Latest / The Neon Show / Ex-Meta VP: ChatGPT Is Just the Trailer for What’s Coming Next

Transcript

Ex-Meta VP: ChatGPT Is Just the Trailer for What’s Coming Next · The Neon Show

0:00
SPEAKER_02: From III T Hyderabad, an intern at Meta to VP at Meta and now founder.
0:06
SPEAKER_02: You shared a very honest take that data centers should not be wasted on trivial things.
0:12
SPEAKER_00: The unprecedented amount of investment that is going in into pushing the frontier of AI capabilities, the reckoning is already here or is coming as to what it means for humanity.
0:30
SPEAKER_00: I think that game will continue and there will be value in the models, there'll be value in the application there.
0:35
SPEAKER_00: What motivated me is how do you push the models even further?
0:42
SPEAKER_00: And there the models need to learn from observations of reality at the atomic level.
0:53
SPEAKER_00: It's getting new drugs to people much much faster than ever possible.
1:02
SPEAKER_00: But underneath it, it's something more nuanced.
1:05
SPEAKER_00: It is.
2:13
SPEAKER_00: We have gotten used to AI.
2:15
SPEAKER_00: Humanity has gotten used to AI all the way from the first time Deep Blue happened, where the best algorithm beat chess, Jeopardy happened where we had amazing search algorithms and Watson was a thing.
2:36
SPEAKER_00: Go recently a chat GPT.
2:40
SPEAKER_00: There's a common theme in all of this, which is you always get to this thing that we think was not feasible.
2:46
SPEAKER_00: And then when it happens, you start thinking about it as an obvious thing.
2:55
SPEAKER_00: The first time you experience Wi-Fi in a plane, you're like, wow, this is amazing.
3:00
SPEAKER_00: And the second time you experience it, you're like, why is this so slow?
3:04
SPEAKER_00: I think AI has this phenomena.
3:06
SPEAKER_00: So for uh for me, the frontier models is basically capabilities that we didn't think was possible recently with all the math uh advances and the coding agents.
3:18
SPEAKER_00: It's just awesome to wake up and have an idea and have an implementation ready within a few minutes.
3:24
SPEAKER_00: So that's frontier models in a nutshell, and you can peel the onion, go all the way to the transformers and the transistors, or stay higher level and say there are language models that are predicting the next tokens, and because of that problem statement, you're able to get intelligence that is generalized and you adapt it to a new problem with a few short examples and it's able to get the job done.
3:43
SPEAKER_00: So it is the center of the universe because a lot of our life has been built around uh digital applications and code, and you're able to use these models to take on new applications and build them.
3:57
SPEAKER_00: It's awesome, it's going really, really well.
4:00
SPEAKER_00: But the thing that motivates personally for me is a big bottleneck that if you don't push on will not get solved.
4:09
SPEAKER_00: And I think the way the frontier models are evolving, um, a lot of the digital applications and applications that you can build around the models, you can do a lot of really good work on the harnesses, you can do a lot of really good work on data, but the nucleus of what you can achieve with the models out of the box because of the general intelligence is getting larger and larger and larger.
4:32
SPEAKER_00: So if you're building an application for an end user or an app with some value, you can take the intelligence out of the box, apply it.
4:40
SPEAKER_00: If it doesn't work, you can do a lot of customization.
4:43
SPEAKER_00: I think that game will continue and there will be value in the models, there'll be value in the application layer.
4:48
SPEAKER_00: But taking a step back, what motivated me is how do you push the models even further?
4:56
SPEAKER_00: And there I feel the models need to learn from observations of reality at the atomic level.
5:04
SPEAKER_00: And that starts getting into the physical intelligence part, where the models can have a hypothesis, act on the hypothesis, change a physical instance in some capacity, observe from it, learn from it.
5:18
SPEAKER_00: And that loop hasn't been built yet.
5:21
SPEAKER_00: To the extent code and math and chat agents are built, enterprise agents, personal assistants.
5:30
SPEAKER_00: Now, that has not been done with physical uh atoms in the loop.
5:35
SPEAKER_00: And that is what is exciting to me.
5:37
SPEAKER_00: So it is less about being at the frontier and saying no to it, and more about what is the fundamental bottleneck in learning.
5:46
SPEAKER_00: And I feel the physical intelligence and bringing that in will actually be very, very exciting and it's a very hard problem.
5:52
SPEAKER_00: And the more people get on it and tackle problems around it, the more exciting the future will be for humanity.
5:59
SPEAKER_02: So, in a nutshell, you are trying to push the frontier to the atoms, not just uh uh let it remain to the bits.
6:06
SPEAKER_00: Yeah, I I I think I think the unprecedented amount of investment that is going in into pushing the frontier of AI capabilities, the reckoning will is already here or is coming as to what it means for humanity.
6:24
SPEAKER_00: What is that unprecedented amount of resources lead to in terms of betterment of our lives?
6:30
SPEAKER_00: And I think that question will lead us to much more in how can the AI agents or AI help in the physical world to make our lives better?
6:42
SPEAKER_00: That can manifest in many ways.
6:44
SPEAKER_00: And one of the things that I am now excited about working on is in the context of that.
6:50
SPEAKER_02: Uh nobody imagined a world, let's say in the last 10-15 years, where the top five companies in the world would be engineering companies.
6:58
SPEAKER_02: So I imagine maybe in a world where the top five companies would be labs.
7:02
SPEAKER_02: Like today we call them labs, right?
7:03
SPEAKER_02: So which are like privately owned but privately most valuable company among also the public companies, they would come in top 10 uh kind of, right?
7:12
SPEAKER_02: But I imagine a world where the top five companies in the world by market cap economically could be could be labs.
7:19
SPEAKER_02: And I think you are right there because if you redirect capital uh towards uh, you know, and what you are trying to do, towards the most important problem a lab can solve, which is you know, for humanity, make life better, right?
7:36
SPEAKER_02: Then it could have a significant outcome.
7:37
SPEAKER_02: There's no in my worldview, there's no doubt in next 10 years, the most five valuable companies would be labs.
7:45
SPEAKER_00: Yeah, uh it's it's very hard to put financial uh predictions.
7:51
SPEAKER_00: Most of the times you're probably right, you're probably wrong, you're mostly wrong.
7:56
SPEAKER_00: Um, but one thing is one thing is certain for me, though.
7:59
SPEAKER_00: What is certain for me is um the maximum learning will come from these models taking actions in the physical space and learning from each interaction and outcome and closing the loop back.
8:17
SPEAKER_00: I think that's where these models are starting to get better, but not there yet.
8:22
SPEAKER_00: We have a big gap from video to sim, real to sim to real.
8:30
SPEAKER_00: These are where the bottlenecks are.
8:32
SPEAKER_00: And there's a reason, right?
8:34
SPEAKER_00: Why did code become so, like, why did these models become so good on code within a year?
8:41
SPEAKER_00: Uh, where a year ago on Thanksgiving, people were like, I see something here, and then by summer, what happened?
8:49
SPEAKER_00: And then by the next Thanksgiving, I don't code anymore.
8:53
SPEAKER_00: I just code in English.
8:55
SPEAKER_00: And some people have said those statements, and when they said it, it felt fine, will happen ever eventually.
9:03
SPEAKER_00: And within a year, you look back, I'm like, man, they were right.
9:07
SPEAKER_00: Why?
9:08
SPEAKER_00: Because the loops were verifiable.
9:10
SPEAKER_00: You could basically unroll the RL loops, you could verify, you could verify, compile the code.
9:16
SPEAKER_00: For math is the same, you have lean, you can verify the proofs.
9:20
SPEAKER_00: It's extremely like it's now a matter of computation and having a reasonable system that just gets better and better by self-play in the context of like a gaming thing, but also more importantly in RL with verification.
9:35
SPEAKER_00: That doesn't exist today for the real world.
9:40
SPEAKER_00: And that loop is bottlenecked by humans, bottlenecked by atoms, bottlenecked by very long lead times where you let's say you take a compound or you create a molecule using some retrosynthesis in case of a small molecule.
9:55
SPEAKER_00: How do you actually verify that that molecule that you have, that retrosynthesis that you have actually leads to that molecule?
10:01
SPEAKER_00: You have to now go to a wet lab, you have to now go in the physical world, you have to create that molecule, follow the 10 steps.
10:07
SPEAKER_00: Who does that?
10:08
SPEAKER_00: Chemist who probably will work for eight hours and then has to sleep.
10:12
SPEAKER_00: And then while they do the manual work, there might be in the context of material sciences, you will have to actually repeat the experiments because even a 1% delta in the material composition leads to a different material, and you have to verify.
10:24
SPEAKER_00: In the case of a small molecule, you have to have precision equipment that you have to actually operate where very few chemists can operate.
10:31
SPEAKER_00: Now, when you do all of that, and after two weeks or one week or whatever, you get a compound and you do the spectrography of that using LCMS machines, then you're like, okay, this works.
10:41
SPEAKER_00: That verification needs to come back.
10:43
SPEAKER_00: That might not even be digitized in most cases.
10:46
SPEAKER_00: You that's the world we live in in the context of the physical, just I'm looking at a specific example in the context of a drug discovery, for example.
10:56
SPEAKER_00: If we change the infrastructure of the world where a lot of this is not human bottleneck, where a company that wants to go for a drug, new drug, is actually able to use the latest and greatest foundation model to identify better targets, is able to use autonomous robots with humans in the loop to be able to carry out the operations 24 cross-7.
11:22
SPEAKER_00: Take more risk care hypothesis because of the efficiency in the system that you have actually unlocked.
11:28
SPEAKER_00: And you put all of this together, and then the learning becomes more verifiable and the data-poor domain becomes a data-rich domain.
11:36
SPEAKER_00: Imagine what kind of a world we will live in if that were true.
11:41
SPEAKER_00: But that's the biggest bottleneck that I at least am excited about and want to go solve.
11:48
SPEAKER_00: Because I mean, if if a person can get a drug three months earlier, it's make or break for them.
11:56
SPEAKER_00: If we can have highly personalized drugs for a particular age and genome and culture, it's hugely valuable, right?
12:03
SPEAKER_00: So those are like fundamental problems of the future.
12:07
SPEAKER_00: That's just one example of it.
12:08
SPEAKER_00: New material, renewable material.
12:11
SPEAKER_00: Uh how so taking a step back, like that's the AI infrastructure that I feel is missing.
12:18
SPEAKER_00: In and as we scale the current frontier models, somebody needs to go, like a good bit of energy needs to go and build that new age infrastructure.
12:29
SPEAKER_00: And that's the goal that I'm most excited about.
12:34
SPEAKER_02: And you shared a very uh honest take, right?
12:38
SPEAKER_02: Uh that data center should not be wasted on trivial things.
12:47
SPEAKER_00: Yeah, I look, I think, I think, I think uh the personal assistant product is non-trivial.
12:53
SPEAKER_00: It's it's hugely valuable.
12:55
SPEAKER_00: Like I use it every day, you use it every day, we all use it every day.
12:58
SPEAKER_00: But I feel like that is this is the beginning of the revolution.
13:04
SPEAKER_00: That's scratching the surface.
13:06
SPEAKER_00: I think being able to shop better, being able to plan your travel better has some value.
13:13
SPEAKER_00: Maybe it has more value for some people than others, but I am not entirely confident that that is the ultimate benefit of AI.
13:23
SPEAKER_00: I think that is one value proposition.
13:25
SPEAKER_00: There's a significantly more important uh and higher value proposition on making our lives much more better, making energy much more renewable, and fundamentally going after discovery.
13:37
SPEAKER_00: Like the thing that flipped my bit recently is um we have this oversimplified view that the language models are learning from each and every interaction, and they are not coming up with novelty, they are basically mixing the innovation of the individual interactions in some novel way and just applying.
14:01
SPEAKER_00: Then you look at some of the math proofs that came, and I'm not a math expert, but clearly I'm trying to follow this space.
14:12
SPEAKER_00: There's a raging debate, but at least one thing is clear.
14:16
SPEAKER_00: Some of the proofs are almost field metal quality.
14:22
SPEAKER_00: So that means if you have enough compute, a good seed start, and can actually make the domain as verifiable as possible, then you can discover new things.
14:35
SPEAKER_00: And that has huge impact, meaningful impact for humanity.
14:39
SPEAKER_00: So I am very, very excited about that possibility of building these systems that can allow models to innovate and discover new things, help discover new things.
14:51
SPEAKER_00: Maybe they will not discover on their own, but who knows?
14:54
SPEAKER_00: Clearly, we were wrong in terms of the rate of progress in some context.
14:58
SPEAKER_00: Um, so I am very, very excited about that because discovery in these life sciences domains has a tremendous amount of positive impact for the world and for humanity, and that is worth fighting for.
15:15
SPEAKER_02: So till now, you know, frontier have become the nucleus, and it's like the nucleus is eating everything around it, right?
15:24
SPEAKER_02: There's a lot of uh, you know, you can say signals, noise, uh right?
15:29
SPEAKER_02: Why do you think the nucleus is just eating everything and is there space for something else outside this nucleus?
15:35
SPEAKER_02: Like, for example, the narrative of the software is getting eaten by the frontier.
15:43
SPEAKER_00: Yeah, um I think I think I think the nucleus getting bigger is real because out of the box, when you look at the next model that is getting launched, all of the customization that you might have done somehow is simplified, which I think is a big win if your goal is not to build a better model or machine learning system, but rather it's in service of a service that you are giving to people and value that you're creating for people.
16:13
SPEAKER_00: Um so to the question of uh in the world where the nucleus is getting larger and larger, um, I think the way to think about it is is the cost and efficiency and speed getting better for the application that you're building?
16:30
SPEAKER_00: And if the answer is yes, you're good.
16:32
SPEAKER_00: And you can always then do say, can I take a dependence on open model?
16:36
SPEAKER_00: Can I take a dependence on closed model?
16:37
SPEAKER_00: How do I architect things in a resilient way?
16:40
SPEAKER_00: And how do I build loops?
16:41
SPEAKER_00: So you're part of the faster loop.
16:44
SPEAKER_00: If you architect your code base, if you architect your company, if you architect your engineering cycles to be much more aligned to this way of development, you will harness the energy, you will harness the speed.
16:56
SPEAKER_00: If you are building in a way that has an impedance mismatch, where each new model will mean you will have to redo all of your engineering, then you're not set up properly, right?
17:06
SPEAKER_00: So that's the how I think about harnessing intelligence for your application.
17:10
SPEAKER_00: Now that there's another important question, which is how do you add to that intelligence?
17:15
SPEAKER_00: And at the end of the day, today it's about data, it's about scale and um iteration speed.
17:24
SPEAKER_00: So if you're building a team and focusing on a problem where you still have a lot of unique value in terms of data, a lot of unique value in terms of uh iteration speed because of the subject matter expertise, you are actually harnessing the energy.
17:41
SPEAKER_00: The nucleus can grow larger and larger, but you're still staying on top of it and creating value for society.
17:46
SPEAKER_00: So that's how I think about it.
17:48
SPEAKER_00: Then there is a completely different paradigm of the current paradigm will not be enough to understand the real world.
17:55
SPEAKER_00: So you then basically move away from pure large language models that are discrete token representations of language tokens or any other embeddings or encodings that you create, whether it's speech or audio or envision, and say, no, I'm going to bet on a completely new paradigm like a world model that can help me simulate the new worlds.
18:13
SPEAKER_00: And then my RL rollout will be that simulations, and then I use that as verification to basically make the next step, which is a different paradigm.
18:23
SPEAKER_00: And then you get all of the nuanced variants in the representation space, in how you train the models and how you use the models, what is a new harness, and so on and so forth.
18:32
SPEAKER_00: So to me, uh it's very, very important to not confuse these two.
18:35
SPEAKER_00: You have to have clarity saying I'm going to use the current wave and the current technology and add value, or I will basically spend a bunch of energy to go counter and exploit, knowing very well that it's going to take a lot of energy, time, and resources, but you're playing the long game.
18:52
SPEAKER_00: So it's totally okay for both, and that's the explore exploit.
18:55
SPEAKER_00: And you always have to balance these two.
18:58
SPEAKER_02: And and where do you stand in the open versus close model debate?
19:01
SPEAKER_02: Because you are seeing it play out, and it almost feels like an arms race.
19:07
SPEAKER_00: Uh so I mean, this this topic is a heavily debated one, and at least my historical stance has been very, very clear.
19:14
SPEAKER_00: And I think fundamentally the principles are the same, which is if you look at the longer arc, longer arc open ecosystems have always won.
19:23
SPEAKER_00: But there's a little bit of a difference between pure open source or open ecosystems to open weights.
19:32
SPEAKER_00: So, what we have going on right now in the foundation model is open weights, not full open source.
19:38
SPEAKER_00: There are obviously some open source stacks, which includes pre-training, RL, uh, like Tinker is an example of uh fine-tuning models.
19:47
SPEAKER_00: Um, Common Crawl is a pre-training data set that people use, GitHub code, maybe some of it is publicly available.
19:56
SPEAKER_00: So those are all things where some components of open.
19:59
SPEAKER_00: So, why did open ecosystems win in compared to closed ecosystems over time?
20:04
SPEAKER_00: It's because when you have the full engine, you could take bits and pieces of it and bring in your innovation that you are trying to tackle.
20:14
SPEAKER_00: So you use all of the advantages of a full engine that has been prepared and you take components of it apart and you bring in your own expertise.
20:26
SPEAKER_00: Now, because it's an open ecosystem, you do it for your benefit, but other people also learn from it.
20:33
SPEAKER_00: And now you have this amazing ecosystem of player where let's take PyTorch.
20:39
SPEAKER_00: PyTorch did not start with all the feature set of TensorFlow, but the number of developers in PyTorch exponentially increased because it's the nucleus of PyTorch started becoming more and more capable.
20:51
SPEAKER_00: It didn't start off with feature parity, but the rate of progress was much larger.
20:56
SPEAKER_00: Linux versus Windows.
20:57
SPEAKER_00: You can take any open ecosystem, most of the open ecosystems where you could take bits and pieces of it and bring innovation across expertise, applications, cultures.
21:09
SPEAKER_00: The open ecosystem thrived because of this exploration parallelism.
21:16
SPEAKER_00: The closed ecosystem has to prioritize because it has limited resources and has to support an existing ecosystem and iterate slowly.
21:25
SPEAKER_00: So the thousand seeds bloom and you then pick some of them, become plants, and some of them become trees, and the forest emerges.
21:35
SPEAKER_00: The nuance of open ecosystems will win is somewhat applicable to open weights, not fully.
21:41
SPEAKER_00: So we need to be very careful.
21:43
SPEAKER_00: With open weights, you still have the weights that you can customize.
21:47
SPEAKER_00: You can make them capable for your application, but you don't necessarily have the full stack.
21:53
SPEAKER_00: So you don't know how to start from scratch and pre train a model with random weights necessarily.
21:59
SPEAKER_00: So So there is that limitation of open weights.
22:02
SPEAKER_00: But having said that, there is absolute place for closed models and open models.
22:10
SPEAKER_00: The closed models potentially will be because of the resources and investment that go in and the amount of talent and expertise that is needed to methodically climb the scaling ladder.
22:20
SPEAKER_00: Will probably and most likely continue to lead in capability, will lead in reliability because it's a service that companies provide for money, will lead in uh in readiness.
22:37
SPEAKER_00: Like you can get your application up and running by just calling an API and everything else is taken care of for you.
22:42
SPEAKER_00: So you assume intelligence and you build your application.
22:45
SPEAKER_00: Then on the other side, something like open weights, you have full controllability, which is awesome.
22:51
SPEAKER_00: You can be much more efficient because you are rate limiting yourself on how to develop applications.
22:57
SPEAKER_00: And uh eventually you feel more like you have a sovereign sovereignty over the entire stack that you're building, which could be very, very useful.
23:06
SPEAKER_00: Uh, but you kind of trade off the capability.
23:09
SPEAKER_00: So to me, that's kind of like the nuanced view of this.
23:12
SPEAKER_00: I am a firm believer that open ecosystems will win and open weights is the beginning.
23:17
SPEAKER_00: Once we get into commoditization, we'll probably see more and more of uh paradox and dichotomy of these two.
23:25
SPEAKER_00: The last thing I'll say is uh in the world where new capabilities can have meaningful positive impact and meaningful negative impact, you need to make sure you have designed the systems with AI in the loop to make things safe.
23:39
SPEAKER_00: So cybersecurity by just humans is not going to work out.
23:42
SPEAKER_00: You will need AI in the loop to build more stronger, secure systems.
23:48
SPEAKER_00: And for that, you would need frontier models.
23:50
SPEAKER_00: So when you have new frontier models that are more capable, we also need to figure out how to navigate the safety before we make it completely open.
23:59
SPEAKER_00: So this is something that I feel very, very strongly about.
24:02
SPEAKER_00: So open ecosystems for the for Win and making sure that we are being extra caution, uh taking more caution about the capability, frontier capability is very, very valuable.
24:12
SPEAKER_02: But historically, if you see, right, on closed ecosystems, Apple the most economic value gets captured by a closed ecosystem, example, Apple, example Tesla, uh right, and and then whatever captures the most economic value ultimately wins.
24:28
SPEAKER_02: Let's say, for example, Linex, Red Hat, how much valuation uh they have in the markets.
24:36
SPEAKER_02: And because it's just not that the movement of talent also is it's a capitalistic world, it's decided by you know the capture of economic value.
24:45
SPEAKER_00: Yeah.
24:46
SPEAKER_00: I think this is a good one.
24:48
SPEAKER_00: It depends on your measure of success.
24:50
SPEAKER_00: So if you look at the number of devices that are running Android versus iOS, isn't the number significantly higher for Android?
24:57
SPEAKER_00: But if you look at the market cap or the economic value, then maybe you are right that closed might have a more business value perspective.
25:06
SPEAKER_00: That's why I think in the longer arc, I still feel if the if the core value proposition of a technology is democratizing it, I think open will win.
25:14
SPEAKER_00: If the core value of the technology is packaging it in a way that is easily usable and commoditized, where I don't have to worry about each and every nuanced.
25:28
SPEAKER_00: That's why Apple is such a powerful ecosystem, right?
25:31
SPEAKER_00: Out of the box, the hardware works seamlessly.
25:34
SPEAKER_00: So there is obviously amazing design aspects to it, but there is also the fact that you can control everything from the hardware layer to the not so functioning CD assistant.
25:46
SPEAKER_00: Uh you control all of those aspects.
25:48
SPEAKER_00: So you get a lot more say in what is in, what is out.
25:53
SPEAKER_00: Uh, but you have an upper limit in terms of adoption.
25:56
SPEAKER_00: Whereas an open ecosystem, the adoption will be much, much more uh in terms of numbers, which to me, I think technology that is highly democratizable is more appealing and more motivating compared to technology that is on the flip side, if I'm using a technology that is very commoditized, I would want it to work reliably.
26:17
SPEAKER_00: So in that world, I'm more than happy to personally, I'm more happy to embrace that and say, even if it's a closed ecosystem, I just don't it's it's something that is super commoditized and I want it to work more reliably.
26:31
SPEAKER_00: So it's much more about ease and reliability than innovation and control.
26:36
SPEAKER_02: So as a summary, uh, you know, what you're seeing, we'll see a world, according to you, where uh open models have far more distribution, but the the value, economic value captured by the frontier would be much more exponential as compared to those.
26:55
SPEAKER_00: I think I think that's a good summary.
26:58
SPEAKER_00: The other thing that I would say is we are climbing the ladders of scale, ladder of scale, and you can see that the model in the loop allows you to compound and build the next version much better and faster.
27:13
SPEAKER_00: So, in that sense, I always feel the closed models because of the investment and because of this compounding loops, will stay ahead.
27:22
SPEAKER_00: The only scenario where that is not going to be the case is we have a new paradigm, and that new paradigm overthrows the current paradigm.
27:30
SPEAKER_00: Then I can totally see the world shifting much more, which has happened in the past, and we'll have to be prepared for it.
27:38
SPEAKER_00: Yeah, I mean, all of computer vision, like back in the 1990s, speech in CMU was considered pretty much solid with MFCCs and spectrograms.
27:51
SPEAKER_00: But if you go and experience the audio assistants at that time and the audio experiences of the latest assistants that we have, it's night and day, but it was still considered fairly solid.
28:04
SPEAKER_00: Uh until 2012, uh all of computer vision was done using handcrafted features and slap on your latest whatever graded boosted dictionaries or SVMs.
28:18
SPEAKER_00: 2012 is the first time when Alex Krzychewski and the folks with AlexNet showed that end-to-end learned features are better for visual tasks.
28:28
SPEAKER_00: So it was a new paradigm.
28:31
SPEAKER_00: Connets were the rave.
28:32
SPEAKER_00: Everybody went after CONNETs.
28:35
SPEAKER_00: Then GANs came, which were more discriminative and uh in nature, and you had two networks that are actually being adversarial to each other, and then you have transformers.
28:46
SPEAKER_00: I mean, in between, I'm skipping a bunch of other innovations like ResNets and so on, but you have transformers, and transformer as a block became this absolute winning strategy for language, and then LLMs became VLMs, and then for robotics, VLMs became VLA.
29:04
SPEAKER_00: Now there is WA world action models, VAMs.
29:08
SPEAKER_00: These are all iterative improvements, not fundamental paradigm shifts.
29:12
SPEAKER_00: But when you have a fundamental paradigm shift, like going from handcrafted features to deep learned features, then you kind of have to throw everything, and there might be new scaling laws, there might be new data sets, there might be new skills.
29:23
SPEAKER_00: I still remember the first uh the first uh GPU that I had to assemble on my desk at Facebook because we couldn't have GPUs in the data center.
29:36
SPEAKER_00: That was not a thing.
29:38
SPEAKER_00: The data center has only CPUs.
29:40
SPEAKER_00: GPUs were not scalable because you could not cool them.
29:43
SPEAKER_00: And you need special infrastructure to cool GPUs.
29:46
SPEAKER_00: So we only had CPUs.
29:48
SPEAKER_00: So to have a GPU training, I had we had to basically start off with assembling the GPU on your desk.
29:54
SPEAKER_00: And just from a pure flops perspective, you could basically that GPU that basically took up my entire desk space, you can have 20 of those.
30:07
SPEAKER_00: You actually have 20 of those running on your phone today.
30:10
SPEAKER_00: 20 of those.
30:11
SPEAKER_00: And that is in the last roughly 12 years.
30:14
SPEAKER_00: So that is the rate at which things change.
30:16
SPEAKER_00: Uh that could that paradigm shift could come on hardware.
30:19
SPEAKER_00: That paradigm shift can come in how the models are learned and built, how we do inference.
30:25
SPEAKER_00: Uh, and it's that's the thing that I am most excited about.
30:28
SPEAKER_00: And that's how I see the world, which is certain things that look very impossible today.
30:35
SPEAKER_00: Uh, you kind of ask, what if?
30:38
SPEAKER_00: What if the hardware is 10x better?
30:41
SPEAKER_00: What if the learning algorithm is 10x better?
30:43
SPEAKER_00: What if serving it is 10x cheaper?
30:46
SPEAKER_00: And then what you have done is you have taken this very turbid problem and you have crystallized and simplified it to a few what if statements that then give you an avenue to, oh, if I push on this bottleneck, it can have a huge impact.
31:01
SPEAKER_00: And that has served very, very well for me in terms of how I make project decisions or career decisions.
31:07
SPEAKER_02: Right now, sovereignty in AI is a big discussion in India, I believe globally also.
31:14
SPEAKER_02: And countries are prioritizing sovereignty uh in AI.
31:18
SPEAKER_02: Let's say India is now prioritizing local language models produced in India, like Sarvum, right?
31:24
SPEAKER_02: And government is incentivizing it, giving them GPUs.
31:27
SPEAKER_02: Yeah, right?
31:29
SPEAKER_02: Uh what's your view on it?
31:32
SPEAKER_00: Man, I'm uh half American, half Indian uh at heart, always Indian at heart, always American at heart as well, more global at heart, actually.
31:40
SPEAKER_00: Um I think maybe maybe the best example I have in my head in terms of my view on this is the following, which is if you go back to Cold War, there was sovereignty, right?
31:55
SPEAKER_00: Basically, Russia was racing towards something, US was racing towards something, they were trying to beat each other.
32:02
SPEAKER_00: Looking back, we can say whether it was healthy, unhealthy, and all of these things, but it led to certain innovations.
32:09
SPEAKER_00: So, my simplified view on this is if sovereignty is about innovation and not about imitation, it actually leads to something good for the world.
32:23
SPEAKER_00: So, and I was trying to think about why that is the case.
32:27
SPEAKER_00: And the way I was thinking about it is the best and the brightest people who are very audacious and they don't really calculate the probability of success.
32:38
SPEAKER_00: They just jump into it and they give it their heart and soul and they grind out every day, are people who are actually excited about innovation and doing something that has never been done before.
32:51
SPEAKER_00: So the strongest teams that are highly motivated, super high on talent, and are able to work hard through all the challenges day by day, are the teams that really are going for something that has never been done before.
33:07
SPEAKER_00: So that's the innovation, which we all know.
33:10
SPEAKER_00: It has nothing to do with sovereignty, right?
33:12
SPEAKER_00: It is purely this has not been done before, I want to do it.
33:16
SPEAKER_00: Now, if you add sovereignty to it as a catalyst, which is we as a country have not done this before, it makes it so much more.
33:26
SPEAKER_00: You wake up with an identity and you want to win.
33:30
SPEAKER_00: But if you mix the two, where someone else is winning and have something, we as a country need to have the same thing.
33:40
SPEAKER_00: I think eventually that fades off.
33:45
SPEAKER_00: And what happens is the financial motivations and the economics win out, and you're like, that is better.
33:52
SPEAKER_00: So building on top of that thing that is better is the right thing.
33:55
SPEAKER_00: So sovereignty will not hold too long.
33:58
SPEAKER_00: It holds in the beginning, but eventually will fade off.
34:01
SPEAKER_00: So the point I'm trying to make is there are two points I'm trying to make.
34:06
SPEAKER_00: One point is sovereignty with an angle of innovation will lead to sustained uh energy and outcomes that are valuable.
34:15
SPEAKER_00: And second, I actually think the way I would measure sovereignty is not whether I own a foundation model or not, but do I have the capability in my country to be able to build things that are needed?
34:29
SPEAKER_00: So do I know how to you can't just pick and peace, uh pick a particular part of the stack and say I need sovereignty.
34:35
SPEAKER_00: Sovereignty on a model will mean nothing if you don't have the hardware to run it.
34:40
SPEAKER_00: You're still dependent on some other country for the hardware, right?
34:43
SPEAKER_00: Sovereignty on hardware means nothing if you don't have the economics to actually where the people are not embracing AI.
34:49
SPEAKER_00: So you need the loops, the product loops to be able to actually add value to society.
34:53
SPEAKER_00: So to me, I think of sovereignty in the context of are you having an angle of innovation that is truly yours, that can harness others' innovations and still have your own say?
35:05
SPEAKER_00: Are you participating?
35:06
SPEAKER_00: And second, is are you building the capability at every part of the stack to be able to say in the world in the scenario where the world is more closed than open, can I build what I need?
35:19
SPEAKER_00: Which I think is a very, very valuable thing to harness to build on and to have the capability in a country.
35:26
SPEAKER_00: That could be any country.
35:27
SPEAKER_00: Uh, and I think that will be that's kind of like the lens I see.
35:30
SPEAKER_00: Hopefully, all these topics are super complicated and the geopolitical topics specifically.
35:36
SPEAKER_00: Uh, but it's very important to be more technology-centric and human value-centric than a country-centric or purely blanket sovereignty, sovereignty.
35:46
SPEAKER_02: I have a take on it.
35:47
SPEAKER_02: Let's say uh uh India since independence till 1991, till the economy got opened, there were usually two cars, only Fiat and Ambassador, and Indian folks.
35:57
SPEAKER_02: Probably you would have also seen while growing up.
35:59
SPEAKER_01: Yeah, I still remember the ambassador.
36:01
SPEAKER_01: Uh it's actually one of the most comfortable cars.
36:05
SPEAKER_02: And then Maruti came in with the Suzuki.
36:08
SPEAKER_02: But in US, also, if you see, like what I have heard is uh US doesn't allow BYD to come in.
36:13
SPEAKER_02: Uh right, so they are allowing Tesla to win in this market, uh, right?
36:17
SPEAKER_02: Obviously, Tesla, one of the amazing companies of the generation, no doubt, but the government is allowing Tesla to win.
36:23
SPEAKER_02: Similarly, China is a complete closed ecosystem, one of the largest markets for the world, right?
36:30
SPEAKER_02: But they don't allow anybody else to win.
36:32
SPEAKER_02: So this is sovereignty in different directions.
36:37
SPEAKER_02: Uh I thinking from an Indian perspective, just discussing with you, uh the more close we went, uh the quality which got afforded to people was less.
36:50
SPEAKER_01: Yeah.
36:50
SPEAKER_01: Yeah.
36:51
SPEAKER_02: But the other, and there is no right or wrong answer.
36:53
SPEAKER_02: The China is another ecosystem, and they are developing their country in a different way.
36:57
SPEAKER_02: Where I don't think US car manufacturers or I don't know how Tesla is big, how much Tesla is big in China.
37:03
SPEAKER_02: But current, like, you know.
37:05
SPEAKER_02: And Southeast Asia, if you see, right?
37:07
SPEAKER_02: Because the BYD economics are cheaper, it's it's ruled by BYD.
37:12
SPEAKER_00: Yeah, I don't know how to think of this equilibrium.
37:16
SPEAKER_00: Uh, I could totally see, I mean, the the the craziest example of this is North Korea is completely closed.
37:22
SPEAKER_01: Yeah.
37:22
SPEAKER_00: I don't know if people are happy there.
37:24
SPEAKER_01: Yes.
37:24
SPEAKER_00: I'm assuming not, but I might be wrong.
37:26
SPEAKER_00: We don't know.
37:27
SPEAKER_00: Yeah.
37:27
SPEAKER_00: We don't know what we don't know.
37:29
SPEAKER_00: But uh I think I think my simplified view on this is we should try and be as global as possible to benefit the end user.
37:41
SPEAKER_00: And second, we should try and create ecosystems that allow you to participate in a healthy way.
37:47
SPEAKER_00: Those are like the two principles that I see.
37:50
SPEAKER_00: And if those two principles dictate for you to be closed, maybe that's the right decision for that moment.
37:58
SPEAKER_00: Or if those two principles dictate for you to be open, then maybe that's the right thing.
38:03
SPEAKER_00: And those are at odds with each other, right?
38:05
SPEAKER_00: One is if the user is like, I just want the cheapest and the fastest car or the most good-looking car, it might be the Italian Lamborghini or Ferrari, which you can have an Indian company make it or not.
38:18
SPEAKER_00: Versus uh we want to make sure that in the world of autonomous cars, we are not at a cybersecurity breach where there are big monstrous half a ton vehicles moving around that can be controlled by an AI of another country.
38:38
SPEAKER_00: So there's just like the doom scenario where you are like more closed, and then there is the user opportunity slash flexibility that is fully open, and you have to kind of like look at the equilibrium at each sector and draw the line.
38:54
SPEAKER_00: I am not an economist, I'm a technologist.
38:56
SPEAKER_00: I think in general, technology has always won when we are more open and when we give more flexibility to the user with garters.
39:05
SPEAKER_00: So that's kind of like my simplified view on it.
39:08
SPEAKER_02: That's that's a hard, a hard topic to take sides on.
39:11
SPEAKER_01: Yeah, yeah.
39:12
unknown: Yeah.
39:12
SPEAKER_02: Sunda Pichai, Mark Zuckerberg, Elon Musk, uh, they're all launching their products on X, right?
39:18
SPEAKER_02: Traditionally, they don't it was like completely traditional media launches and newspapers, conferences, blogs.
39:26
SPEAKER_02: It seems the world is shying away from that and just X has become the battleground.
39:34
SPEAKER_00: Yeah.
39:35
SPEAKER_00: Um so I have two takes on it.
39:39
SPEAKER_00: Number one, I think I think the core at the core of it, what is happening is you are trying to control the perception, and the best way to control the perception of a launch is to adapt it.
39:55
SPEAKER_00: So if you look at how traditional launches have happened, or how people used to do product launches, or even like going all the way back to announcements of novel things, maybe the oldest thing is having publicity in a newspaper that you have a new restaurant and it's in page three, or what I don't even know what the numbers exactly are, but then obviously broadcast television, where that's due to branding there and consistently reach more and more, and the exact time of when that where which channel the ad is and so on will dictate.
40:30
SPEAKER_00: And then obviously the online advertising came where it's much more personalized and you're like targeting to the location, to the user, to the age, to the gender.
40:39
SPEAKER_00: To me, that's the evolution of it, which is you are you have something awesome, you want to bring it to the world, but like there is this vibe.
40:51
SPEAKER_00: What's the vibe of the model?
40:52
SPEAKER_00: What's the vibe of the launch?
40:54
SPEAKER_00: You start controlling the vibe of the launch.
40:55
SPEAKER_00: Or how do you control the vibe of the launch?
40:58
SPEAKER_00: You basically put off, put forth, you do step one, and then you see the reaction of the Twitter RTR X folks, and a lot of AI folks across all Frontier Labs, across all the Frontier applications across the world are there.
41:12
SPEAKER_00: They immediately give you the first set of feedback and vibe, and then you adapt to it.
41:18
SPEAKER_00: And the recent, I mean, now I'm not at Meta, but Muse Wonder launch, it looks like uh there were some inference issues in a typical product launch.
41:28
SPEAKER_00: Probably you put it out there, and if there is some issue, then it goes through some channels, and after a week you collate it and then you figure out.
41:35
SPEAKER_00: But you see Alex Wang and a bunch of folks on the team real-time reacting to, oh, we we found some inference issues, we're going to go fix it.
41:44
SPEAKER_00: And then the folks are happy, then they're like, okay, it's not the model quality that's a problem, it's actually something else.
41:50
SPEAKER_00: So the first thing that I that I I kind of see what's happening in X is you are meeting the audience and adapting and making sure the perception of what you have built is positive.
42:05
SPEAKER_00: There's a very important second thing that I that I think is very valuable is perception leads reality, but reality catches up.
42:15
SPEAKER_00: What does it mean?
42:16
SPEAKER_00: So if at the end of the day the product, the model, whatever you're building is not good, no matter how much you control the perception, it'll catch up, right?
42:27
SPEAKER_00: So I actually think of this as fine-tuning your product launch rather than defining your product launch.
42:32
SPEAKER_00: So another thing I was thinking about, and uh this is just an example, is if you look at all the product launches, model launches, there was a lot of uh heated debate of probably very one-sided of the math launches that happened because the math community was not necessarily fully tuned in into X world.
42:58
SPEAKER_00: But when you have a math innovation where you have models solving field level problems and millennium problems, field field metal like problems and millennium problems.
43:08
SPEAKER_00: But if you don't have the audience there, it doesn't fall in the right place, right?
43:12
SPEAKER_00: Like the the debate is very one-sided or or very myopic.
43:16
SPEAKER_00: So you have to be very, very careful, and that is something that I was thinking a lot in terms of making sure you meet the audience, making sure you adapt and control the perception.
43:26
SPEAKER_00: But most important is making sure that you have the meat on the bone, you have your awesome product or model that you're launching that is actually really, really good.
43:35
SPEAKER_00: Because at the end of the day, the reality will catch up to the perception.
43:38
SPEAKER_02: Your journey is like phenomenal, you know, from triple IT Hyderabad, then an intern at Meta to VP at Meta, like very few people uh, you know, reach that.
43:49
SPEAKER_02: And you worked with Mark directly, you worked with Jan Likon directly, and now you're a founder.
43:55
SPEAKER_02: So, what philosophy has guided throughout your life in all these?
44:00
SPEAKER_02: Pivots, decisions.
44:02
SPEAKER_00: I mean I have been very, very fortunate, number one.
44:06
SPEAKER_00: I think very, very, very few people none of it is planned, to be brutally honest.
44:14
SPEAKER_00: And I feel fortunate when I look back.
44:18
SPEAKER_00: But then the people that you mentioned, and many, many more actually, have shaped me as an individual and have shaped my ambition.
44:32
SPEAKER_00: The common thread in all of those risks and decisions has been trying to identify a bottleneck of the current present, right?
44:44
SPEAKER_00: Which is what is not working and why is it not working.
44:49
SPEAKER_00: So it starts with spotting something that I feel is like there is something that's not working.
44:56
SPEAKER_00: And slowly and steadily, methodically reaching out to the experts of that area and learning more about which parts and components are working, which are not, what are the assumptions, why are the assumptions right or wrong, what can change.
45:17
SPEAKER_00: And once you have built enough conviction around that area, so that's the think of it as a learning part of it.
45:24
SPEAKER_00: So there's a spotting part, there's a learning part.
45:26
SPEAKER_00: And the learning is not done by you alone.
45:28
SPEAKER_00: Learning is done with being very, very humble and going to the experts of the world.
45:32
SPEAKER_00: And the best part is most of the experts in the world are always very generous.
45:37
SPEAKER_00: So when you approach them and ask why they are so passionate about that space that they will spend hours describing their expertise.
45:46
SPEAKER_00: So I'll give you a concrete example.
45:47
SPEAKER_00: Just a few months ago, I was really looking into robotics and I was like, oh man, I mean, perception is solved, locomotion is solved, planning has been there for a long time, control is, you know, I mean, it's mathematically very, very strenuous, but most of control was figured out.
46:04
SPEAKER_00: And then you very quickly, after a few hours of conversation, come to the very come to a conclusion from talking to world experts that dexterous manipulation is the biggest bottleneck of robotics, and hand is the ultimate frontier.
46:18
SPEAKER_00: I could get that from a chat GPT conversation or any conversation with an agent, but it is not that statement.
46:25
SPEAKER_00: It is all of the assumptions of these world experts that they're leaving that help you culminate not just the final result, but the whole nucleus of the space, right?
46:36
SPEAKER_00: So that's the learning part where you build conviction.
46:39
SPEAKER_00: And then you will never get a perfect answer of like do X, Y, and Z, and there is success.
46:46
SPEAKER_00: You have to take a leap.
46:47
SPEAKER_00: And that's the leap.
46:48
SPEAKER_00: You once you take a leap, there's no second guessing.
46:51
SPEAKER_00: You take a leap and then you make it work.
46:53
SPEAKER_00: You build.
46:54
SPEAKER_00: You build your own on your own, you build with an amazing team.
46:58
SPEAKER_00: And the most important thing in the build stage for me is having a very high bar of excellence.
47:03
SPEAKER_00: You want to surround yourself with exceptional people, and that exceptional people come either because of the talent or how mission-oriented they are, and they're willing to work tirelessly and relentlessly towards that mission.
47:17
SPEAKER_00: It can come in both ways.
47:18
SPEAKER_00: So once you've you yourself are building or you have built a team around it with that kind of a talent and hunger and exceptional expertise, then you patiently build.
47:30
SPEAKER_00: And the last part, which I some people, it's it's a very different mindset, is to always judge the outcome in terms of our real world, our people in the real world using it.
47:42
SPEAKER_00: So it's not a demo, it's not a paper, it's not a patent, it's not a blog post, it's just not a number on a table, but it's a true manifestation of something in the real world that people are experiencing.
47:53
SPEAKER_00: And you hold yourself to that bar.
47:56
SPEAKER_00: So that's the cycle that has helped me time and again, which is spot, learn and build conviction, leap, build a team and expertise with humility and deliver.
48:10
SPEAKER_00: Like the delivery is not a prototype or a research artifact, it's people in the real world using it.
48:17
SPEAKER_00: And then learn being humble and learning through the entire phase, where you're not approaching anyone of this with, I have done this before, I know how to do it.
48:25
SPEAKER_00: You're approaching with, I've done this before, I'm going to leverage that experience, but I'm again starting from ground zero and I'm going to follow this whole recipe again.
48:33
SPEAKER_00: This has really, really served me as a student in triple IT, as a uh first-time employee in SRI, Stanford Research Institute, which that was my first job, as a student again in Georgitech, as an intern in Meta, and then at the end of it, whatever VP and any other title that you can think of.
48:51
SPEAKER_00: And now as a founder CTO of a new adventure that I'm embarking on.
48:55
SPEAKER_02: Can you share like what was working with Jan Li Kun and Mark?
49:00
SPEAKER_02: And what are their superpowers that make them such big figures in the world today?
49:05
SPEAKER_00: I mean, I think the world has talked enough about them and has seen enough.
49:09
SPEAKER_00: Maybe the the thing that the luxury that I have had is working more closely with them on certain projects and for longer durations since for Jan specifically, since we established FAIR, and even the to date, even though we are not working in the same company, we are still in touch and we talk.
49:35
SPEAKER_00: I think Jan is a scientist at heart, and he is extremely he's he has a fundamental belief that the existing paradigms will have limitations and there needs to be new paradigms, and we all like we have to assume and be prepared for those new paradigms, and very rarely do you see a person who has been so categorically aligned on one thing for a deck two decades, and then the entire community is basically saying that's wrong, that's wrong, that's wrong, that's not the way, and then it's right.
50:15
SPEAKER_00: And even with these world models and embedding Jappa and all of these models that he's I don't think at a at a technical level uh I would necessarily bet against Jan or for Jan.
50:32
SPEAKER_00: It's so that's that has been very, very uh valuable for me.
50:37
SPEAKER_00: And one of the other things that Jan has been very thoughtful and has helped me shape is he I used to fundamentally believe research and engineering, I still fundamentally believe research and engineering coming together will have a tremendous amount of impact.
50:53
SPEAKER_00: And this is where I and Jan don't necessarily agree, but it's it's awesome to actually not agree with some people who are luminarists like Jan.
51:01
SPEAKER_00: And he is like, leave research alone and they will produce amazing outcomes, which is also true.
51:07
SPEAKER_00: So that dichotomy is real.
51:09
SPEAKER_00: Uh have a research lab, give them absolute freedom, give them absolute resources, and leave them alone, you will get innovation.
51:16
SPEAKER_00: And then the other end of the spectrum is you take a very, very hard problem, you bring in a multidisciplinary team of researchers and engineers, you set goals, audacious goals, and you execute, will also lead to amazing systems.
51:30
SPEAKER_00: So that's in a nutshell.
51:33
SPEAKER_00: Like there is an amazing amount of overlap and respect where I have learned so much from Jan in terms of building really, really strong research teams, giving space and time and resources to explore for the research teams.
51:45
SPEAKER_00: On the flip side, my own view, worldview, which is don't isolate research and have this two-tier system of research and engineers, but bring them together in a multidisciplinary way and give them audacious goals and uh set milestones to make progress towards those goals, will lead to better outcomes.
52:04
SPEAKER_00: So that's Anyan.
52:06
SPEAKER_00: Mark, I I I've been fortunate to work closely, uh, but of course, there were many leaders that were always a buffer for me.
52:13
SPEAKER_00: Um they were more working closely with Mark, and I had the opportunity to work closely in the context of technical stuff, which is how do we build the models, how do we think about open source, how do we think about the products on these models and what technology, what investments.
52:30
SPEAKER_00: At the heart of it, I have not seen a more fearless leader.
52:34
SPEAKER_00: Uh and uh I think I think two unique things about Mark that probably the world misunderstands because of how um I guess the media narrative is and meta or Facebook narrative is, and now I'm not an employee, so I can more talk more freely, probably.
52:56
SPEAKER_00: But uh I would say the same thing even as an employee, I guess, is Mark genuinely wants to build and uh connect technology to value for people.
53:13
SPEAKER_00: So that's the driving force for him.
53:15
SPEAKER_00: And second, he is a very uh he is one of the fastest learners in terms of adapting to new waves and taking bold risks, where even if you understand the technology at first, you probably are circumspect, but he has this absolute clarity where he's willing to take the risks.
53:38
SPEAKER_00: And that is something that I have not seen in other founders uh yet, to the extent that Mark has done.
53:44
SPEAKER_00: And I still remember the Instagram acquisition.
53:46
SPEAKER_00: I was an intern at that time, and everybody in the company would be like, why would we spend a billion dollars for an app that is a camera app?
53:54
SPEAKER_00: And Mark was like, Show me an app that is growing at this rate after a few weeks of development, uh, week over week, month over month, and nobody has an answer.
54:03
SPEAKER_00: And now you look back and everybody's like, of course, we should have done this.
54:06
SPEAKER_00: So there's so many instances of it where you don't necessarily need to be super close.
54:11
SPEAKER_00: Uh, but I was very, very fortunate because when the AI group was created, he used to sit with the AI group, and I was there and I was an intern and then I became a full-time engineer, and he would want to understand every little detail of things to help shape his mental model.
54:29
SPEAKER_00: And then the last thing I would say is once you like once he understood an area, he would try and simplify it to the core to make strategic decisions.
54:45
SPEAKER_00: That is something that I really valued and I learned by observing.
54:49
SPEAKER_00: So I used to always think about if I understand a space, I will understand the full complexity of it.
54:57
SPEAKER_00: But actually, then I realized expertise is not about understanding the complexity alone.
55:02
SPEAKER_00: That's half of it.
55:04
SPEAKER_00: The other half is using that to simplify as much as possible and get to the core tenants that then help you make strategic decisions.
55:14
SPEAKER_00: And I think Mark, for me, I have not witnessed the better person there.
55:19
SPEAKER_02: Yeah, and I think even uh every time I'm surprised, let's say the scalia acquisition, it just the surprises, the kind of bold, big risks that that he takes.
55:32
SPEAKER_02: And ultimately, I think we said, you know, leaders are right often.
55:38
SPEAKER_00: Yeah, I mean, look, I think it's always easy to celebrate the wins.
55:43
SPEAKER_00: Uh a lot of people have criticism on the investment that went into reality labs, but now the glasses is a very uh hit amazing product that I use personally a lot, but also has a lot of controversy associated with it.
55:57
SPEAKER_00: So I think it's really, really valuable, as you said, to judge based on the outcomes that are highly successful and the asymmetric outcomes that are successful.
56:09
SPEAKER_00: But it's also valuable to see what is the framework an individual who has been successful repeatedly, what is the framework they're using rather than the outcome.
56:21
SPEAKER_00: Because sometimes the outcomes are good, sometimes the outcomes are bad.
56:24
SPEAKER_00: But if the framework is resilient, then you benefit from it, right?
56:27
SPEAKER_00: You look at a leader, you're at least for me, the realization is when I look at Jan, when I look at Mark, or I mean, I have never worked with Elon or many of Bezos and so on.
56:39
SPEAKER_00: What I appreciate is when they share the frameworks they use, that I feel is more valuable for humanity and for other people to understand and then shape it into their own view or their own style.
56:56
SPEAKER_00: That is something that I would encourage more people.
56:58
SPEAKER_00: So that's what I try to do.
57:00
SPEAKER_00: So for me, it's not about emulating Mark or emulating Jan and so on.
57:03
SPEAKER_00: Like I said, I disagree in some cases.
57:05
SPEAKER_00: Uh, I think the boldness comes with risk appetite and uh could have ramifications on evolving the product much faster than it's ready, for example.
57:15
SPEAKER_00: You can take a different view on it.
57:18
SPEAKER_00: How you build teams, research teams, where research and engineers are two separate classes of people versus it's intermingled.
57:26
SPEAKER_00: I disagree.
57:27
SPEAKER_00: Uh, and I think that's where you have to stand out.
57:30
SPEAKER_00: And there are a bunch of things I'm hoping the next generation or uh the current generation will disagree and build.
57:37
SPEAKER_00: Uh, and that entropy in the system is a good thing.
57:40
SPEAKER_00: Otherwise, you have a very monocular culture at a at a at an industry level, which is not good.
57:45
SPEAKER_02: What are the things that you disagree with, Mark on?
57:49
SPEAKER_00: I have I am in no position to disagree or agree uh there as a at an individual because the amount of things he manages and what the burden that he carries is very different.
58:03
SPEAKER_00: Um it's it's less about disagreement.
58:07
SPEAKER_00: I my my current focus or what I am personally motivated is the the amount of investment that is going into AI, we need to work tirelessly over the next few years to connect it directly to value for humanity, not in an indirect way where it will get better because you have better tools.
58:36
SPEAKER_00: I think that is where my motivation is right now, and I'm very excited about that.
58:42
SPEAKER_00: And there are many, many problems to be solved there, and uh that's kind of like the focus area for me personally.
58:49
SPEAKER_02: You know, tell us uh more about you quitting at the peak of your career, uh, right?
58:55
SPEAKER_02: And and you have a team that worked at Frontier and played a role in AI advancements as much as it could have, right, for any other team, right?
59:04
SPEAKER_02: So introduce us to the idea, vision, team, and challenges.
59:10
SPEAKER_00: So uh I'm super, super excited about Warp uh because of the people and the mission.
59:18
SPEAKER_00: Um I had a very comfortable job uh even before I went to my PhD.
59:30
SPEAKER_00: I was actually in SRI, it was one of the best research places in the world doing RD.
59:36
SPEAKER_00: I was just an undergrad.
59:38
SPEAKER_00: I was able to do RD, build amazing systems.
59:40
SPEAKER_00: One thing in SRI that I learned is you have to build real-world systems.
59:46
SPEAKER_00: Papers are an artifact of it.
59:48
SPEAKER_00: And that was something that stuck with me.
59:50
SPEAKER_00: And then uh I wanted to sharpen my skills and I went to PhD, and uh I made this very crazy mistake where uh somebody told me that you always want to do PhD in a domain that is not uh raving right now, but has the potential to be mainstream by the time you graduate.
1:00:08
SPEAKER_00: So I thought robotics is very rudimentary back in 2009.
1:00:12
SPEAKER_00: So I'll take robotics.
1:00:13
SPEAKER_00: By 2014, robotics will be mainstream.
1:00:16
SPEAKER_00: Completely wrong by almost a decade, more than a decade, right?
1:00:20
SPEAKER_00: Uh but but I did jump into robotics and I learned very quickly that the hardware was never ready and the software was catching up, and AI or perception systems were not at all were the big bottleneck.
1:00:32
SPEAKER_00: So I jumped into perception and then deep learning happened, and I just stuck with deep learning.
1:00:36
SPEAKER_00: Then I made a switch to generative AI.
1:00:38
SPEAKER_00: I'm only giving the switches that worked.
1:00:40
SPEAKER_00: There are probably many, many things that didn't in the in that path.
1:00:45
SPEAKER_00: But the the underlying assumptions on the arc have been the same thing, which is spot the bottleneck, build conviction by talking to the experts, by leaning on the experts, leap, build, build with excellence in mind, and always land the outcome, right?
1:01:08
SPEAKER_00: As I said, I think the frontier model progress will happen, but we need to connect that to the physical world.
1:01:15
SPEAKER_00: That was the thread that I was trying to pull.
1:01:17
SPEAKER_00: And as part of that, I got in touch with Kevin Weil, who um was the ex-CPO of OpenAI.
1:01:26
SPEAKER_00: And uh he left OpenAI as a VP and uh was doing life sciences there.
1:01:32
SPEAKER_00: And he and I started talking, and he told me about the fundamental bottleneck of these models because of how data-poor the environment is, is the wet lab automation and how we need to make sure that that infrastructure needs to be built quickly.
1:01:52
SPEAKER_00: That will then lead to discovery.
1:01:54
SPEAKER_00: And then very quickly it became a robotics problem and a foundation model problem and has huge impact for humanity.
1:02:00
SPEAKER_00: And then I was like, but wait, this is not a technology problem alone.
1:02:06
SPEAKER_00: This needs science, and we need a team of scientists on the chemistry and biology side to be with us, guiding us, because I can't, if I wake up and I say there are a bunch of arms that are actually automating some part of it, which part?
1:02:21
SPEAKER_00: Why?
1:02:22
SPEAKER_00: Is that the true bottleneck?
1:02:24
SPEAKER_00: So I then met and spent a lot of time with Saul Riesberg, who was a Terremoto before and Schrödinger and so on, and cannot imagine a better chief science officer that I can work with.
1:02:37
SPEAKER_00: And he had this absolute clarity that if we can focus on small molecules and help there, and it's a very small part of the full pipeline of discovery and lead optimization.
1:02:48
SPEAKER_00: But if we can help the drug discovery companies make it more efficient, make it faster with each iteration, we can have a huge impact on the final mission.
1:02:58
SPEAKER_00: Then I met DJ Partel, who unequivocally basically took his career and optimized for health and devoted health, is where he was.
1:03:09
SPEAKER_00: He is, and now he'll be the COO of our company.
1:03:12
SPEAKER_00: And then they asked me, can you be the CTO?
1:03:15
SPEAKER_00: We cannot find a better CTO who has robotics background, foundation model background, who has shipped and built real-world systems.
1:03:23
SPEAKER_00: I mean, you look at the set of people and the motivation, and then you can be part of that.
1:03:29
SPEAKER_00: That's one part.
1:03:30
SPEAKER_00: And second, you look at the mission, which is it's a fundamental frontier of learning that if even if you have 1% improvement, we'll have a crazy amount of impact.
1:03:42
SPEAKER_00: And I'm not a person who will go for 1% improvement.
1:03:45
SPEAKER_00: I'll go for solving the space and lifting the tide.
1:03:50
SPEAKER_00: This is how I have approached all my career.
1:03:52
SPEAKER_00: I think those two things were so compelling that it was never about leaving Meta or leaving VP or moving away from frontier.
1:04:01
SPEAKER_00: And it was more about getting out of my comfort zone and going back to day zero of building with the team with this set of people on that mission.
1:04:10
SPEAKER_00: And that is a once-in-a-lifetime opportunity.
1:04:14
SPEAKER_00: And if we are successful, then hopefully we can help the whole space and all of the drug discovery companies to go towards their mission much, much more faster.
1:04:25
SPEAKER_00: I think that's kind of what is like absolutely exciting.
1:04:29
SPEAKER_00: And uh, I I it wasn't it never occurred to me about what am I leaving, what is the opportunity cost.
1:04:34
SPEAKER_00: It was always like, what is the opportunity cost of not doing it?
1:04:37
SPEAKER_00: And that is something that I could not live with.
1:04:40
SPEAKER_02: So let's say what would the success metric be?
1:04:43
SPEAKER_02: If WAP is successful, what would the world look like?
1:04:58
SPEAKER_00: Getting new drugs to people much, much faster than ever possible.
1:05:08
SPEAKER_00: But underneath it, it's something more nuanced.
1:05:12
SPEAKER_00: It is producing new compounds per week per chemist at an unprecedented scale, that lead optimization and drug discovery is not bottlenecked by the chemistry part of it.
1:05:32
SPEAKER_00: That is the beginning.
1:05:33
SPEAKER_00: I actually think the eventual goal for Warp is making discovery compound with each new experiment that we do in the real world.
1:05:46
SPEAKER_00: Done by humans, done by robots, done by in a coexistence where it compounds the learning.
1:05:52
SPEAKER_00: And the next discovery is faster.
1:05:56
SPEAKER_00: And that is what is the true metric of success.
1:05:59
SPEAKER_00: So warp.
1:06:00
SPEAKER_00: Wants to make discovery compound, make discovery faster.
1:06:03
SPEAKER_00: And we definitely want to start with a very focus of small molecules in lead optimization to help drug discovery companies achieve their goals faster.
1:06:13
SPEAKER_00: And the reason for that focus is robotics and physical work, physical intelligence is in a much more rudimentary stage, but is very fast, like becoming mature very, very fast.
1:06:26
SPEAKER_00: So in our lifetime and over the next few years, we will see an explosion of automation in this space.
1:06:33
SPEAKER_00: And I want Warp to play a part, lead the part, and make it a reality.
1:06:39
SPEAKER_00: So either we or many, many other companies motivated by us and the progress that we are making are actually bringing and having that impact for humanity.
1:06:48
SPEAKER_00: I think that's what makes me wake up every day with that energy and optimism.
1:06:53
SPEAKER_02: But you're also dependent on pharmaceutical companies to move at pace that you guys will be moving at.
1:06:58
SPEAKER_02: How will you make, let's say, the elephants dance?
1:07:04
SPEAKER_00: I like the elephants dance.
1:07:06
SPEAKER_00: Um look, we could be moving at warp speed, but to your point, I think we need to bring along everybody in the community.
1:07:15
SPEAKER_00: That is why I think of warp as less about doing a particular thing and more about what is the fundamental bottleneck.
1:07:25
SPEAKER_00: And if I can go and solve that bottleneck, there is an industry-wide incentive to basically now say I can do more because you have unlocked the bottleneck.
1:07:38
SPEAKER_00: You can see actually the evolution of how foundation models penetrated code, penetrated math.
1:07:49
SPEAKER_00: It is eventual where it will penetrate chemistry and biology.
1:07:53
SPEAKER_00: And I think the hard conversations have to happen at various levels.
1:07:57
SPEAKER_00: There is a geopolitical hard conversation.
1:08:00
SPEAKER_00: Are we always going to be dependent on wet labs that are physically run by some parts of the world?
1:08:05
SPEAKER_00: Or do we need to onshore it?
1:08:07
SPEAKER_00: There is an existing industry that will basically be very wary of using foundation models or automation, but would rather want somebody else to take the risk and depend on that because it's highly risky.
1:08:21
SPEAKER_00: There could be challenges around how do we compound the loops?
1:08:26
SPEAKER_00: Can we do the optimization within small bodies or do we need small molecules or do we need to add data from small molecules and antibodies and so on and so forth?
1:08:34
SPEAKER_00: So, like I said, the framework is simple.
1:08:38
SPEAKER_00: You identify the bottleneck, you have built enough conviction because you assemble a team of world experts.
1:08:43
SPEAKER_00: I cannot think of a better team than Kevin or Saul or DJ or myself and the set of people that we have assembled even in the first few weeks of our formation of the company.
1:08:52
SPEAKER_00: The founding team I look at and I almost walk into the room thinking, am I the one slowing them down?
1:08:58
SPEAKER_00: That's a good feeling.
1:08:59
SPEAKER_00: You want that imposter syndrome because then you know you have built a rock star team.
1:09:03
SPEAKER_00: And then you take each and every challenge.
1:09:06
SPEAKER_00: And whether it's making the elephants dance or you yourself becoming an elephant over time, everything is possible.
1:09:13
SPEAKER_00: You have to believe in it.
1:09:15
SPEAKER_00: And then showing up every day with each challenge and finding a solution, that's the builder mindset.
1:09:20
SPEAKER_00: Uh, somehow there is this notion that because I was a VP of a bigger company for so long, it was more like about growing the ranks.
1:09:29
SPEAKER_00: But deep down, many of the decisions and many of the lessons that I've learned because of how Meta as a company is and how Mark as a leader is, is the builder mindset.
1:09:37
SPEAKER_00: You basically wake up and you take the challenge and you navigate the challenge and you become better.
1:09:44
SPEAKER_00: And there is value that gets created because of that.
1:09:46
SPEAKER_00: And when there is value, people flock and people want to use you.
1:09:50
SPEAKER_00: So that is how I think about it.
1:09:52
SPEAKER_00: I am less worried about the existing way of doing things because everybody has now rationalized that AI, when done properly, with the right data, with the right methodology, um, with the right engineering scaffolding around it and talent that is actually very optimistic, will lead to intelligence.
1:10:18
SPEAKER_00: And at warp, we are trying to go after scientific superintelligence.
1:10:22
SPEAKER_00: Not anything less.
1:10:24
SPEAKER_00: It is scientific superintelligence.
1:10:26
SPEAKER_00: That means we believe AI will unlock scientific superintelligence, and that will unlock huge value for humanity.
1:10:33
SPEAKER_00: And we want to make it happen.
1:10:35
SPEAKER_00: There's no second thoughts about it.
1:10:37
SPEAKER_00: We want to make it happen with everybody.
1:10:40
SPEAKER_00: We want to make it happen by bringing in all of the companies that we would want to partner with, including existing companies, new companies that fund, because it is valuable for humanity and it can have a tremendous amount of positive impact for society.
1:10:53
SPEAKER_00: That is what I believe in personally, and the team believes in.
1:11:03
SPEAKER_00: No, I mean, look, uh you you guys have backed so many amazing founders and uh such amazing missions.
1:11:09
SPEAKER_00: Um, at Warp, one of our goals was to actually have a wide set of people who believed in this mission.
1:11:19
SPEAKER_00: Uh obviously, in the current environment, raising funds is not the challenge.
1:11:25
SPEAKER_00: It's raising funds with the right set of people who are willing to back us for our mission and play the long game.
1:11:31
SPEAKER_00: Because this space will need consistent belief, even when there is an upturn, downturn, things are working, things are not working.
1:11:40
SPEAKER_00: So I am very grateful both for you and for the whole uh neon fund and uh and the set of people because from the onset of an idea to the formation of the team and to the execution, uh, the consistency with which you guys have backed us the idea and obviously even the future, uh, that's very awesome.
1:12:04
SPEAKER_00: And that's what we are looking for as we are thinking about assembling a set of people that will actually help us achieve this mission.
1:12:12
SPEAKER_00: Uh, so I'm I'm grateful that you are part of it, and I'm grateful that uh you will continue to be part of it.
1:12:19
SPEAKER_02: Thank you.
1:12:20
SPEAKER_02: Um thing that I'm always curious about, right?
1:12:23
SPEAKER_02: And I want to uh, you know, it's not like Silicon Valley is better than the rest of the world, it's different from the rest of the world.
1:12:32
SPEAKER_02: And we often compare Bangalore or other parts of the world to Silicon Valley.
1:12:37
SPEAKER_02: What makes it, right?
1:12:38
SPEAKER_02: What gives us the risk-taking ability?
1:12:40
SPEAKER_02: What gives us fearlessness?
1:12:43
SPEAKER_02: What gives us like nobody could imagine a five-year-old company could be trillion dollars in Silicon Valley?
1:12:49
SPEAKER_02: It's possible.
1:12:50
SPEAKER_02: What gives us the art of possibility?
1:12:56
SPEAKER_00: I have been here since 2012.
1:13:00
SPEAKER_00: It predates me in terms of what all happened.
1:13:03
SPEAKER_00: But we have read about it, we're living the outcome of it, and we are trying to shape the future of it, I guess.
1:13:09
SPEAKER_00: The simplistic view I have is there is enough critical mass of believers who are pretty much undaunted on the impossible, and it is a self-fulfilling prophecy of sorts where something that is crazy a day is normal the next day, and is almost abnormal to not be normal the next day after.
1:13:51
SPEAKER_00: That rate of adaptation and the crazy ambitious belief in the concentration of people is almost unprecedented.
1:14:05
SPEAKER_00: So it it is, do you know?
1:14:06
SPEAKER_00: I mean, you you have you taken a Mumbai local train?
1:14:08
SPEAKER_02: Yeah, yeah, yeah.
1:14:09
SPEAKER_02: You almost get pulled out.
1:14:10
SPEAKER_02: You don't take it, just hold on to it and wait for it also to get pushed out.
1:14:15
SPEAKER_00: So, did you ever plan on getting into the train that way?
1:14:18
SPEAKER_01: No, no, no.
1:14:19
SPEAKER_00: You stand there, but you know you get in there.
1:14:21
SPEAKER_00: What is so you just want to survive the train.
1:14:24
SPEAKER_00: Correct.
1:14:24
SPEAKER_00: But once you get in, what is the feeling?
1:14:27
SPEAKER_00: You made it.
1:14:28
SPEAKER_00: You made it.
1:14:29
SPEAKER_00: Did how much of it was you versus how much of it was the inertia of everybody else goal to getting in there?
1:14:33
SPEAKER_00: But individually, everybody's feeling that.
1:14:35
SPEAKER_00: Yeah.
1:14:36
SPEAKER_00: Collectively, everybody's in, right?
1:14:39
SPEAKER_00: I almost think of that.
1:14:41
SPEAKER_00: Maybe I grew up in Mumbai and that's the analogy that comes to my mind.
1:14:44
SPEAKER_00: But it's it's that.
1:14:46
SPEAKER_00: It it is that critical mass of insanity at an individual level, everybody's thinking everybody else is insane.
1:14:53
SPEAKER_00: But an electric collect at a collective, everybody's insane to go after those ambitious things.
1:14:58
SPEAKER_00: And I think the resources that are needed support that ambition.
1:15:06
SPEAKER_00: So the fact that Neon is helping us and so many other folks, including the many, many angels and uh uh Lightspeed, that is our main investor.
1:15:17
SPEAKER_00: Uh what do they see?
1:15:19
SPEAKER_00: They see the ambition, they see the talent of the team, and they see this track record of navigating challenges.
1:15:26
SPEAKER_00: And that multiplied by a 10,000 times, and the density of people and the ideas across this 30-mile radius is something that is so unique and it's perpetuating, which is the previous success leads to the next one, the next one leads to the next one, and so on and so forth.
1:15:44
SPEAKER_00: So I haven't seen that ever.
1:15:47
SPEAKER_00: And to me, I'm really, really glad that I came here in 2012, and I'm really glad after 15 years, that 14 years I'm still here, and hopefully, over the next 14 years, I'll see a few more waves of it.
1:15:58
SPEAKER_00: And that's that's the simplified view I have.
1:16:00
SPEAKER_00: Uh, I don't have a good answer for it, but uh that's that's my view.
1:16:04
SPEAKER_02: So, in in our interactions, what you know, my observation is your superpowers are you know less about technical, though you are technically so much, you know, you know, uh amazing, right?
1:16:16
SPEAKER_02: Among let's say the top 50 people in your field or top 10, I don't know.
1:16:19
SPEAKER_02: Uh but you have the gift of you know looking into the future and saying that I'll solve this problem no matter how hard it is.
1:16:27
SPEAKER_02: And and the gift of storytelling.
1:16:29
SPEAKER_02: How how did you build these skills?
1:16:31
SPEAKER_02: Because for our audience, these are like look like almost impossible skills.
1:16:35
SPEAKER_02: And if you see today kids in India, kids in Hyderabad, Mumbai who are looking at this podcast and imagining a future where they aspire to be you, right?
1:16:47
SPEAKER_02: How do they get there?
1:16:49
SPEAKER_00: I hope they are aspiring to be more than me and much better versions.
1:16:53
SPEAKER_00: Um but uh the the simplified view I have is okay, number one, make sure that you are always optimistic.
1:17:12
SPEAKER_00: I think the world will beat pessimism into you.
1:17:15
SPEAKER_00: Your family, your friends, they all want good stuff for you.
1:17:20
SPEAKER_00: That's why they will be possibly pessimistic and try to make sure you take less risk in life.
1:17:25
SPEAKER_00: What are they trying to do?
1:17:26
SPEAKER_00: They're actually trying to make sure you're safe, you play safe.
1:17:32
SPEAKER_00: But your inherent energy should not fight them, but fight that conservatism and negativity and surround yourself with more people and inherently bring that optimistic view.
1:17:46
SPEAKER_00: So I can find 99 reasons to fail.
1:17:49
SPEAKER_00: But if there is one reason that it's successful and you can make that one much more higher probability over time with your effort, it's worth betting on.
1:17:57
SPEAKER_00: I think that's one thing that I have learned and I would love for the next generation.
1:18:01
SPEAKER_00: And I actually think the next generation is probably way better at this than at least what I was and how I was upbringing, what my upbringing was.
1:18:08
SPEAKER_00: So that's number one for me, which is be very optimistic.
1:18:12
SPEAKER_00: And take the sovereignty example, take the open or closed model examples, take the physical intelligence, leaning into physical intelligence, whichever one it is.
1:18:20
SPEAKER_00: All of that story starts with what is the optimistic view of this for me, for the people that I care about, for the society that I want to build or I want to be part of.
1:18:30
SPEAKER_00: That's number one.
1:18:32
SPEAKER_00: Number two is always have a learning and growth mindset, which means what you know is minuscule.
1:18:42
SPEAKER_00: What others will teach you is the whole world, is the universe.
1:18:47
SPEAKER_00: So if you have a learning and growth mindset to it, then the unknown will look like an opportunity to learn than a fear of I don't know that.
1:18:59
SPEAKER_00: So I don't know how to do it, I don't know how to learn it, I don't know how to use it.
1:19:03
SPEAKER_00: So a lot of people are afraid of AI taking jobs or making whatever you're learning today a commodity.
1:19:09
SPEAKER_00: I think that's reasonably warranted.
1:19:12
SPEAKER_00: But if you flip the coin and say everybody will have to navigate it.
1:19:16
SPEAKER_00: So if I can navigate it better by learning better, then I'm going to be in a better position.
1:19:21
SPEAKER_00: And if everybody does it, then the whole society gets better, right?
1:19:24
SPEAKER_00: So that's the second part.
1:19:26
SPEAKER_00: And the third is a fundamental belief that you will find a way.
1:19:31
SPEAKER_00: No matter what.
1:19:33
SPEAKER_00: Because you have people that you are surrounding yourself with, because you have well-wishers and you pay it forward.
1:19:38
SPEAKER_00: So if someone has done a help for you, then you go and help.
1:19:42
SPEAKER_00: And if you have these three the belief, the optimism, the learning mindset, and the belief that it will work out, it will work out.
1:19:52
SPEAKER_00: I can assure you.
1:19:54
SPEAKER_00: I have never been disappointed.
1:19:55
SPEAKER_00: I'm one data point on it.
1:19:58
SPEAKER_00: Uh, I have every other data point that I've seen has worked.
1:20:02
SPEAKER_00: And the underlying part is showing up every day and working hard because nothing is free.
1:20:07
SPEAKER_00: So if you can make those ingredients a way you live your life with, always be optimistic, always be learning, always have the belief and show up every day.
1:20:18
SPEAKER_00: Keep working at it.
1:20:19
SPEAKER_00: I don't think you can fail.
1:20:21
SPEAKER_00: There is no room for failure.
1:20:24
SPEAKER_02: My last question before we conclude this amazing podcast, Manal, is uh we we know your success, right?
1:20:30
SPEAKER_02: Through this podcast, through what you have done previously, everything, right?
1:20:35
SPEAKER_02: Can you share with the audience the struggles that you have been to?
1:20:38
SPEAKER_02: Because you have been normal kids like everybody else will be listening to this podcast, right?
1:20:42
SPEAKER_02: Grew up in middle class India.
1:20:45
SPEAKER_02: Uh right.
1:20:47
SPEAKER_00: So I have been very, very fortunate because my parents have instilled education in me from day zero on.
1:20:53
SPEAKER_00: I have been very, very fortunate uh to have an amazing sibling.
1:20:57
SPEAKER_00: She uh so she's a doctor now, and uh she and I had a bet on who will become doctor first.
1:21:03
SPEAKER_00: I never became one.
1:21:04
SPEAKER_00: So she definitely won that.
1:21:06
SPEAKER_00: Now I'm actually taking a revenge by doing warp.
1:21:08
SPEAKER_00: I kept I I keep telling her that you know, I've never done a medical, I have I've not been a doctorate from a PhD or a medical perspective, but for the first time, maybe I can build a company that can actually make it happen.
1:21:18
SPEAKER_00: I've been very I I lucked out with an amazing wife who basically puts up with all my limitations, and we have two kids.
1:21:26
SPEAKER_00: Um the struggles are first and foremost disbelief.
1:21:31
SPEAKER_00: You always have an imposter syndrome in your life when you're trying to do something big.
1:21:36
SPEAKER_00: The first two decades or three decades of my life, I was very insecure of myself, of what I can achieve, of my ability to learn, of my ability to stand out and be in a room where I am the, I guess, the underdog.
1:21:52
SPEAKER_00: But I kind of harness that.
1:21:54
SPEAKER_00: Now I approach every room as if I'm the underdog, I'm doing the right thing.
1:22:00
SPEAKER_00: That means I get to learn from everybody in the room.
1:22:03
unknown: Right?
1:22:03
SPEAKER_00: So the struggle is that imposter syndrome and uh just working very, very hard and giving more than what you can think you can give.
1:22:15
SPEAKER_00: And in especially in those days where uh things don't work out, a particular launch didn't work out, a particular team member doesn't think highly of you, or you're not able to bring your A game, but you still show up, you still try to push through because the next day might be a little better.
1:22:33
SPEAKER_00: I think that has been a very important struggle that is much more self-centered.
1:22:38
SPEAKER_00: Uh, I don't know how much, I mean, uh, in India we don't wake up with health as a core thing, running and fitness and so on.
1:22:47
SPEAKER_00: Uh, I was very, very fortunate to basically during my PhD have a lot of friends who were more European and who were European and who were visiting, and they would teach me some survival skills of running and being fit.
1:23:02
SPEAKER_00: And then I took on meditation and fitness as a core goal of my life.
1:23:08
SPEAKER_00: So I went on a 10-day water fast.
1:23:11
SPEAKER_00: That was a struggle, but it's mental fortitude that you build.
1:23:15
SPEAKER_00: I ran half marathons.
1:23:17
SPEAKER_00: I've never run a half marathon, so that's something I have failed yet.
1:23:21
SPEAKER_00: I have run half marathons to stay fit because now I can go multiple days or hours uh without food or without any break, obsessed about a problem.
1:23:32
SPEAKER_00: So I have done back-to-back red eyes, six cities, five, seven cities over five days and given talks, and a lot of people are like, How do you do it?
1:23:40
SPEAKER_00: It's like, because I'm fit.
1:23:42
SPEAKER_00: So the struggles are very like so the way I was thinking about it is there is, and Nat Friedman actually gets credit for it.
1:23:51
SPEAKER_00: He's like, uh, I want to be macro optimistic and micro pessimistic.
1:23:56
SPEAKER_00: So macro-optimistic is having this belief that things will work out.
1:24:01
SPEAKER_00: So if you have a fitness goal, having that belief that things will work out, I'll be fit.
1:24:06
SPEAKER_00: If you have a product goal or if you have a technology goal or if you have a personal goal, whatever it is, you want to be always macro optimistic, which is you believe that it will work out.
1:24:15
SPEAKER_00: And you want to be micro-pessimistic.
1:24:17
SPEAKER_00: So every day you wake up, you're like, this is unhealthy.
1:24:21
SPEAKER_00: I'm not going to basically eat it.
1:24:25
SPEAKER_00: I should not take a break.
1:24:27
SPEAKER_00: I should learn this.
1:24:28
SPEAKER_00: I am at a conference, I don't feel like socializing, but I need to socialize to understand what everybody's thinking, how they are approaching the problem.
1:24:35
SPEAKER_00: So you want to wake up with that hunger every day, every moment, where you're pessimistic that you're not doing enough.
1:24:42
SPEAKER_00: You have to do a little bit more, you have to do a little bit more, you have to actually aspire for a little bit more excellence.
1:24:46
SPEAKER_00: If you combine that, you eventually uh end up uh each struggle becomes a good memory.
1:24:54
SPEAKER_00: That's kind of how I see it.
1:24:59
SPEAKER_00: Thank you for having me and uh thank you for doing this.
1:25:02
SPEAKER_00: Uh it it it really helped me see who I am better.
1:25:06
SPEAKER_02: Thank you so much.