I think AI may be the biggest change of them all. It's not actually technology. The best way to understand it, it is a new form of labor. How has venture really changed from your perspective? I think venture has changed in every technological way. There's obviously this gap of like, you know, visionary founders not even gonna make things happen. Right? Disproportionate to the opportunity that's out there right now. The ability that founders need to have to look ahead a little bit and find enough believers to make that into a reality, that continues to be pretty scarce. I think people get too trapped in near term realities. You're gonna massively devalue things that were previously valuable. So what still has value? So one thing we clearly see today, I would say in this moment, in this AI moment is This is AI twenty thirty. Today I have Rohit Sharma. He's a partner at True Ventures. And I have Dhruv Kanekar. He's a founding member at CADRE and a partner at SDX. And I'm Chad, founder at CADRE and SDX. We're excited to dig in. Yeah. Awesome. So Rohit, you mentioned that, you know, your background is very awesome. You've gone through being an engineer, a founder, an operator. So with that being said, you know, what is this notion of what you kind of talk about with falling backwards? And now that you're in, you know, true in VC Yeah. How have you kind of walked through that path throughout your journey? I think, just quickly, so falling backwards is shorthand for not being able to plan it, right, in the sense that when I came to the Valley I had no idea what startups are, what venture capital is, what equity fundraising, any of that means. But I think I was very lucky to be with a lot of very good people and that's a theme that I think I've kept going back to the last three decades I've been here now. If you are with good people, you will figure out how to do very good meaningful things and I was lucky to benefit from that. So within six months of me coming here, was able to go on a path of starting a company, a company that eventually became quite big, I would say everything I have learned about venture, learned by being in that company, from two people to seven fifty ish at the peak, and then we got acquired by a very large company. So a lot of lot of lessons from that era that has stuck with me and that's I I think I've been very fortunate to learn from these experiences along the way. So it was not a design, I never thought I would have a startup or I would be in a company of my own or I would do the things I'm doing today, but I think I've been lucky to see some opportunities ahead of me at every juncture and and I think I've been very fortunate with timing and luck and good people to capitalize on that. Yeah. Yeah. For sure. It reminds me of something that I heard from a Steve Jobs commencement where it's like, you won't see, you know, the dots connecting while you're there, but when you look back, it makes sense. And do you think, you know, your, I guess, path and journey would have been any different if you had tried to plan it very thoroughly rather than just seeing where it goes? Look, I could give you a very biased answer that of course planning wouldn't work because that's my lived experience, but that's by no means the only way to do this. I think if you're true to what you want to do with not trying to be narrowing your thinking, you would have good outcomes, So could you get to this with planning? Probably. Would you have a objectively better outcome? Maybe. But experientially the outcome I think that I've had living through these, finding some of these things, working on them has been just amazing, right? But I don't wanna say this is the only optimal way of doing it. There are probably multiple paths to it. It is just the one that I have been on has been tremendously rewarding for me And I work by learning continuously. And actually, this is a weird way of saying I have finally found like your sort of person to career fit. That venture, early stage venture, especially with the phenomenal team here at Tru, allows me to do a very diverse set of things and be engaged in companies that are making and launching and operating satellites to deepen human biology with computational intelligence to industrial autonomy to robots to name your next ten things. And I don't think I coulda done that in a single career, in the conventional notion of a career, I think I'm very thankful that I ended up here because I this feels like no work to me, which I think is the best way of telling if you're in something you really truly innately enjoy, I I do absolutely enjoy it. So I think that's just, that's just what the trajectory happened to be. I'm sure there are better ways to do that. I I I have not optimized it for anything, not for money, not for credentials and title, and not for any other outcome. But that's not a good way of doing things. It's probably a bad way of doing these things. So you wouldn't recommend the fire, fire path for young folks? I don't know if I can recommend much or anything these days. I think paths are so different and so much more possible now that I think having an individual take on everything is actually more possible than it was maybe thirty years ago. So I would advise everybody, listen to a lot of things, including like people like me who opine on stuff. Listen to a lot of stuff. You don't have to follow any of this And and you build your own, I think, in the end because that that reward is unique. You know, think I think a lot of, you know, when I was when I was younger, I'm like heroes. Right? And they they sort of kind of guide your path. Like, I wanna do this. I wanna do that. Right? And some people are like, you shouldn't get attached to any one person or any one idea. In your journey, did you have any of those sort of heroes or you know, philosophies that you gravitated towards? I think learning and relying on your ability to learn has been pretty continuous and with me from a very early age and I think having this somewhat stubborn streak to not be going by what others are telling you at the same time. A little bit of contrarian in you? Yeah, I think that's been there. Yeah. And this ability that no, things, no matter what mix of things I end up with, I should be able to work through them somehow, I've had this belief I think for a while. Not that I'll be the best through it, this is not in that sense, it's just that, no, I'll I'll manage. And then looking back, no regrets. And I think I actually get that maybe from my aunt, goes, if you walk through a lot of gates in life, you ever just look back and think, oh, what if I did that or that, that's not useful. And I think that was just, in hindsight, just really good advice. It's like, no, you cannot do things with some regret. So you do things with a move forward mindset all the time. Sure. That that, yes, I've chosen what the best thing is for me and for folks around me and we will make it work. Not like we could've done it fifty other ways. I think engineers sometimes have this tendency to look back and optimize everything and I don't think you can do that in real life. Yeah. I I forget, there's there's just like one engineering law and it says like, the worst thing you can do is is over optimize early on. Right? Yeah. If you overfit your curve, you know you're gonna end up in bad scenarios, so let's not do that. Yeah. I know you briefly mentioned it, but but walk us through your journey to True Ventures. You mentioned you sold your company and it and it eventually brought you here, but could you kind of double click into that? Venture is really bad at designing good obvious pathways to venture. I guess just the nature of this business. It relies on asymmetries everywhere. So there aren't like very focused ways of getting into venture. I think, of course, I had the privilege of working in startups, working with a lot of venture firms, being in this area physically and then being connected to so many networks, so it puts you in touch with people who eventually sort of figured out there may be something here. And I think I was lucky to meet the folks at and I'd known so many of them actually prior to coming to Tru. And and then just just like any startup experiment, okay, go hang out here and see what you can do. So a lot of injuries around that, not like, oh, we have these three positions coming in these five years and this is how venture doesn't get that organized, think by design. So it needs to be a bit chaotic to discover the non obvious all the time, so it's fine and structurally it's like that. But I think I was, again, lucky and and the right place at the right time to to be able to trust some of these people and say, okay, see let's see if there's something here that works for for both sides. And I I I think looking back now, thirteen years, maybe more, withdrew or fourteen years, Yeah. There are lots of things that are working. I've been very happy working here. I think folks have in general liked me being here. They tolerated me for for so long. And I think it's it's been a great relationship. So it's it's good in life when you end up working with people where one, it doesn't feel like work. Two, are in an unstructured enough environment for a person like me to be able to just do things before they have to make sense. And I think that is really rare in venture and it's rare in early stage investing, which is what we do, but I think it really does enable you to work with founders very early, work with them in a very honest relationship. Let's go see if something is there, like it's less about executing on a business plan or two specific milestones. It's a lot more about we believe what you might do, right? And it's okay if you fail, right? That's part of our business model. But it's more important that you try it in the way that you see it. So I think that that sense it's worked really well. Yeah. Well, I'm gonna skip ahead here just a little bit since we're we're on the venture topic, but you know, obviously things have have changed quite a lot in the last few years here in all industries economically, politically, socioeconomically. How has venture really changed from your perspective? I mean you you have this purview, this seat, right, where you can talk about how AI has has definitely changed things but I think venture has changed in every technological wave that has happened in the valley going back to the beginning of the valley, like I mean, few blocks over there is literally the place where this all began, the HP garage, where it started initially with semiconductors, right, and and and then eventually got to systems, or actually going back for semiconductor, you could say HP was kind of like this non obvious company that had to exist, making tiny oscillators for labs, acoustic oscillators for labs actually. So the part that is really meaningful here is the change sometimes precedes the technical wave and sometimes comes after it, but it is continuous. Venture today is no longer what it was when I started and when I came to the valley. I I still have paper checks from my seed round, right? That's not a thing anymore, obviously. The amounts are different, the way you build teams is different, the kind of markets you can go after is very different. And I think AI may be the biggest change of them all because I think it opens up so many more pathways to driving meaningful change in just about everything we do. And it has this element which is very unique, it's not actually a technology, I think I've said that multiple times now. The best way to understand it, it is a new form of labor, we don't quite know how to tangle with it just yet. And the outcome isn't just better productivity or gains here and there, the outcome is a very different way of making things we want to make, products we want to make, workflows we wanna change, what we do interpersonally will change. So that, I don't think is understood just yet. This is a very big opportunity and this should drive a lot of change on our side of the fence as well, right. It's not like, oh, everybody else changes and venture stays constant, that's not obviously the case. So I think things are very different already in the last couple years, they will continue to change over the next few years. So you have to sort of, one lens to look through it is what becomes more abundant and what's still scarce, right, because that's the kind of asymmetry we always deal with in venture. I think abundance is on the side of can I dream of something and implement it with nearly zero cost, friction, and time? Yes. Right, I I can literally think of any complicated end to end system and have Anthropic or OpenAI or Gemini just reduce it to a certainty for me within seconds. We've never had that. Yeah. Because there has been friction and time and momentum involved in creating these things, and that was scarce, and that had value. Now parts of it no longer have value because of their abundance. So what still has value? So one thing we clearly see today, I would say in this moment, in this AI moment is the ability that founders need to have to look ahead a little bit and find enough believers to make that into a reality. That hasn't exploded ten thousand fold, right? That is not an algorithmic path up the way code is or the way technical capabilities are. So we know that this founder set of will and determination and your your intent to go do something and drive this change and make it real in the world, that continues to be pretty scarce. Right? Namely, you speak of of like the the ability to distribute what's built, right, and get it to market, and like that's No. I think also scarcity in the sense that a founder's ability to conceive what must be different and bring that to reality. Sure. It could be a hardware product, it could be a software capability, it could be a system, it could be a new way of doing something, that still seems to be there, at least so far, twenty twenty six. Right? So that's interesting. Now I think we have a magnifier on it because of our early stage position in the ecosystem, right? We really do want to be the very first to believe something before there is data, before there's proof, before there's hard evidence that, yeah, this is a thing. There's a product and a customer and margins in a company, right? We we are at a step just prior to that, right? Somewhere between the inception of that and the first implementation is where we live. So we continue to see that the number of founders isn't going up logarithmically like I said. The opportunities might be it. So there's an imbalance again, which I think means more opportunities for the ones who do wanna go build something, bigger opportunities for them because the tools they need are at a dramatically lower cost, at a dramatically lower timeframe of implementation, and their ability to have a very large surface area of impact is clearly there. So it's interesting. I think venture's gonna go through massive change over the next few few years. Like you said, I'm curious on this too, where there's obviously this gap of, like, you know, visionary founders not even gonna make things happen, right, disproportionate to the opportunity that's out there right now. But I'm wondering, you know, like a lot of folks that I see right now trying to start companies, they will put together some Vibe coated solution and go and make it a product, right, and go as fast as they can but maybe not consider exactly how durable or long term the business might be or the idea. Right? So is that something you see as like a pattern or what do you tell folks like, let's step back and think a bit on what will work out? I think there is a lot of noise and froth in the market clearly. And we see that there are a lot of people solving what seems to be a problem today, right? I think that's a fallacy that got developed over the last fifteen years in in the Valley which is, oh, solve a problem that you have or that you see, it's like, yes, but my expectations for you as a founder, especially if you're an engineering founder, are just higher, right? Solve a problem that as an engineer you can think of scaling to an impact that is much beyond where you are, right? Don't just solve, like if you have x y z problem today in software, yeah, trust that the system solves it fairly easily, like that's that's not the right thing to do. I think the right thing to do is hopefully much bigger now because with AI on your side, you can take on very very large challenges. Like my expectations are, again, using the same word, logarithmically higher than they were a few years ago. In terms of what a two person team can do, like no, I want you to be able to do a lot before you sort of commit yourself to a specific single product or project. So I think that's That is changing. If you look around today and say, I will do something, or OpenAI doesn't do this today or Anthropic doesn't do it, so I'll do x, that's not meaningful. Like, People forget that the gravity around those three giant projects, Gemini, Anthropic, and OpenAI, is very, very It will suck in most things that are really near. And forgot a fourth. And I'll let you build a fourth one, but these three are are big enough to, I think, include a lot of things that are within a few weeks, few months of instantiation, so I don't think there's value there. Now there's still value in getting to the customers and getting that impedance mismatch sorted out, because customers' ability to intake this and actually put it to work is still going to be limited. Yes. And I think that's a very large opportunity. The hard part is we're still calling these things with the terminology of the last twenty years, so we don't truly understand where some of these things are. Like you mentioned the word services, like services meant something very different five years ago, ten years ago than what it does today. So I think that could be a very big market where you're trying to see how people can adopt what's possible now, and that gap is big. Right? And it's not like all companies disappear tomorrow so there's no need for it. That's less less likely, and I think people have taken a very shallow, maybe cartoonish take of this that, oh, there's no more software or no more software companies. No, that's clearly not the case. In fact, I could argue AI means more software in more places continuously now, so actually, one, you need more engineers. You need more real engineers who know distributed systems, for example, right? There's just no way around it because you're now required to think at a very high level. The second part is, the way data goes back and forth anywhere gets used and synthetic data plus real data plus other data gets generated and moved around to train, to do inference, to continuously learn, any software that attaches itself to that likely has a great next ten years, right? So let's not sort of use the same brush to say, you know, all software is irrelevant. No, that's not the case. So I think lots of those fall in the category of changes that are beginning, but their end state is not clear yet. But that's fine. I think at at any point in the Valley's history, there's been that kind of uncertainty all the time. Uncertainty is good for us. Again, it means that the bigger companies won't just figure out everything and there would be no startups. So obviously, that's not a case. Yeah. And and not a good case for us and startup founders. I think we are exceptionally well aligned in looking at what uncertain there may be and finding creative ways to solve it with with good founders. Yeah. I I want to lean in a bit to diffusing this technology, right? So obviously there's these these big three or four are moving extremely fast and there's just gravitational changes almost weekly, right, sometimes daily in terms of the speed to development, the speed to product. And it's tough for companies to keep up, let alone large enterprises with, you know, many people that they need to manage change across and so forth. That's one of the reasons that you know we started this company is is we saw this opportunity with AI to to be a catalyst to diffuse this, right? But you mentioned like services have taken kind of a different different shape. It's taking a different shape for us in the sense of we are building a lot more software than we thought we would and that is helping us with sales, with strategy, with solutions, with product and research. Right? Talk to me a little bit about what you mean by services are are changing and do you see this model becoming like services as a software? How does that evolve? I I think I'm not sure what terminology sticks in the end but I'm a big fan of figuring out from first principles what the needs are and where my customers in in this construct benefit. Right? I wanna have that value very clear. Now, does it take the form of one or multiple companies or not? I I think that parses itself out over the next few years. But it's very clear that the needs are, I have so many incentives and reasons to adopt this new technology so that whatever product I might have, physical or virtual or digital, I can make this product at a lower cost with higher degree of fit with my customer and deliver them the value that they are really finding with my product, right? I wanna continue doing that. The notion of what a product actually is is also now worth reconsidering, right? So product used to be, okay, we sort of settle on this minimum common definition of what capabilities it should have, and then we fight like hell on which features it should have that address some segment of customers and ignore some others, right? That's a continuous trade off for any product company. Now if you'd add the cost of building a Brazilian integrations is zero, my product actually is just temporary in its instantiation, it could fit every possible customer that ever was, which is not a thing I could do in the last generation of software, for example. I still think I can't do in hardware at all, but pure software? Yeah. So can I conceive a product today that has a hundred fifty customers and a hundred fifty critically different forms of it? Yes. Can I main because it used to be, well, reason you wouldn't do it is you can't maintain that many whatevers, right? And your regression matrix for testing is gonna look like hell because you have a few billion commentorial possibilities you gotta test out before you release it. But if all of that is zero effort and zero friction and very high quality, why wouldn't I just do that, right? So that's one what if, which is, okay, does the product need to exist in the forms it has in the past? Or does it disintegrate into these streams that are really owned by what you call service, I might call something else, right, which is no, I have a capability, I'm gonna adapt it to your use, so you get value out of it. Whether we call it a product or service, should not really matter, right? So that's one critical difference coming. We are gonna redefine what products mean in this generation, in this era of AI. Then the other thing is, well, this model we've had of creating product and quite literally handing it to a customer is also up for reinvention, so why do I do that? In the old days of software, before cloud existed, that was it, right? You actually had to install a software product on your premises, on your laptop, desktop, server, whatever it was. Then with cloud, we could have it somewhere else and operate it from multiple places, which is sort of the status quo today. But I could turn all that around, like I should never need to ship anything to you or the cloud. What I really wanna understand is what do your customers want and whether I do it on my end without you ever having a notion of what this product is in today's slice of time, may not mean anything. But I'm fulfilling what your customers need and they see value in it for which they pay you and I enable that, right? So I think this whole construct also changes. So what a product means and where it has to be and how it gets exercised should be different. Like to me, the forward deployed engineers part is really interesting thing to dig into more, right? FTEs are really discovering what product ought to be, what the capabilities have to be, Alright? So if they can then solve it on the spot without ever having to create a permanence on the product axis, why wouldn't I do that? I would do that all day long, So that's I think a trend that is not going to really reverse is is my belief. I think we keep going with it, we solve more problems, whether or not a product is required in the conventional sense may not matter at all. So on that point, I'm actually it's an interesting take you had there where the forward deploy engineers come kind of come something that everyone talks about now, right, where it's like we want more FTEs and we need to deploy this and really get them hands on. What do you think makes them so potent as a role where you have the the side of we understand engineering fundamentally, but you're not constrained by like we're not able to talk to clients and have that sales motion as well to be able to understand what they actually need? I think FDE is is is shorthand today for bridging that impedance mismatch or that gap that I talked about earlier, right? So I'm giving you a engineering entity, one person or ten people or a hundred, however many it takes, to solve the problem that you have. Right? We have the last version of FDEs, which is they come with a platform and what they really do is to match your workflow or your context, and the context was workflow plus your problem to the platform they sort of had and that then made the product work. I think we are already in the next innings of that which is they understand your workflow in the context and maybe specific data sets you have and put it together with a logic that should work for you, whether it requires a platform or not is immaterial now. So I think that's the next generation already. The FDs are really solving problems that you have, hopefully in a way that is repeatable, so once they've done one thing, that gets committed everywhere, so you can continuously sort of have this roll forward motion of taking on more more things and committing the last thing that was done to what we may call a product. Sure. Right? With FDEs it's obviously an overloaded term, right? And and as AI gets gets more potent you start to see the convergence of of roles, right, and FDE being more and more prominent, right, you're seeing more and more of this convergence. What are other types of roles that might see a convergence like in FDE? I think you could do the mental experiment that maybe the FDE thing is already out of date because what I really should have are four deployed agents, ensembles of agents, armies of agents. Yeah. Right? Do they need an FDE at the tip of the spear? Maybe. But right after them is an ability for you to cast the workflow that you really have into a system of agents, into classes of agents that I instantiate that then carry on learning, carry on implementing, carry on creating the solution that you need. So I think that also changes very quickly. So I was playing or, you know, having this thought experiment the other day and let's not consider hardware for a second but just a pure software play. Given you have a perfect SOP, perfect access to context whether they're APIs or you know MCPs or whatever, what are some problems that agents cannot solve today? I think to the extent the problems are computational or they can be computationally represented, I I I think that is a limitless kind of scenario. Should be able to Attractability. They should be able to solve just about everything. If I can represent it digitally, they should be able to solve it. If I can model it, they should be able to solve it. Now, where do these things fall to? So today's LLMs, and LLMs are not magic machines, right? These are Shannon's machines, information processing machines, I should know why they behave the way they do. It's a massive universal approximator machine, it can give me the relationship between one thing and every other thing on as many dimensions as I would like, which is great. But they also are not de novo creating information that doesn't exist, Right? So I think you have to respect what it is that these agents and the LLMs they back into, and every agent is is like a small LLM, if you will, right? What is it that they're capable of, what is it not capable of? So if their problem that they're dealing with is entirely accurately represented in a digital fashion, so you have good data for that, you have the logic that you need to operate with, and you need You can tell it what the right outcomes are. So say you can construct a complex or simple reward function, like that's good and that is bad, they should be able to do that, right? Where it begins to fall apart, if it's a physical product, we haven't yet put the physics and chemistry of this universe into the LLMs just yet, that's really hard, right. These are not machines that understand quantum mechanical behavior so they can do like a field analysis of everything that's around us. We are pretty far from that. Yeah. The other thing is LLNs are really good at the digital representation of data they have from the last thirty to forty years, and then domains where you don't have that digital data, they're going to really struggle, right. So we're beginning to see that in domains like we are deep in biology, we see that firsthand in biology, where there aren't sufficient data sets to really predict or to come up with so called biology foundation models. They're very very narrow data sets, so your inferences are gonna be very limited and highly full of errors, if you will. So the real world, worlds that are running on laws of physics and chemistry and sparsity biology are not quite contained in in this yet, but that's not that far. We now have a way of getting there. But LLNs themselves will also transform a lot over the next few years. That's the other thing. They're not going to be just the LLMs we have today. Yeah. I'm curious on that. Do you think LLMs, you know, are just that architecture is gonna take us all the way to AGI or you mentioned obviously there's a lot of data that we can't understand well. Yeah. Do you think there's going to be a new shift coming soon before we get finally to the point of I think there are shifts already. And you can say, I mean, what you may have thought of is AGI five years ago, we already have that with with certainly the Opus class of models in the last four months, right? Because they do a lot of these things sufficiently well. The other thing is then you're deep in the philosophical domain of trying to judge where these machines should go rather than a rigorous framework of evaluating them. We don't really have that framework, right? So the the philosophical question is also not something we can ignore, is we have tended to think of these things as external to us humans, and somehow the intelligence we have is special or different and internal to us, right? I think these boundaries go away over time. The way an LLM might behave in order to change and to evolve the way I think is very much internal to me. It may start out as an external, but it becomes more collaborative over time. So how is it external to me at all is the question I ask, right? So if it's not, then the AGI question is sort of a moot point, like, don't care. It has become an intelligent entity which is interfacing with humans at scale. It's great, but it's not human like in any other way, like, that's our flaw of thinking it like that, it's not sanctioned, it's not gonna sort of jump up and try and do these things people claim in science fiction all the time. But it's a really useful manner of intelligence, and I think as we blur some of those boundaries, we redefine some of these existing questions and formulate new ones. AGI yes or no, I don't think is a relevant question anymore. Are these things very useful for us to go imagine the next few steps? Yes, absolutely. And I do think there was a step change over the last few months because things were then mostly good. So you can have these concatenation of agents suddenly become very productive. Right? You could have tried the same experiment two years ago and it would have been just a dead end chaos. Right? But it's not anymore. And and you have these wonderful things like, what's that Carthazi's last thing from a few weeks ago by running agents for five minutes and then Auto research? Yeah, auto research. Was like, why does auto research work? Because the individual atomic constituents of each one of these is not good enough. So if it's good enough most of the time, yeah, it will more or less work every time, and you have this beautiful error function that goes to zero after twenty four hours. Yes, right, that should happen. But that wasn't possible two years ago, right? So I think that notions of these that are expanding very asymmetrically across the board, I think the more useful question to ask is, is it useful enough in the domain I'm interested in? Right? And then those answers are different. Like, coding, is it at ninety percent plus all the time? Yes. In discovering new atomic properties of something, probably at two percent. That's fine too. Right? Well, let's let's talk about those domains that LLMs kind of suffer in currently. Right? Like you mentioned, like biology, there's not enough sufficient data sets. Right? So that's that's a problem to solve for. Right? They cannot think outside of their training set, right, and and create net new data, right, outside of their distribution. Now there's a lot of labs that are trying to work on new techniques outside of LLMs, there's diffusion that's kind of coming back into vogue, right, which is really exciting. You've got the Arc AGI team, you've got Ilia with SSI. Right? Do you have any ideas on on on where things are going? Are you guys investing in in this sector? So we are looking at all of these things is the shorthand answer, but also we are skeptical about a lot of these things because I think humans keep making this mistake over and over. We think we are somehow very special. So, to me, there is no speciality to a human way of thinking, so I must mimic it in a machine setup. Yes. Right? To me that that equivalence just doesn't work. Well, it's good to mention that. So I I think that's a religious point of view, not a technical point of view, which is fine. Yeah. It might advance us, I fully granted that, that it may give us new ways of LLM behaviors, yes. So can I have better ways of sort of figuring out short term, long term memory, yes? Can I have a better way of continuously learning as a result? Probably yes. But is that by itself the whole answer? Well, no. Like, preferably mimicking a human in an atomic form is kinda useless, sort of like the humanoid things, like they are good to look at and very useless. The reason I laugh is is I'm speaking to a friend later and he has this blog called Neurons Don't Back Prop, right, or something like this, right? And I think maybe one of the things that we've gravitated towards too much is trying to like anthropomorphize everything and you know, treat everything like a real neural network versus you know, things might be very different. Things may be very different. I think the nature of this intelligence is different. It's not just, oh, and in fact, that's why the AGI formulations have sort of always bugged me. It sort of sets this very artificial standard in place that it has to be human like to be, so it's like, no. We've had intelligence in different forms forever, right? The the wheel rolling on a road, whatever, fifty thousand years ago was intelligence and it helped us a lot. This is yet another different kind of intelligence, it's gonna help a lot. But the human equivalence part is really sort of, I think useful for public imagination, not so useful in technical terms maybe. And then the thing is, then you go borderline into more philosophic ways of understanding what intelligence might be, what language might mean, does language encompass all meaning, or is there meaning outside language? Like, you could drift forever on these dorm room discussions for days and create some other ways of understanding it, which actually are very useful, I don't wanna denigrate that, I actually think that's useful. I think the philosophy part is maybe the most useful to understand what AI may be, not technical terms. Definitely not talking to engineers about it, right? They have a very shallow view of that. But none of those are complete answers, is sort of my Right? They are all partial answers to something that we should really honestly admit we don't know yet. It's very early. It is a form of labor intelligence sort of rolled into one. It would upset all our previous notions of how commodities and scarcities and abundance actually works, it's like a live social experiment in that. I think we're gonna massively devalue things that were previously valuable, and we're going to give rise to points of value and intelligence we had nonexisted before. But to me, that is default a very interesting scenario. Yeah. And not just because shallowly as, oh, venture will do well in that scenario, but as a human, like, I actually like that because I think it has a chance to break through maybe some real limitations we have in our world and society and we may have a different way of doing it. Because if you think about it, if you look at the world these days, last two months, it seems like we are playing out the aftermath of geological formations and oil and fossil fuel. Right? So we're not quite in this new intelligent world yet but I think we have the beginnings of a redistribution of what value and the ability to generate that value means and that's mostly a good thing. Yes. Going back to to like trends in in VC, one thing that I've noticed is there's a lot more investment into research, right? I think, I don't know if it was YC or or or another group called this like the age of research, right, for venture capital. And when you look back, you know, LLM scaled because there was a clear scaling law, right? It said, you know, more data, more compute, right? And that's very easy for venture to understand, right? I invest in this, we get better models, we distribute those and so forth. Now you've got these new research lab with uncertain ideas, unproven, and is it now time to give a chance of those ideas to potentially find the next scaling law that gets us you know, LLNs or the current the current state of the art models? I think scaling laws have a long ways to run just yet. I don't think we are at the end of that. So could I do two hundred billion, two trillion parameter models now cost permitting, would I get better results? Yes. Should I do that? Probably. But should I abandon research on the other side? Well, no, I think we gotta do both at the same time. And I think there are multiple ways. Elements just happened to bring together a really elegant math computation with its implementation, like the matrix vector multiplication and the way I can instantiate it happened to really fit. It's not that GPUs were designed for that or that was designed for GPUs, this happened to be a really good match for it, right. Can I do better? Yes, likely I can do better. Can I do better at scale? I don't know, right? But all these additional sidebar ways of making it better and having it done differently, think will make us as a whole better, is is my firm belief. And we'll see this come around, I think, most clearly and short term on the inference side of things. As we get going in massive scale inference of many different kinds across many different workflows, you need many different kinds of intelligence now, right? And it's not gonna go back to this where I train a model, I run inference with it, I go back and wholesale retrain it, right? That's really not the way it's gonna work. I need to continuously train it and learn. So while I will suck up all the data I have, I will also need to generate synthetic data sets that don't have zero information, so don't just collapse model behavior, but I need a new way of doing that, right? So I need to have all of those ways come up, and yes, some of them will be maybe at the same stature as LLMs have today to do better information processing, and maybe not. But I think in all ways they're augmented, augmented to what we have today, and they should help us do better. It's sort of the way I look at it. So again, not a either or scenario, right, it's like it's a very firm and between all the, I want this and everything else. And are you investing in in that and? I think we invest in wherever we see for now a set of humans that have insights that deploy machines but they're not just machine like in their insight. So does that include the entirety of this field? Yes. And for us, very selfishly, it also has to fit the investment model. Can we invest in something that's gonna need a billion dollars tomorrow? Probably not. But can we invest in something that has two million and two people and eighteen months experimentation to see something which is now visible? Absolutely. We do that across biology, across robotics, across so called sort of physical intelligence, and we are very, very early on those things, right. No matter what the research and the publications have been there, the reality is you can have a bunch of cameras look at, let's say, ten bouncing balls and we cannot yet deduce like all the laws of the universe from that one picture, although all the physics of the universe is in play in front of you, right? So we have a lot of gaps between where we may be in a few years, and I'm very excited because I think all those are great areas to invest in. The information that is around us in any square meter of the planet, we are looking at a tiny, tiny fraction of that today, whether it's plants or microbes or the things we have built on top. Like, these things, materials, like cements and and engineered materials and natural materials, these are very primitive ways of doing these things. Let's just be honest about that. Right? So so I expect all of these to improve. Yeah. You know what? Not not to go back to to the the whole AGI term, I thought it was interesting. I think Demis from from Debye had mentioned, you know, one way to tell if the models are truly intelligent, forget the the nomenclature, is give them all the information up until, you know, general relativity and see if it can deduce, you know, or Yes. Figure that out. And I thought that was an interesting experiment. I think any of those experiments are good, right? Given the starting conditions for what we know happen and can it at least duplicate that, I think some of it probably yes, all of it certainly not. Yeah. Right. Well going going to experiments, right, for for models, you know, I think there's a lot of room to improve in benchmarks. I think one of the cooler benchmarks that we've seen is is Rprise and you know, those are friends and so you know I've seen kind of their development for Mark one, R two, R three and so forth. But something that's interesting is a lot of these benchmarks are static, They don't measure the continuous nature of these models and know sometimes you go on Twitter and you're like, it seems like Opus has been lobotomized this week, right? And and find out that it has been, right? And so Yeah. You're getting this benchmark that's static but over time there is this more subjective benchmark, right, that is evolving. What what are your thoughts on on general benchmarking techniques? Benchmarks or frameworks are are useful because they give us a common way of comparing a few things. They're not necessarily benchmarks to assess, let's say, the bulk intelligence factor in any one approach or any one model, that's not really useful. But they're useful in comparative across multiple things because that's the engineering analysis I should do to best fit what models I might use for a particular problem. So I think benchmarks to me, if they were more problem guided, would be more useful, right? So should I have a benchmark which is physics special or chemistry or biology? Yeah, it's really time that we have that. I think because we started defining them in general intelligence terms, we have sort of gone away from optimality there, but that's okay. It's been useful. Yeah. But I I would expect they split into domain specific rubrics pretty quickly here and dynamic because you you do wanna assess the behavior of the model with data, which is not something I can do today at all. Right? I sort of have the canned behavior out of the end of the training runs that I can characterize, but then in in presence of different kinds of data sets, what is that? And that, all of that leads me to just, as an engineer, I look at these things and think, we are very early in this. For example, one thing we don't do yet is these are information processing models, right? I know from Shannon's theory and laws, I don't yet say, here's a dataset that represents reality, and here are the noise, characteristic and biases of this dataset. If I knew that, can I structure the LLM better to ingest that data at a much higher degree of accuracy? Yes, like we know, we do that in communications every single day, like any WiFi or mobile communicating device is doing that in real time twenty four hours a day, but we're not doing that yet with the LM because we don't need to. The primary gate is so high, I don't yet need to deploy those strategies that I know I've worked through in signal processing for decades. So as we begin to do that, I also expect a massive change in frameworks to evaluate the function. Right? The the frameworks should change and they need to not just be dynamic, as you say, but also I think a better representation of the kind of problems we are asking these models to solve or model plus data complex to solve. Yeah. Yeah. For sure. Looking forward to two thousand thirty, what do you hope is a thing that people don't neglect? Obviously, there's people that have a bias towards a certain set of ideas or problems. What do you hope people actually spend time to focus on maybe isn't being pointed at right now? I think I mean, I'm impatient about this. I think people get too trapped in near term realities. People forget like two years ago, three years ago at the chat GPT moment, there were so many business plans we filled it up, oh, these models don't have memory, they can't read PDFs, they can't throw IO real files, blah blah blah, work on this. Like, that's not a good way of thinking about this. So I continuously apply that framework to even the reality today. It's like, oh, I can have hundred agents stand up and do something, but there's really no good state from handoff, messaging frameworks don't really work, they don't scale. So I'm impatient about all of these that let's kind of look past these, and I think people get too trapped by the shiny new things of today. Like, oh, OpenClock came in, and look, this is all new. It's like, no, this is not new. This has been worked on for a year, and it should not surprise you in the way it has. Right. Right? So we should expect a lot more of these. My expectations from AI are very, very high. And I think they continue to scale ahead of where the reality is right now. So twenty thirty, four years from now, I, one, I I hope we stop counting these somewhat misleading indicators of quantifying this tech, like people counted tokens and cost and cost per token, none of that is relevant in the long term. And in fact, every tech lead has shown us when we stop counting these somewhat misleading early indicators is when we get to the real gains from that. Like, there was a time when we counted transistors, right? Actually, four blocks from here is Shockley's house, right? First few transistor things all happened here in first couple square miles around where we're sitting right now. And that was the thing, till we figured out how to put down thousands, then tens of thousands, then millions, and today, a typical engineer fresh out of school can probably put down billions of transistors on a chip without even thinking about it, right, because we're not thinking about that optimization anymore, but we are using the gain that is now possible from all of those things. Same thing with servers, there was a time you knew how many computing things you had, it was on your desktop, below your desktop, in the hallway, then we got to the cloud, then we had virtualization and now Exactly, any engineer can spin up tens of thousands of servers, they die off, nobody cares. So when we stop counting servers, the real point of gain that starts. So today we are trapped in this misleading set of metrics. The tokens are not a good way to characterize LLMs. Yeah. Billions of parameters are not a good way, that's just the thing we can count, right? So it's necessary to do that, but it's not very useful over a period of time to do that. Does this tend towards like Goodhart's Law where, you know, when a measure becomes a target, it ceases to be a good measure? It's Well, I think it's already all these measures are not good measures. Like, we've had so many teams come up and say, here's how you use less tokens, like, that's entirely irrelevant. Like, ten percent fewer token costs because I can have some model autonomously route queries is not useful, it's not useful today, it has a half life of like two weeks. So, and and the thing there that is not that they shouldn't try it, the real high cost in these trial is actually the human time, focus and attention. Yes. Right. Yes. So that's now going to an endeavor that is not gonna pay off. That's what bugs me. Right? Not that they're trying it, of course they should try all these things, but if you're gonna burn two years of your lives doing something, that remains like the most useful measure or metric. So don't don't use it up in a trivial way. Yeah. That makes sense. Make it go. That's awesome. Well, I appreciate the time, Rohit, and and Chad, and this was great. Hey, come by and chat anytime. We love UCSD, we love working with SDX, and hopefully Yeah. We'll have more context there. I guess like one last question just on SDX while while we're here. You know, I started this community three, four years ago, graduated from UCSD, met Dhruv. Dhruv started the chapter at UCSD and we've seen just so much talent. Obviously you you have some of that talent now here at Trueventure which is very exciting. Just to end it off and and round off the discussion, in in light of experimentation and new ideas, what advice would you give to these you know entrepreneurial potential founders at UCSD looking for the next thing to spend potentially the next decade Experiment every single moment of the day with some way, manner or shape of putting AI to work for yourself or a problem you think needs to be solved or something. This is not going to be intellectually understood because it's not something where you have an asymmetry in people who understand it, people who don't. Most of us engaged in this do not yet understand it, right? So it's a great level or it's a great sort of equalizer of sorts. So I would say a student finishing up today in CS or AI or data science at UCSD is no more or less equipped than a person who spent twenty years at Google, right? Those levels are very close now. They're probably not exactly level, but they're not twenty to one, hundred to one different the way they were decades ago, right? So use that. And the way to use it is you experiment a lot with every way of putting it to work for you. And it really is useful in every single domain you can think of. I mean, I I am an engineer, but I I was in Greece last week, I used all the models constantly to tell me better about where I am, and regardless of the guidebook or whatever it sat on the plaque in front of me, these historical sites that have been around for five thousand years, no, tell me the shape of that thing, go find every parallel that existed at that time or trade routes that was like, you can construct an experience that is now deeply more meaningful because this exists at such a low cost, right, and low friction. So I struggle to think of a domain where it's not helpful. So who are the best native experimenters that would define the next generation of it are all the younger people in schools and just out of schools. This is not gonna be, this wave is not gonna be prosecuted by people who've spent forty years in the valley and they'll be on top of this. This is going to get reshaped by people who are coming up from a different way of using it, and allocating their attention and critical thinking in ways that we cannot conceive of today. And that is tremendously energizing for me because one, it means nearly everything should change. Two, that change will come from non obvious directions, like that is a given. Yes. Just like IBM in the nineties did not predict the next thirty years of Internet, we have the same scenario today. So while you have the Google and Metas and Microsoft command so much of our attention, we also know by default, like that is not the next thirty years. The next thirty years of what we, or our lives are gonna be, is going to be determined by people who are probably uniformly today between eighteen and twenty five years old. Right? That's a certainty. Yeah. So I wanna lean more into that when it comes to people experimenting with it, especially early in their careers or in the middle of their education, that's what you do. And the other thing I'd push for is just get to primary data and primary sort of values all the time. So what's the most useful thing today to learn or study in universities? Like, basic sciences. Like, learn physics and chemistry and basic, like, computing. Computing is not programming, by the which is something people conflate here all the time. No. Computing, which is what does a Turing machine consist of? How does it work the way we have constructed it to? So you understand this. You understand Fourier transforms and signal processing to understand diffusion. You're not gonna understand diffusion if you don't know signal processing, right? So those kind of things just need to be basic tools that you have, plus then AI to do a bazillion what if experiments at no cost. Right? Which is, I think, an amazing act. I would have loved to be graduating today. I I I I would have too. So just to summarize, the asymmetry is is closing. You know, go towards primary data All the time. Learn the sciences. Learn sciences not because it tells you the answer, it tells you a better way of asking questions. Like, engineers have long had, what do engineers learn after going through engineering school? They learn two things, how to observe something, how to analyze something, that's it. And if you can continue to do that, you should be able to ask that question of everything. Not that I sit here and say, hey, I'm an electrical engineer, I could figure out fiber optic communication at terabits per second. Yes, I can do that. But I should ask the same kind of question with the same criticality of like this table in front of me. Right? Why is it so inefficient to make this? Why is it two hundred bucks? Like, what's so special about it? Is this support over engineered? Is it aluminum? Like is it cast or made with water, whatever it is, right? So that's what engineers have, right? They should have this ability in a weird way to live an unhappy lives because they're questioning everything around them all the time, Right? Like, why are things the way they are and how do I improve them? How do I make them better and different is a question that I think you ask all the time. I think your ability to ask and answer those questions today with AI is just phenomenally different than every generation in the past And I think that's a very, very exciting opportunity. I love that. Well, with that Cool. Thank you so much. Awesome. Thanks, Thank going well. Take care. Nice seeing you, Chad, and nice seeing you through as well.