Is Your Startup Model Proof? Vijay Krishnan Turing Founder & CTO On What Startups Should Build to Win

What if the company quietly making OpenAI, Google, Anthropic, and Meta's models smarter was founded in India?

Turing is one of the companies shaping how AI is advancing. It started in 2018 as a talent platform that found the top 1% of the world's engineers, became a unicorn in 2021, and then made a bet almost nobody understood at the time. 

In early 2022, long before ChatGPT existed, Turing began helping OpenAI improve its models by feeding human expertise directly into training. Today its network of more than four million vetted engineers and domain experts powers the post-training and evaluation work behind the frontier labs, and the company crossed roughly 300 million dollars in revenue while staying profitable, at a 2.2 billion dollar valuation.

Vijay Krishnan is the co-founder and CTO. He was an NLP researcher at Stanford back when almost no one believed that predicting the next word could ever turn into reasoning, and he explains why running a modern model company without human-in-the-loop data is like entering a race with three tyres instead of four. He walks through how a model is actually taught to use software like Salesforce, why coding became the beachhead for every lab, and what changed for Turing the moment Scale AI was absorbed into Meta.

The conversation then turns to the question every founder is now asked in the room. What is your moat against Claude? Vijay's answer is to go deeper than the frontier labs can reach, into the outcome you own and the context that lives inside an enterprise.

If you are excited about how AI actually gets built, who really trains the models, and how to build a company that survives the labs, this episode is for you.

00:00 - Trailer
02:20 - The Indian company quietly behind OpenAI, Google, and Anthropic
04:50 - How Turing went from a talent platform to a unicorn to an AI research partner
08:20 - The bet nobody understood
11:50 - Why a model company without human data is "racing on three tyres"
15:50 - Why coding became the beachhead for every frontier lab
19:50 - What changed for Turing the day Meta bought Scale AI
23:50 - "What is your moat against Claude?"
29:50 - Will AI create more lawyers, not fewer?
34:50 - The teams where engineers haven't written code in six months
39:50 - 16% of the Philippines' GDP is under threat from AI?
43:50 - How you actually teach a model to use Salesforce
49:50 - How robots are taught real-world work
54:50 - Why million-dollar researcher packages are breaking startup hiring
59:50 - The one kind of AI company that gets stronger as the models improve
1:04:50 - Product vs. services, and the trap that quietly kills AI startups

-------------
India’s talent has built the world’s tech—now it’s time to lead it.
This mission goes beyond startups. It’s about shifting the centre of gravity in global tech to include the brilliance rising from India.

What is Neon Fund?
We invest in seed and early-stage founders from India and the diaspora building world-class enterprise AI companies. We bring capital, conviction, and a community that’s done it before.
Subscribe for real founder stories, investor perspectives, economist breakdowns, and a behind-the-scenes look at how we’re doing it all at Neon.
-------------
Check us out on:
Website: https://neon.fund/
Instagram: https://www.instagram.com/theneonshoww/
LinkedIn: https://www.linkedin.com/company/beneon/
Twitter: https://x.com/TheNeonShoww

Connect with Siddhartha on:
LinkedIn: https://www.linkedin.com/in/siddharthaahluwalia/
Twitter: https://x.com/siddharthaa7
-------------
This video is for informational purposes only. The views expressed are those of the individuals quoted and do not constitute professional advice.

Transcript

SPEAKER_00 0:00

You are as close to frontier model lab as possible. AI is going to eliminate millions of jobs.

SPEAKER_02 0:06

The human should be able to do at least one thing better than the model. At least one thing. It cannot be zero things.

SPEAKER_00 0:10

One reason I also feel slightly bearish about certain kinds of roles is that they're trying to create real-world data using factory workers in India, Bangladesh, Vietnam, and then selling this data to labs and physical intelligence.

SPEAKER_02 0:25

Physically, yeah, it's still early. We are in the early innings of this where best practices, etc., are also not known.

SPEAKER_00 0:31

Paul, we see the different stages. Don't have an answer to this question. At least, what is your mote against clouds?

SPEAKER_02 0:37

The number one uh thing before you get into defensibility is make sure that whatever it is you're building was impossible to build uh four years ago.

SPEAKER_00 0:48

Hi, this is Sadhaat Aluwaliya. Your host at Neon Show and also managing partner at Neon Fund, a fund that has invested in some of the best enterprise AI companies that started from India for the globe, including Atomic Works, Podraft, CloudSec. Today I have with me Vijay Krishna from Turing. Vijay, uh Turing is one of the most defining companies, I would say the top 20 companies that are defining how AI is moving forward today. Alright. So proud uh you know to see what Turing has accomplished. Congratulations on all the success. Thank you very much, Sidar. Right. And uh Turing is helping foundation models become better. Right? That is one part, but also you are also creating custom models for some of the largest enterprises globally across pharma, financial, and so many different domains. Right. So I believe it's uh it's uh uh in various ways you are shaping how AI is getting used uh today.

SPEAKER_02 1:46

Um yeah, uh thank you, uh Siddharth. Yes, uh that's right. Um a small um uh this um I would I would make a small edit though. Uh the I would say uh yes, uh you're correct about what we are doing on the uh foundation model side. Uh uh by putting together uh uh teams of uh human experts in various uh domains, uh we ultimately look to uh transfer uh human uh knowledge and expertise onto the models, and this uh you know varies across a range of areas. On the enterprise side, um custom models are occasionally done, but um on the enterprise side, we take much more of a sort of uh pragmatic approach toward uh solving the customer problem by whatever means necessary. And often that means uh uh building the right uh um the simplest possible uh agent tick product, um being very mindful of uh who the end customer is, that will deliver the maximum value to them. Um uh yeah. Uh and very often we do find that uh uh the right kind of uh agenc products built on top of existing models is uh m the right high ROI decision for most enterprises.

SPEAKER_00 2:53

Got it. So so you build for enterprises both uh agents built on the existing large models, as well as you help them whatever required if it's required to achieve a certain outcome, building custom models or small language models. Do you help?

SPEAKER_02 3:09

Correct. Correct. That's exactly right. Yeah, yeah. On the um correct. Uh I think uh I would say um uh when it comes to um frontier model companies, right? The um the the the goals are uh sort of within a certain uh uh range of uh one another, right? Like everybody more or less wants to build the world's best model in uh certain uh they want to achieve certain capabilities, etc. However, the enterprise side is very, very different, right? Uh um we work with enterprises in uh banking and financial services, in uh drug discovery and pharma, retail, all sorts of things. There are different companies in drastically different uh stages of uh maturity. Some of them uh have uh never built a single uh sort of uh AI application, some of them have built uh hundreds of uh uh AI applications but want to play bigger. So um yes, on the enterprise side, uh we I mean we we meet the customer wherever uh they are. Uh our goal is uh that uh we want to um we want to ultimately ensure that we first of all uh help them find problems that are at the intersection of uh large business value and you know what models can uh achieve with uh reasonable effort, and uh then um uh help them uh with uh uh you know creating the simplest possible solutions that uh create the maximum business impact.

SPEAKER_00 4:25

Got it. And and tell me about you know uh the journey of Turing and how Turing evolved with the evolution of the AI market.

SPEAKER_02 4:35

Yeah, yeah, great, great question. So uh when we founded uh when uh yeah, Jonathan and I founded uh Turing uh back in 2018, about uh you know it's about eight years ago, uh the um uh we were uh primarily founded as a um talent platform uh that uh attracts talent from uh all over the world, uh have uh has automatic ways to source um VET to identify the top uh 1% or so of the world's talent pool and also uh help um uh large uh fully distributed teams uh function uh very well uh and very effectively. Uh uh we uh so for the first uh three or four years uh that was uh the sort of entirety of our value proposition. Uh and uh we when we became uh unicorn in uh 2021, that was more or less the entirety of our value proposition. So while we were using AI to uh make many of these things uh actually uh happen, uh AI labs themselves were not our uh customers at uh that point. This whole thing uh started when uh yeah, when we started helping uh OpenAI um you know a little over uh about four years ago, uh like uh since from early 2022, long before uh uh Chat GPT uh launched. And uh the uh that is how we got started in this particular area. So I'd say uh I mean OpenAI wrote this uh instruct uh this famous instruct GPT paper in early 2022, um uh that uh showed that uh um uh relatively small uh volumes of uh high quality um human gener uh uh generated data or high quality data of some kind can um substantially move uh model performance, provided the base model is sufficiently strong. And uh uh that in many ways I think has set the stage for uh this uh human-in-the-loop uh methods um uh that have followed from the years since then. Of course, you know, OpenAI became uh wildly successful, and yeah, uh in the in the coming years, uh this uh this um uh these uh specific human-in-the-loop methods involving experts in different areas and uh uh range of ways to ultimately ensure that uh that expertise and um knowledge uh efficiently transfers onto the models, this became pretty much like a uh like a must-have. Uh it be it it sort of became like uh if you're running a uh model company without this component, it's like you know, uh trying to uh enter a racetrack with uh three tires instead of four, uh basically.

SPEAKER_00 7:05

And can you also tell us about the evolution of these model companies, right? Because uh uh transformer models were built by Google, but really popularized by OpenAI. Right? And today uh OpenAI and Claude are today the most popular uh model companies. Philanthropica and OpenAI are today the most popular model companies. And it happened very fast in the last six years already.

SPEAKER_02 7:30

I mean, look, yeah, yeah. I mean, uh so uh I mean I at the moment I would I would uh I I would say Gemini's uh pretty close too. Uh I mean each one's uh ends up having their own uh um like uh uh I I I do think uh um these uh you know models of open AI anthropic Google uh have definitely gotten to the point where uh they are within uh sort of hairs uh breadth of each other ultimately. Yes, uh there are the in domain A, one um model might be better, in domain B B, the other model may be better. If uh uh if we have the same conversation two months later, uh the ordering may again be different. So ultimately I feel it has become pretty close. But but yes, you um I I get the spirit of your question. Yes, obviously a lot of these key uh key um advances came out of Google, the Transformer came out of Google, not just Transformers, but even uh this uh BERT and uh this uh um the BERT models uh very much came out of uh Google also, uh which were yet another sort of key building block. I f yeah, I feel like uh the um so I I mean I I I suppose there are different levels of insiders here, but uh I mean I used to be an NLP researcher, right? Like and uh the um uh even uh uh if I look at uh Stanford faculty members uh um uh that I uh worked with who are NLP researchers, I don't think most people quite saw um improving uh performance of language models as a method to create uh emergent reasoning properties. Like I worked on language models many, many years ago, almost two decades ago. At that time, the thinking always was um if you're able to engage in uh good quality next token prediction, you can do certain things like uh better machine translation, you can do better speech to text, etc., by not uh putting uh clumsy word orderings uh together. Um, when it comes to say speech to text, let us say today uh you're talking to somebody over Zoom or on the phone, you don't hear some things, you'll still auto-compute that properly because uh you understand the nature of language and concepts and everything, basically. So the whole idea was uh you could help these models get better at that. I think I'm sure you know there were some slightly early adopters into this, some later adopters, but I'm but I'm willing to bet, for example, at least at the time Transformers came out, Bert came out and all that, nobody really thought this was a uh this was a path to um uh emergent reasoning capabilities, uh a path to AGI itself. In fact, um uh this uh even that uh the original Transformer paper um uh it was written with the goal of uh sort of uh improving on uh um some of the more classic problems, you know, getting some 10% accuracy gain on something. No uh people might not have fully seen uh how it would become uh uh foundational to everything. Yeah, I mean clearly, you know, OpenAI sort of uh uh this uh saw those opportunities, they leaned in on it pretty uh quickly. They yeah, they went all in on it, and yes, uh some of the other companies were slower. But again, now I do think we have reached a point where everyone understands the importance of this and the companies with the capabilities like uh like Google are very much doing uh what it takes uh to um to have one of the most competitive models on the market.

SPEAKER_00 10:42

When Frontier Labs started working with uh data providers or uh providers who could provide them AI expert, AI researchers globally, they started working with many labs like Turing, right? There was Kale AI, Turing. So how did each lab or uh differ from itself?

SPEAKER_02 11:03

So this is a um uh so I would say some uh you know some uh this is uh the um uh the the frontier AI facing uh data business is one that's been rapidly evolving. Uh it uh you know every few months it looks uh radically different uh than what it did uh the you know six months uh prior or 12 months uh prior. Um I would say there is a there's a little bit of okay. First, um maybe there's uh there's a few characteristics to this. And you know, whatever whatever I'm saying right now, um it may look very different six or twelve months from today. But yes, historically there have been a few necessary ingredients. Uh first of all, um this uh uh uh a company should have the uh I mean should have uh high quality talent, should have uh diverse talent across a range of areas, should uh uh this uh um should have very large volumes of uh talent, should be able to sort of uh um move very rapidly in terms of uh marshalling uh uh really large teams. You should uh have the right uh sort of uh uh controls in place to ensure uh that you filter and get the very best talent. You should have the right uh controls by way of product and processes to ensure that you produce good quality data because I mean all things uh machine learning or AI are like garbage in, garbage out. You produce bad data, it'll uh you know um the the models will uh just degrade. And uh so uh so one needs to do that too, and one needs to rapidly uh adapt to the uh changing needs in the market. In fact, um the more one can even engage in various kinds of uh proactive investment in new areas that you think the the models ought to get better in, and uh uh you um and if you call correctly on that, that that too can be a huge competitive advantage. Um the one thing I think that makes this business a bit of a moving target is uh that the models do learn pretty rapidly. Like if you're offering a certain type of uh this thing, uh a certain kind of a partnership, a certain kind of a collaboration to labs with a certain, let's say, type of talent and a certain type of task, that is going to get deprecated six to nine months from now because all the models would have learned that. And uh the um the real question becomes uh can you then help them in the next stage of the journey uh from there? So I would say these have probably been some of the key things. Now, in terms of the players themselves, rightly, uh uh yes, uh uh us scale and various others uh have definitely been uh competitive. Um all of us had you know different levels, uh different types of expertise at different points of time. I'm sure all of us um uh expanded our reach of offerings and our uh um our extent of expertise over time. So uh yeah, this is a dynamic thing. I it's um uh yeah. The when when we started uh some years ago, our uh um pure forte was uh coding and all things uh all things uh in and around uh helping the models get better at coding, but we have uh grown too. Some of the other players uh I'm sure started off in uh with uh with a different uh sort of core expertise, but then rapidly got better at other things also.

SPEAKER_00 14:08

And let's say to get less for Claude to get better at coding, what kind of data Claude required to get better?

SPEAKER_02 14:16

Yeah, yeah. So at different points in the journey, it's been very, very different. Like in the in the earliest days, right? Like uh I want to say four say uh four years ago, uh I don't think the uh most of the models were uh that good at even reliably solving say uh medium to hard lead code kind of problems properly or being able to explain them properly and so on. So at that time, even data along those lines um uh was uh valuable. Of course, the models got better and better, so that meant uh the the um uh either uh the uh data had to involve harder problems or they had to ultimately I I guess close the gap. Um I mean at the end of the day, maybe let's um um you know understand what we are working toward, right? And then uh talk about what are intermediate steps here. Well uh obviously what we are working toward uh here is uh that uh to to get to you know artificial general intelligence, get to super intelligence, that means ideally uh you want this uh you want the models to be on par with the very best uh uh knowledge workers, maybe even uh surpass them in uh in that capability. Now, um in the very early stages, there may be a lot of very basic uh things uh that have to be solved. Just like for example, a human software engineer. Yes, if they can't solve a medium uh uh level lead code problem, well, at least historically uh they it might have been difficult for them to be valuable. So, in many ways, that was like a step one there. Then, of course, you uh uh the um uh we um uh uh this uh we worked on helping evaluate and um improve uh model performance increasingly in a lot of uh real life settings. Um got uh uh particularly instances where you had to get the model to um you know understand some really large code base that could be in uh you know uh that could have hundreds of millions of lines of code across different repos, different uh files, all sorts of uh things, and uh be able to plan, uh reason, uh debug, do all those uh things. So, yes, one uh one one had to get better at uh putting together the right kind of harnesses, the right kind of uh problems, um, the right kind of environments that actually test and uh um help improve model performance in these uh uh in these different uh scenarios. So um yeah, in the in the early days, obviously it's much simpler to uh you know generate those lead code kind of uh problems because uh and uh so that's why people start with that, because uh under the uh with the thinking that if uh you know if the model doesn't even do this, why even waste time on harder things? But increasingly, as uh the model performance gets better, the um the complexity in even designing the task properly, doing the right measurements, etc., goes up, but that is ultimately necessary.

SPEAKER_00 16:56

And was it a big moment for Turing when this is scale got acquired by Meta? Uh right, then all the model providers would have wanted to work with a neutral pair. I think at that point of time I remember scale was like just at uh uh number two position.

SPEAKER_02 17:14

Yeah, yeah, yeah. I think uh yeah, I I think that is um uh yes, that that is exactly right. Um uh those um uh you know that sort of thing uh did happen in the uh in the aftermath of that uh of the deal with uh scale. But I mean, yeah, yes, that was that that was a temporary moment. But honestly, in the in the you know, medium to long run, um uh the labs do want uh um I mean it's a um on one hand, maybe they don't want to be working with uh hundreds or thousands of uh partners, but on the other hand, uh probably they don't want over dependence on one or two. They do want to uh uh have a uh sufficiently wide uh latitude in terms of uh um the the options at their disposal uh on the data side, since that is a powerful lever and ultimately um it is good for them as well if they actively engage um uh a number of uh uh capable uh players since that'll just mean that the range of capabilities available to them is also more.

SPEAKER_00 18:14

Today uh almost everyone is afraid that you know uh what as the entrepreneur, what they are building, Claude might come up next in their domain. And also all VCs at different stages don't have an answer to this question while funding the portfolio or asking the next set of companies what is your moat against Cloud? So, what what do you think?

SPEAKER_02 18:35

Uh no, no, so this is a very good question, right? Like I have I've seen people uh um this uh fixate on this a little bit in a vacuum. Uh um like um I'll first tell you what I do do not think is the solution. I've seen some uh VCs even uh literally almost talk like Luddites that uh you know I don't want to touch uh anything that uh uh that can be impacted at all. But then uh in many of the cases, the sorts of things they are talking about uh involve uh throwing out the baby with the bathwater because on one hand it is true, you don't want to be rendered uh moot by the next version of uh Claude Code, but you also want to take advantage of the fact that we have some uh totally unbelievable capabilities at our disposal that allows us to create certain uh kind of uh uh companies and products which would have been impossible four years ago. You want to take advantage of the latter, like when people say that you know AI companies are growing at unprecedented rates, it's only because they are they have suddenly been able to do certain things that are of such a disproportionate value uh that uh you know customers are um uh willing to uh sign on uh at such rapid rates. So I think one should one should keep that in mind as well, uh that while this is a consideration, so is the other one. I think on the um I I uh I think the more so uh um you know, just like everyone else, I I don't think even the frontier uh VCs know some sort of uh airtight answer to this uh question. But there are definitely a bunch of uh um you know, let's call it heuristics that are helpful, uh basically. So um I think one of the things that uh that is definitely helpful is the more sort of uh deeply you verticalize in a particular uh domain, uh the better uh uh it is. Um uh that is that is number one. Um like um coding in general tends to be a somewhat horizontal problem and far uh far too much in the uh you know in the uh immediate uh roadmap of all the model companies, just because not just because of a bunch of reasons. Um like one is of course, well, one is of course it is a valuable industry. Second is that because code is inherently auto-verifiable, it is easier to make more rapid improvements there. Um number three is that uh I would say coding in general uh has nice spillover effects in terms of improving uh reasoning capabilities of models as a whole. Whereas, you know, let us say uh the model companies prioritized uh something uh like uh the models getting better at medicine. I'm not too sure if that would uh help improve reasoning as a whole, basically. But yeah, coding and mathematics have been known to have that spillover effect. So yeah, so the areas like coding have generally been, at least general coding, um, have uh are definitely more susceptible to this. Um something that I think uh has a um you know will need some significant FD kind of uh uh angle again could be uh you know um could definitely be something uh uh that creates some nice defensibility. Again, um you uh you probably want to be uh um careful to ensure uh that uh there is actually a product at hand uh there uh as opposed to you know uh maybe uh going 100% uh FD here and having some thesis around what you want to uh be solving for your clients. I think there could also uh the I think form factor also probably matters. Like for um, yeah, form factor slash uh You know, uh to what extent you're taking responsibility for the outcome, that also matters. So, for example, in a world where you're going and uh automating a whole bunch of uh call center operations for some particular uh thing or uh um or doing something else, like um um yeah, when it comes to some product or service, if you're able to go and uh directly tell uh clients that you can do certain things a lot uh cheaper and faster, uh that sort of thing could be valuable. Like uh um Foundation Capital um uh recently um you know invested in Tessera, and I know Tessera uh raised a round from Andreessen there about making some uh SAP migrations a lot cheaper and faster. So here they are actually in a sense putting their money where their mouth is. They are saying, you know, um I'll uh uh I'll do these things uh some uh 5x cheaper and faster uh compared to all the traditional methods due to uh uh the specialized things we are doing around AI. Now, when you're doing this kind of end-to-end thing, you are less likely to get disrupted. Interrupted if uh, you know, if anything, if the um if the cloud models or codecs or even Gemini models or any of these other models uh get a lot better at uh coding, uh these people only benefit. They don't uh because because uh there is this end-to-end uh ownership. So I think all these are valuable. I guess some other things uh again, the more maybe you're taking some kind of ownership of uh, you know, uh cases where people are paranoid about this last mile and all that, you're taking responsibility and ownership again. Uh that can give you some uh nice uh defensibility. Uh all things equal, I suspect um that uh the uh the more you're sort of selling to non-technical people or less technical people, again you may have a better shot. Like, for example, the um the thing with uh with with uh software development team is that they are perfectly you know capable of uh playing with a lot of things that are largely self-serve in a manner which a lot of others may not be. So I think there's a number of these heuristics. There's of course others, right? Like uh uh Foundation Capital has been famously popularizing uh uh the idea of a context graph, an idea that I've seen getting a lot of uh traction. Uh this uh products that sort of uh compound in the extent to which they uh collect a lot of context about the enterprise, not just enterprise data, but also um as a result of uh the product use, uh uh collect more and more data about why various decisions were made, all those exceptions, all those things that were historically sitting in people's heads, all these also can, I think, serve as a nice uh mode of defensibility. Yeah, I uh I I suspect that there is no you know clear uh uh clear science to this, but generally speaking, the more of uh these various things you have, the better. But but again, uh I think uh I would I would really sort of emphasize you you don't want to throw the sort of baby out with the bathwater. The number one uh thing before you you know you get into defensibility is make sure um that whatever it is you're building is uh uh is uh was impossible to build uh four years ago. Make sure you know it is 10x or 100x better than the status quo, make sure that uh it has a capability to uh you know uh grow rapidly. Like nobody cares about uh company uh uh that is a slow growing and defensible God. Yeah.

SPEAKER_00 25:10

And uh also in the last four years, uh the infra opportunity that got created, let's say the picks and showels that we say, uh, have been amazing because Turing is an example of that. Superbase, click house, all are examples of that. What are the things that you see will now be picks and showels for tomorrow? What are the other opportunities that are now uh not being captured, but can be the picks and showels of tomorrow?

SPEAKER_02 25:37

It's a it's an in yeah, it's a it's a very interesting question. So there are I'm I'm I I guarantee there are all sorts of things here, you know, uh around uh um uh around this uh data centers, data center infra uh power and uh this uh chips and um this one, um these uh GPU marketplaces and all sorts of things. A number of them have been doing quite uh well as well. Um I'm willing to bet there's a lot of uh open problems there. I I keep hearing about a lot of uh um exciting opportunities, even there. Things like, for example, um uh various ways to um this uh um make things uh more efficient by um uh like today. Uh I'm uh I'm sure in a lot of uh training runs, inference, etc., uh your resource allocation may not be the best because of uh bottlenecks on memory uh versus GPU and things like this. And various uh I'm willing to bet various companies are solving interesting things there. On the software layer, yeah, I think the more um you have some of these uh carefully thought out uh uh vertical solutions, right? Uh the uh I think the more uh the more effective uh this would be. Like uh the um like uh clearly cursor wrote a very nice uh nice way. Uh now uh now of course uh the cloud code methods are uh proving quite effective. But there's probably a lot of areas where uh um where the um where the simple chat interface may or may not be the right uh this thing, uh may not be the right way to elicit inputs and uh uh and solve a problem. I can totally imagine one um uh one could build various uh sort of uh uh picks and shovels version if uh one uh focused on some uh sub-industries. Like I'm willing to bet um if one were to build something uh that uh some uh some sort some sort of an appropriate platform to help uh this drug discovery better or something, it may look very, very different. Um I think uh yeah, I think it's an evolving field. I think when it comes to this uh data angle itself, it is it is a very, very rapidly evolving field, right? Like in in in some very important ways, Turing and many other companies. Yes, we were in in a broad sense, we have been in the data space for four years, but in a very important sense, we are almost radically different companies every every two years because uh of uh what this thing uh demands. Um right, like for example, uh this um moving from uh this SFT or RLHF data to these RL environments with auto-verifiers is a is a somewhat almost markedly different game. This the game itself could look very, very uh different in the future. Um yeah, it's a it's a good question. These are the only things that come to mind at the moment.

SPEAKER_00 28:33

And what are the latest uh things that you can share your working at during at uh which have only been possible this year?

SPEAKER_02 28:46

So um the sorry uh on the lab side or the enterprise side or both? Both both. Very interesting. So well, obviously as the uh you know as the model capabilities have got uh better and better, uh the uh and particular particularly when it comes to um their agency capabilities for solving uh um not just these one-shot problems, but even these uh medium horizon problems and hopefully soon uh long horizon problems. I think uh the it's become a lot more practical uh to uh put it to use for many valuable um applications in the enterprise. Like we have built auditor copilots, we have built insurance underwriting co-pilots. Um like in the past, uh like uh maybe two, three years ago, yes, um uh you know, clients that were early adopters wanted to tinker with this and uh wanted to get a better feel, could have been uh using it. But it it is well, it is quite likely that many of those things uh would have been some distance from going to production and adding uh tremendous value. Now I think it is the uh uh this is very much the case in a number of uh different uh domains. Um the yeah, obviously there's a lot of uh surrounding infrastructure that is constantly being built around this uh when it comes to uh agent orchestration and uh uh RAG and various other sorts of things as well. We yeah, we are dealing with a much, much more uh mature ecosystem here. Um yeah, uh when I mean when it comes to uh look when it comes to even uh how software development uh teams are run, right? Uh this is uh this is quite an unprecedented thing, right? There are uh so many companies now where um uh software engineers have not had to write a single line of code in uh say more than six months, which is uh I mean quite unprecedented, right? Like it almost uh reminds me um like um uh yes, during my uh I mean during our undergraduate days, uh I'm sure we heard uh that in the 60s or 70s or whenever it is, people probably used to write machine code or used to write assembly code uh directly. Um yeah, now it now it almost feels like as big a paradigm uh jump, if not uh bigger. So uh yeah, this is this is definitely quite uh uh uh novel. I I I I'm willing to bet that in a lot of things, uh even in you know, use in the private equity, use in uh investment banking and various other things. I think the reasoning capabilities of models have just uh uh gone up so much that uh just the quality of the decision making, quality of the analysis, etc., is probably a lot stronger to the point. I mean, ultimately, right, like there is a certain tipping point that is uh um when it comes to using these models. There's a certain sort of level of accuracy where you know you have to second guess almost everything. There is a there is a certain thing where okay, it is like 95% right. And then there is this uh point where you're almost willing to trust it completely. I think in many areas it is coming, it's coming awfully uh close there, which uh which is I think uh quite uh novel. And um even now I think there's a lot of uh domains where uh uh which are still far from where we are with uh coding. Um like with coding, you know, literally, I I I know so many folks now at uh big tech uh producing thousands or tens of thousands of lines of code every week. They don't have time to review it, not as a person who's supposedly doing the code review and so on. Definitely, this is not the case with uh law and uh uh and uh number of other law or taxes and a number of other things, but it's only a matter of time. I think the models will get there in the next uh probably year or two.

SPEAKER_00 32:16

So there are two schools of thought, uh right. One is uh that AI is going to eliminate millions of jobs, which is happening at PC, right? At the tipping point, like Amazon, big tech cutting thousands of jobs. The other way is people are saying is it can lead to Gemon's paradox where if you make something cheaper, more of it gets produced. Thereby we might have more number of lawyers rather than in in the future rather than less numbers. It was just that law was so complicated uh that it took somebody five to seven years to to do it. If if AI can help make a person lawyer in two, three years, then we'll see more number of lawyers and more like legal litigations getting faster. Let's say litigation taking years, for example, in India or months to years in the US, it might compress to a few days. What's your view on it?

SPEAKER_02 33:08

I generally tend to throw in with the Jevens paradox side uh more um for a few for a few reasons. Um is yeah, one is yes, I think I do think the reasoning is sound. I do think uh, for example, uh access to legal services is incredibly expensive in the in the US, and I'm I'm sure uh this thing will play out uh um in terms of increased uh demand, etc. Additionally, I'll also just point out that you know people have been um uh people have been uh talking about this uh idea for centuries uh and uh it's all it's always been wrong. It's uh uh in uh um in uh in the past. But that said, I gu I guess the uh I guess there is there is one you know obvious um sort of you can say key key ingredient that has to play out here, right? Which is that uh the human that is engaged when it comes to the specific profession at hand should be able to do at least one thing better than the model. At least one thing, right? Uh it cannot be zero things, it it does not need to be a lot of things. But uh but uh uh if the human can do at least one thing uh better than the model, yes, uh they are going to add value. Uh that is where the cost will go, all the other costs will go to zero, this will become, you know, uh at least as far as uh um microeconomics goes, the uh this is where the value where the cost center goes and such. But um now if I think of something like coding, right? One reason I also feel slightly slightly bearish about certain kinds of roles is uh that the uh I mean I I uh yeah, uh for example, when it comes to a lot of uh the junior roles, a lot of the roles that uh simply sort of um uh involve all involve oper operating at an extremely narrow task uh scope, it is difficult to see how those could stand, basically, as opposed to you know uh roles like let's say in the software engineering uh world, it is uh much easier for me to understand how uh uh how roles that involve uh say being uh both uh being uh quite customer-centric and understanding business and understanding product and being responsible for you know decisions. I can I can see how those may stand for a long, long time. But yeah, today uh teams are very much organized in this entire pyramid uh kind of uh structure. There's a lot uh I'm I'm afraid to say that uh uh a big chunk, maybe over half the uh you know uh software uh uh engineers in the world operate at a relatively narrow task scope. And maybe when it comes to a lot of junior engineers, uh that may be the only place you only way you can start, also. So uh I mean there there are very much uh I would say real uh short-term uh problems here in terms of uh who is going to sort of um invest in uh um uh in those individuals going up the uh career ladder, etc. How is the next generation of uh folks uh even to get produced? And uh I think um yeah, there's a there's a lot of fundamental questions that I think uh are still uh and this I think very closely even ties into the question of uh um how education itself should evolve and all that uh stuff, uh which I don't think anyone uh quite has uh um answer to.

SPEAKER_00 36:28

Yeah, like like we discussed the some degrees which required five to seven years or five to eight years, uh those those setups have been 40-50 years old. So now when everything is questioned, then the education, those those compression of those degrees can also be questioned, right? Why can't you get the same?

SPEAKER_02 36:48

Right, right, right. Yeah, a couple of other things, right? In uh you can say a little bit toward this bare cases that yeah, um I'm I'm sure you know uh um various uh types of uh work that a software engineer is doing today uh is not going to exist in the future, just like uh you know what software engineers did 50 years back is not what they do today. Um there can absolutely be various sub-industries where um you just have a reallocation of uh talent. Like, for example, in the US, uh, from I'm I'm not a history expert, but to the best of my knowledge, some when the US was founded, 70% of the country was into farming and agriculture. Today it's like 2%. But okay, the uh uh it's not like uh everyone was uh some farming adjacent, etc. So uh things like that could uh happen too, and uh talent gets reallocated uh elsewhere. But but I suspect there is enough meat uh uh broadly within uh this one, uh within the uh umbrella of software where the role will drastically evolve.

SPEAKER_00 37:44

Yeah, even uh if we take only history of California, the first largest wealth was created in oil, then railroads, yeah, yeah, yeah, yeah. Then tech came much, silicon and tech came much later.

SPEAKER_02 37:55

Right, right, right, right. That's right. That's right.

SPEAKER_01 37:57

Yeah, yeah.

SPEAKER_00 37:58

The history and again silicon led to tech, and again, tech is leading to wealth getting created through silicon.

SPEAKER_01 38:04

Right, right.

SPEAKER_00 38:07

But uh interesting thing is that uh discussion of definitely the role uh and the number of software developers required uh the role of a single software developer is increasing, whereas the number of software developers required in a single company is shrinking by magnitudes. So that is the first uh area where uh AI is you know decreasing the number of headcounts in a company.

SPEAKER_02 38:33

Yeah, I I found one thing kind of remarkable, right? Um one prediction on we uh uh on which I've I'm just I I turned out to be wrong was I would have expected uh this kind of productivity gains to come uh all the way on the stack, uh along the stack. Like I would have expected that, for example, the number of uh uh let's say good venture fundable companies created in uh 2025 would have looked like 10x or 100x of 2023. I would have I genuinely thought that, but it's it's not grown all that much. And at least at the moment, it it almost seems to be like if I split uh this into, you know, um if I split uh uh creating value or creating technology into just for um crudely speaking, let's call it invention versus execution, something like this, uh in invention slash uh uh innovation versus uh execution. Yes, while this has uh done a lot uh to the execution part, it does not really seem to have speeded up the uh innovation uh and the invention part. Because theoretically, imagine you speed up everything 10x, then who cares? Uh, there are 10 times the number of venture funding startups. Uh yes, each startup can be smaller. Uh, there is still you know valuable things for people to do. Arguably the value they're creating is even more, etc. So it would be interesting to see what happens uh um over there. Like uh right now, for example, uh obviously there are various uh, I mean, the frontier AI companies itself and including some specialized companies, are doing narrow versions of this, things like uh you know, AI for science, AI for materials discovery, but that that is still narrow in the in the space of scientific research. Like a lot of the um real-world innovation is not uh that fundamental. It's a business model innovation, it is uh some uh it is various something else that is a little bit more uh grounded in more uh uh mundane knowledge about the world, like what is the demand in a particular place, uh those sorts of uh things. Um so yeah, it it remains to be seen what if anything uh um is the speed up uh that can be obtained there.

SPEAKER_00 40:31

Got and uh there are other fundamental risks which are economic risks globally. I said today 16% of the Philippines GDP come through call centers. And today companies like Decagon, Sierra, Retail, there are so many other companies in CX space which are automating the entire work of a call center. Then 16% of a GDP of a company is under country is in that threat. True.

SPEAKER_02 40:57

No, no, true, true. I mean it's uh yeah, it's a dynamic world we uh you know live in. Uh um obviously, while uh you know um uh while uh economic uh uh progress and new technological advances uh increase uh GDP as a whole, that doesn't mean each and every individual's uh income is uh increasing, uh right? Um yeah, yeah, absolutely, absolutely. I mean individuals, uh I'm sure uh um the um like uh creative uh uh the the process of uh creative destruction of uh cap uh capitalism is always uh kind of uh uh painful uh for a lot of the uh individuals that are uh impacted. But yeah, without that, uh you know you get into this zero growth uh regime and uh uh which is uh which is sort of even uh worse.

SPEAKER_00 41:42

So you cannot be socialist.

SPEAKER_02 41:44

Yeah, yeah, yeah.

SPEAKER_00 41:45

And why was Turing such well placed when you know at the advent of these frontier models? Like what had you done that you could capture the moment?

SPEAKER_02 41:57

Yeah, yeah. So in our case, uh it was helpful that we were um um uh we were probably one of the um uh largest high-skill talent platforms at the time, if not uh the largest at the uh at the time we had the ability to um, you know, while uh at that time at least, I mean now it's a little different. Obviously, everyone has adapted, but like let's say four years ago, right? Like while they were definitely data annotation companies that had uh relatively more skill uh low-skill labelers, which is what a lot of the traditional machine learning pipelines uh needed. Um the um what uh these uh um uh LLMs and the modern foundation models needed was just something very different because they had the capability to engage in um uh let's just say a more cognitively complex reasoning. So the kinds of experts they needed was uh uh very different. We we had the ability to put together um uh uh large volumes of uh fully wetted uh this uh high skill knowledge workers and you know get them to row in the same direction and everything. This was not uh yeah, the we I mean we were fortunate that uh we had this capability already as a result of our uh this thing, uh of our talent business. But uh yeah, that that is exactly what I think uh put us in the right position to seize the moment.

SPEAKER_00 43:15

Right place, right time.

SPEAKER_02 43:16

Of course, of course. That's all that is life.

SPEAKER_00 43:19

So rise of like synthetic data, like what's your view on synthetic data? Unless if at your ring, how much the data that you are producing is is from uh real world versus synthetic.

SPEAKER_02 43:29

Yeah, yeah. So synthetic data definitely has a place, right? Uh let's uh split uh uh synth uh synthetic data into two parts, really. There is one type of synthetic data where uh there is no human feedback needed at all anywhere, right? Uh generally speaking, we don't end up engaging in this uh ton simply because we feel like whatever we can do, so can uh uh at least the frontier labs. We may do synthetic data for uh for a classic enterprise when it comes to you know building uh Custom model for retail or something like this. But uh but uh yeah with the Frontier Lab, um we we really don't think there is uh uh uh too much uh uh differentiation uh brought to the table by uh that means. And second, there is also I think limits to how much uh value can be added there because ultimately there is just not too much new information. If you're telling the model itself to kind of generate, yeah, there is uh I mean there is all sorts of things about uh uh how uh um beyond the point it is uh really not going to help. The the kind of um synthetic data we are engaged in, which uh I think we are quite excited by, and we think there is a tremendous uh um you know um potential in it uh as well in terms of uh the um the ways and means it can help improve the models is uh some of these human-in-the-loop methods. Yeah. The most uh uh popular one uh of all, I think, would be these uh the creation of these uh uh RL environments with auto verifiers and uh uh this one. Um yeah, and uh realistic scenarios. Now uh um would would you maybe like me to explain as an example?

SPEAKER_00 45:04

Yeah.

SPEAKER_02 45:05

Yeah. So let us maybe take an example of uh um imagine you want to train a model uh to use uh Salesforce. I'll give a very narrow example, but you know, it'll from my example it'll be clearer uh what is the broader uh picture here.

SPEAKER_00 45:19

Autonomously or with human in the autonomously.

SPEAKER_02 45:22

You want to train it to uh like uh I you want to train it so that uh you know I say that uh uh I got this PDF from a client, go and update Salesforce if you do the right thing. So let's say let's say you want it you want to train it to do such uh activities. So the um one way to produce trained data for this is you actually produce a lot of uh different trajectory data. You record your screen, you record the mouse movements, you you know, you do all these kind of things. The difficulty, of course, with this is this is not a very high uh ROI of human feedback because again, uh human um uh collecting human feedback uh data uh takes a bunch of human effort, all sorts of things are needed for uh quality control, and the uh the ROI of something like this is not the best, basically. Uh but there's an alternate method. The alternate method is uh you uh create a uh some kind of a simulated version of uh Salesforce with uh the key uh let's say UI ingredients or uh uh whatever else, you put together some kind of realistic company databases, maybe create 10 different uh uh company sample company databases, uh, you know, one of uh um uh some uh SMB, one mid-market, one large, uh do uh then uh create a bunch of uh different um scenarios and auto verifiers. Like one uh one sample prompt and uh um verifier uh tuple might look like the following. I may say, yes, here is a PDF of a purchase order I got from my client, uh please update Salesforce. Now, what I'll directly do is uh uh instead of giving a sample example, I can um uh define various events. I can say uh okay, I'll let the model do its thing. After it does its thing, I'll check whether an account got created by some name uh that matches uh something in the PDF. I'll check whether an opportunity uh got created within uh the uh uh account that matches something else in the PDF. I'll check whether that opportunity had a budget value that matches something else in the PDF, etc. Now I can give all these uh uh so uh I can give all these uh various uh things uh different kinds of credit. I can even write some rubric-based verifier. I can say, you know, um the uh check what it wrote under description and notes, see if that makes sense and that is an accurate summary or not. I can fit this into an LLM intern for scoring that uh uh point number four. And so now the thing is uh now if I go and create uh a number of different scenarios. Uh from each one of these scenarios, uh the one is you can run reinforcement learning directly. If you don't want to, you want to keep it simple, you can generate synthetic data too. You can uh very much uh uh run the model multiple times on these. There may be uh the um uh there may be multiple ways to create this. It is uh it need not be that everyone has to go uh uh navigate exactly to some place and click. There may be some other way to click in order to create an opportunity to do something else uh to uh you know uh fill this uh data and everything. So you can you you you could basically from this uh one single action, you could get thousands of pieces of synthetic data. So the uh this method is uh one I think which has a lot of legs in uh the um particularly the more you can uh um operate in this particular paradigm of uh prompt and uh then uh these rule or rubric-based verifier uh tuples. They uh that is a very nice way for uh sort of uh um uh creating synthetic uh data, but uh where the human feedback was a crucial element. Yeah, in a lot of these areas we find the um uh the human feedback ends up being crucial in terms of creating the right uh judgment around uh um what the different rules or the rubrics uh should be, what uh um um ensuring that you have a sufficient uh diversity of coverage also across all the possible company scenarios and everything. Uh if you don't have enough of that, again the model will fail will struggle in a lot of real-world use.

SPEAKER_00 49:10

And today, let's say the amount, what's the percentage uh roughly the the model companies consume in real versus synthetic data to train the models?

SPEAKER_02 49:23

See, it is very difficult to say. I'm um I'm I'm willing. If you if you look in terms of volume, obviously I I'll be very surprised if the if the latter is not uh bigger, right? Because it's it is ultimately cheaper. If I if I get something that is uh cheaper and lower quality, I will take I will obviously do more of that. It's almost a little bit like asking uh what is bigger pre-training data or post-training data. Pre-training data is obviously bigger because that has trillions of tokens. I can get it for free, I will do what I want with it. But if it turns uh but uh yes, this real world data, uh, you know, uh human data, post-training data is also of disproportionate value. Like def uh definitely the uh what shall I say, the um uh the um value the models get out of uh human data is much, much more, except that this costs uh more also, it is slower to produce, etc. But in practice, it's a balance. I think uh yes, if a cheap commodity can be bought at large uh volume or produced at large volume and it'll it'll help up to a point, you'll do that also. If if something else is a scarce resource, that is also helpful, you'll you'll get that also.

SPEAKER_00 50:23

So so today uh uh for these model companies, uh right, uh a new era has become physical intelligence. So are you also into physical intelligence? Yes, yes, yes. And what what is physical intelligence for you?

SPEAKER_02 50:38

So this is all things around uh um it involves uh multimodality, involves uh sort of uh navigating the real world, building uh world models, building uh models for uh robotics, uh perception, uh the the whole thing.

SPEAKER_00 50:54

So so there are several set of new companies that have emerged and growing in physical intelligence. So they're saying the original model companies and their data providers won't be able to match them because let's say they're trying to create real-world data using factory workers in India, Bangladesh, Vietnam, where they're all wearing cameras, they're wearing gloves, uh right. And then they are selling this data to to labs and physical intelligence. What's your view around that?

SPEAKER_02 51:25

No, no, it is uh I mean, yes, it is it is an exciting area. We have done a lot of work in that area uh as well. Uh uh yes, it uh it involves a different uh kind of thing. Uh um it involves a different kind of uh talent, it's uh it involves a different kind of uh uh task and uh everything else. But uh yes, we have done a lot of uh work around this kind of egocentric uh uh data capture that involves uh uh individuals wearing uh wearing the right devices like a GoPro or some other uh uh thing, um uh some other piece of equipment and uh performing various tasks. There are uh yeah, the uh there are uh different variations. Look, the there is a lot of diversity of tasks. Uh, I don't think there's any um I mean when when some novel tasks come, it's perfectly possible for some player to really double down in that, specialize in that, etc. That is also uh what uh I think keeps uh space somewhat uh dynamic. Uh I don't think any company that's uh doing well in this uh space can like rest on their laurels uh either because the space is rapidly changing. Uh yeah, but that also keeps on creating openings for uh others.

SPEAKER_00 52:34

And uh data labeling as a form factor has been evolving very fast, right? What do you think is future in uh the the like the the new companies that will evolve in in data learning? What would they look like?

SPEAKER_02 52:47

Sometimes they could be these uh you know entirely different, like emerging uh uh, you know, uh let's just say emerging areas, like uh physical AI. I think physical uh in physical AI, it is still early. We are enough in the early innings of this uh game where uh a lot of best practices, etc., are also not known. Uh um, us included in uh in in various uh um areas uh are uh yeah, um I mean we are helping our uh sort of uh clients experiment with various things, not like some mature areas where uh the market has uh cleanly converged on the best way to do uh different things. So obviously, new area could be one. Uh when it comes to the models themselves, the uh uh um uh generally speaking, uh there's less of a room for simpler tasks. The more uh um you're doing expert-level tasks, uh the more uh room uh might be there. The the more you're into some uh uh you get closer and closer to these long horizon tasks, the kind of thing that would take a human uh one day or one week to complete, the better uh again the more headroom there is because those are sorts of the uh the kinds of uh areas where the models tend to fail the most. They generally end up doing better at the uh short-term ones. There's a lot of gaps, like even in computer use uh agents, etc., there are definitely uh gaps. There is uh definitely gaps when it comes to uh real-world domain expertise uh and so on, and and within coding itself, where the models have made rapid advancements, also there are uh gaps.

SPEAKER_00 54:14

And among the Fortune 500 enterprises that you work with, uh what is the most common uh or I would say uh the sector in which there is the highest AI adoption right now?

SPEAKER_02 54:28

I mean, I don't have a you know proper random sample, right? Like our uh um uh I I can tell you of our experience, but uh the that would only tell you um uh some interesting sectors where we see a lot of AI adoption. Uh that is not to say that some uh some other company may not be able to uh pull this off. So uh so first I would say there is there is definitely a uh very interesting adoption across uh all across the board. There is a lot of uh interesting white-collar work, judgments, audit, compliance, uh uh um yeah, various uh kinds of uh problems where this exists. I think in the adoption tends to be a bit more in uh uh industry, uh even more in industries where which are almost entirely white collar, like for example, banking, financial services, etc., which all pretty much everything is virtual. Whereas um certainly in uh things that involve manufacturing uh and other things, construction, oil and gas. Yeah, construction, oil and gas. There may be certain things, but there's a but but there are big parts of the job that are not uh that are very much kind of on the ground uh today. I think given that um uh the model's physical AI capabilities are still meaningfully uh behind, of course, that uh narrows the set of opportunities. The lowest hanging, I mean, even if uh uh you know when we work with a company in manufacturing, we would try to restrict ourselves to those kind of uh problems that uh involve more of the white-collar work, some sort of judgment, some sort of uh uh helping them uh create uh new mechanical drawings, all sorts of uh things, but something that uh is uh virtual.

SPEAKER_00 55:56

Understood. And today we have been discussing, you know, uh like the market in general. Risks from open claw. What do you see are risks and opportunities like invention like open claw have bought?

SPEAKER_02 56:16

Well, obviously the risks with a lot of the uh you know this level of autonomous thing is that it is you know it is tempting, uh right, like and uh people may sort of uh seed uh uh people may cede more and more control. I'm sure uh these things uh uh create uh potentially create a lot of security holes. Uh the um uh this uh there are yeah, I'm I'm sure you know various kind of novel cyber attacks are possible, um cyber attacks of different kind, like attacking uh somebody else's uh uh this uh uh um AI, or uh you know, um uh trying trying a lot of different uh variations uh when uh experimenting with a lot of different variations to sort of uh um this uh launch various kind of security attacks on uh humans itself. I think uh yes, obviously, uh you know, given the rate of uh all of this, I mean not just uh even the likes of open claw, right? Like even if I look at uh, for example, uh the kind of uh capabilities that even uh Claude code is uh unleashing, right? Where uh people are suddenly able to produce many tens of thousands of lines of uh code, uh nobody has time to review, nobody uh uh has this one. At best, people can do a sanity check and move on. I I guarantee it is going to create a lot of uh uh security holes and uh various other sorts of uh things uh down the line. But yes, people's you know, counter-argument is uh that uh human developers write a lot of bad code also, so this may not really be worse, which is an interesting uh um counter.

SPEAKER_00 57:51

And what do you think is the future of IT services company? Because like you know, the traditional IT services companies which made big from India are today under a lot of pressure.

SPEAKER_02 58:01

Yeah, yeah. No, um, so they will uh they will definitely have to have to get substantially reinvented, right? Like today, very much uh uh the models in these companies have been uh entirely about uh this uh uh entirely uh headcount related. Uh uh that is uh yeah, that model is definitely going to get a lot uh uh come under uh severe. I mean, it's already coming under pressure. Uh yeah, the the more uh some of them are able to uh adapt to some palantier kind of model and do do do that successfully. I mean, everybody talks about it, don't you? Everybody is doing MD. Yeah, everybody talks about it. But yes, uh the the individuals who are able to do it successfully, I think will will of course come out uh you know uh come come out uh really really strong. Um because arguably some uh some of the very significant advantages uh these people have is uh their uh deep understanding of a lot of uh different uh um industry problems, uh having uh um uh teams of uh individuals that are uh this thing uh that are both technical and understand those problems, etc. So theoretically, uh you know uh if they are able to embrace the opportunity properly, they could they could move into that model. But uh yeah, it it is very much clear you you know they will uh um they will have to uh reinvent uh themselves to uh go more toward this uh uh taking ownership of outcomes, the more uh they are able to be this um uh software plus uh FDE rather than uh you know pure uh consulting, um the better it would be for them.

SPEAKER_00 59:33

So when you are today approaching to enterprises, are you competing with the likes of Essenture? Because they are also promising agents, transformation.

SPEAKER_02 59:44

I mean, theoretically, look, uh theoretically, I'm sure a lot of people are are uh um them and and various others, I'm I'm sure, are promising uh promising very uh similar things on the uh on on this side. The reason uh I think uh enterprises still like working with us is that we are probably the only uh this one uh only company that can help them with AI transformation that has also had a significant role to play in all the model advances, all the way starting from GPT-3 until present. Um we probably have a better uh perspective than most in terms of what the unreleased models are doing because we are actively helping improve them uh um and uh uh you know um what capabilities are likely to come uh uh say uh later this year. And if you're building something, what sort of thing is maybe not worth building today because it's guaranteed to get deprecated six months or 12 months from today, uh and and so on. Uh and some some of our clients um and some of our clients who have uh bigger ambitions also beyond building simple simple use cases. Um Ideally like the fact that uh uh there is uh there is more depth to us. And uh in fact, with some of our partners who have built many uh hundreds of uh agent products in the organization already and who are trying to play bigger, they find that we are maybe the only uh uh partner who's uh actually able to uh help them scale to their next level of uh uh ambitions. Things like you know building custom RL environments for uh um this to improve agents in their industry and all that. A lot of which uh starts looking awfully close to the kind of uh frontier uh research work done at the labs.

SPEAKER_00 1:01:17

But this is close to custom services for them, right? Uh for these enterprises, no?

SPEAKER_02 1:01:23

Yes, yes, uh, yes, yes. There's a lot of, I mean, there is uh some in uh uh industries in which we see we have started to see some very interesting reputability, like for example, in the financial services, we have uh we have built custom products and we have a lot of uh uh reputability in in terms of certain kinds of uh uh thing, uh certain kinds of products and solutions that benefit uh um the uh you know uh the research analysts at these uh uh investment firms and various others. But yes, there is there there are absolutely uh things that are uh custom. Yeah.

SPEAKER_00 1:01:55

And like a lot of times now today founders are getting questions in AI specifically. How much from next round we see how much of this is product versus services? So what's your opinion? What's what should a balanced approach look like? And or is it a yeah, yeah, sorry? Is it a SaaS era question? Product versus Yeah, yeah, yeah.

SPEAKER_02 1:02:16

So I think uh look, I think the answer is uh somewhat different for uh for a company that is starting today from scratch versus uh for us, right? Like in our in our case, obviously one uh one big big variable also is uh how can I um uh better uh leverage the strengths I already have. Whereas if you're a company who's starting from zero, you have no strengths. You have to you have to look at uh how to um you know what is uh um uh optimal line of attack for uh you. So uh the um uh yeah, so for a well, for a new company. I mean it is obviously if one is able to engage in uh this kind of you know uh high-end AI consulting and everything, you're able to get the right experts in, etc. It can be a useful wedge. It's my I mean uh even among uh consulting or services, uh high-end uh sort of cutting-edge consulting is a lot more differentiated than uh relatively more mundane consulting. But that said, yes, you you absolutely want uh um you know uh you absolutely want uh um uh uh a rapid path to prioritization. The sooner you can do it, the better. In fact, if you're starting a new company, maybe ideally to ideally start from there instead of uh just uh wandering in the dark uh for uh you know do doing a bunch of consulting and hope something productizes. Uh yeah, if one is starting a new company, I would definitely uh advise them to think uh carefully about where the uh product experience lies.

unknown 1:03:37

Yeah.

SPEAKER_00 1:03:38

Would you advise that today 50% services and 50% product is okay if it offers more growth than just plain pure product?

SPEAKER_02 1:03:48

I would I would say so, because um, you know, one of the issues also uh but but again, right? Like yes, uh yeah, yeah, I I absolutely think uh the and this is routinely done uh uh done by a lot of uh good uh AI uh companies as well. I feel like uh and in yeah the the real question becomes one of you know is uh is this 50% services, 50% uh product uh like a story? Is it really you know 99% services? That uh that is often many people's uh concern. But yes, having uh having the need for a meaningful services component uh does uh for example, uh uh I mean it serves a useful purpose also uh today, right? Like in a manner which uh might not have been uh um valued as much uh a few years ago, which is that uh it it absolutely gives you a certain uh moat against the likes of uh next cloud code advance and everything. If your thing is um genuinely uh you know difficult to uh do in this uh pure uh sort of a self-serve manner, then uh arguably even uh model advances, etc., are helping you rather than hurting you.

SPEAKER_00 1:04:49

And hiring ML researchers today is brutal, right? With labs paying them million dollar packages. What would be your strategy for or advice to startups on hiring?

SPEAKER_02 1:05:01

I would say it depends entirely on what the startup is trying to do, right? Like uh look, uh, if it's an application startup, I don't think you even need an ML researcher to begin with, right? And even an ML researcher will get frustrated. I have I've definitely seen um mismatches, right? Like uh um there are uh certain people uh who um yeah uh like so certain people are excited about uh moving the frontier of this uh uh this uh like literally I've had uh some of these elite researchers who who are friends of mine uh telling me things like uh uh I care about moving the froder frontier model capabilities. I don't care about automating some uh silly form uh that may be used uh millions of times in the insurance industry. So the reality is yes, one has to think about uh what is it your uh you know company even needs, where the edge comes, etc. Right? Like uh yes, you um for a certain type of application uh company, uh the um uh your all you may need is uh your you know individuals that. Are extremely AI forward and quite smart, and you be a certain level of uh uh this if uh if you're putting some uh significant uh model investments to work, you're playing a somewhat different game, but then you better also be sure that you're not your you're not uh your first round is not five million. You better be sure your uh your founding team is such that uh people on a slide deck will write your 50 million to 100 million check. Then you can uh you can play a different game. Of course, if uh if uh if uh if uh this uh if you're an all-star team, you're you know uh this uh if if an ex OpenAI and an ex deep and uh head of research come together and uh raise uh they they'll probably raise a billion dollar round. Uh yes, uh they can do anything they like. They can go after uh they can try to compete with the three big labs as well. So, yes, I think it's important to know uh what game you're playing. Um uh how are you sufficiently resourced for that uh particular game? Like, yeah, like the one kind of game which uh I I think is impossible to win is uh one where uh you know you went and raised a three million pre-seed, you hope to raise a series A in 18 months, and then uh now you're saying you'll make a create a major foundation model play, it'll just not work.

SPEAKER_00 1:07:08

Have you tried yourself making a foundation model at Turing?

SPEAKER_02 1:07:11

We at the I mean we don't have uh uh any goals around that. I mean, we uh we we absolutely do uh research in terms of uh demonstrating the uh the value of our data and environments, even uh to inform ourselves uh internally. Like we frequently, for example, do research that demonstrates uh that uh our data, our environments, whatever else, uh has helped uh improve some of these uh mid-sized models, like uh 50 billion parameter model and so on. But uh no, we uh we have no interest in getting into that game. But yes, we do we do find that some kind of active uh internal uh vibrant research is necessary to be an effective uh research partner to our clients. At the end of the day, um yeah, uh the labs don't want um to be working with somebody that they have to babysit or somebody who's uh pure data vendor. Like ideally, they want people to engage with them at the level of saying, like let's say a researcher tells us I want want to improve the model in areas A, B, C, or ideally, even a research partner who tells them, I feel this is where your models are weak. This is uh this is of a lot of commercial relevance to many of our clients. Would you like to improve it? Would you like us to give you a path? Ideally, that is the kind of partner they want. So uh toward toward those goals, we engage in research, but no, we we have uh no goals uh of uh straight up making a foundation and and do you provide data across all modalities like voice, video, yes, yes, all of them, all of them. Yeah.

SPEAKER_00 1:08:32

So I I think you are as close to frontier model labs as as any company could be possible. Uh where are they themselves going, right? Because one theory is that they want to eat up both forward and backward. Backward means they'll try to develop their own cloud, and forward means they'll try to eat up application layer as much as possible, thereby has occurps and everything. So what is happening in reality? Because these are all speculations.

SPEAKER_02 1:09:03

Uh no, um I I don't think uh so I don't think it is wrong to say that uh conceptually this is correct. Um conceptually this is correct. Now, as a practical matter, uh it is uh yeah, as a practical matter, some things are more susceptible to falling, some things are uh kind of uh less susceptible to uh falling. Um so uh yeah, the I think uh um the more uh sort of uh you can play in the space of uh problems uh uh which are uh this thing which are less susceptible to falling, the better. Like for example, even two years. I mean, honestly, uh this uh SASA apocalypse type of problems are one. But for example, I uh even a couple years ago, I always thought uh certain categories of startups which were doing some uh almost uh like a band-aid on what is uh today's uh a hole in today's model is really not going to survive because the next version doesn't need that band-aid, basically. Uh like a lot of uh uh yeah, like like quite a few uh I think startups in the space of this uh uh AI safety security, not all, but at least some of them uh were very much in uh in this mode. Yeah, so one has to be, I mean, yeah, uh the uh you know uh yeah, there are there are a lot of obvious problems where it's very clear you're not gonna have uh uh this one. Uh um this uh unless you're extremely well capitalized and you're directly intending to compete with these uh people, uh you shouldn't be playing uh there. But uh yeah, um but well it's a nature of life that uh despite uh despite all of these, uh we we may still get surprised. Uh uh I'm sure uh you know none of us have quite predicted the last uh few years uh anywhere close to perfection. So but but at least we can take the best measures at our hand so that uh uh there is uh enough of a chance of uh victory exists. Like for example, Palantir is uh I feel a beautiful example of uh one where uh they have built so much of this uh deep uh verticalized uh uh stuff uh that uh um model advances have only uh made them a lot uh stronger and stronger.

SPEAKER_00 1:11:08

Yeah. There's very uh their their motives around uh their customers, right? Government contracts, defense contracts.

SPEAKER_02 1:11:15

But it's not just that, they they also do very well in commercial. Again, uh the uh you know historically people used to say that uh for Palantia, uh I'm not exactly sure, but uh to the best of my knowledge, about half of their uh uh business is uh straight up uh Fortune 500. And uh the point is I think deep understanding of that and uh and and really, really be maniac being maniacal about uh this uh productizing and you know figuring out uh uh you know uh how to build a product where uh even if OpenAI's models, uh, anthropic models, Gemini models, all of them get 10x better, how uh building it in such a way that uh you know uh your product offering only gets stronger, not weaker. Like I I remember Sam Altman saying this in interviews, even you know, two, two, three years back, uh uh long before the SaaS apocalypse and everything, that uh there are two types of companies you can build. There's one kind you can build which is going to get worse as the models get better. There's no sense in building that. There's another kind which will get better if the models get a lot better. Build the second.

SPEAKER_00 1:12:12

Yeah. And and do you believe that the old uh the the time for the old SaaS companies is over? There were hundreds of millions of revenues on SaaS tools.

SPEAKER_02 1:12:24

So there is no question at all, they will have to rapidly adapt, right? Uh the um yes, uh in a world where uh they don't adapt, it's uh it's very difficult to see uh you know how they have uh I mean, yes, you may you may have legacy customers, etc. But ultimately um uh a lot of these companies get uh good valuations on uh this uh on the back of uh uh future growth projections, right? Like if some of them don't adapt at uh uh at all, they will absolutely contract. Yes, they may still have terminal value, but it'll be much, much smaller terminal value uh compared to uh this one. Um uh yeah, uh it'll be a lot less valuable uh if it is uh that kind of shrinking uh thing compared to uh one where it is growing. The interesting question is, of course, you know, whether a bunch of them do manage to uh successfully adapt. It is uh I feel it's an open question. That is uh part of the issue uh b uh over here uh becomes uh that in many of the cases there is no easy way to just uh you know slightly um do some patchwork by sticking some AI on top of uh what already exists. They may have to radically uh redo the product, radically rethink. Um, and in many cases that can create all this uh innovators' dilemma and all that uh stuff. But uh but I'm willing to bet again, they all, you know, uh uh I mean they are all smart people. I'm I'm sure uh uh a number of them will uh uh this one uh you know will uh uh will make some investments in that direction. I think uh um the their future value will be entirely uh um I mean the biggest variable in the future value is how rapidly they do that. Like whoever completely reinvents themselves and becomes like the palantir of that particular sub-industry will probably totally start killing it uh because obviously they have a lot of significant advantages too that uh that a new incumbent doesn't. But yes, they have innovate as dilemma and all that, which are very, very real things.

SPEAKER_00 1:14:15

And uh currently, you know, every startup uh that is coming out today wants to go FDE model. But uh some of the drawbacks that I saw today, let's say in FD model, the first few clients, it's very high cost driven. Unless your clients are able to absorb that kind of a cost, uh it can almost like kill a startup. So, what's your advice on how to position yourself or your first few pilots or first few customers on the FD model? Like how to take include that in price.

SPEAKER_02 1:14:49

No, no, no, no, this is uh look, this is a very good question, right? Um the obviously the the uh the model works uh better uh if you're dealing with large clients where you're creating a lot of value for them and uh consequently you're uh able to um meaningfully price it uh um more as such. Uh so um I do think I mean I do think just like as uh it happens in the entire spread of startups, uh there are definitely people building things that are uh um that are in the big scheme of things incremental of over what is already there. Uh if you're incremental over what is already there and you have an FD model and uh you know bad cost structure, you you literally have the worst of all worlds, basically. So um yeah, I I uh I I think a few things one can maybe consider to mitigate is okay. Well uh definitely one one practice I'm not a big fan of is just uh uh you know do uh uh an arbitrary bunch of FD work and hope you uh stumble into some gold mine. I don't think that is uh ideal. Uh I I do think it's worth uh worth uh you know talking to a lot of customers, uh being pretty thoughtful about what can even uh uh this uh you know what can even work. Let us say if you want to um like maybe you know, maybe one very crude thumb rule, right? Like uh if one wishes to build a venture-funded startup, ultimately uh what is it you uh the thing you want to convince uh yourself and also uh investors is that uh um the uh you have a plausible path to a few hundred million ARR in uh say you know in uh uh four or five years and a billion ARR in uh 10. Uh so then um uh in in a lot of such combinations I know of, uh you necessarily have to have a mechanism uh to uh scale uh uh at least your uh um let's say you know Fortune 100 customers to say 20-30 million in revenue over time, maybe uh somebody who's a little lower down the list to 10. That means you should have a mechanism uh to uh create uh say uh 500 million in value ultimately for uh for a Fortune hundred. So uh start maybe start from there. Maybe uh you know uh maybe uh Fortune hundred uh or Fortune 500 customers are your not your main thing, maybe it is uh somewhere else. But yeah, certainly make sure that your your ultimate value prop is uh sort of uh plausible. Um if um some SMB can uh sometimes be a good like training wheel if uh provided whatever it is you're building there can port to enterprises uh to iterate and you know get some bunch of design partners. But again, it can if um uh there may be whole uh classes of uh problems uh that the Fortune 100 or the Fortune 500 care about, which uh for which there's no SMB equivalent, even and therefore you you may be shutting yourself out of those entirely. The making the jump could be relatively um difficult too. So yeah, I think the um I mean the key point to make is again, right, uh the people are customers are not just paying for AI, they are paying because uh the the there's an unbelievably um uh unbelievable step function uh jump in uh terms of the value proposition coming from these capabilities. If you yeah, you I suppose if you throw it at the right problem, if you go after a large enough uh customer, you uh ideally have done the work where yes, you can do some uh polishing of the uh value prop and you know discover things deeper with uh FD, etc. But it's not like uh you're completely wandering in the dark. Um all these I think can improve your prospects of getting FD, right?

SPEAKER_00 1:18:19

So thank you so much, Vijay. I enjoyed a lot our discussion. Thank you for being so candid. Uh I could have gone for hours and hours, right? But want to make sure that you know our conversation is packeted well for the audience, and we might give them a teaser for the next conversation between us.

SPEAKER_02 1:18:37

Great, great, awesome. Uh great speaking with you, sir.

SPEAKER_00 1:18:40

Thank you.