Venture Capitalist Warning: What You Should Know About AI (But Most Don’t)

If my AI is unpredictable and I cannot verify its behavior, that's gonna be a challenge where you have this missile on the road that is unpredictable in what it's doing and it could kill somebody. I can say that about healthcare, industrial AI. Would you want an AI doctor where you could not verify how it arrived at the conclusions that it did or to be able to say, yep, those are the right conclusions and we can verify them? Artificial intelligence is moving faster than any technology in modern history. But do we truly understand how it's shaping what we see, what we trust, and what we believe? Today's guest is a venture capitalist and long-time technology investor who works at the front lines of generative AI, cybersecurity, and deep tech. In this conversation, we examine where Gen AI really stands today, how values and incentives shape these models, the rising risks of fraud and deepfakes, and what it would take to build a future of AI that is not just powerful but trustworthy. This is Bay Area Innovators and I'm your host, Steve Esposito. Samir, welcome to our program. Thank you for having me. Excited to chat with you today. Fantastic for you to be here. So AI, there's so much talk, so much hype, and generative AI, and I want to see if you can get us a sense of where we are today with generative AI. But before that, to give us an explanation for our audience, what exactly that means. Yeah, absolutely. I think when you say the word generative AI, what does that actually mean and how is it different than other kinds of AI? Because clearly, we've had AI for decades and we've had neural networks for a long time. There was an initial, well, not initial, but the 2012, so more than a decade ago, is when there was this explosion of interest in neural networks. And if you think about what's different about generative AI versus if you go back to 2012 when we had breakthroughs like ImageNet and in computer vision, in speech processing, in speech recognition, you can think of generative AI and discriminative AI. And the way to think about this is discriminative AI is basically classifying things. You present it with an example and it will tell you if it's a cat, it's a dog, it's a person, if you've seen that show Silicon Valley, hot dog or not hot dog, that's an example of discriminative AI. The way to think about generative AI is that it can do all of the things that discriminative AI does, make means classify whether it's this or that, but it can also generate new examples. So if you say to a generative AI, create a dog, it can create many examples of what a dog looks like. So right now, for example, everyone's very excited about Nano Banana, the Google image generation model. And what that is doing is you write a prompt and it is generating visual examples of what you're trying to describe. So it's not just classifying things, but coming up with new examples that would be complying with the description that if you were to look at it and I were to look at it, we would both say, yes, that's the correct description of this image or this content that's been created. Now, those are not really new. They're not really created. There's a database where they're being. This is the thing where generative AI or these generative AI models are trained on very large datasets. Even discriminative AI was also trained on large datasets, but obviously the volume of data that we're talking about now is much larger. So it's seen a lot and it can come up with novel combinations of what it's seen versus just regurgitating things that it's seen in its training data exactly as they were. So to some extent it is novel, but it is based on seeing a lot of existing examples, whether it's text or images or video or music or any of that, and then coming up with novel representations that align with what you're looking for, but it is derived from the training data that it's seen previously, even though that example itself may not be in its training data and it's generated and it's truly novel. Now, generative AI is not autonomous AI. And I want to clarify this because there's some discussion and confusion about that. Tell us what autonomous AI is and where are we when it comes to that? Yeah, so I think if you think of the discriminative AI, which is making classifications, generative AI, which is generating new examples of whatever modality you're looking at, and then people tend to use the idea of autonomous AI intermingled with any of these. And the reason there's so much excitement about generative AI and also concern about job loss or it taking over functions or things that humans do, it's because people are also thinking about AI having agency and being able to perform tasks or complete workflows on its own. And that's autonomous. And I think the best example of autonomous AI that at this point, many people are going to be familiar with is if, let's say you've taken a ride in a Waymo in San Francisco or an autonomous vehicle. That's an example of autonomous AI. And I think we have to recognize that we've been talking about autonomous cars now for decades, but it's only now we're starting to see them in the real world, in real commercial deployments where you and I can hail a taxi in San Francisco and there really is no driver there. That's an example of autonomous AI. So that's considered autonomous AI? That is autonomous AI. And autonomous AI does not have to be generative AI. Autonomous AI does not have to be discriminative AI. Autonomous AI could be a combination of those things. And there's of course, another class of machine learning and AI called reinforcement learning, which also has a role to play when you think about building an autonomous system or autonomous AI agent. And so yes, generative AI is distinct from the idea of autonomous AI. It can enable autonomous AI, but there are different concepts. So it simply means it can make decisions on its own. Like if someone jumps in front of the car, it doesn't really need, probably still it's in a database or whatever system it has, but it makes a decision what to do at that point. Yes, exactly. So autonomous AI is making decisions based on the data that it's seeing, and then using those decisions to take new actions. That's how you should think about autonomous AI. Well, let's talk about some of the challenges that come up with AI in general. And you already touched on a few, the job loss and other things. But the first thing I wanna talk about is, so we have these algorithms, we have these models, and there's rules that create these models. And depending on the company, usually it would be that company that creates the rules. So the question probably that many people have is, how do we feel like ChatGPT, for example, probably has its own 80% of the market in chat for this. So in so many ways, what Sam Altman may decide on a model, on a rule, will influence 80% of the people who use that tool, how they find things, how they think. So what are your thoughts on that? Yeah, so I think when we use the term rule, it can have multifaceted explanations behind it. And I think what you're alluding to is, if you're open AI, and you're Sam Altman, leading open AI, you've aligned your model in a certain way. And even before you've aligned it, you've trained it on a certain amount of data. Now, if we assume, whether you're open AI, or Anthropic, or Google, or whoever, there's probably a base set of pre-training data that everyone is going to use, common crawl, scraping the entire internet and the web. But after that, you may start to get more distinctions in the kinds of training data, and the sources of the training data that the different organizations training these very large models are using. That will have an effect on the performance of the model. Some models will be better at certain tasks than others. Right now, if you look at the excitement around Google and Gemini 3, Google claims they are the best at multimodality. And you'll say, well, that makes sense. If you think about Google at large, and the data assets they have, whether it's YouTube, or Google Docs, Google Search, Google Images, they fundamentally have an advantage in their own data that they can bring to train their models and make them very good at doing multimodal tasks. It's not to say that open AI can't do that, and Anthropic can't do that, but it will take more effort for them to acquire data where Google has naturally accumulated that data over a long period of time, given their core businesses. So that's one element of, call it rules, and it shows up organically based on the type of data that these models are being trained on, fine-tuned on, to make them perform better, or achieve certain kind of performance on certain benchmarks and also real-world tasks. The other part of it is, we cannot ignore the culture of an organization, where the organization is, what country they're in, what are the rules of that country, what is the culture of that organization, what are their values, what are their social norms, and those will also make their way into how these models respond, how these models will or will not say certain things based on user queries. So this is a great example, is if you think about, initially when ChatGPT first came out, there were countries in which, obviously they did not have their own AI models in-house, so many people were using ChatGPT, and you would find that ChatGPT's responses did not align with the local in-country social norms or politics or rules and regulations. So that's one example of Silicon Valley's norms and mores and culture now showing up in, for a long time, what was the best model, that's why everybody used it, but hey, that doesn't mean that that model is representing the cultural values, the social norms, the politics of that particular country. And I think to some extent that will continue to happen. And I think another maybe counterintuitive example is if you think right now, so many people are looking at China and saying, wow, they really have taken the lead in open-source models, like in terms of truly opening up models, so therefore, China's being the most transparent. However, if you then think about, well, the usage of models inside of China, and by that, I mean mainland China, how could you have models that are not going to align with the values of the Chinese Communist Party, right? And if those models, if they've been tuned in such a way to adhere to the policy of the Chinese Communist Party, well, then you don't want those models to be just taken as is and then being used in the United States or in European countries or other parts of the world. And so I think when we were all very excited about DeepSeek at the beginning of the year, what you saw was, oh, great, there's this amazing open-source model, we're gonna take it, and we are going to align it with our values and our norms, and any guardrails that were put on these models for their uses in China, we're gonna take those off. So that's an example of when you think about rules and how these rules have to be adapted based on where a model's gonna be used and by whom, you have to, it's not a one-size-fits-all, and therefore, I think whether it's sovereign AI in countries or a healthy ecosystem of both open models and proprietary models, that will be a fit depending on use case and where they're being used. Now, I know one of the topics that you are pretty familiar with and an expert in many ways is about verifying, or how do we get to trust these models? Yeah, so this is clearly a persistent and ongoing effort, and it's not a solved problem. So when you think about every, I'm sure you're familiar with hallucinations. Of course. When models output things, and they sound very confident in what they're saying, but they're not grounded in reality. They're either made-up facts, or they are, in fact, completely wrong, in that the facts are something else, and they're saying something else. Those are the easy ones to figure out. Those are the easy ones, right? And I think it's also important to recognize is these models become more and more capable. This problem is gonna get harder, because if you're talking about, everyone's excited about, oh, we've got a PhD-level model in my pocket or available to me, whether it's open source or proprietary, and we're talking about superhuman performance at some point, how are we gonna verify that? How are we gonna verify superhuman performance is accurate, or even PhD-level performance is accurate? Most of us don't have a PhD, and so we would then have to basically resort to verifying information, looking up citations, and that's what we're doing today, is basically part of the reason Perplexity was getting so much excitement or so much buzz was they made a stake in that all of the outputs in Perplexity would have citations that you could go and look up. Well, that sounds great. That means I can trust what this model is saying, except even the citations can be hallucinated, and we've seen examples of this whether it was in the legal domain or in other domains where models will confidently output certain citations, and so you're like, okay, well, this is facts from the real world, but what you find is that even those are made up. So this is a real challenge, and I think this is where right now if you think about truly autonomous AI versus human-in-the-loop AI, the role of a human is to verify the output from a model, ensure its correctness, and I think at least for the foreseeable future we're gonna be in this state. Hopefully humans will have to do less and less touch, but ultimately they are still in the loop, and I think there isn't yet a, call it canonical way of saying that's it, hallucinations have been solved, nothing for us to worry about. We can put autonomous AI into production, and it will not make up things or hallucinate information and take bad actions, but it's an ongoing area of research, and there's lots of ideas around how you deal with uncertainty. This is one of the root causes. If you think about how large language models are built, they're not designed to say, I don't know. They're designed to give you an answer, and when they don't know, they are going to tend to confabulate things to be able to give you an answer even when their own internal confidence is low about the output that they are presenting to you. So learning the inner mechanisms of how these models work, making sure they're less of a black box, being able to interpret, and being able to see when you give it a certain prompt and it gives an answer, what parts of the model are lighting up? What training data are they resorting to to come up with the answer? These are in the direction of solving this challenge, but it's not a fully solved problem, and I think it is a persistent challenge. It has certainly played a role in perhaps putting a little bit of a damper on how fast, let's say, AI is being adopted, especially in B2B and in the enterprise. Now, I know you're a big fan of Star Trek, and I want to touch a bit about some of the technology, some of the amazing things that Star Trek has, and kind of where we are today, and what would you like to see happen as far as putting those technologies into reality? The top two, and one is, I think they're both now feel more tangible than maybe they did five years ago, but one is certainly much closer in to where you and I will be able to experience it sooner than later. So one of the ideas in Star Trek was this idea of the universal translator, and the ability to kind of break down walls in how you're able to communicate with anyone, and being able to do universal language translation from whether it's from English to Klingon, or from English to French or Spanish, even whether it's human languages or alien languages. And this always felt like, wow, it would be amazing if we ever achieved this, and it was truly transparent where you weren't thinking about it, and it just worked. And I think we are taking pretty good steps towards that direction with, first of all, the performance we've achieved. Meta has something, right? That's right, Meta has it. Even Google, which has been doing Google Translate for a long time, Apple, I think all of the big players, and also people that have specialized in language translation have benefited from generative AI, large language models to be able to achieve much more effective language translation, and even on pairs of languages that you were not explicitly training them on, and emergent properties where you're able to see, well, we never really trained it on this language, but it seems to have an understanding and is able to translate into that language. I think the other thing that we're seeing that's enabling this is we're seeing more and more capable hardware, right? If you've been following Apple and AirPods Pro 3, one of the selling points is live language translation in your AirPods, right? To me, that's the closest we've come to the idea of the Star Trek Universal Translator. Now, it still connects to your phone, it still has to rely on both a cloud and on-device AI capabilities, but you can extrapolate from here where it's not hard to imagine, at some point, we may be able to do it standalone with just AirPods that don't need to be tethered to a device, and you could still have a reliable, robust user experience where I could be traveling to a foreign country, and I can at least hear the other person in their native language, and they can hear me in my native language, and we can actually exchange ideas and information, and I think that's super exciting, and for me, that makes me very happy, and an idea in Star Trek is becoming real in my lifetime. I think the other one that is just so fascinating is the idea of the holodeck in Star Trek, where basically the ultimate simulation, if you will, an immersive reality where all of your senses, your ability to basically imagine anything and make it physically real, and you can immerse yourself in it, and maybe others can immerse themselves in it. Now, obviously, we've had mixed reality. We've had AR and VR. We've had VR headsets. From a very narrow aperture, we've been able to suspend reality and immerse ourselves in these experiences, but it's been time-consuming and tedious to create these virtual worlds, and especially if you wanna create photorealistic virtual worlds that you're immersed in that are fully immersive, and I think it's costly to make these things, and obviously, we have limitations, and we have to wear headsets, and they have a certain field of view, but I think what we're seeing now with world models and generative AI, and again, if you think about something like Genie 3 from Google. A new frontier for world models. With Genie 3, you can use natural language to generate a variety of worlds and explore them interactively, all with a single text prompt. If you think about World Labs, Fei-Fei Li's company, the idea of a generative model that is able to basically, you write a prompt, and it's able to create a 3D virtual world, one that you can walk through, and it's basically creating it as you go, and it's very customized. It could be very personal to you. Maybe the near-term next step is, now it's super easy for me to put this into my Apple Vision Pro or my MetaQuest 3, but I can also imagine that at some point, the physical immersive reality, it could be a room-scale thing like this, and I can start to actually suspend my view on reality and start to feel like I'm in a virtual environment that feels very real. That no longer feels like science fiction and fantasy. It feels like, hey, AI is going to be a key unlock for really democratizing what it takes to create these virtual environments on demand and their realism, the photo. Imagine Nano Banana for a video, for a virtual world, and you won't be able to tell just by looking at it to say, well, oh yeah, I see some artifacts or I see something. It looks very real, and assuming the immersion is real, I may completely lose myself in that environment, and I think that's pretty exciting. Now, just like in Star Trek, you also have to worry about what are the downsides of doing this? People maybe losing touch with reality and getting lost in these virtual worlds, this environment which is disconnected from the physical world, and I think that has some social implications as well. Now, what about things like warp speed and maybe even more interesting, the teleportation that happens? Do you see those on the horizon in any way? Yeah, so I think warp speed is an interesting one where if you look at some of the, there's been actual research that research scientists, NASA have done on what would be a warp drive if we were ever to have the capabilities to build one, how would it be instantiated, and it actually has a pretty good connection to what was imagined in Star Trek. Now, there are some other challenges with realizing that in terms of the kinds of energy that's required, energy density, and being able to control that, but from a theoretical standpoint, I think the idea is that, hey, there's nothing theoretically that is telling us that this is impossible, right? And I think that is very exciting. I'm sure AI will play a role there in terms of how we model this, how we simulate what a warp drive would look like. I think AI will play a very big role there. I think that's exciting. And then your other one about teleportation, the transporter. Clearly, at a very, very, very miniscule scale, if you look at some of the advances in quantum mechanics, this idea of being able to teleport particles from particles here, and we're able to have a particle with the same properties that shows up in another place without ever going in between, but I think we have to be honest with ourselves that the scale that that is happening, I think there's a lot more, both science and engineering, and I'm sure AI will play a role there too in helping scientists and engineers do the R&D to figure out how to scale up that concept and those capabilities. Now, I'm going to ask you soon about your favorite or your ideal company that you'll invest in, but before I do that, the challenge I see with all of these is that all the things we talk about, like ChatGPT or all these functions that, you know, you have billions of users, right? These are much more likely to bring profit to companies that build them because there's so many people use them. How do you turn that around, or do we want to turn it around, you know, because that makes the difference between getting something amazing versus something that everyone can use. You know, I don't want to say a challenge, but this is a good lens to put on. When you're doing fundamental science and R&D, you know, as venture capitalists, you're always like, well, is that a science project or is this something that has commercial application? But I don't think these things are disconnected because so many things are grounded in fundamental science before they cross some threshold and all of a sudden there's this aha moment about commercial applications that may be possible. And what you also see is a lot of times things are incubated in national science labs, in research organizations, in universities, and the goal was, you know, fundamental scientific R&D. There was no concept or no idea what will be the commercial benefit of this down the line, but it does happen. And I think the most exciting is when it's unexpected, when you weren't planning on it. And so I think we have to recognize that, but I think back to your point, if you're taking private capital and you're backing a company and a set of founders, you have to be very clear-eyed about what is the return on investment for that capital that you're providing. If you are saying, hey, they're gonna write some amazing papers, that's great. They're gonna have some fundamental breakthrough, that's fantastic. But then you have to be very honest about, well, what does that mean for this organization's ability to generate revenue and in what timeframe? So I do believe that there is a role for venture capital to back deep science, deep tech organizations, but also recognizing what belongs in a national science lab, what belongs in university, and what is now an endeavor that has enough pre-work done that it should be something that private markets, venture capitalists should be backing so that you can take it to the next stage. And what is your ideal company to invest in? Yeah, so in Turing Capital, I think as you look at what we do, our focus has been on the application layer and looking at, in the application layer, verticalized place, meaning you're looking at specific vertical industries. You're not just looking at large language models and generative AI or a hammer and everything looks like a nail. It's more like, hey, where do we have some serious challenges whether it's the amount of time, the amount of capital, the effort it takes to solve a problem, or a problem has been historically intractable in a particular industry. But we know if that problem was solved, a lot of people are willing to pay for that solution because they will immediately get the ROI from it. And so the ideal company for us is really, we look at what we call indispensable industries across different verticals. These are the things that are not the front and center shiny object that every consumer is getting excited about. These are the things that are running the world. They may be behind the wall. They may be in the basement, but they're super critical to everything that we do and we rely on them. We almost take them for granted. And looking in those industries to say, well, where are the hard problems that have not been solved yet? That things are either really expensive, things that have not been done yet because they would be too expensive, or things that are being done, but clearly there's some inefficiencies there, but we don't know how to do it any better. And then if we can take that and say, okay, well, here's a set of entrepreneurs who have some deep insights about a particular industry, have either lived and breathed those problems by being immersed in them. They know the pain points. They've also assembled a team where they have machine learning and AI expertise, but they have this deep domain expertise about that industry. And they can marry those in a very thoughtful way where they can come up with a novel solution and they're building some amount of defensibility. It's not just about, hey, I'm gonna execute really fast and run hard and I'm gonna be able to be successful that way. We wanna see that. But we also wanna see, to your point, that it's not gonna be like tomorrow, Anthropic and OpenAI are gonna wake up and say, oh, that looks like an interesting problem. Why don't we just flip a switch here? Now we can say we can also solve that. And all of a sudden, this company is no longer relevant. We've seen examples of, since the advent of ChatGPT and even in the prior years, where a big platform provider will turn on some feature or some capability and you have people that have been filling gaps. And all of a sudden, those gap fillers have an existential crisis as to, well, what is our relevance now? So we don't wanna do that. But I think if you focus on verticalized plays with deep domain expertise married with thoughtful applications of AI, and then ideally, that deep domain expertise also means you have some path to collecting very hard-to-find data and proprietary data to fine-tune your models. And you're being able to do that in such a way that again, tomorrow, the big platform guys can't just wake up and say, okay, we're gonna go hoover up that data too. Then that would probably check all of the boxes of the kinds of company that we would get excited about. Deep fakes and fraud is increasing. Cybersecurity becomes an issue with the advance of AI because it just becomes easier for the bad guys to do these things. What are your thoughts on that? How can this be addressed? Because it's pretty much, it's a cat-and-mouse game, right? Yeah, this is something that I actually worry about quite a bit. And I think this has profound implications across society, not just in tech. And so, to your point, it is a cat-and-mouse game. And I think if you go back even before ChatGPT and large language models, and if you go back to 2017, 2018, 2019, there's a lot of excitement about GANs, generative adversarial networks. And that was maybe the first examples of deep fakes that we were seeing, whether they were in the image domain. Video was much harder. Video was always seen as the frontier. But images, maybe some things in obviously text, maybe some in audio. But it was like, oh yeah, this is something to worry about, but it's not imminent. Like, this is gonna take some time. I mean, I still remember having conversations with some cutting-edge researchers working on generative AI back in 2017, 2018. And we would take bets as to how long would it be before we would see photorealistic videos that you would not be able to tell were not shot with a real camera, because they're so realistic, you have no idea how that could just generate it. And I think our general belief was, well, it's possible it could happen by 2030. But I would not expect it to happen in early 2020s. And here we are, right? And if you look at VO3, and you look at Sora, and you look at other models, even the open-source models, I think there's a lot of reason for us to be concerned. Whether it's deep fakes or increasingly referred to as AI slop, yes, it's fun. Yes, we can do cool, cute things like, you know. And they get millions of views or whatever. Millions of views, because it's entertaining. As humans, as animals, our brains see novelty. Concepts that we can only imagine, now we can see in film or on a video, and it's entertaining. And that's fine. But I think, like with many things in AI, it's dual use. And I think the potential for real harm is still ahead of us. We are starting to see examples of it. We've seen plenty of examples of deep fake audio that's been used to basically swindle companies into releasing payments in the millions of dollars because, hey, I thought I was talking to my CFO. I thought I was talking to my director. That's already happening. That's a thing today. And there are companies that are being created to defend against these kinds of deep fakes. Now, if you think about multimodal deep fakes, which are also a thing, you know, when you think about what you can do with generative video, this problem will become even more pronounced and profound. And I think there's a few things we have to do. One, in some ways, it's AI fighting AI. So the defensive AI and making that more robust to be able to catch things. But we also have to recognize in a cat and mouse game, if it's about one up leveling the other. So as your defensive AI gets better, well, that will also make the offensive capabilities better and on and on we go. So we need to think about other methodologies to make us all aware and savvy as to what is the content that we're looking at? Where did it come from? Is this real or not? That's the fundamental question. And I think there are other techniques like content provenance, content watermarking, being able to attest to the integrity of when a piece of content was captured. And from that time to when your eyeballs were on it, how is it altered or manipulated along the way? And is there a chain or a record of those alterations that were made? And if someone tried to alter it that wasn't supposed to, then you also have a record of that. So I think, and then the biggest thing is a social awareness, right? Making people aware that don't believe everything you see online. Increasingly, you have to be worried about that. You have to come at it with a healthy dose of skepticism. And there has to be a role for regulations and the law. The law has to catch up. I think what we've seen so far is there's been laws passed around what happens if you have deep fake porn and you can cause reputational harm to somebody by taking their likeness and creating a fake porn video. And there's now laws on the books that would cause you to suffer either being penalized or having to pay huge penalties if you're caught doing things like that. But I think we need stronger rules there. And I think the people have been worried for the last few years, well, what will this mean for elections? And what kind of political content is disseminated during an election? We haven't had that complete pandemonium with people just getting sucked into fake content yet here in the US, but we can't rule that out. And we should be worried about what happens in the next election in 2026 because we've already seen examples and not just one side or the other of people creating deep fake content and people getting fooled by it. So I think there's risk for financial harm. There's risk for reputational loss. I think there's risk for political influence. And the one that we're just at the beginning of is the impact on cybersecurity. And I think we have to recognize that while AI for cybersecurity can help us catch all kinds of attacks and malware and issues with computer code, because guess what? AI can write code and understand code. So it should be even better at catching those things than we are. But at the same time, I think we have to recognize that that same capability means it can also write malicious code and do it at a level that we haven't seen before. And I think we're starting to see the early alarms that that is going to be a thing. And I think the worst of this is when you take these potential capabilities and you combine them with a state actor, which has the resources and maybe a mission, that could be an extremely dangerous combination of the kinds of deep fakes that we're going to see in the near future. Yeah, thanks for clarifying that because that's definitely a concern that probably a lot of our audience has. I want to ask you something a bit more philosophical. Where do you see, let's go into the future, maybe I don't want to say 10 years, but let's go into the future when AI in so many ways, it's optimized for everything we need. We have robots doing the work and everything else. And I almost can't think of many jobs, again in the future, that AI won't be able to do. So what would be the vision, your vision or maybe others, of what people will do when we get to that point? I think this is the question we're all wrestling with, right? I can't tell you, I have a very clean cut, dry answer to this. I think there's a nuance to it. I think, if I were to be cheeky for a minute, I would say, well, right now, if you think about everyone's worried about robots taking jobs, imagine we have very large robot fleets at one point. Well, there's going to be a whole new industry around how do we service these robots? How do we fix them? How do we, if you're leasing robots, how do we run companies that are leasing your robots and then taking them back, deploying robots? There could be a whole industry around the logistics of large robots and companies that are making robot services available to all of us. So that's maybe the optimistic view. Now, the cynical view would be, well, there's probably robots that will be used for that too, and then other robots for those robots, and down the chain you go. So we have to accept that there will be new kinds of jobs, new kinds of professions that will be created in a highly AI automated economy, both in the physical world and in the digital world. And some of those, in fact, a lot of them, we don't even know what those are yet. And I think as certainly as adoption takes off, that's when you'll start to see the signal around, well, what are these new economy jobs that are relevant for humans to take on? And what are those? So that's the optimistic part of me that says we just don't know yet. The other part of it is, well, as I said, as much as we like to believe in autonomous AI, there's still some fundamental things we haven't yet cracked for us to say, well, autonomous AI is here, it's gonna be deployed at scale, hands off, and AI is going to make decisions and complete actions completely on its own, and by the way, come up with its own goals, right? And that has its whole another set of things we should talk about is when AI is autonomous, can define its own goals, and then go and execute its own tasks. So I do see a role for human in the loop. Now, humans, I think will, going from very low level tasks to higher and higher tasks, call it basically orchestrating what robots or AI agents are doing. In some ways, you can think of humans becoming maestros. You have an army of agents or a family of agents, and you're basically giving them the goals. You're telling them what to do and what not to do, what to focus on and what not to focus on. Is it possible that at some point, we also reach a point where they don't, we don't need humans to give AI goals, that AI is completely self-directed? AI can do research and experiments and hypotheses, and that then results in coming up with novel products and services. It's certainly possible, but I don't believe it's imminent yet. If you talk to all of the leading researchers, they believe we'll get there, but it's gonna take time, and there's more fundamental breakthroughs that are required. But I think the idea of cognitive labor, right? Cognitive labor and a lot of it getting commoditized or democratized, commoditized, and that has profound implications for society, and I think the biggest worry is not that it will happen, it's the pace at which it's happening, and the ability for us as societies to have both the resourcefulness and the appetite to adapt our socioeconomic fabric to this change and do it in a time scale that conforms with the pace of change. That, I think, is something to be worried about in terms of social upheaval that will result in. Well, I completely agree with the fact that there will be new jobs for people to do, but what I have a hard time understanding is the numbers that they're looking at. They have, what, eight billion people, and so the people that you mentioned is gonna be a handful. You know, I'm thinking about the lady in the grocery store or the truck drivers or the people in the factory, right? That's gonna be replaced. So again, the speed of how we're gonna get to that may be a challenge. I think it's also speed, again, to what we talked about in the beginning, which is it's not one size fits all, right? If there's parts of this world where human labor is so much cheaper than it is here in the West, that if you say to them, like, well, we can bring in this $2,000, $3,000 robot and it can do X, Y, and Z, I'm like, that's great, but I can hire 10 people and they can do that tasks as well, and it's cheaper. So the question is, at what point does it really kind of profuse through all of civilization at such a scale that even places where, today you can't make dollars and cents of deploying a lot of AI and automation, whether it's physical or in the digital realm, that you will even be able to do it then? Then I think, yes, it has some profound implications. And I think the question that really puts this out, and if you speak to other guests in the future, you should ask them this question. What do you tell your children? What should they study in school and what career prospects are they going to have? I have yet to hear from anybody an answer, whether it's a well-known venture capitalist or a titan of industry, that is truly satisfying. And I think the reality is we just don't know. And I think you can come back to first principles that, look, critical thinking, the ability to problem solve, the ability to think across domains and stitch ideas together to think creatively, those naturally feel like, hey, if there's some profession that banks on those capabilities, that will be the place that you should focus on. But the reality is we just don't know. What is it that motivates you to do your work every day? What it makes you go forward to go to work and engage with people? Yeah, I mean, I think if you look at the job of VC and anytime there's a big platform shift like we're going through right now, I was not an investor when we had the internet platform shift. And let's say this is something like that, maybe even more profound. I think that's a really exciting time because it challenges the status quo. It challenges whatever pattern matching you built up to say, well, okay, I lived through the cloud era and the SaaS era and the mobile era. I know the pattern of how these things will play out. And therefore, I'm gonna be able to have a crystal ball to know where to place my bets. I think you're fooling yourself if you 100% believe that. You can learn from the past. And I think what satisfies me in VC is you can form a hypothesis about where is an industry going? Where is the market going? Why these entrepreneurs versus those? Why this company versus that? You can use capital as a instrument to test out that hypothesis. And then you can become a board member to be on the journey to see, was your hypothesis correct? What is the data telling you in terms of what assumptions you made? And then closing the loop to see was it a successful investment or not? And I think for me, what I find very satisfying is that there are so many variables that no matter what anyone tells you or how they pontificate about, well, I think this is what's gonna happen in five years. This is what's gonna happen in 10 years. There's so many variables and many of those are hidden variables that frankly, don't believe those people. Talk to the people that are the builders, the engineers, the scientists, that are in the trenches, innovating, coming up with new capabilities. At least in their specific domain, they probably have a clear-eyed view on what's gonna work and what's not. But for me, what's satisfying is you get to talk to those people, you get to talk to titans of industry, you get to talk to other investors, you get to talk to entrepreneurs and you have to synthesize a view. And guess what? It's not like you synthesize a view and then, well, we'll see what happens. You can close the loop. You can test that hypothesis and then, hopefully, you learn from what you got wrong, what you missed, what was a blind spot and then, when you do it again, hopefully, you're gonna get better at it. But the world will constantly throw new, unexpected information. So as a animal with a biological brain with a large neural net, there's a lot of novelty in the world of what I do and I think that's extremely exciting. Would you say in that whole process, being humble plays a role? I 100% believe that. I think one of the ethoses of Turing Capital is having humility, at least recognizing that in the world, there's things you don't know and there's things you don't even know you don't know and as long as you're aware of that versus feeling like I have some God's eye view or I have some oracle which lets me predict the future or I have some really asymmetric access to information and therefore, I'm better at predicting the future, I think humility is extremely important. I think it's just, it comes back to also being a good human being and how you relate to others and their journeys. So for me, it would be difficult, whether it's VC or anything else and not come at it with at least some sense of humility on how to approach it. You mentioned before that AI's next chapter is not just about speed and how powerful and functional it is but also about how we can verify it, verifiability. Tell us more about that. You've been talking about it. Yeah, we touched on autonomous AI and some of the challenges to putting that in production and that's one of them, right? If you think about safety critical systems, right? If you think about why has it taken us so long to get autonomous vehicles on the road without a driver? Because it's a safety critical scenario. If my AI is unpredictable and I cannot verify its behavior, then that's gonna be a challenge where you have this missile on the road that is unpredictable in what it's doing and it could kill somebody, right? And I would say it's not just autonomous vehicles. I can say that about healthcare. I can say that about industrial AI. And so I think we have to recognize that if historically deep learning neural networks, oh, it's a black box. We don't know why it does what it does. That's not good enough. We cannot put these systems into production in safety critical scenarios where we cannot make reliability guarantees. Reliability guarantees means verifying their behavior. And so there's a lot of research in this area and verification at different levels. Everything from the physical layer, like what's called mechanistic interpretability, like really looking at a neural network and its outputs and trying to figure out like what is it doing in terms of how it's learning, how it's inferring things, whole active area of research. This idea of citations like we talked about where a model will output an answer to your query and it will cite certain resources. Well, I wanna be able to trust those citations as being real, right? So I think the idea of verification comes in different flavors. It's an active area of research. I think in some places we figured out how to put guardrails and modularized systems so that even if they are unreliable, stochastic, we have, it's one module and we have other guardrails that can account for that stochasticity or that unreliability and mitigate it or put what we call error bounds to say, well, even if it makes a mistake, but if the mistake is in this range, we can tolerate it and we can build around it. So that's one technique. But I think it's an ongoing area of very important R&D. And I think there's more to come in terms of what will it take to put verification that's so robust that it brings us closer to saying, yep, we can put autonomous AI, deploy it, and whether it's for healthcare or self-driving cars or whatever it is. Just give you an example. People, oh, we're gonna have an AI doctor by 2030. Yeah, yeah. Would you want an AI doctor where you could not verify how it arrived at the conclusions that it did or to be able to say, yep, those are the right conclusions and we can verify them? Or a pilot for that matter. Or a pilot. Those are the two. And right now, the real answer to how we're doing it is human in the loop. And that's why I think human in the loop is not going away at least anytime in the near future. Now, another aspect of this is we talk about making the distinguishing between reality and fiction. And when you have so much AI-generated content and you cannot distinguish it, the question is, does it even matter anymore? That's a great question. And one of the things I think about is having a discussion that, imagine you saw a deepfake video and it was of a politician doing some act that, oh my God, how could they do that? Immediately after you saw it, or even a little while after you saw it, you were informed, well, that was a fake. But it still left the impression. Exactly. It's left an impression on that emotional brain. You saw something and even if later on the fact, the factual verification is, oh, it's not real, that seed of doubt is implanted in your brain. And I think that is a real problem. And the implications of that, where more and more of us are exposed to synthetic content, AI slop, deepfakes, some of it is entertaining, some of it is meant to shape our views of the world. And even if you figure out it's fake, it still leaves an impression. The other situation, which is also something to worry about, there's gonna be real things that happen. And then there will be deepfakes that are created. And if there's a real video, and someone who doesn't wanna take responsibility says, oh, that's a deepfake. That never happened, right? And then the onus is on you or the, or wherever that video came from to prove that it was a real video, right? Versus denying all responsibility. Both of these things are gonna happen. And I think this is where creating more awareness, more education for the public, or how to think about a healthy dose of skepticism for whatever you see online, verifying information, trusted sources, and also the responsibility of the content creators and the massive networks in this idea of content provenance and watermarking, and of course, rules, regulations, and legal frameworks, all of these things have to come together to be able to deal with this problem. Samir, thank you so much for your time. We really appreciate it. It's great to be here. Thank you. Thank you so much. Thank you for watching another episode of Bay Area Innovators. If you enjoyed this episode, and you know someone who should be on our show, please send us an email at sfbayatepochtimes.com. We would love to hear from you. Thank you, and see you next week.