#334 The State of Data & AI with Tom Tunguz, VC at Theory Ventures

Generative AI is transforming industries at an unprecedented pace. But as AI changes how you or your team work, one thing is clear. Your skills also need to evolve. At datacamp, we offer everything you or your team need to adapt and thrive with AI. Whether it's business users looking to get the Most out of ChatGPT and Copilot, or developers and data scientists looking to fine tune models. You can learn the entire AI skills spectrum on Datacamp. Power your AI transformation today. Start learning@datacamp.com.

We thought that pre training was dead and all the techniques about bringing in the total of human knowledge and compressing it into an LLM, we'd solve that problem. But the launch of Gemini 3 basically shattered that illusion. The nimblest companies, the companies that are relentlessly customer focused, will be the ones who win because they need to anticipate and hear from their customers how those buyer preferences are changing. How do you start to create bigger, more complex documents that describe an entire project in a way an agent can understand and work on? While you are hosting a podcast, while you are attending a meeting, while you're having lunch, the status symbol now will be how many agents are working for me while I'm having tea.

Welcome to Data Framed. This is Richie. It's that time of year where we start to wonder what just happened and what's coming next. I want to know if we're in an AI bubble and what's going to happen with employment and how to keep up with constant change and whether or not product managers are still useful and how is the AI ecosystem going to change and what are the hot new startups up to and should I trust AI with my credit card details? I have so many questions and there's a lot to talk about, so I need a guest with their finger on the pulse. That guest is Tom Tungus, the founder at Theory Ventures. Tom's a highly data driven venture capitalist, having cut his teeth at Google on the AdSense program before moving to Redpoint as a partner. On top of that, he's a repeat Data Frame guest. So let's welcome Tom back and find out what's going to happen in 2026. So, hi, Tom, welcome to the show. Or should I say, welcome back.

Oh, thrilled to be back. Thanks for having me on, Richie.

Yeah, yeah, Good to see you again. Uh, so I want to kick off with perhaps the biggest question at the moment is are we in an AI bubble? Uh, do you think this is true?

Yes, I would say we are in a period of incredibly rapid investment in AI and I Mean just to kind of. So I think it's. We, we ran this analysis of comparing AI to other boom cycles or other bubbles and the, the most relevant one, I think only for scale, is railroads in America. So at the peak of railroads, we were spending. The U.S. was spending about a little bit more than 6% of our GDP on railroad construction. And it took about 40 years to get there. So the railroad started like 1810, 1820, and the peak decade was 1860. And we were spending the equivalent of what would be today of about five, five and a half trillion per year. Right. I mean, tells you like, wow. I mean, can you imagine how many people were chopping trees down? And so today AI is about 1.5, 1.6% of US GDP. Next year it'll probably be closer to 2 and a half to 3. So it will be half of the scale of railroads.

Okay. I have no idea railroads were so expensive. I mean, I suppose it makes sense. It's like it's a lot of infrastructure. Then I guess that's the same thing with AI, right, Is it requires a load of infrastructure in order to make it work.

Right. Huge amount of infrastructure. And so anyway, so it's not. Our spending is profligate, but it's not as bananas as say it was 200 years ago. The other dynamic here is we thought there were scaling limits. We thought that pre training was dead. And all the techniques about bringing in the total of human knowledge and compressing it into an LLM, we had solved that problem. But the launch of Gemini 3 through basically shattered that illusion. There was a tweet two days ago, I think, when they launched Gemini 3, that the improvement from two and a half to three was the single largest improvement in performance of any model in Google's history.

Oh, wow.

Right. Which is, I mean that'll make you sit up and be evaluated. And what does that tell us? Well, you know, within the last six months, the team at Google have figured out how to completely break the perceived law that we thought or the scaling limits associated with pre training. And that means that when the Blackwell chips ship and they're, they're three to five times more powerful than the current Hopper chips, that we should see improved performance. So you have two components contributing to performance. You have an ability to improve algorithms both on pre training and post training. And post training was the theme of the summer. And then you have basically, you know, significant improvement on the chip side. And so yes. Are we spending amount of money on all this? No doubt. Are we seeing dramatic improvements in performance yes. And then the last thing is, I think as a result of those performance improvements, the productivity gains that we'll see should be real and significant. And then will we overspend? No doubt. Look, we will. We will build way more data centers than we need. Nobody knows what that point is. We will overshoot.

Okay, yeah, that. That's really like, optimistic or at least very good news that we haven't. We're not hitting these limits around scaling just yet. There are still innovations being made. And do you have any sense of, like, where all these performance games came from in Gemini? Like you mentioned that post training was a big theme recently. Is that where we're getting all the gains from now?

Well, yeah, the theme of the summer was around post training, so reinforcement learning environments. And the way I think about that, and this is not totally true, but the way I think about it is if you have a robotic vacuum that enters in a room, it needs to do two different things. It needs to make a plan for how to clean around tables and the carpet, and then it needs to reward itself. So just the way that you and I might take a vacation after a long stint of work or grade ourselves at our work, what our objectives and key results are. OKRs. A Roomba has to create OKRs for cleaning a room. So 10 points for every square foot or meter, and then, you know, 50 points for every chair and then a hundred points for a full bin. Over the summer, we have been focused on the automation of both of those. The first one is nearly fully automated. The second one is still research. And so that was a big push. And the theme coming out of Q1 was all the pre training, all the compression of human knowledge is done. My guess, and I don't have any knowledge here, but my guess is there's been some pretty significant use of synthetic data to drive the pre training gains. And that's really exciting because if we can have AI generate data that is then used to improve the AI systems, we have this virtuous cycle that will last for some period of time. The naysayers will say, and I think with some reasonable caution, that at some point, if you just keep generating synthetic data, you have what's called model collapse, or if you just keep building fake data and fake data and fake data, all of a sudden your foundation is not very strong. But Gemini 3, and again, I don't know the details suggests that they're meaning, I mean, huge, biggest gains ever to be had from intelligent applications of synthetic data.

That's very cool. And I have to say I like your idea about okrs as basically like a reinforcement learning system for humans then I never really looked at that.

Yeah. And if you can create. Right now we're at a place where, if you can create those loops and we can do it in coding, where you can spin up a cloud agent and tell it keep iterating until you satisfy these conditions and satisfy these tests, then it works. And now we're trying to figure out what do those reinforcement learning loops look like for non coding applications. Okay.

Yeah, it certainly seems like coding applications have been one of the big areas of over the last year or two years really about. This is maybe the most important use case for generative AI. But you think actually some of the techniques applied from that domain, they're going to be used in other, in other fields. Is that your suggestion?

We've been deploying them inside of theory. So manage most of my email with an agent, manage most of my tasks. Customer relationship management system that we have is all through AI. We have agents developing tech algorithms to extract information from podcasts and lots of other applications. The challenge historically has been what's called tool calling. So read an email and then find the company URLs in there and then look them up in the CRM and then enrich them. There has been a significant amount of error at each step. When you have a multi step process, the error explodes the longer the number of conditional probabilities. And what's been amazing, and I have been using Gemini 3 here for the last three days, is the tool calling ability. I had looked at the benchmark. There's a benchmark that was created by Sierra called Tao. And the benchmarks would lead you to believe that the improvement in tool calling ability of Gemini 3 compared to Sonnet 4. 5 are relatively modest. But in practice it's unbelievable the difference. And so we're, we're getting there, we're getting very, very close. Okay.

Yeah, certainly, I think, I mean all year I feel like I've been talking about AI agents non stop and at the start of the year it was like AI agents are coming and then it's gradually got to the point, well, they're kind of here and. But because every company is now selling an AI agent, some of them are very, very simple things that I guess agentic is probably the maybe the politest way of saying it. And some of them are incredibly like powerful things. Do you have a sense of where the sweet spot is in terms of like powerful agents that actually do what they promise to do?

Well, yeah, summarization agents, I think were the first ones to work very, very well. Take large volumes of information. We're investors in a company called DropZone that summarizes security information for people who analyze the incidences inside of security operation centers inside of enterprises. And there the agents will query a database and they'll go and read online about some particular vulnerability. They might read previous documentation and summarize it. That works. This brilliant application of it, it's sort of the same thing, right? Go and research within a particular domain. I think we'll get to a place in early 2026 where you will start to see some much more sophisticated applications. I mean, and we're starting, there's some evidence of it within the world of customer support. So a typical customer support call might have 20 to 25 agentic tool calls. That's not to say it's a single call with 25 trains pulled by a locomotive. It's just, it's 25 short base pairs combinations. I'm mixing metaphors here, but so we're starting to see some of these workflows with, with pretty significant tool calls. I hope, if anything should give us hope. OpenAI released the updated version of Codex this week where they said they had broken their record of like two months ago of a seven hour continuous job and now it's closer to eight and a half, nine hours. Right. So if we can do it in coding, well, there's no reason that we can't do it with marketing or customer support or sales.

Okay, that's interesting. So, yeah, I remember chatting someone recently. They're saying, okay, if you're trying to build an agent, then you want five tool calls max and then otherwise it doesn't work. And you're saying like 25 is now the sort of limit of what's good. So I guess this means like it's like moving a file or looking up something from a database, or it's opening or changing something on it on a ticket somewhere. Is that the idea?

Yeah, that's right. Yeah, exactly right. Exactly right.

Okay, and so the idea with the long running Codex thing, then this is like what constitutes like a good thing you can do with agent. So I get the feeling for AI coding, if you've got a very simple task, you get the bot to do it for you. You don't need to touch it yourself. If it's like I want to build an enterprise application that something takes days or something, it doesn't work. And so eight hours is like the sweet spot for bot can do this rather than Human.

I'd say eight hours is the upper.

Upper bound. Upper bound, yeah.

Upper bound. I think we consistently see agents operating for 15 to 30 minutes with success. And that will be the sweet spot today.

15 to 30 minutes.

Okay, on coding task.

On coding task, does that mean you need to change your workflows in terms of, like, trying to break things down into smaller chunks so then you can hand it off to an agent?

Well, I think it's the opposite. I think it's. How do you start to create bigger, more complex documents that describe an entire project in a way an agent can understand and work on? While you are hosting a podcast, while you are attending a meeting, while you're having lunch, I think the status symbol now will be how many agents are working for me while I'm having tea?

Okay. Yeah. How much can I get done while I'm busy doing the stuff that I need to do as a human?

How much passive work? Right. There was a wave, I think, in Covid of like side hustles, and now it's actually, no. I. I want all these agents working for me. Can I have 10 of them simultaneously? Can I fill up a queue, a hopper that. Where all of the tasks are sufficiently well defined that the agent can go and. And just grind through it? Okay.

I do love the idea of just like, outsourcing a lot of my work in this way. Do you have a sense of how to get started on this? Like, how do you get your first agents going?

Oh, I think the. Okay, so a very simple agent is an upcoming meeting summarizer. That's a very simple one for a lot of people. You know, Podcast Listener was one of the first ones that we built. Newsletter summarizer. So I have an agent that goes through my inbox and looks at any email that has the word unsubscribe in it, archives all of those emails, and then summarizes and tries to find anything that could be interesting, and then sends me a compendium of all that. Those are the really simple agents or actually the. The. Okay, so I have this. I have this question in my mind, which is, what is the difference between an agent and a cron job?

Oh, go on, talk through. I presume the difference is like some kind of smart thinking or.

Yeah, I think. I think the main difference is that the. The agent should be self improving, but that's not necessarily always the case. Like, a coding agent does not necessarily improve. A coding agent is just constantly checking to see is there work for me to code? And if so, I code and if not, I don't. And a CRON job, well, it's kind of the same thing, right? Check to see if I have new emails. Or I should be checking to see if I have new emails and if I do, I process it and if I don't. So maybe the only difference is that there's some form of non deterministic programming within an agent. As I've developed all these different tools internally. My CRONTAB file is enormously long because there are all these agents running at different intervals. Check every one minute, check every three minutes. And in fact, a tip for anybody else building out there, don't use CRON on Mac. You should lose use launchd. It's far more robust. CRON won't restart after you put your Mac to sleep. Okay. That.

I feel like there's a horror story.

Trying to figure something out. No, no, no. It's just, I was just like, what. I just built all these agents and they, they're so lazy. Every time I put the computer to sleep, they don't start up. And so, and it was, it was Claude code that taught me that CRON is historically, and has been for a really long time, incredibly brittle.

Okay. No CRON on Mac. That feels like you've a useful thing to know about. All right, so I think, I mean, we're talking a bit about outsourcing a lot of your work to agents here. I know recently the job market's been pretty tough, particularly for recent graduates. Do you have any career advice like, particularly for people like you're just coming out of college or you're just trying to get started on your career. What do you do?

Yeah, unemployment rates are for college graduates are approaching 10%, which is quite high. The. I think the most important thing is there's a tremendous amount of demand for people who, irrespective of their age or seniority, can reinvent workflows with AI. Nobody knows. There's no one in the valley who has a secret for the best way of working with a coding agent, that there's no one who has discovered the new optimal way of designing a marketing campaign or the best way of structuring a customer support organization. It's all new. And in addition to it being all new, it seems to be all new every three months because the capabilities of the underlying models advance to such an extent that previous assumptions no longer hold. And consequently, if you are a person with ambition and curiosity, demonstrating a reinvention of a workflow in a domain you want to pursue is very likely to help you Get a job because a manager, a hiring manager will look at that implementation and say, this may or may not be perfect. It is a meaningful step forward. But more importantly, I'm looking at someone who has both the curiosity and the ambition to iterate to what could be a meaningful differentiator for me internally and as a hiring manager, I want to be promoted and then ultimately for the business as a differentiator.

I love that. Just being curious and trying new things because it's kind of wild west. Everyone's equal in this sense.

Yeah, that's right. We reached, I mean with SaaS we had reached, this is what the modern data stack looked like and this is the salesforce marketo stack. We don't have any of that. It's all wide open. I like that.

And so actually I've heard it both ways from hiring managers. So I think some companies are very much like, we only want to hire very senior people because they've got domain knowledge and we don't care about junior people at all. And other companies are, we want young people because they're much more likely to be coming to the fresh mind about how can I use AI to improve my productivity here. So I don't know whether you have a take on this. It sounds like you're leaning a bit towards the latter.

Yeah. So you definitely see. I think you're right. We used to, used to see engineering organizations structured as pyramids where you'd have one senior person and then two architects and a bunch of tech leads and then a lot of software engineers. And now we see them structured more like rocket ships where the wide part of the pyramid is smaller. I, I bet that is a temporary phenomenon. And I think it's a temporary phenomenon because the rate of code generation by those senior people will very quickly outstrip their ability to review the code. There's just no way. There's just no way. And we still need human in the loop to be able to review the code. And that will initially be some of the work of people coming out of school. We'll be evaluating the code and effectively QA on top of the AI quality assurance on top of the AI generated software. And the other dynamic here is I think we're in this little nadir of people generating a tremendous amount of productivity while the rest of the market catches up. But as the rest of the market catches up and we'll go back to higher levels of burn, higher level of investment because the overall productivity to remain competitive will rise just the way the sea level rises. Everyone needs to Respond.

Okay, all right. So you think it might be a temporary phenomenon then, that you need more senior people at the moment, but at some point it's going to change again. So actually you mentioned code review is going to be one thing that's going to become, I guess, bottleneck in the near future. I have heard an idea that actually product teams need to change because engineers are more productive. You need to hire more product managers in order to manage the extra stuff that the engineers are creating. Do you believe that story?

I think it's true, yeah. I think at some point. Well, so within epd, within engineering, Product and design, an engineer can now code design a product and then design the UX themselves. And the same is true for a designer. A designer can design the product, create a PRD and then ultimately vive code it. So you have a skill set expansion that blurs the lines across the EPD departments. And so some people take, I mean, I've heard some perspectives where, and even Claire Vaux will say this AI is the death of product management.

Oh, so going the opposite direction then?

Yeah. Why?

Why?

If you're an engineering manager, why would you need a pm? A PM is just orchestrating communication across multiple stakeholders, both internal and external. What does the customer want? How do we build it? When do we prioritize it? Can we validate it? Who are the initial design partners? Well, you could argue AI can do all of that. A product manager can look at an engineer's role and say, particularly for front end work. Well, an engineer is just implementing a specific. Well, what does the coding agent do? We need far fewer agents. I think we'll see short term aberrations, but I don't think we'll see pretty. I don't. I think the AI engineer EPD teams of the future will look very similar to the ones that we have today. The ratios might change, the productivity will be hugely increased. But I don't know if any one role increases in scope or is obviated. I think they remain relatively similar.

Okay, all right. Yeah, I guess you can't get rid of like one role entirely, but there's going to be some sort of blur between the different roles. I mean, do you think any other teams are going to be restructured or they're going to be affected by the AI productivity changes?

Well, customer success is probably the one that has been the most acutely and immediately impacted with the advent of the forward deployed engineer.

Oh, tell me about this role. I've heard about it. What does it involve?

Yeah, for deployed engineer is exactly what it sounds like, it's a salesperson who is extreme, who's an engineer. And they typically report to the VP of engineering, but they are deployed or they are sent to the customer. And the primary responsibility is taking an AI product and making sure it works really well at a customer. And so you need somebody who has typically has an engineering degree to do that. Historically, customer success has been not so technical. And if we think about the Boom and the SaaS ecosystem, if you're configuring HubSpot or if you're configuring Salesforce, you don't need an engineer to be able to do that. In fact, you probably want somebody who is more comfortable or looks more like the buyer so the buyer has a certain level of comfort. Many of these AI systems are bought by very sophisticated engineering teams and as a result you want engineers on both sides. And when the buyer is not a technical person, there's another need which is the configuration of the software to drive the intended outcome. Particularly because many of these pieces of software charge on a resolution basis actually do the work that's really important. And so the combination of those two things has basically reinvented customer success for many companies to be these forward deployed engineers which are, they're typically much more expensive and the contract sizes have to go up and they have, but so that's, that's I think the team that has changed the most.

That's fascinating. I have to say the combination of sales skills and engineering skills, that seems to be pretty rare or certainly like I guess post sales and engineering. Not many people with that. So I guess is that like a good high salary role then if it's such a rare thing?

Yeah, yeah. I mean I think the number of FTE job requisites is the beginning of the year up 10x.

Oh wow. Okay. All right. So yeah, maybe a career talk out for some of the audience then if you've got both those sides of the skills, I'd love to think about like how you go about dealing with all these like rapid changes. So I mean, because there's so much going on. I know you talked a bit about how product market fit for a lot of products just changes so rapidly. Can you talk with me like why is it happening and how'd you keep up?

Yeah, so I think so. Steve Blank wrote this book called Four Steps to the Epiphany, which the design philosophy for companies around establishing product market fit. I remember reading it in 2008. It was a brilliant book. And then it kind of created this movement called the lean startup movement. The idea was remain as small as possible until you've achieved pmf. And at that point you can scale. And then there was a counter movement, countercultural movement, which is called a fat startup. Raise as much money as possible, hire a lot of salespeople, and then after you sold, figure the product out, which actually is the way Salesforce was built. If you read Betty Off's book, he had a tremendous amount of salespeople and was burning a huge amount of money very early on. And during that era it was kind of taken for granted that once you had crossed the PMF threshold, once you were able to, a startup was able to grow from founder led sales to account executive led sales. The founder could train a salesperson to sell the product and that they could demonstrate repeatability, that it was all about scaling. And that's because the underlying infrastructure really didn't change that much and buyer demands really didn't change that much. And today, as we've talked about Richie, the capabilities of the models are changing so fast, the buyers don't really know what they want because nobody knows what's possible. The definition of PMF is no longer constant. It's this like you have to constantly reestablish pmf. So you might have a foundation model company moving up and competing with one of the clients, one of its customers. And so all of a sudden you might have had PMF as a business, an agentic company building on top of say Anthropic or OpenAI. But if OpenAI or Anthropic move into your market and they have the underlying model where you've lost PMF or there have been categories of software where initially the startup was developing a content generation prompts that we're very, very good at generating prompts. And then a new model iteration comes around and that model has intuited all its prompts and so there's no value in those additional prompts and you've lost product market fit. So I think that's, that's a dynamic where we're seeing, and it's happening so fast, faster than I've ever seen it.

Wow. Yeah. I mean certainly when the technology is changing so fast and like everyone's kind of jostling for position. Yeah, the, there's, there's chaos going on. So do, do you think, like, do you have a sense of who going to be the big winners from all this change and who are going to be the losers? Like how do you make sure that you're going to be the winner?

I guess Well, I think customers will be the winner because they have so much choice and the, the models are becoming that much better and, and the price for performance is significantly better. The, it's the nimblest companies. The companies that are relentlessly customer focused will be the ones who win because they need to anticipate and hear from their customers how those buyer preferences are changing.

Okay. Oh well, maybe there is still use for product managers and it's like speaking to your customers.

That's right.

Deciding what you want. Okay. Yeah, so I love that idea of like being customer focused is the way to win. So you mentioned that in general customers are winning because there's so much competition. And I have to say like anytime like I think of like what's an AI use case? You know, you do a search for that and there's already like at least half a dozen companies trying to solve that problem with AI. Do you think this sort of overlap is going to persist? Like the fact that like almost any possible AI use case, we've got six startups at the moment.

Yeah. Well the prize is so big. Yes. And then the amount of venture capital dollars coming in are just ramping. I think 70%, 75% of venture dollars this year are already in AI and asymptote to a hundred percent a total dollars coming into the venture capital asset class are growing pretty enormously. I think there's a very reasonable likelihood. Just to kind of give everybody a sense. The US venture capital industry was about 10 billion when I started in 2008 and it's currently around 200, let's say 200, 225. I think by the end of the decade you might see approx. Approach 500 billion and there's lots of different reasons for that. But if that's the case, then you can think about the number of startups that will be funded and that's twice as many today.

That's impressive numbers. Like my mom always bothers with all these things. I have to start with like world GDP I think is about $100 trillion. So you can go like, okay, 1 trillion. That's 1% of world GDP. So 500 billion, that's like, that's a good chunk of like the amount of money in the world.

Yeah, well, I mean, yeah, so half a trillion dollars in the US is about, is about 1/3 of 1%. Oh no, it's about 3%. It's about 3% of US GDP.

Okay, all right. Yeah, yeah.

Oh it's a little 1 1/2%. Yeah.

There's probably an AI can with the math there.

I'm just trying to do the math. Yeah. It's about one and a half percent of US gdp that would be then going into startups, which if we think about, I mean, I remember looking at the National Venture Capital association analysis of publicly traded companies, something like a quarter to a third of publicly traded companies at the time were venture backed and so. And the total number of jobs created as a result of them was something similar. So it's actually an unbelievable, unbelievably efficient machine.

Yeah, incredibly so. So I'm curious. So since we've got quite a lot of overlap with, with AI startup and I guess even bigger companies like what they're selling, do you think we're going to see a round of like mergers or acquisitions going on next year?

I do. The FTC just lost their lawsuit against Meta Facebook for a monopoly within social media. And I really hope that that opens up M and A in a very big way. One of my hopes is that we stop seeing these Frankenstein acquisitions where the cap tables are renegotiated at acquisition time, where only a fraction of the team is going to. Not the totality of the team.

Oh, can you give an example of this?

Yeah, I mean, I think the scale AI Meta acquisition where the business was valued at 28 billion, Meta paid 14 billion and a small number of the Skale team went over to Meta. But Skale is continuing to operate as a separate entity. It's not really M and A.

This was to like, I guess skirt around all the rules on mergers and acquisitions.

Yeah. And the hard part about all that is that it breaks the social contract. It breaks the social contract that early employees make when they join a startup and they say, I'm willing to take below market compensation because I believe in this business, in the case of an extraordinary outcome, I will be rewarded commensurately. And if that doesn't happen and that loss of trust exists, well then Silicon Valley as a concept that cannot, cannot thrive because there's distrust. And so anyway, I, I think we had this and I hope it's a temporary contortion around legislation and I hope that the FTC or The lawsuit that MetaOne really opens up the M and A environment in a very meaningful way. So I think we'll see that in 26.

Okay. Yeah, so I guess some interesting corporate transitions gonna happen. Actually, the, the biggest sort of recent example of this, it was on the data side rather than the AI side with 5tr buying DBT labs. I don't know. Do you Have a take on is this going to be good for the, the data industry as all.

I think modern data stack is in a wave of consolidation. No doubt about that. The 5tran DBT merger, very strategically important. It's a way for, I guess the way I phrased it was you have two unicorns competing with two Decacorns with Snowflake and Databricks and by joining hands they now are a viable third alternative. And there are three. Just simplifying for a second. There are three components that we're talking about. We're talking about the etl. So bringing data from many different places into a database. There's the database itself or the cloud data warehouse and then there's the transformation of the data within that, within that database. And so DBT +5tran represent two of those three and Snowflake and Databricks also represent two of those three although they're not identical. And so I would expect to see DBT5 trend push into compute and develop their own cloud data warehouse because that's where all the margin is to be able to compete. And then maybe they can create a third viable alternative to Snowflake and Databricks. That would be the ambition I suspect behind that merger.

That'll be fascinating because it's been really for the last few years it has been Snowflake and Databricks have been the two main players in the sort of cloud data warehouse. So yeah, having having a third option rather than being a duopoly would be very interesting.

It would be very interesting. Yeah.

Okay. And I guess where does that leave business intelligence then? Because that's, that's the part that those companies don't really touch.

Yeah. Well we are investors in a company called Omni that's growing incredibly quickly, also riding the wave of consolidation and they, it's an xlooker team and they're having a tremendous amount of success consolidating existing BI spend and rationalizing onto a single product. Product BI I think will remain a separate category even though it is adjacent. Many times customers buy BI plus data warehouses at the same time to have an end to end solution. But both products are sufficient, they're just different. Right. Like a database system has its own level of complexity, much more infrastructure focused and a BI system has a not obvious level of complexity underneath that. A lot of people think it's just charts, but it's actually far more complicated than that. And so I think they'll remain relatively separate here for forever.

That's interesting. And actually Khan Zima, the CEO Omni, he's a former Data Frame guest as well. So, yeah, he had a very interesting take on how BI is changing. And a lot of it's about just getting the data right and how you think about what's going to feed into my dashboard, rather than just like anyone can create a dashboard at this point, but it's like, how do you get the data right first?

Yeah, that's right. There's this pendulum within the world of BI that swings between governance and freedom or anarchy. And so I guess the way I characterize it is during the first wave, you've had Cognos, Business Objects, Hyperion and Microstrategy sold it. It defined all of the reports. Took a really long time to develop those reports. There was a rebellion in the form of tableau in 2003 that allowed anybody to download visualization software and make really beautiful charts, and that created a sort of chaos. And then cloud wave came and Looker developed a system of governance on top of cloud. And then Omni is trying to. And I think doing very well, being able to straddle both the individual creator freedom as well as the governance associated with standardizing metrics through an organization.

Absolutely. Like, there's very much a difference between, like, okay, I'm creating a dashboard to analyze, like podcast listens or something like that. And this is a dashboard that is like the official metric of, like, how much money does the company have? Like, they need very, very different levels of governance and freedom there.

Exactly, exactly. Right.

Okay, wonderful. All right. Are there any other trends that you're particularly interested in at the moment or you're keen to talk about?

Well, I think we're seeing a lot in the world of AI data engineering and the automation of AI, of data engineering, the dynamic assembly of pipelines at runtime. One of the questions we've been asking ourselves is there's been a wave of companies developing AI specific languages, programming languages, and we can debate whether or not that has merit. But it is interesting to ask ourselves a question. What if SQL were rewritten without the need for humans to understand what it was doing?

Definitely. That's fascinating. SQL's been around more than 50 years now. It's not changed an awful lot, actually. I quite like some of the changes in DuckDB. Like, they've cleaned up SQL a lot now.

Yeah. And then the LLM calls where you can call an LLM as a column. Yeah, yeah.

But tell me about these other. These new programming languages specifically for AI.

Yeah. Many of them are still in stealth. And the idea is many of the programming languages have affordances for humans so that they're legible. Right. We went from assembly to Python. Python looks like a poem, and assembly looks like somebody's banging their hands on a keyboard. And so as we move higher and higher in the world of interpreted languages, it becomes more and more prose like. But now we can go in the opposite direction, which is why can't an AI write Assembly directly and then only transpile to an English interpretable language during debug?

Okay, I guess this was kind of the idea behind Microsoft. Net, right? You had these sort of, you got C and whatever, which is vaguely human readable, but then it compiles, not compile. You had this intermediate. Net language underneath. And then so there was, I guess a human readable. Sorry, machine readable, but not human readable. So is this a similar idea?

It's a similar idea, yeah. But instead of. The idea is that a machine would just intake English and then output assembly or machine code by code. And I don't know if it will work. I mean, the reality is, for a lot of these models, the reasons they're so phenomenal at coding is just because of GitHub, the amount of open source code and the examples and the design patterns associated with these languages. It's always fascinating thought experiment to say, let us remove the initial parameters that really determine the path that this particular technology has progressed over the last 50 years. Right. It's total first principles thinking. And that's. That's always exciting to people who love ideas.

Absolutely, yeah. I mean, so I remember there was a thing last year. So, I mean, most AI code seems to be. I mean, there's a lot of stuff here in Pytorch which is like this sort of nice layer of. This is how I build models. But then Andrej Karpathy was like, okay, let's REWRITE, I think GP2 in C. And it's like it runs much faster and it's like a much smaller code base because, you know, all these like Python dependencies. So this seems like a similar sort of idea. It's like, can we go a bit closer to the metal there?

That's right, exactly. Very well summarized.

All right, wonderful. Okay, so before we wrap up, I need to talk about one of the. One of the things you built. So I know you've created your own podcast processor, so tell me all about this. Like, what did it do and how did you build it?

Yeah, I don't have time. I mean, there are so many wonderful podcasts and, and there's a Lot of phenomenal information within those podcasts. And so I have time for a few per week. But wouldn't it be great if a robot could listen to the podcast on my behalf and identify the key quotes? So we built an internal system that collects all the different podcasts, listens to all of them, and then identifies interesting statistics or companies or people that we should know about. And it runs in the background. And, and it's been working really well, I think, for the last three or four months. We're able to process. I think my list is now around 55 or 60 different podcasts. And it helps us with diligence and figuring out memos and what other data points might be interesting to a particular company. So it's becoming a compounding data set for us.

That's very interesting. I always think of, like, the data trained audiences being humans, but maybe like, at the future it's gonna be like some of my audience's bots as well.

Yeah, I mean, it's, I wonder if the speech sound will be different. Right. We're, we're starting to talk about generating websites that are optimized for robots. And certain websites, more than 90% of their traffic is for AI. Will it be different for speech, I wonder?

Absolutely. I mean, yeah, I see. Someone's trying to pitch me the idea that you need to redesign your retail experience for bots, because shopping agents are going to be the future at some point in the. Well, I guess maybe 2026, maybe 2027.

Yeah, it's right. I mean, so Nike, when they drop a new pair of Jordans, more than 99% of the traffic that day is robot. There are B2B robots, wholesalers who are looking to buy that shoe and then resell it. There are customers that Nike would love for them to buy the shoe. And then there are customers that Nike would prefer that they don't. And so how do, how does Nike.com tell them apart?

Yeah, that's interesting. I have to sneak ahead. Culture is a bit alien to me.

But. Right, but do you understand the dynamic? Right. Or let's say no, it's the same.

With, with concert tickets. Right. Is that popular? Is having a show and then there's a limp number of space in the auditorium.

So.

Yeah, there's a lot of bots trying to buy tickets immediately.

Great. Parallel. That's probably a much more tangible parallel for.

Wait, are you into sneakers by any chance? I'm just wondering whether that's where you.

No, no, no, no. I, I, I have the ones that come in the store. Not the ones that. That I need an agent to bid for.

Okay. Yeah, but that does actually sound like quite a fun project. If you want to build an agent. Okay, go and build your own shopping agent. It's maybe high risk because, you know, you're giving it your credit card details, but if you're really into sneakers or.

Yeah. Well, you can create a virtual card with a. With a limit on it, so it can only spend a certain amount of money. And so you can. But, yeah, I think this is the way. I mean, I wonder if this is the way. I don't. We were researching pickleball rackets yesterday, and I just asked for the four top Amazon links and then clicked on it and pressed by. I didn't even read the reviews. I just asked AI to read. Read them for me. I thought we're close. We're really close.

Yeah, it depends on, like, the. The risk of purchase a pickleball racket. It's like, okay, I guess you can't go too far wrong and you've only wasted, like, what, tens, maybe. Maybe hundreds of dollars. But, like, would you do that for, like, a car or would you, like, how. How big do you want to go on this? I guess.

Well, okay, this would be awesome because we're not far from this, but I think if you're in Texas, you could actually have a flow where you ask AI, which is the best Tesla to buy for you, and then you can go to the Tesla website and press buy, and then a few hours later, the car drives itself to your house. That's really living in the future.

That is very much in the future.

Nice.

Okay, I guess for the brave souls, like, putting a lot of trust in AI, then.

Yeah, go for it.

I might test drive it first. Like, I hope there's an option to, like, go back on your word at that point.

Yeah.

Uh, okay, cool. All right, so just to wrap up, because I always need new ideas for people to follow. Whose work are you most interested in at the moment?

Oh, there's this wonderful guy, Gavin Baker. Okay. Gavin runs a hedge fund called Atreides, and he is a wonderful analyst of AI and the public markets. So he had a wonderful tweet, I think, earlier this week, just about what's happening within the world of AI and so he's very much on point. So I'll recommend Gavin.

Gavin Mega. All right. Okay, One to watch out for then. Super. All right, thank you so much for your time. Tom.

Richie. Pleasure was mine always. Great to see you.