models. And if you're one of the leading players in, you know, helping, um, uh, make, uh, tokens of those, uh, models, uh, accessible, and you're both growing fast and the market does not, uh, expect, uh, any ceiling, uh, to be hit anytime soon, but given, uh, where this is, uh, um, likely to, uh, grow, uh, you may be valued very highly.
- NKNikhil Kamath
In the- in these layers, is there any other insight as you-- You guys are in four parts of the ecosystem. Everything outside frontier in a way, right? Actually, no inference and no frontier, but you're in every other part of this ecosystem I explained.
- AMAzalia Mirhoseini
Well, we, we build chips for inference, but yes.
- NKNikhil Kamath
Right Can you tell me some unique insights that you as insiders have that I could learn from? Anybody can take that up.
- VKVijay Krishnan
So maybe I'll, I'll share m-more of a business insight than a technical insight at the, at the moment. So I think it-- what is interesting is, uh, that, uh, you know, uh, historically there was a lot more e-uh, emphasis, uh, placed on, uh, defensibility. I mean, now it's not like it, uh, people don't place an emphasis on it. But in a world where it's been possible to grow at unprecedented rates, like, you know, eight to 10X a year, and, uh, you're able to keep that going for, uh, three, four years, uh, straight, grow a 1000X and so on, the market seems a b-bit more forgiving of, uh, some-something where, um, the, the mode may not look as, uh, robust. So for, uh, the tokens are one example. I'm sure, you know, implicit in your question was, "Hey, this doesn't seem as defensible," but the market, uh, does... Interestingly, it is-- it seems to be a lot of such cases where, uh, the opportunity to grow eight to 10X also happens per year.
- 37:09 – 41:48
The Gap Between AI Capability and Adoption
- AKAkshay Kothari
I would say, like, the big thing I'm noticing, um, I did a small trip around the world over the summer, is, um, I think the-- we're definitely in a bubble here [chuckles] in terms of how we use AI. Um, I think most people around the world that I met, um, in different countries, I think when they say they're using AI, they're essentially using it as, like, a better search, uh, tool. Like maybe better Google search, like, you know, just Q&A. I think a lot of companies in the Valley, and I think there's a few companies around the world, are really using, um, agents as, uh, as like a multiplier, right? So, like, you could have, I don't know, fifty employees, and you could have five hundred agents doing meaningful work. Um, and that agentic sort of, um, setup and capability, I feel like the diffusion is-- hasn't happened at all. I think there's a study I think I saw where, like, the top, like the top one percent is spending ten times on AI than the top ten percent, right? And so, so I think, uh, yeah, the capability is on an exponential. Diffusion is much slower.
- NKNikhil Kamath
Why is that?
- AKAkshay Kothari
I think historically, every new technology has taken like a decade or two to be absorbed by society. Uh, like two interesting examples is that I think when we got the printing press, we started to print like handwritten things, and only after a while we realized, oh, actually typography can like make it print much better. Or when we got the film camera, we started to take, um, films of the stage, uh, and only after a while we decided that movies could be moving, the camera could be moving. And so in some ways, this is like new technology that I think we're still learning how to absorb it. Um, even the agents that we're building today is very much like trying to replicate how humans do work, which is the first phase. But at some point, we'll realize, wait, that's not how they should work. [chuckles] They should probably work more natively. And I think it will only-- I think it's over incrementally we will learn that. And then I think the other reason is just pace. I think as you s- as you shared as well, like every six months, I think the Valley is shifting, sort of exponentially growing. And, um, even the Valley is barely keeping up with the research and the capability. Like, it's so dynamic. It's like the... You think this capability is extremely good at high cost, and you're like, "Wait, can I serve this to my customers?" And then you sort of have to lean in because that thing will just be like a tenth of the cost six months from now. Uh, and it'll be a new capability that you are trying to keep up with. And so I think part of it is, yeah, I think we're all figuring out, like, well, how do we work in this new world? Uh, and then research is just continues to... I don't know, every six months there's, there's a big unlock. And as, as Ilya said, there's more unlocks coming. I was just doing the math actually this morning, like Luna, which is highly underrated, it's like one-twentieth of the cost of the Sol models. So you can run Luna now two times an hour. So every thirty minutes you can run Luna for like one sweep, uh, for an entire month for five dollars. And so roughly at hundred dollars, you can have AI take all your context and run proactively for you as a heartbeat, right? And so all the stuff we've been doing with AI has been reactive. It's like you ask something, then it does something for you. But the cost is getting to a point where you can now just have this thing every minute look for work that it can do for you, where it, it gets to almost proactively just telling you, like, "Hey, these are the tasks that you probably should be thinking about." And not only thinking about, I'll go do it unless you a- unless you tell me to stop. Um, so only for hundred dollars, and it's like a really powerful model. And if you think another year from now that cost goes down by a factor of twenty, then it's like five bucks. Someone's [chuckles] literally checking for you every minute what it can do for you. And so now I think it's just the, the biggest barrier is just imagination. Like, what do we imagine this new world of abundance? Like, what kinds of work could you do? There's no reason to, uh, not be able to leverage it, right? It's, it's like, it's mostly blocked by our own imagination for what we want it to do.
- 41:48 – 46:04
Is the AI Investment Boom a Bubble?
- NKNikhil Kamath
When I speak to people here on the ground, I'm a stock market guy. Uh, I've spent my life in investing in stock markets and stuff like that, finance guy. Right now, every 20-year-old has started a company six months ago which is worth a billion dollars. And there are 22-year-olds who are trading secondaries of these billion dollars com-companies charging a carry on top of it. It all sounds like a bubble. It sounds crazy. But like every bubble, people say this time is different. If I were to go out on the street today and look at how cities have changed or how normal people have changed by virtue of AI, not that much has changed. Like you said, maybe the diffusion is going to take a long, long time. But does it scare you guys who, who are living in this world when you see this around you, all the young people coming up with billion-dollar companies from last weekend?
- AKAkshay Kothari
I think one thing that is definitely hurts in this, uh, is that it's so easy to raise that, um, I think the talent is diffused perhaps. Um, and it's, uh, definitely funny money, like funny paper money on what people sort of believe this new startup to be, right? And so I think you sort of have to compete with that. And I guess on the flip side, it's, it's amazing that you can actually go start a new company and build so much with so few people, and that there's so much dollars available. So I think there's a trade-off there where I think there is crazy amount of dollars going in, and everything looks like the next big outcome. Uh, and, and the downside is that it's diffused.
- NKNikhil Kamath
I met a few people this last week, and I was asking them, why is it not showing up in the GDP, for example? If productivity were truly to be going up in the manner that everybody's talking about it, A, why is it not showing up in the GDP, and B, whose money is it? All this money coming in, where is it coming from? Like, it's coming from allocation that could have gone to a whole host of industries and is getting allocated in one bucket per se. So it does sound scary to me, but I'm not the most knowledgeable on this, so I might be wrong and you guys might be right.
- AMAzalia Mirhoseini
Yeah, I think it's, it's definitely, uh, the case that a lot of the resources are just funneling into AI. But AI is like a... into the AI stack, like the entire stack that you were just talking about earlier from like services, applications, all the way to hardware, data centers, fabs, and so on. So, so we are definitely seeing that. But at the same time, AI is a technology unlike any other technology that can go in and consume pretty much every other application service by and large, and it clearly has this potential to create exponential value. So a lot of this kind of like assumptions and valuation and such because we are seeing for the first time the kind of like revenue that frontier labs are showing and reporting and the, the pace at which it's accelerating. We have not seen that before in any other tech- with any other technology. So it just, just like, it makes a lot of sense that this powerful technology can exponentially create a lot of value, and that is going to help everyone, like, like raise, uh, all the waters. But at the s- uh, at the same time, um... Uh, and, and things are gonna happen real quick. Uh, we don't see the diffu-diffusion everywhere, but just like in our work a year ago versus now, we, we will see how things have changed, and it's just only a matter of time how it's gonna change.
- 46:04 – 53:00
What Happens When AI Can Improve Itself?
- NKNikhil Kamath
If I were to ask you, Azalia, RSI or recursive self-improvement, can you explain that concept to me?
- AMAzalia Mirhoseini
So RSI, uh, is the idea that if AI becomes a-- capable of improving itself rather than having teams of human experts in the loop trying to improve AI to get it to next generation, then we get into this flywheel self-improvement cycle where, uh, technology improves at a pace that we have never seen before.
- NKNikhil Kamath
Can you dumb it down for me further? When you say AI improves, improves itself, how does that happen?
- AMAzalia Mirhoseini
Uh, for example, uh, instead of having teams of engineers and researchers develop the next architecture or curate the next set of training data or, uh, how you map these architectures onto the hardware, AI itself does that job. Uh, and it can do it better and faster than any single or teams of humans that can, can do. Uh, so imagine the next generation of OpenAI model is entirely trained, trained, uh, everything from data recipe, methodology, optimization, fine-tuning, all of that by the current model that OpenAI has. Um, and once we turn this process into a process that just becomes a function of how much compute you have and how effective your AI is using that compute, then things can accelerate a lot faster 'cause you're not bounded by like two teams of humans meeting with each other on a Tuesday or like experiments going right or wrong and you-- every- things can happen a lot in parallel, a lot faster, and so on.
- AKAkshay Kothari
Is it essentially basically creating the researcher?
- AMAzalia Mirhoseini
That's like one, one kind of real- realization of self-improvement. Like, but imagine it could be everything, right? Like if we have AIs that can design their own chips, they can design their own software stack, their own model stack. Um, given an application, they can cu- curate their own data and environments to optimize for. Then it's just everything is like in every stage of this, every component of AI stack, we have AI improving itself, and it can co-optimize across these stacks, and that's, that- That's gonna be, uh, there's gonna be a flywheel effect.
- AKAkshay Kothari
Do you have a point of view of, like, what the world looks like in that?
- AMAzalia Mirhoseini
Um, I don't-
- AKAkshay Kothari
Like, I mean, what does, what does knowledge work look like?
- AMAzalia Mirhoseini
At this-- Um, it's gonna be very interesting. [chuckles] Um, I don't think it's like-- Again, I don't think this RSI is like a binary, like one day RSI starts, and then we are just like... Um, 'cause it's already happening, right? AI is already in the loop of everything we do. Like, at Recursive, AI is, like, the main kind of driver of, um, all sort of, like, development that we do. Um, it, it-- And when I say AI, it's not just LLMs. There are other, other kinds of AI and optimization that we can use. Um, but I feel like in, in that future, a lot of, like, a lot of, like, human kind of involvement in a lot of tasks would be, like, in their creativity. Like proposing new problems, new applications, new ideas, and then in also enabling ways to evaluate the outcomes, and everything from that, like the problem all the way to the solution, will be automated. Um...
- NKNikhil Kamath
If I were to ask to what end when a AI improves, is there an answer in order to do what?
- AMAzalia Mirhoseini
Sorry, I didn't catch that.
- NKNikhil Kamath
Like when a, say, a AI is recursively improving itself, improvement kinda suggests that there is a definitive goal in mind. What is it improving towards?
- AMAzalia Mirhoseini
Uh, and, and that could be create different kind of, uh, the objective basically.
- NKNikhil Kamath
Yeah.
- AMAzalia Mirhoseini
For example, um, the cost of intelligence, I would say that's the metric that is probably the most important metric that we all should be chasing, right? How-- 'Cause it seems like for any level of intelligence, we get to that very quickly, and then the, we push that barrier. So the metric for RSI could be cost of intelligence. So at each level of intelligence, how we can reduce the cost, but also for any given cost, how we can maximize and really push the frontier of intelligence.
- NKNikhil Kamath
And how would you define intelligence?
- AMAzalia Mirhoseini
So intelligence is the success rate or accuracy or ability to solve tasks. So intelligence in coding means something. Intelligence in, uh, designing a new vaccine means your ability to create those successfully, and so on.
- AKAkshay Kothari
I think the models have, like, a bunch of benchmarks, and then there's new benchmarks that get added, uh, as, as they get to, like, hundred percent accuracy, right? And so, for example, Harvey added a legal benchmark. Uh, it's probably going to be like a biology medicine benchmark that... And, and, and a lot of those ones, I think current models are not capable of-
- NKNikhil Kamath
Right
- AKAkshay Kothari
... doing certain level of tasks. And so I guess in a recursive world, the AI will just keep working on those objective functions.
- NKNikhil Kamath
So this is a conversation between all of us, so please feel free to ask each other stuff. You're from three interesting sub-sectors of this.
- AKAkshay Kothari
Right. Right. Right.
- VKVijay Krishnan
Yeah. Yeah. I think one thing I'll, I'll add, uh, also, right, like, just like, uh, for example, um, uh, when it comes to, um, psychometric, uh, testing of, uh, humans and, uh, the research around intelligence, a lot of it ultimately taps into the fact that, um, the, um, what shall I say? Uh, the skill or ranking on different skills are, that are seemingly disparate, uh, end up being still, uh, substantially correlated, which allows you to, uh, isolate some common factor. To an extent, I would say such a thing, uh, does happen with the m-models too, in that, uh, you know, um, getting the models to be better at mathematics, uh, appears to improve, uh, uh, their reasoning capabilities across the board. Um, um, uh, I have, uh, I've heard from friends that, uh, you know, getting the models to be better at, uh, Olympiad-level pro-math problems, uh, um, e-ended up creating nice spillover effects by way of, uh, improvements on factuality and, uh, things like, uh, this. Um, so the, uh, yeah, I think as a result of this, uh, that is what I think makes this a portfolio of these benchmarks, uh, um, a reasonably robust comparison, at least when big improvements are made.
- 53:00 – 56:35
The Changing Work of Training AI
- AKAkshay Kothari
Do you-- The-- I'm gonna ask a question for you. Like, you said every nine months that you sort of have to, like, redo the work, the data, I guess, whatever the data labeling or the problems they are, uh, are, they're solving, right?
- VKVijay Krishnan
Yes.
- AKAkshay Kothari
Like five million people.
- VKVijay Krishnan
Yes.
- AKAkshay Kothari
I'm curious, how, how do you think about, like, okay, what's the next thing? Is it coming from what labs are asking for, or is it you sort of feel like where the puck is going? Like, how do you decide?
- VKVijay Krishnan
Uh, I think, I think both are, both are quite necessary. If you, you know, just wait for explicit, uh, asks, uh, from the labs, uh, that can be, uh, too late, uh, basically, and that makes, uh, makes it difficult to be a competitive, uh, player in the space. No, one has to, um... I mean, for us, again, uh, fortunately, we, uh, have a substantial, uh, enterprise deployment business as well, where we work with a number of, uh, Fortune five hundreds, uh, observe a lot of AI failures out there in the wild, et cetera. We try to triage, uh, our, our knowledge coming from there, coming from, uh, what we see in the latest research, our own, uh, in-house, uh, uh, research, our, uh, uh feel for, uh, where the boundaries of different models are, and what we hear from, you know, different researchers, uh, regarding, uh, likely priorities, et cetera, in future. But yes, generally speaking, we find that, uh, it is important to, uh, proactively invest, uh, to get ahead of, uh, things. Uh, like if, uh, uh, you know, speaking of competition in this space, right? Like certainly one player were to proactively invest and build exactly the sorts of things labs want, while, uh, the others are just waiting for instructions, the other person is, uh, going to starve and never, uh, you know, get any business.
- AKAkshay Kothari
So for example, right now, like, how do you decide if you should invest in, like, biomedicine more versus, let's say, legal?
- VKVijay Krishnan
I, I think a big p- uh, amount of, uh, it will, uh, definitely come, uh, based on our view on, um, what, uh, you know, how much, uh, the models can improve, how low-hanging fruit is it, what have we heard from the d- different labs, how... Y- yeah, to, to what extent is something a sufficiently common priority versus, uh, something that only maybe one or two labs might be interested in. Obviously, proactive investments are easier to justify if we feel pr- quite confident that there are, you know, uh, many people, five, 10 labs or more, uh, that, uh, uh, will likely want to, uh, pursue a specific area. So these are judgment calls. Obviously, we don't have anything like perfect accuracy on this.
- NKNikhil Kamath
What are you prioritizing right now?
- VKVijay Krishnan
So, uh, uh, definitely a lot of things around, uh, coding. Uh, I think, uh, real enterprise knowledge work, right? Like bringing a lot of the improvements, uh, that have, uh, happened with the software engineering, uh, profession, uh, to, uh, various other, uh, functions in the enterprise, uh, particularly white collar work is, uh, I would say a big part of what we do. Lot of things around, uh, frontier STEM as well. Frontier, uh, mathematics, physics, chemistry, uh, engineering disciplines and things like this.
- NKNikhil Kamath
Will this become a big job creator?
- VKVijay Krishnan
Uh-
- NKNikhil Kamath
Training AI
- VKVijay Krishnan
... I'm, I'm, I'm sure this is a... this will be a big part of it, yeah. I'm, I'm sure it's-
- NKNikhil Kamath
Do you employ a lot of people?
- VKVijay Krishnan
Yes, yes, absolutely. Absolutely. Close to 10,000 full-time active, uh, equivalents all over the world, yes.
- NKNikhil Kamath
When you say active equivalents, what is that?
- VKVijay Krishnan
No, no. So the point is, uh, uh, I'm, uh, I, I would count, uh, maybe two part, uh, two part-timers as one, basically. That's what, that's what I mean.
- NKNikhil Kamath
Right.
- VKVijay Krishnan
Yeah, yeah. Because, uh, uh, depending on the kind of task, there are some tasks where, um, one might be, um, working full-time on something, uh, for, uh, many months on end. There may be a lot of other things that are bursty. Like, for example, uh, the kind of examples you gave me, right? Like, uh, some biomedical and legal. These may be the sorts of things that are, uh, perhaps a bit more bursty, not, uh, something, uh, um, specific companies may be very interested in these things, but, um, if you're a frontier lab, this would be one of many priorities.
- 56:35 – 1:00:38
Why Silicon Valley and China Lead the Race
- NKNikhil Kamath
Can I ask a stupid question? I mean, all of them are stupid, but this one is a little bit more stupid. Like, back home, where I come from in India, I always wonder why is all the AI innovation happening here and in China. Like, I've interviewed a lot of these guys, like Dario and Sam and Elon and all of them. I don't know if they're aliens or exceptionally s- like superhuman intelligence or anything like that. I don't know. I don't think so. But... Or for that matter, you guys. Why are you able to build these New Age companies sitting in this region? Whereas, say for me back home, I'm told you can't do it, or for most of the world outside of maybe the, maybe the Valley and China. Why is that?
- AKAkshay Kothari
I think one, one anecdote that I feel like is very true is that, um, I think you take a lot of these people, and you, let's say, you know, crazy idea. You're anywhere in the world, you're at a party, and you say this crazy idea, and people will be like, "What is this guy smoking?" and sort of move on. I feel like Silicon Valley is one of the only places [chuckles] where you have this idea, and even at a party, people will be like, "Okay, tell me more." And I think there's the curiosity that anything is possible, and then there's also the ability for investors to willing to fund crazy ideas that is quite unique.
- NKNikhil Kamath
You think that's the main reason, risk capital?
- AKAkshay Kothari
I think it's... Yeah, it's, it's a huge enabler. Um, I mean-
- NKNikhil Kamath
There are many people willing to risk capital here versus rest of the world?
- AKAkshay Kothari
I remember when I was at Stanford, um, one of the things that what struck me, I, I know we, we have that in common. Uh, but one of the things that struck me when I was doing my, uh, master's was that, uh, like starting a company was everyone's plan A.
- NKNikhil Kamath
Mm-hmm.
- AKAkshay Kothari
Like, if you got a job, like people would look at you as if like, "Sorry it didn't work out for you." [chuckles]
- NKNikhil Kamath
[chuckles]
- AKAkshay Kothari
And so it's something in the air. Like, I think people are wired for it, and part of it is like you're surrounded by people who are willing to support you. Like, even like failing in the Valley is, is like means that you're probably increasing your probability of being able to raise again.
- NKNikhil Kamath
So it's an attitude thing. Like I was in China earlier this year. I stayed there for about a month. There is different. There, in a way, the government will subsidize you for a decade if your goals align with theirs.
- AKAkshay Kothari
Mm-hmm.
- NKNikhil Kamath
And the government is the v- venture capital firm backing innovation. Here, it's more private capital.
- AKAkshay Kothari
Top-down, bottom-up.
- NKNikhil Kamath
Yeah, but it's both coming down to risk capital. Is that the differentiation, somebody willing to put money behind what might sound like a stupid idea initially?
- AMAzalia Mirhoseini
Uh, that's definitely a big factor because, uh, a lot of the work we do is a function of how much compute we have. Some of, some of us might be better at, at how we use compute, but having this foundation of this expensive, very valuable compute is gonna be very important. Um, and just like we don't see frontier labs like in the US outside of Silicon Valley, so maybe there's just that proximity to your friend or who has worked in a frontier lab or maybe do have worked in a frontier lab before and now you want to have your own lab and your own bet. And just like the, the, the culture and the vibes are gonna also be an import-- are, are also matter and then is an important factor. But it's like, less like, "Oh, you can't do it of- wi-without this," but the capital and the compute
- 1:00:38 – 1:07:23
Where Would You Build an AI Business?
- AMAzalia Mirhoseini
is gonna matter.
- NKNikhil Kamath
If I were to draw up the courage to start something and take a chance, let's say I have $100 million and I'm gonna go raise the rest. Would I try and start-- I know I can't start a frontier lab with that amount of money or even-
- AMAzalia Mirhoseini
A hundred billion dollar?
- NKNikhil Kamath
Hundred million.
- AMAzalia Mirhoseini
Oh, okay.
- NKNikhil Kamath
I-- Here everything is billion, but [laughing]
- AMAzalia Mirhoseini
[laughing]
- NKNikhil Kamath
I wish I had a hundred billion. Unfortunately, no. But if I had $100 million to spend and start a business, or $100 million to raise and start a business, do you think I should be in the chips and data centers, the power plants, the frontier labs, the Neo lab, the new labs, the training, the inference guys, or the application or service like you rightly added? Where would I start a company? If these are all the hot sectors, and this is, this is largely the entire ecosystem, right?
- AKAkshay Kothari
Yeah. Yeah. Mm-hmm.
- AMAzalia Mirhoseini
I, I think it all depends on, like, what, what cool idea you have. Like, if you have an idea, say, for a new model, like, that gives you a compute efficiency multiplier over existing, like, very traditional neural architectures, then by all means, um, that-
- NKNikhil Kamath
Is that everything, compute efficiency?
- AMAzalia Mirhoseini
It's gonna be a, a huge factor, especially in AI training and AI development. Because if you're spending 100 million, uh, on a model, but you're saying, "I'm 10X more efficient than my competition," then this is as if you're spending a billion dollar on your model. So then you're gonna have same quality as a lab that has a billion dollar in capital to train their model. Uh, so that's on the modeling side. And on the, on the-- So the way I think about, like, how things are moving, um, frontier models, like, are a few months, maybe six months, maybe one year ahead of, like, open models. And this ga-- And this has continued to be true. For every new benchmark, for every new level of intelligence defined by those benchmarks, other models are gonna catch up. Uh, and what's gonna matter, uh, is that, again, going back to the intelligence per dollar or intelligence per cost, and if you can serve and train at that, at a better efficiency for that, that's all you need to have an edge over your competition. Now, for this intelligence per dollar, you can come in at different places. You can come in in the chip layer, in the software layer, in how you design your racks in a data center, or how you curate your data or your model. Like, all of these can shift that frontier for you, and that could be... A-and that's how you can win.
- NKNikhil Kamath
In the last internet boom, the application layer seemed to capture a significant portion of the value. You wouldn't focus on that today? Like, if I were to build user interfaces at the very end?
- AMAzalia Mirhoseini
I, I thi- I think that's also a great a-area. That's not as, like, if you build an application that becomes really popular, just like we were seeing just recently with some startups, I think that, that distribution and that, like, uh, user base is gonna matter even now. Um-
- NKNikhil Kamath
Does that worry you, Akshay? Like, if Notion is sitting on top of I don't know what it's sitting on top of. Say it's sitting on top of Anthropic, and there's a connector where I use Notion on Anthropic. It doesn't worry you that all these frontier models are building competing products with pretty much anything that's doing well?
- AKAkshay Kothari
I think the interesting thing right now is, like, I feel like everybody is competition with everybody else. Like, 'cause, 'cause the things are converging. Like, what you build is converging. Um, so in our case, actually, we built, like, the, you know, we built our MCP and our APIs, CLI roughly, like, a year ago. And, uh, there was a lot of debate internally. It's like, "Wait, if we give up all of this, like, is that a risk to us then?" There's a little bit of a leap of faith, for sure, because I think if we don't do it, somebody else will build that with better support for that. Um, what we found actually is interesting is that, um, the people who are using us, even without actually visiting Notion, are probably two, three times more active, uh, and actually, like, many times more valuable in terms of being able to converting, uh, to, to a paid customer. And I think one of the reasons that's driving that is because I think people are just using so many different agents. Uh, and so if you're using Claude, if you get too close to that ecosystem, and if your company decides that now I actually wanna use the new ChatGPT product, like, you can't talk to that. You can't leverage all these different models. You can't pull that context out easily. And so essentially, like, we found ourselves in an interesting spot where people are looking for a neutral playground-
- NKNikhil Kamath
Right
- AKAkshay Kothari
... where they can put their stuff, and then they can use all these different agents.
- NKNikhil Kamath
Right.
- AKAkshay Kothari
And so-
- NKNikhil Kamath
That's a bit like blockchain territory, right?
- AKAkshay Kothari
It's a bit of, like, it's a, it's-- infrastructure is-
- NKNikhil Kamath
Yeah
- AKAkshay Kothari
... what we're building, where, uh, in a world where you use multi-models and multi-agents-
- NKNikhil Kamath
You take your context with you wherever you go.
- AKAkshay Kothari
Yeah. It's like a passport. You just show up-
- NKNikhil Kamath
Mm
- AKAkshay Kothari
... with an agent, and then you can start using that. Um, but I think to your, to your previous question, I actually feel like, uh, probably all-- I can probably make a bull case for all of those, uh, layers of the stack. Um, I don't know. I feel like-
- NKNikhil Kamath
Which would you make the most bull case for?
- AKAkshay Kothari
I think we're probably all biased in our own, uh, layer where-
- 1:07:23 – 1:17:40
AI Agents and the Productivity Debate
- AKAkshay Kothari
for what they do.
- NKNikhil Kamath
The demand is almost not going up by virtue of people using it more, it's going up by virtue of agentic outcomes, right?
- AKAkshay Kothari
Mm-hmm.
- NKNikhil Kamath
It's not that me as a end user is using... clicking more on a AI product.
- AKAkshay Kothari
You're exactly right. I think one interesting data point we found is that, um, I think the starting of the year, maybe like, I don't know, 1% of our, uh, users were agents, uh, interacting with Notion. I think that number is 30%-
- NKNikhil Kamath
Yeah
- AKAkshay Kothari
... today. Like a year from now there'll be more agents than humans.
- NKNikhil Kamath
I met someone for breakfast this morning who was saying as much as 50% of all clicks now are agenting-
- AKAkshay Kothari
Yeah
- NKNikhil Kamath
... on the internet, and it'll go up to 90%.
- AMAzalia Mirhoseini
Yeah.
- AKAkshay Kothari
And I'm really excited about that world 'cause I, I think there's many more agents we can serve than humans, uh, in terms of like the capability, the infrastructure we provide.
- NKNikhil Kamath
Why does it excite you though?
- AKAkshay Kothari
'Cause I think like if you were building a software application here in the past, you'd be like, "Oh, maybe I can get to 100 million knowledge workers." Now I can get to a billion agents serving all sorts of different things that they could-
- NKNikhil Kamath
But your margins will go down because the user might not have been comparing Notion with something which is very similar, but maybe an agent will, at price.
- AKAkshay Kothari
True. And I think w- uh, again, like we're providing the infrastructure and I think, I think the customers will choose whether agent is doing useful work. And we provide agents as well, but I think you could bring your own agent, you can leverage all these different parts. Uh, but it... I think it's already true for a lot of companies, there's probably like five times more agents at work than humans. And it's not actually replacing humans. It's very much like augmenting human intelligence.
- NKNikhil Kamath
Yeah.
- AKAkshay Kothari
Like each person is able to now go from doing the work to, I think as you were saying, reviewing the work and approving the work and imagining new things to do, which is quite exciting. Why it doesn't show up in the GDP numbers is also interesting, but I, I think it's probably because only the very few companies are using it that way. Like the vast majority are still not leveraging it.
- NKNikhil Kamath
You know, last year I was interviewing Sam, OpenAI Sam, and asked him what stock to buy. He said productivity is going to go up so much you should buy companies like Hermès and LVMH because as people have more means, they'll want to differentiate by aligning with brands which set them apart.
- AKAkshay Kothari
That seems to have gone down actually. [laughs]
- NKNikhil Kamath
No. They've crashed. [laughs]
- AMAzalia Mirhoseini
[laughs] Did you buy those?
- NKNikhil Kamath
They're down like 50%. So while we have this debate and everything in theory should be doing more and more productivity, the world seems to be in a really terrible place with not that much productivity gain at all.
- VKVijay Krishnan
Yeah, yeah. Let me in fact bring some counterpoint to the table and maybe this will also, uh, uh, um, though I feel, I, I do feel, uh, genuinely that, uh, there are exciting opportunities across the entire, uh, set of, uh, problems you mentioned. Y-yes, uh, in fact I would, uh, actually, uh, extend what you said about GDP to one more, uh, thing.
- NKNikhil Kamath
Mm.
- VKVijay Krishnan
Right? Um, point me one, uh, Fortune 500 company that, uh, I, I mean, not the people who are selling AI infra, but you know, somebody who's an end consumer, like a Pepsi or a Coke or a manufacturing company that's meaningfully mo- uh, grown top line revenue or reduced costs. And if you've done neither, you know, have you really, uh, created, uh, value at all, right? And I do think there, um, uh, there is, um, uh, likely going to be a lot of, uh, work and a lot of opportunity and the kind of products and solutions that, uh, actually take these, uh, uh, AI capabilities, uh, to these enterprises. And actually, um, I mean, we are talking about, uh, AI can do 90% of the work, okay? Help the companies become leaner, help their, or h-help them with their same, uh, uh, team and everything, uh, grow their, uh, top line, uh, substantially if, uh... Um, a-and, uh, a-a-and I think a lot of that will require work because I, uh, I do believe that, uh, and this is something we observe as well, um, it, it's, it's very easy to, you know, uh, when you're a tech person, you're a software developer, you're AI built, et cetera, to, uh, not have a good, uh, theory of mind for what the rest of the economy, uh, uh, is like. Uh, I think we are in a situation where first, a lot of these problems are not easily auto verifiable, so the models are just not that good, uh, there to begin with. And, uh, a lot of work is, uh, needed to, uh, get that, uh, better. Second, uh, the individuals are just, uh, fum- uh, fundamentally less, uh, technical. Uh, it's not like a lawyer is going to, uh, every week, uh, do evals between all the different models, which is what they have to do today, uh, right? Like, uh, somebody, [laughs] uh, 10 new models are dropped this week. You're really, uh, expecting, uh, them to do evals between, uh, 10 different models, compare small and large models, draw, uh, [laughs] of, uh, you know, where, uh, something, uh, their workflow is on the frontier. It's just not gonna happen, and this is where I do believe a lot of application companies, a lot of AI transformation companies, I expect some of these, uh, uh, this, uh, P style, uh, plays, uh, to do with, uh, um, total AI transformation of many companies, et cetera, have, uh, very exciting opportunities for them.
- NKNikhil Kamath
It does seem a bit circular, no?
- AMAzalia Mirhoseini
And that's what? Circular?
- NKNikhil Kamath
Circular. Like one company in this world itself, one company selling to another, booking revenues on another. It seems to be going around, the money that's being raised.
- AMAzalia Mirhoseini
There's definitely, uh, some of that, for sure. But there is more, it's more than that, right? When, um- I think the use of AI across like all sort of industries we are seeing that's rising, and it's like, it's just the beginning of it. Like once you use it and you understand how good it is, you're not gonna stop. You're gonna-- You just wanna get-- use more, use more and more of it.
- 1:17:40 – 1:24:23
Are We Losing Our Own Thinking Skills?
- VKVijay Krishnan
uh, industries, uh, adopt AI, uh, appropriately.
- NKNikhil Kamath
You know, when it comes to context specifically for what you do, I feel like tomorrow if-- I used to write a lot. Like I used to write articles. I used to-
- VKVijay Krishnan
Mm-hmm
- NKNikhil Kamath
... I had to give a speech every now and then. I used to be able to write it. Strangely enough, now I've become stupider.
- VKVijay Krishnan
Mm-hmm.
- NKNikhil Kamath
Uh, I think Socrates said this once, that writing is bad because then memory will go away.
- VKVijay Krishnan
Mm.
- NKNikhil Kamath
I feel like that's happening in its own way to me, where even if you ask me what is forty-five into forty-five, like maybe five years ago I knew how to calculate. Right now I don't. In five years from now, I will be stupider from that lens than I am today. But in the future, if we, we adopt all these crutches along the way, do you think the core skills of a human will go down? Like evolution happened because we needed to remember more, we needed to kinda do more, we needed to use our brains more. As we progressively use our brains lesser, what happens?
- VKVijay Krishnan
So I think, uh, so I would maybe split this into two parts. Uh, a- again, this is very, very speculative, uh, right? Uh, there's different, uh, views on this, uh, subject Uh, so there's no question. Obviously, if you use it less, uh, uh, specific skills you're not gonna use. Um, like I, I remember my dad, uh, still has, uh, the, the slide rule and things like that that used to be used in the, uh, '70s. Uh, nobody uses that. I'm, I'm sure various, uh, skills like that have, uh, probably been, uh, uh, lost and, and probably are not even, uh, relevant. As adults, it is, uh, less of a concern to me. You know, this sort of, uh... I mean, you're very much talking about some of these loss of the skills as adults. I'm pro- more concerned about that in the, in the context of, uh, uh, pedagogy, because, I mean, the reality is, in an economy, you get paid if you add some marginal value. Now, um, a child that is growing up, if they don't pick up any, uh, basic skills whatsoever, it is very difficult for me to see what value they can add over an, uh, um, over a model, uh, uh, at all. Uh, there has to be some, uh, non-zero, um, place, uh, in the chain where, uh, you have to be better than the models. And I do think in pedagogy, this is becoming a, a severe, uh, i- issue for su- uh, sure.
- NKNikhil Kamath
And another thing, if I ever were to write something now-
- VKVijay Krishnan
Mm-hmm
- NKNikhil Kamath
... I would make sure it's behind a paywall of some sort.
- VKVijay Krishnan
Hmm.
- NKNikhil Kamath
And not allow for a model to consume and regurgitate it. Because whatever you write that it's able to consume freely is kind of wasting the time I took to write it, and I think more people will think like that, no?
- AKAkshay Kothari
Why, why do you say that? Like, wouldn't it be better if your idea traveled?
- NKNikhil Kamath
I think it used to be better for my idea to travel because I would benefit from my idea having traveled. Right now, my idea is not traveling as much as the principle of my idea with the anonymized author that will travel if a model, model were to consume and regurgitate. So I don't benefit so much from it.
- SPSpeaker
I'm gonna bring Anna in.
- NKNikhil Kamath
Yeah, yeah.
- SPSpeaker
Yeah, I think it was-
- NKNikhil Kamath
You can exchange whenever you like.
- SPSpeaker
Okay, okay. I'll see how this goes.
- NKNikhil Kamath
But you guys can continue.
- VKVijay Krishnan
Yeah, yeah.
- NKNikhil Kamath
Hmm.
- VKVijay Krishnan
I mean, I suspect though there is some element of, uh, um, genuine, uh, almost a tragedy of the commons, uh, kind of a thing, right? Like, you may feel like, uh, merely keeping out this doesn't help because other, uh, people will have, uh, somewhat similar ideas anyway, and this doesn't, uh, get me too much. Like, uh, um, in the past, right, like people, uh, talked about, uh, how, uh, these social media companies benefit from The New York Times and whatnot.
- SPSpeaker
Hello. Hello, guys. [laughs]
- VKVijay Krishnan
Yeah. I suspect, uh, it-
- SPSpeaker
What was that?
- AKAkshay Kothari
It's like you guys are-- You can replace each other.
- SPSpeaker
[laughs] That's right.
- 1:24:23 – 1:30:21
Meet Anna Goldie
- AKAkshay Kothari
way.
- NKNikhil Kamath
Anna, you look slightly different from you did f-five minutes ago. [laughs]
- SPSpeaker
[laughs] Costume change.
- NKNikhil Kamath
You want to start by telling us a bit about yourself?
- SPSpeaker
Yeah, sure.
- NKNikhil Kamath
Or context.
- SPSpeaker
I'm happy to. So I guess my background, I studied computer science and linguistics at MIT. My goal was always to do language modeling. I think at that time, it was a different paradigm. Um, I did my master's in, like, dialogue systems at MIT, and I, I did my PhD at Stanford in, like, reinforcement learning for chip design and also for LLM post-training. Um, I think-
- NKNikhil Kamath
The one thing I'm learning is if I had a child, I should send him to Stanford.
- SPSpeaker
Why Stanford? Why not MIT?
- NKNikhil Kamath
So all of you have gone there.
- SPSpeaker
Yeah?
- NKNikhil Kamath
You like MIT more?
- SPSpeaker
I think they're both good.
- AKAkshay Kothari
[laughs]
- SPSpeaker
I think, um, Stanford is more entrepreneurial.
- AKAkshay Kothari
Right.
- SPSpeaker
Which I think is also valuable.
- AKAkshay Kothari
I went to Stanford 'cause MIT didn't take me, so.
- SPSpeaker
[laughs]
- AKAkshay Kothari
[laughs]
- SPSpeaker
That's a good-
- NKNikhil Kamath
Is that how it is? MIT's harder to get, get into?
- SPSpeaker
I think... I remember when I, when I was visiting MIT for the PhD admissions weekend, they were saying that if it snowed during the visit days, yield went down 60%. So I think MIT is spikier. Like, we had tunnels that you would go into under the, in the winter to get between the buildings. Stanford is a nicer place.
- AKAkshay Kothari
That might be interesting stat for people who get into both, what they pick.
- SPSpeaker
[laughs]
- AKAkshay Kothari
I wonder what the preference. I guess it depends on the weather.
- SPSpeaker
I think it's like MIT is about optimizing for pain, so, in all dimensions.
- NKNikhil Kamath
Stupid-
- SPSpeaker
If you wanna suffer
- NKNikhil Kamath
... stupid person asking very smart people a question. Were you guys always smart?
- 1:30:21 – 1:34:48
The $100 Million Business Question
- NKNikhil Kamath
And I'm gonna ask you a question. I am a unintelligent-
- SPSpeaker
[laughs]
- NKNikhil Kamath
... man- [laughs]
- SPSpeaker
[laughs]
- AKAkshay Kothari
[laughs]
- NKNikhil Kamath
... if that's a word. If it's not, you can correct me. And I have managed to borrow $100 million from somebody.
- SPSpeaker
Okay.
- NKNikhil Kamath
And I wanna start a business around the AI ecos- ecosystem. It could be, uh, what he does, what he does, what you guys do. I can't do a frontier lab because I only have 100 million. What should I do? What should I focus on? Which sector?
- SPSpeaker
I think it's a really a function of you. Like, what are, what is-
- NKNikhil Kamath
Yeah
- SPSpeaker
... what are you deeply passionate about?
- NKNikhil Kamath
I said nothing.
- SPSpeaker
What do you deeply understand?
- NKNikhil Kamath
Nothing. Nothing. I'm passionate about-
- SPSpeaker
[laughs]
- NKNikhil Kamath
... turning a return on the capital employed. [laughs]
- SPSpeaker
I guess, like, I don't know. For me, the way I thought about it was, like, okay, what is something that only, like, I have to... Like, has to happen, and I have to do it, and do I have the right co-founder? And all of those things have to be true, and then I do it.
- NKNikhil Kamath
What if I don't know the answer to that question?
- SPSpeaker
Maybe you shouldn't start a company then. [laughs]
- AKAkshay Kothari
[laughs]
- NKNikhil Kamath
Good advice. [laughs] I think it's such an interesting way to-
- AKAkshay Kothari
Become a VC?
- NKNikhil Kamath
Uh-huh.
- AKAkshay Kothari
Become a VC?
- NKNikhil Kamath
[laughs] I feel like the VC bubble is so bad right now, it's, it's getting to the point of ridiculous. I know VCs who are trading secondary like they're selling potato chips. Like, I buy from you today, sell for double to Anna tomorrow. Anna then sells it back to you, then you sell to Akshay day after tomorrow.
- AKAkshay Kothari
[laughs]
- NKNikhil Kamath
It's just, like, ridiculous.
- AKAkshay Kothari
I'm left with, I'm holding the bag. [laughs]
- NKNikhil Kamath
You know the greater fool theory? I don't know when it ends.
- AKAkshay Kothari
Mm-hmm.
- 1:34:48 – 1:37:18
What Google Got Right—and What It Missed
- SPSpeaker
people were collaborative with me, but...
- NKNikhil Kamath
What is Google missing out on? They seem to have... I mean, Google is easy to track because they're listed. Alphabet is a listed company.
- SPSpeaker
Mm-hmm.
- NKNikhil Kamath
What did they miss out on in this last decade that they could have done better?
- SPSpeaker
I think they did a lot of things right, so I'll say that first. I think they have brilliant people there. They have really strong researchers, really strong engineers. They have compute. They made the bet on TPU really early, and I, I think that was smart. They have this distribution, they have this ecosystem, they have data, um, so all of that is great. I do think that culturally, sometimes things move more slowly than would be ideal, and I think it's sometimes a function of, like, if you're a 200,000-person company, the communication costs, the, the ability to mobilize, it's much, much harder. And I, I remember as, you know, as Ali and I were there when Transformer was discovered, like, we were there in the middle of the night, we were working, and we heard, like, Lukasz Kaiser, like, shout because they got a high Blue score in the middle of the night, like, high machine translation capability. Like, you know what I mean? Really special things happened at Google, at Google Brain. But then the question is, like, you know, leveraging that. And we also saw that internally, people trained LLMs inside of Google. Like, there was this project called Meena that was a precursor to LaMDA, which was a precursor to Gemini, and Noam Shazeer and Daniel de Freitas, they wanted to release Meena, and Google was concerned about the PR, basically. Like, that there are journalists who are dedicated, their entire careers are about looking for mistakes that Google made and writing articles about those. So Google was terrified that if they released the model and it said anything wrong at all, that they would be crucified. And so they, they didn't release the model. And then Noam and, um, Daniel left, and they started Character, and then they came back, and then I guess they left again. But-
- NKNikhil Kamath
While it might be bad for Google, maybe it's good for entrepreneurship, right?
- SPSpeaker
Exactly. I think that the... I think Google's philosophy here is that by bringing smart people together, by treating them well, by letting them, you know, basically share their innovations with the world, that, like, it raises all the boats, including Google's.
- NKNikhil Kamath
Did AlphaChip have something to do? Was it, like, a precursor to TPUs?
- SPSpeaker
Um, I think TPUs became, existed before AlphaChip, but AlphaChip was used to train, to, to generate chip layouts for the next version of TPU, and then we trained AlphaChip on those TPUs, and then we, we've been taped out now in five generations of TPU, Axion TPUs, Pixel phones, Waymo chips, external adoption
- 1:37:18 – 1:39:59
AI Chips and NVIDIA’s Advantage
- SPSpeaker
by MediaTek.
- NKNikhil Kamath
What is NVIDIA's moat and- I mean, I, I don't understand the tech at all, but I, I can understand that they have operating system, CUDA or whatever, which has context, significant amount of context.
- SPSpeaker
Mm.
- NKNikhil Kamath
No?
- SPSpeaker
I would say that their moat has evolved over time, so I don't think their moat currently is CUDA. Like, I don't think Jensen said anything about a CUDA moat in more than a year now. And that, that's kind of... I- it's interesting that at this time when models are allowing smaller teams to move more quickly, also models are allowing, you know... Models can write directly in assembly.
- NKNikhil Kamath
Mm.
- SPSpeaker
And so that means that maybe this, this, these abstractions that were created for humans aren't as important anymore. And so that moat for this particular incumbent has weakened. At the same time, they have an entire ecosystem. They've created racks. They have this NVLink fusion program. They have created, like, you know, open source models like Nemotron. And, and so like, you know, CUDA is just one part of the overall system.
- NKNikhil Kamath
So you're saying in a way, like, backward and forward integration across the value chain, that's the moat now?
- SPSpeaker
Yeah.
- NKNikhil Kamath
And scale.
- SPSpeaker
Yeah, scale, and, like, just reliability. People know that n- the GPUs, NVIDIA GPUs work well. They can trust them.
- NKNikhil Kamath
If you were to be a betting person, would you buy them at 30, 35 times forward earning today based on this moat?
- SPSpeaker
I, I think NVIDIA's gonna do well.
- NKNikhil Kamath
Yeah? And if you had to make a binary pick between Alphabet or NVIDIA? [laughs]
- SPSpeaker
[laughs] Um.
- NKNikhil Kamath
At today's valuation.
- SPSpeaker
I mean, I think I, I think NVIDIA's gonna do well. I also believe in Alphabet, but...
- NKNikhil Kamath
But do you pick NVIDIA?
- SPSpeaker
I, I, I, I think... I, I, I won't, I won't say. [laughs]
- AKAkshay Kothari
I think we did get the answer.
- NKNikhil Kamath
I think you said it. [laughs]
- AKAkshay Kothari
[laughs]
- SPSpeaker
[laughs]
- NKNikhil Kamath
That's okay. Do you have a view on the dumb person, me building something with 100 million, like something more specific? If I'm not passionate about something and I'm starting from, like, scratch, is there a business opportunity if you had the time you would do, but you don't have the time?
- SPSpeaker
Oh, there, I mean, there's-
- NKNikhil Kamath
Yeah
- SPSpeaker
... so many opportunities.
- NKNikhil Kamath
Tell me one so I can steal the idea of.
- SPSpeaker
[laughs]
- AKAkshay Kothari
[laughs]
- 1:39:59 – 1:42:54
AI Shares, Secondaries and the Wealth Boom
- SPSpeaker
[laughs]
- NKNikhil Kamath
Are you guys raising a round?
- SPSpeaker
We're not right now, but...
- NKNikhil Kamath
Are you selling secondaries?
- SPSpeaker
[laughs] No.
- AKAkshay Kothari
[laughs]
- NKNikhil Kamath
[laughs] Is a VC charging a 10% carry and 2% management fee and selling your secondaries?
- SPSpeaker
No.
- NKNikhil Kamath
No?
- SPSpeaker
No. [laughs]
- NKNikhil Kamath
[laughs]
- SPSpeaker
I don't think we... We don't allow that.
- NKNikhil Kamath
No?
- SPSpeaker
No.
- NKNikhil Kamath
No SPVs on Recurse?
- SPSpeaker
No. [laughs]
- NKNikhil Kamath
[laughs]
- AKAkshay Kothari
[laughs]
- NKNikhil Kamath
Everyone else seems to have one right now. Yeah. I know so many people who have gotten rich just out of, uh, selling Anthropic stock-
- SPSpeaker
That's interesting
- NKNikhil Kamath
... or, like, funding their rounds. People who took money from other people into a SPV of Anthropic and charged some kind of carry have become madly wealthy.
- SPSpeaker
[laughs]
- AKAkshay Kothari
But that money's probably done really well too.
- NKNikhil Kamath
Is it... Huh?
- SPSpeaker
I-
- AKAkshay Kothari
But that money probably did really well too.
- NKNikhil Kamath
Yeah, that money did great.
- AKAkshay Kothari
Yeah.
- SPSpeaker
I think that Anthropic doesn't allow any transfers, so you don't recognize them either, so I'm not sure what happens with all those SPVs.
- AKAkshay Kothari
But, um, uh-
- 1:42:54 – 1:45:07
Can Countries and Companies Own Their AI?
- NKNikhil Kamath
Do you guys have a view on sovereign AI? As a foreigner in this country-
- SPSpeaker
Mm
- NKNikhil Kamath
... notwithstanding tariffs and all the other things, I mean, you guys are natively from the same place I am, but you're American now. A lot more people are looking upon China as a place to replicate, where, you know, you could have products which are very American product-like, but- They're replicas that are run domestically
- SPSpeaker
You know what's interesting? Actually, my real first job was I was on the China team for TripAdvisor, and we were making a heavily localized version of TripAdvisor for the Chinese market
- NKNikhil Kamath
But I think that'll happen to pretty much everything, including social media
- SPSpeaker
Yeah.
- NKNikhil Kamath
So I was spending this week meeting social media people
- SPSpeaker
Yeah
- NKNikhil Kamath
And increasingly, I think governments across the world are concerned about a foreign company controlling so much mind share of their citizens. And in a lot of cases, there have been coups in different pockets of the world. And I think governments are rightly so thinking about it that way. But that'll happen to social media. It'll probably happen to the AI models, the frontiers will have replicas. Uh, is that something to spend time on? Like, become a inference company, sit-- become like a base ten, sit on top of a open sourced, open weight model. Wait for the day that the government decides to block foreign models.
- SPSpeaker
I mean, there's probably opportunity there. Yeah.
- NKNikhil Kamath
And capture some context and sit on it.
- SPSpeaker
The other thing is it's not just at the country level. Individual companies like to-
- NKNikhil Kamath
Mm-hmm
- SPSpeaker
... feel ownership over the models that are doing really critical tasks for them, really sensitive tasks.
- NKNikhil Kamath
Yeah, yeah. Absolutely. That could be my business idea with hundred million dollars.
- SPSpeaker
Okay, we figured it out. [laughing]
- NKNikhil Kamath
Yeah, absolutely. Absolutely.
- AKAkshay Kothari
Data centers.
- NKNikhil Kamath
Yeah.
- AKAkshay Kothari
Yes. Uh, like a JP Morgan does not want, uh, Goldman Sachs to, you know-
- SPSpeaker
Yeah
- AKAkshay Kothari
... get to know its, uh, insights, but it does want to leverage AI maximally.
- NKNikhil Kamath
The same is true for countries.
- AKAkshay Kothari
Yeah.
- NKNikhil Kamath
And ethnicities and different people.
- 1:45:07 – 1:52:07
Open Models, Scaling Laws and Local AI
- NKNikhil Kamath
What do you think about the open weight Chinese models, which might be a few months behind American ones? How do you think that'll play out long term? Anyone can take that question.
- AKAkshay Kothari
I mean, it's bound to be a significant source of, um, um, competitive pressure for sure. I, I, I do expect in, in terms of, uh, raw inference tokens and everything, a lot of it, uh, to, uh, uh, pass through these, uh, open weight models over time. I expect people will be using the absolute, uh, frontier ones in the cases where, uh, it matters to have that, uh, level of intelligence. Uh, I do expect if you're a frontier model, uh, key to your edge will be, uh, staying ahead of, uh, m-m, the, uh, you know, the rest of the world by a few months at least.
- NKNikhil Kamath
You know, you-- she's, Azalia spoke about scaling laws earlier. At some point they stop.
- SPSpeaker
Do they?
- NKNikhil Kamath
Don't they?
- AKAkshay Kothari
[laughing]
- SPSpeaker
I remember at Anthropic, like we would see like scaling laws, scaling laws, they keep going many orders of magnitude and then it seemed like it stopped. And we're like, "Oh, there was actually a bug. Like our precision on the numerics wasn't right. Oh, it kept going another. Oh, now it's stopped. Oh, they fixed some other issue." I don't, I don't think it stops. I-
- NKNikhil Kamath
Do you think the frontier models will stop at some point of time?
- SPSpeaker
In what sense? Like people stop spending the money? I think that that's the-
- AKAkshay Kothari
As in they will stop being in the frontier or-
- SPSpeaker
No
- AKAkshay Kothari
... like the open weight models will catch up.
- NKNikhil Kamath
Yeah. In a way, the open weight models will catch up.
- SPSpeaker
I mean, it's all about... Like if you have this like ecosystem of open models, to some extent, like th-there is a effect of like people building on each other, all the boats are, are rising. At the same time, scaling laws are real, so if you've diffused the compute resources, each individual model, there's a limit to how intelligent it can be. If like the frontier labs have the same like, you know, strength of like they've been as just as systematic in their training recipes and their model architectures and data and all of that, there's an amortization. I also think that the US can do better, as in like, you know, we should-- more American companies to be releasing open source models and we can build on each other.
- AKAkshay Kothari
Mm.
- SPSpeaker
I don't think China or any country has a monopoly on that.
- AKAkshay Kothari
I couldn't agree more. Like I wish there were actually like dominant US open weight models.
- SPSpeaker
Exactly. Why not? There's no reason. I think, I think actually there is a reason. It's like in a forest there's trees that shadow the, the plants on the forest floor like don't grow. In Chi-- like in the US like OpenAI Anthropic they have such strong models people feel like, "Oh, it's silly for me to train my own. Why would I release that?"
- AKAkshay Kothari
Is it that or is it mostly because distilling is illegal and so you can't do it here?
- SPSpeaker
Oh, I mean, I think that's a factor as well.
- AKAkshay Kothari
Yeah.
- SPSpeaker
Because if you, yeah, you can't, no American company would distill any of the frontier models. They will adhere to the terms of service.
- AKAkshay Kothari
Yeah.
- NKNikhil Kamath
Distilling, just to explain, can you tell me how that works?
- AKAkshay Kothari
She might be the expert. [laughing]
- SPSpeaker
[laughing]
- AKAkshay Kothari
Yeah.
- SPSpeaker
Basically peop-people can take outputs from the strongest models and they can use that to train, to take a smaller model and train it to be sort of approximating that larger model.
- NKNikhil Kamath
Mm-hmm.
- SPSpeaker
If you have access to the activations you can do even better, but even just with the o-
- 1:52:07 – 2:01:05
AI Risks, Security and Human Control
- NKNikhil Kamath
A lot was in the news about 10% of humanity dying and all of that. I'm sure everybody read it or saw it or heard somebody speak about it. As somebody who's not in this world, it scares me. As people who are in this world, how do you think about it?
- VKVijay Krishnan
I mean, I feel like, look, we are already, uh, getting to this, uh, point, right, where, um, uh, I mean, you-- we are not even talking some, you know, crazy, totally sci-fi scenarios. We are talking about, uh, uh, various kind of autonomous wea-weapons. We are talking about the fact that you should not have hu-human in the, in the loop because, uh, if, uh, you do and you slow it down and your enemy does not, then you are at a competitive disadvantage. Uh, I can totally imagine o-once you go down this, uh, thing, I mean, yeah, I, I, I feel like a lot of this, you know, comes down to, um, I imagine a lot of the, the game theoretic, uh, uh, analysis, uh, that was, uh, there with the nuclear weapons. There was all... I, I mean, everyone had their own theory. There were theories about annihilation. Uh, there was a theory of this, uh, mutual, um, mutually assured, uh, destruction, uh, or MAD in short. Uh, yeah. Here, here I feel this is, uh, this is such a, a, yeah, it, it, it is genuinely difficult to, um, analyze what happens in those limits, in my opinion. But, but it's obvious that, you know, you know, uh, obviously there are giant cybersecurity risks. It's, it's very obvious that, uh, for example, uh, just like a lot of good can be done with the, these incredibly powerful tools, a lot of bad can be done, uh, too. That is, uh, clear, yeah, at the limit.
- NKNikhil Kamath
You know, in nuclear bombs there is a human kill switch.
- VKVijay Krishnan
Mm-hmm.
- NKNikhil Kamath
A computer can't deploy a bomb. There needs to be a human who turns two keys and does whatever, right? Like, at least in theory, it's like that. Do you think the same will apply to AI everywhere? Like, humans will find jobs in just being intermediaries between a network and the end task. Like say, for example, my bank will become such that tomorrow a, a computer can't instruct the bank where to transfer money. There will have to be a human on the other side who's doing a FaceTime call with me, making sure it's me instructing him. In that way, the bank needs more employees.
- VKVijay Krishnan
Yeah.
- NKNikhil Kamath
Like, I'm trying to extrapolate to a situation where, say, AI in some form hacks the entire banking system, and all kind of fund transfers and transactions stop. What is the solution for it? Is having a human intermediary?
- SPSpeaker
I guess, are humans even better equipped to detect-
- VKVijay Krishnan
Yeah, yeah
- SPSpeaker
... this? That, that would be my answer in there.
- NKNikhil Kamath
I think humans are better equipped to detect that the instruction is a real person on the other end.
- SPSpeaker
I guess people fall for these, like, phishing attempts all the time. People get catfished. Like, I don't, I don't know if I think that we're that well equipped to do this.
- VKVijay Krishnan
Yeah. And, and I feel, right, like the essence of operations teams, I mean, and we have now run lo- large numbers of them, is that, uh, if-- to get a large scale operation going, you really can't have, uh, to, uh, totally autonomous thinkers. People have to follow simple if-then-else instructions, much simpler than what even, uh, you know, a pre-AI co-computers could, uh, solve. So, uh, yeah, I, I'm, I'm with, uh, An-Anna on this. It is, uh... I, I feel like in the past, the comparative advantages between, uh, humans and machines were somewhat more clear. Like in a world where, okay, you could write, uh, clear rules and everything, it w- it was sort of clear what is, uh, the domain of the machine and what is the domain of the human. Um- Uh, clearly, uh, um, but yes, obviously that is blurring
- NKNikhil Kamath
Like I've been using Instinct for a bit.
- VKVijay Krishnan
Mm-hmm.
- NKNikhil Kamath
And I think the kid, he's so articulate. I heard him today in the morning at the gym. I was playing his first interview.
- VKVijay Krishnan
Mm.
- NKNikhil Kamath
Neo Theo. Uh, he's so young and so articulate, but he was sharing data on Instinct, and he said something like forty percent of users-- He's growing at ten percent a day, which is crazy to begin with. But something like forty percent of users are giving him access to a credit card.
- AKAkshay Kothari
Mm-hmm.
- NKNikhil Kamath
Now, that sounds like a recipe to some kinda like disaster, which is upcoming.
- AKAkshay Kothari
Revenue growing a lot. [laughing]
- NKNikhil Kamath
[laughing] I'm trying to like extrapolate or wager a guess as to what happens post that.
- VKVijay Krishnan
Mm-hmm.
- NKNikhil Kamath
Beyond that. Beyond that hack where the banking system shuts down for a day or two, what happens?
- AKAkshay Kothari
There's an interesting-- Another one that I saw was, um, uh, you know, so many, so much of our funds are lying in like point five percent terrible savings account. And I think, I think there's another one where like imagine if everyone's personal agent took that money, uh, that is currently sitting at point five percent and actually can get you four or five percent treasuries, like it would collapse the bank system. [chuckles]
- NKNikhil Kamath
But then the treasury yields would probably have to change as well. A bank is able to survive because of that point five percent. If you take off that margin, he'll have to charge you somewhere else.
- AKAkshay Kothari
Mm-hmm.
- NKNikhil Kamath
Much like the productivity debate of software companies.
- AKAkshay Kothari
Mm-hmm.
- NKNikhil Kamath
Banks will not make more money or less money.
- 2:01:05 – 2:07:10
What Should India Build in AI?
- NKNikhil Kamath
Sovereign AI, before, I don't know, the world becomes multipolar and everybody services their own country, much like China, before that world, right now, what's the use case? You guys are native Indians. What should the Indian government or Indians be doing around AI without the risk capital that is available here or in China?
- AKAkshay Kothari
Build compute.
- VKVijay Krishnan
Yeah, that's a no-brainer.
- AKAkshay Kothari
I think that one, it's like as much as you can.
- VKVijay Krishnan
Yeah.
- AKAkshay Kothari
I feel like there's probably gonna be infinite demand, and I think the country should Be able to force the companies to tap into local compute so that they have that, uh, I think open weight models for India is a-
- NKNikhil Kamath
That is small market. See, at a GDP of, say, 4 trillion, how much compute requirement will be in India even if they were to be forced to use context and compute domestically? Like, what can we do as a country to earn... I'm not-- I'm just taking India as an example because we are from there, but what can countries do to have a pie of the AI boom? Generally it'll have to come from another country paying this country to do a certain thing. Also, we have the problem of the IT outsourcing, which has to be... That, that hole has to be filled in with something else, right?
- AKAkshay Kothari
But actually, like, I mean, my understanding of sovereign AI is, like, the countries want certain things to be served from their local neighborhood, right? I think, uh... or local country. And if that is something that the country wants to push for security reasons or whatever, um, I feel like, I don't know, how much compute will Indians use over the next five years seems infinite. Uh, and I think it's a question of, like-
- NKNikhil Kamath
But it'll be a part of their GDP unless it goes up.
- AKAkshay Kothari
Yeah. I think there's also, on the application layer, it's interesting that, um, the IT companies have not become the Mercor, Turings of the world already. Like, that seems like a missed opportunity, uh, because a lot of the people who are actually solving these problems or, you know, labeling data are actually in India. And so the fact that companies here are tapping to that labor and creating companies here seems kind of odd, that I feel like we've missed the boat on that. Yeah, like, I think what is China doing in terms of creating open weight models that India is not doing? Is it just, like, still expensive, or is it that we're not being ambitious enough to go build it and have a local model too? Or at least competitively have a local open weight model.
- NKNikhil Kamath
I think that risk capital, like, I wouldn't undermine how big of a problem that is. Like, the biggest attempt India has at AI arguably could be that Sarvam company, and I think they have about a couple of hundred million at hand.
- AKAkshay Kothari
Yeah. It's kind of like we have risk capital here, and then [chuckles] China has, has capital in the government.
- NKNikhil Kamath
Yeah.
- AKAkshay Kothari
And then everybody's in between-
- NKNikhil Kamath
Yeah
- AKAkshay Kothari
... trying to figure out, like, "Okay, I don't have this or this, so what do I do?"
- NKNikhil Kamath
Yeah, correct. That's the question, actually.
- AKAkshay Kothari
That's the hard-
- NKNikhil Kamath
I don't have this or that, what do I do? [laughing] Do you have an idea, Anna?
- SPSpeaker
I, I guess it's... I think it's an ecosystem issue to some extent. Like, actually one, one of my, um, collaborators at Stanford, like, I think he was selected by the Indian government to start a company, and they were giving him a significant number of, like, thousands of GPUs for his company, and he's training, you know, a speech model, like, on Indian languages.
- NKNikhil Kamath
Mm.
- SPSpeaker
And I think that's amazing. So I think that that's something that helps, right? And that's something that the Indian government is doing right now. And I think they're rolling it out with a small, like, really elite cohort, and then they're gonna scale it up. But I think, like, a VC ecosystem also helps.
- NKNikhil Kamath
Right.
- SPSpeaker
But I don't know how to make that happen.
- NKNikhil Kamath
Yeah.
- SPSpeaker
But...
- AKAkshay Kothari
I actually have one more idea. Um, uh, building on that and building on the fact that Anna talked about how small teams can now build really impressive companies, uh, it's still, like, awfully hard to start a business in India. Like, uh, I talked to an entrepreneur in India today, and I say, "Yes, I'm willing to angel invest. I'll get the paperwork three months from now." Uh, uh, and then it still requires notarization. Like, I, I, I sort of feel like I should stop investing for all the sort of headache. Uh, so it's just, like, make... I, I know the government's trying, but I feel like there's still so much room to just enable the youth to, you know, start companies, small businesses. Like, maybe not everybody needs to start an AI models company. I think there's going to be a proliferation of small businesses that feels, like, ready to be built, and can we do anything to remove the barriers for, for people to be able to do that?
- NKNikhil Kamath
Today, after a long time, I was having lunch with, uh, Neeraj, the GC guys, and all of that. After a while, I heard Americans are fighting to invest into an Indian company, Skyroot Space. I think Benchmark wants in, everybody wants in. I think, uh, the prime minister back home has done a lot to make the red tape lesser and allow for Skyroot to really thrive. So maybe space is something where they could... There's some-
- AKAkshay Kothari
And space is actually a great example of what India can do, right? Which is, I think India can take a, um, take something that the world is doing and actually do it at, like, a hundredth of the cost-
- NKNikhil Kamath
Yeah
- 2:07:10 – 2:09:04
Are AI Safety Warnings Sincere?
- AKAkshay Kothari
make it better.
- NKNikhil Kamath
What about all this regulatory arbitrage stuff? Like, everything you hear about the doom, the doomsday sayers are doing it because they want a regulatory arb which makes it harder for new people to come in and compete. Are you on that side of the argument, or are you on the side that believes that conscious is kicking in and these people are being authentic about what they're saying?
- VKVijay Krishnan
I think they say what they believe. I do not believe, uh, uh, y- I mean, I have, I have generally met enough Anthropic, uh, researchers that said, "I would, uh, uh, uh, if I had to choose between, uh, uh, you know, uh, uh, this AI safety versus, uh, sort of my stock going to zero, I would ha-- I'm happy with my stock going to zero." I have not seen, uh, too many co- [chuckles]
- NKNikhil Kamath
Yeah
- VKVijay Krishnan
... people working at these companies holding that view. So I-
- NKNikhil Kamath
As ex-Anthropic.
- SPSpeaker
Yeah. I, I think I actually reported to Dario when I first joined. I, I think he's sincere about it.
- NKNikhil Kamath
Yeah.
- AKAkshay Kothari
I think I'll, yeah, I would say the same, um, r- for, for a lo- It's actually not just one person. There's actually several researchers who are seeing that as a true risk. Um, I don't know if we can quantify it to be ten percent or higher or lower, but I, I think that's the challenges, right? I think it's almost like the communications has been tough. It's, uh... I, I, I think that, that ten percent post went to, like, a local Indian newspaper. Uh, and so-
- NKNikhil Kamath
Oh, it didn't blow up here so much?
- AKAkshay Kothari
No, no, it went everywhere. I mean, my point is, like-
- NKNikhil Kamath
Fair
- AKAkshay Kothari
... it traveled so much-
- NKNikhil Kamath
Yeah
- AKAkshay Kothari
... that-
- NKNikhil Kamath
I mean, if you, if anyone relevant is going to say ten percent of the world is going to die, it is going to travel. [chuckles]
- SPSpeaker
[chuckles]
- AKAkshay Kothari
Yeah.
- NKNikhil Kamath
Right?
- AKAkshay Kothari
And so I think it's, uh, yeah. I guess my, my point is, like, I don't know how productive that episode
- 2:09:04 – 2:14:41
How Would You Spend Your Last Month?
- AKAkshay Kothari
was versus the risks are real.
- NKNikhil Kamath
Anna, if the world were going to die, what would you spend your last one month doing?
- SPSpeaker
I think I would keep doing what, exactly what I'm doing right now.
- NKNikhil Kamath
Really?
- SPSpeaker
Yeah.
- NKNikhil Kamath
You wouldn't wanna go on holiday, you know, touch grass? Like, [chuckles] look at the-
- SPSpeaker
I mean-
- NKNikhil Kamath
... sunset?
- SPSpeaker
... I wanna spend time with, like, the people I really like. Like, you know, my, my co-founder, my team, but also, like, my family, my husband, and my friends, so I want that.
- NKNikhil Kamath
Yeah.
- SPSpeaker
But I think that that's what I'm doing now, so.
- NKNikhil Kamath
Hmm.
- SPSpeaker
How, how about you?
- NKNikhil Kamath
I don't know. I think I would, like, go, like, full-on hedonistic, like, just-
- SPSpeaker
[chuckles]
- NKNikhil Kamath
... self-abuse.
- SPSpeaker
[chuckles] What, what is self-abuse?
- NKNikhil Kamath
I don't know, like, get intoxicated. Like [chuckles] buy a car and then break it down. [laughs]
- SPSpeaker
[chuckles] Just like smash it.
- NKNikhil Kamath
Like, I don't know. You would do the things you wouldn't normally do.
- SPSpeaker
Maybe, maybe you should just go buy a car and smash it right now. [laughs]
- AKAkshay Kothari
Get it out of my system.
- SPSpeaker
Why wait till the end of the world, right?
- NKNikhil Kamath
I think there's so much you do not do because you worry about the long-term ramifications of those acts.
- SPSpeaker
Mm-hmm.
- NKNikhil Kamath
I have a feeling for society which knows that tomorrow might not be there, they will care lesser about those long-term ra-
- AKAkshay Kothari
Mm-hmm
- NKNikhil Kamath
... ramifications. I think we all-
- SPSpeaker
Yeah
- NKNikhil Kamath
... inherently organic- organically optimize for pleasure in the short term. It's very hard to delay gratification. It's a very evolved thing to do for humanity per se.
- 2:14:41 – 2:17:48
What Comes After the Smartphone?
- NKNikhil Kamath
You know the one thing that's not changing? AI is changing and blowing up, but the means of interacting with AI, like the devices, the hardware that I use, like a iPhone is still an iPhone. Why is that not changing?
- AKAkshay Kothari
It's coming.
- SPSpeaker
I, I think-
- NKNikhil Kamath
You think?
- SPSpeaker
... that's what I was thinking too.
- NKNikhil Kamath
Yeah?
- AKAkshay Kothari
There's gonna be, like, an explosion of hardware devices.
- SPSpeaker
I think so.
- AKAkshay Kothari
Yeah. I feel like it's just taking time 'cause I think, it's like it blew up four years ago. I think people are iterating, but there's a lot of ... Like, I, I don't know. It feels like the renaissance of hardware, in the Valley at least.
- NKNikhil Kamath
I think that's, if I were a venture capital guy, I think that's the thing one would probably bet on, right?
- SPSpeaker
I, I think, like, part of the reason that there isn't a bigger vari- like, a wider variety of hardware is just because it's, it takes so long to design a new chip. And so for example, like between generations of TPU, the team would try to make very, like, conservative, minimal changes because otherwise it will take, they, even in a year or two years, they couldn't close that chip.
- NKNikhil Kamath
Right.
- SPSpeaker
So I think that if you can compress these design cycles using, you know, AI methods, then you can be way more creative in the ter- types of hardware that you can tape out and deploy into the world, and maybe iPhones can look very different.
- NKNikhil Kamath
Do you have a view on that? Like, how would you interact with hardware, said hardware, whatever were to come about? Would it be voice? Would it be something you hold in your f- hand like a phone? What would it be?
- SPSpeaker
I mean, there even like brain-
- NKNikhil Kamath
You said the cost of, like, capturing all context is now whatever. Luna, how much did you say it was?
- AKAkshay Kothari
$100 a month can give you, I guess, a query every minute.
- NKNikhil Kamath
Yeah, so would it be that? Like, you have a device and-
- SPSpeaker
I mean, I think, like, like, glasses are actually a pretty reasonable one, right? You could have a, uh, the model see and hear everything that you see and hear. And there are interesting things like, you know, brain computer interfaces or you can, you know, if someone is, like, not speaking out loud, but they're, they're thinking about saying a particular thing, that can be picked up by hardware devices. That's an interesting way of having a more private communication with your device. I think, like, typing on your phone probably isn't the, the future way of interacting with these intelligences.
- NKNikhil Kamath
Voice seems quite efficient, but I just feel weird, like, like-
- SPSpeaker
It's not private
- NKNikhil Kamath
... yeah, like, me and Aryan were driving here, and we were both on our laptops. But I don't know, speaking to it while you're sitting with somebody feels weird, no?
- SPSpeaker
I think there are ways of picking up just as you think about s- saying a particular thing, it makes your, your vocal cords move in a particular way. It can be detectable. So that could be a way of, like, speaking to the device while having it not be weird that you're there with someone else.
- NKNikhil Kamath
Or, like, neurally encode it.
- SPSpeaker
Yeah.
- AKAkshay Kothari
Yeah, these things are sort of advancing quite rapidly. It's just a question of, like, when does the, when do they hit mainstream?
- SPSpeaker
Mm-hmm.
- NKNikhil Kamath
Right.
- AKAkshay Kothari
Um, I think it would be a wonderful world where we don't have to look at devices and can interact with the real world more.
- 2:17:48 – 2:21:45
The Future of Social Media and Entertainment
- SPSpeaker
That's right.
- NKNikhil Kamath
How do you think real world stuff changes? What happens to social media? This is very selfish of me to ask-
- SPSpeaker
[laughs]
- NKNikhil Kamath
... because I want to build something in it. What do you think is the future of social media?
- SPSpeaker
So is that, that, that's where you're h- doing your $100 million venture?
- NKNikhil Kamath
Yeah.
- SPSpeaker
That makes sense. I guess, why are you so passionate about social media, or what do you want to...?
- NKNikhil Kamath
I feel like incumbent social media, today's social media, is great for dopamine, but what was its initial promise is not being delivered. I don't think anybody's making connections on social media anymore. I, I personally feel that it'll go back to, you know, the whole touching grass offline world-
- SPSpeaker
[laughs]
- NKNikhil Kamath
... and people who can orchestrate real world stuff. The new social media will almost be like, "Get off your device," in my opinion.
- SPSpeaker
The Po- the Pokemon Go-
- AKAkshay Kothari
I think it could go the other way.
- NKNikhil Kamath
Yeah?
- AKAkshay Kothari
I, yeah, because, uh, you-
- NKNikhil Kamath
You think, like, that virtual world with those devices-
- AKAkshay Kothari
Or you'll be dealing-
- NKNikhil Kamath
... what Meta tried
- AKAkshay Kothari
... yeah, you'll be dealing with a lot of-
- SPSpeaker
Yeah
- AKAkshay Kothari
... agents and, uh, everything, uh, which is even one step removed ra- uh, from contacting your friends on, uh, social media.
- NKNikhil Kamath
But usage on incumbent social media seems to be going down in the youngest of cohorts, which is strange. Like, the 16 to 20 year olds or 25 year olds are using social media lesser today than they were two years ago. They're logging in less number of times and stuff like that.
- SPSpeaker
How are they spending their time?
- AKAkshay Kothari
AI companies, I don't know.
- SPSpeaker
At Touching Grass, like, what are they doing?
- NKNikhil Kamath
They're, I'm-
- AKAkshay Kothari
I, I believe everybody company has a, a server
- SPSpeaker
They're founding agents. Yeah
- NKNikhil Kamath
Like, you know, even here, when I meet really cool kids in-
- SPSpeaker
[laughs]
- NKNikhil Kamath
... all the kids who are raising billion dollars and started the company two weekends ago, a flex today is no longer, "I have a billion followers on social media." They're all talking about, "I have this app which blocks my social media usage time from the f- from more than 30 minutes. I am not on TikTok," or whatever is the cool new thing. Generally, whatever these cool kids do, I think kind of, like, trickles down and-
- 2:21:45 – 2:25:51
Will AI Replace the Apps We Use?
- NKNikhil Kamath
Do you have a view on all the user interface applications? Like, I don't know, Uber or all the stuff that agents or even Instinct, for that matter, seem to be eliminating from somebody's life in a way, right? Like, if I can just tell Instinct what food I want, why do I need to go on Uber Eats or DoorDash, or actually, for that matter, my business?
- VKVijay Krishnan
So I hold a little bit of a counter view on this. I don't think, uh, you know, this, uh, talking to something or, uh, chatting with something is the right, uh, UI for everything. Um, like in the-- in case of food, um, uh, you may want to get ideas. You may wanna browse, uh, what is available, uh, you know, what are different options, et cetera. And, uh, talking to you is not the most efficient way to convey that to you. I think it's the same with, uh, uh, online, uh, shopping and, uh, so on. Maybe you wanna get some ideas. You wanna compare. Uh, yes, I'm, I'm sure more intelligence will-- of this sort will help you, uh, compare. And same thing, right? Like, uh, I, I keep hearing about how, uh, someone can, uh, book, uh, tickets using AI. I really do not, uh, think it's the best, uh, possible, uh, use case at all. Uh [laughs] , like you really couldn't, uh, be a lot more efficient than, uh, what you can, uh, uh, do, uh, with, uh, online booking now. Uh, I think there are legitimately, uh, a number of, um, use cases where, uh, a novel UI will improve things. There's a lot of other where you may not be crazy far off from the, um, from the frontier.
- NKNikhil Kamath
Right.
- AKAkshay Kothari
Yeah, I agree with that. I sort of feel like it's so early. Like, I think we're in like the MS-DOS era. [laughs]
- SPSpeaker
[laughs] Yeah.
- AKAkshay Kothari
Just like text back and forth.
- NKNikhil Kamath
Mm.
- AKAkshay Kothari
That can't be, that can't be the future.
- NKNikhil Kamath
Right.
- AKAkshay Kothari
It feels-
- SPSpeaker
And these models are really good at generating interfaces.
- AKAkshay Kothari
Yeah. Like, I feel like text is so cumbersome, and then voice is somewhat delightful, but then it's so intrusive and like, uh... So I feel like some new UI will emerge for the, for the use case you're trying to do.
- NKNikhil Kamath
Actually, the glasses didn't take off, right? Not-
- SPSpeaker
No, they didn't take off. I mean, I don't know maybe yours are smart glasses, but I don't think so, right? I mean, I think-
- AKAkshay Kothari
I think like the, the really good ones that they have are currently still, like, too pricey, right? Uh, like I actually tried the, the Meta where it's literally maybe thirty percent more weight than this, and, um, I think that's the first device I felt I could give up my phone.
- SPSpeaker
Oh, wow.
- NKNikhil Kamath
Mm.
- AKAkshay Kothari
Yeah. I-
- SPSpeaker
So why, why didn't you give up your phone then?
- AKAkshay Kothari
[laughs] It's not available yet.
- SPSpeaker
Okay, that makes sense.
- AKAkshay Kothari
I think like they have the technology, but it's probably like twenty thousand dollars or something. Like it's, it's too expensive. So I think if you track the, the r- the new glasses that they release every year, they're adding a little bit more functionality. It started with just taking a photo and maybe listening to a podcast, but I think it's gonna keep adding to a point where I feel like-- I do think that's a-- I mean, for me, it's like I'm already wearing glasses. [laughs]
- SPSpeaker
I, That's what I was gonna say. What about p- us people who don't wear glasses? What are we gonna do?
- AKAkshay Kothari
Yeah. Uh, but I think eventually-
- NKNikhil Kamath
Yeah, I don't, I don't like it, having something on my eye all the time.
- SPSpeaker
[laughs]
- VKVijay Krishnan
Honestly, the only reason, uh, I wear my smart glasses is, uh, because, uh, to reduce the, uh, friction in, uh, picture taking. If I go with my kids out somewhere, yes, I end up taking more, uh, pictures. Uh, that's it. But, uh, um, that's not using any AI whatsoever.
- SPSpeaker
I remember it was like a dog fooder for Google Glass originally. Like I-- it was just not the right time because there was just-- there was-- the applications that could be served at that time were just almost nothing. Like you could take pictures-
- AKAkshay Kothari
Mm-hmm
- SPSpeaker
... or you could do like some live translation, or you could do Google Maps, and that was pretty much it.
- 2:25:51 – 2:28:57
Final Ideas: Product Design and Better Learning
- NKNikhil Kamath
Any fleeting thought, last thought for the day?
- AKAkshay Kothari
You should take the $100 million and, um-
- NKNikhil Kamath
Yeah, let's make this the thing
- AKAkshay Kothari
... open the-
- NKNikhil Kamath
Last part, what should I do with $100 million?
- SPSpeaker
Yeah, that's a Ricursive Intelligence center here.
- AKAkshay Kothari
Uh, open a, um, university for product design in India.
- NKNikhil Kamath
Hmm. Okay. I'm gonna borrow $100 million from a bank, and then what should I do with it? You can sell like 0.1% of your Anthropic stake and make it anyway.
- SPSpeaker
[laughs]
- NKNikhil Kamath
Much more easier for you. [laughs] But what is your advice to me to do with my $100 million?
- SPSpeaker
I mean, it seems like you're passionate about social media, so I think you should do that.
- NKNikhil Kamath
Yeah.
- SPSpeaker
Yeah.
- NKNikhil Kamath
Maybe we'll chat, and you can tell me what to build.
- SPSpeaker
Okay, sounds-
- NKNikhil Kamath
Yeah
- SPSpeaker
... sounds good. [laughs]
- AKAkshay Kothari
Recursive social media. [laughs]
- NKNikhil Kamath
Yeah. And nowadays, the people I met in social media are saying IRL, Inverted Recursive Learning.
- SPSpeaker
[laughs]
- NKNikhil Kamath
Like, this is a new word I've learned in the last few days.
- SPSpeaker
A different sense of IRL.
- NKNikhil Kamath
Yeah.
- VKVijay Krishnan
I think it would be interesting if there could be an AI, uh, tool that substantially is able to, um, accelerate, uh, uh, learning.
- NKNikhil Kamath
Mm-hmm.
- VKVijay Krishnan
Um, I don't know, suppo- uh, suppose a child could literally learn at 10X the rate or something like this, right? Like, uh, with, uh, rapid, uh, personalization and, um, being sufficiently mindful of the UI, uh, as well of this, uh, uh, whole-- I mean, the, the entire UX of this in, uh, terms of there is a sense in which you have to pay your dues as well and, and expend thinking tokens to actually master some material. Uh, uh, I'm not sure how, uh, how precisely that whole thing will, uh, um, come together, but it s- it seems to me, th- yes, uh, it, uh, the models seem to have the intelligence, uh, necessary, uh, to, uh, you know, um, beat, uh, what happens in a classroom substantially. But yeah, I, I think, uh, I think some product, some, uh, one will need to think carefully about how to make this work. Today, I, I, I don't even think, uh, these things are particularly good for a- adults, uh, who presumably have some motivation and everything. And, um, yeah, for, um, for children, I'm, I'm curious how this would be and would that unlock a lot of human capital along the way of, um, you know, billions of, uh, people around the world? Uh, that, that would be interesting. Though, though always, uh, challenges with monetizing and all that, but, uh, it's, it's always struck me as, uh, an open and interesting idea.
- NKNikhil Kamath
Yeah. Anything else? Well, thank you guys for doing this.
- AKAkshay Kothari
Mm.
- NKNikhil Kamath
Uh, it was fun for me. I learned a lot. Hopefully, it was fun for you. I know you did not learn a lot, [laughs] but thank you for taking the time and doing this.
- AKAkshay Kothari
Thank you for bringing us together.
Episode duration: 2:28:57
Install uListen for AI-powered chat & search across the full episode — Get Full Transcript
Transcript of episode YU-RrDdyx34
