Interviewer (5:29)

AI agents were supposed to be the big rage in 2025 and they're supposed to be the year of AI agents. Do you see 2026 finally being the year that AI agents have a big effect on the economy?

Richard Socher (5:39)

Yeah, it's really interesting. A lot of people overestimate. I see AI is kind of struggling. A lot of people are like, it's black, it's like overhyped and nothing. Like it's just kind of bubble burst. And the other people are like, oh, it's going to change everything. And next year we have 20% higher GDP. And it's the, the truth as always, is somewhere in, in the gray middle, it doesn't create as many fun buzzy headlines. But the reality is that AI is already changing different industries. And we're seeing that like if you were an illustrator that has one style and that style is fairly common in the world and it can be trained on, on the Internet, AI has disrupted that entire field. Now the illustrator industry is not as big right as the movie industry or the music industry is. For instance, they don't have as many strong copyrights as the music industry. And so we, there's, there's, we don't see as much sort of disruption in terms of overall GDP. But I believe that the largest GDP driver for developed economies in the world will be AI. I don't think it'll be 10% for now. Even the Internet or you know, Electricity and so on. It takes usually years to really get into every industry, every company and so on, to adopt it. And we'll see that with AI too. But what we are seeing certainly now is that the people and the organizations that use AI and really lean in are slowly starting to pull away from the people and organizations that don't. And if you work in any kind of job that requires intellectual work, which is, you know, more and more jobs in the future, as physical automation has already happened, you will not be able to say, I'm not good with this agent thing in five years from now. Just like right now, if you work in a high paid job, you cannot say, I'm not good with this Internet thing, or I'm not good at this computer thing that used to be a thing that maybe 20, 30 years ago you could say. And people like, okay, whatever, you know, I'll just use a tax machine or something like, you just can't say that anymore if you want to be taken seriously in the workplace. And I think that agents are an obvious one. Why? Because you're just so much more productive with it. And we're seeing this in companies like in our customers.

Interviewer (7:34)

What are some early use cases for AI agents that you see being deployed in enterprises today?

Richard Socher (7:39)

Windserve is a customer of ours for programming. Programming is a massive application we see in the enterprise in a lot of different places. Legal work, Harvey is a customer. So like automating more and more legal work, there's a lot of interesting investments that we're making. Also at AIX Ventures, in AI legal tech, we see companies like Iloca that automate architecture and design for architecture firms. There are a lot of painstaking, slow work. Um, we see consumer apps obviously changing a lot and giving us answers more quickly. We see journalism being massively disrupted on both sides. People don't read the original news as much anymore. But also journalists are becoming more productive. And you can now have an AI journalist in almost every little town and city that takes in data from different places and then writes like articles for just that, you know, 5,000 person town, the way you couldn't do with, without AI. So yeah, we're seeing it in almost every industry.

Interviewer (8:35)

Full circle back to the local newspapers. We had local news and then we had the Internet and now it's AI local news.

Richard Socher (8:40)

Yeah.

Interviewer (8:42)

Besides running you.com, which I mentioned recently raised at a $1.5 billion valuation, you also have a $250 million AI fund. Tell me about that fund.

Richard Socher (8:50)

Aix Ventures. It was really exciting. Started sort of from my angel investing and has since really grown with an incredible team. Um, we've been very fortunate to invest in a bunch of companies in their seed rounds like Hugging Face, Perplexity, weights and biases flow like Whisper Flow that changes how people interact and talk to their phones and their computers. Ambience that helps how doctors keep notes. You know, it's such a frustrating job. Imagine you're a doctor, you really want to work with patients and you spend 20, 30, 50% of your time typing up notes and working on some computer system to keep it up to date because you have to do that for reimbursement and so on. Ambience and others help automate that. Massively invested early in Windsurf, which is, you know, improving AI coding and many other edible unicorn companies. Grown a lot. Yeah, we're mostly focused on the I X. What is that X? It's again different apps, consumer apps. It's some of that infrastructure that AI needs like Hugging Face. All these different examples from architecture to legal healthcare. I think bio is also an incredibly exciting space right now. Tech bio is really will be changed massively because of AI. What essentially calculus did for physics, AI will do for biology. It's the right language, the right way of thinking about large scale complex systems. And so from first principles we're very.

Interviewer (10:11)

Excited about, do you think there that's inevitable? A lot of people in biotech believe that it's just different systems and that's naive. To plug in AI into something like.

Richard Socher (10:19)

Biology, I think that's very wrong. We're seeing it. We're seeing very interesting companies, companies like Parallel Bio that use AI to build and track organoids and you know, have full FDA approval to essentially test immunotherapy treatments and drugs in these organoids instead of in live animals. You know, it's just like incredible. There are millions of animal lives will be saved and times to get through. FDA will be cut down by years with this one company. So the impact is there, it's undeniable and it'll just get bigger and bigger.

Interviewer (10:56)

How do you go about deciding what to invest in your fund given AI is evolving so quickly? What are your first principles for investing in early stage?

Richard Socher (11:04)

You really have to look at the team a lot. Just like the intellectual horsepower of the founders. Their willingness to work, to grind, to not give up. You know, running a startup, starting a company is a huge like emotional roller coaster.

Interviewer (11:17)

Right.

Richard Socher (11:18)

Someday you think you're the next biggest thing in place or company in your space, and another day you think maybe it's all dead and it's not going to work out. And like, you have to just like work super duper hard. And so we look at strong founders that have those that sort of not giving up attitude and are really smart and work on the right things. It's always a balance between stubbornness and adaptability.

Interviewer (11:38)

I mentioned you founded your first company in 2014. Obviously you've gotten through a lot of highs and lows. And especially within AI, has it been easier for you to stay even guild or is it still very, very bipolar days?

Richard Socher (11:50)

Humans are definitely sort of baseline creatures. So if your average day is pretty crazy, you get used to crazy days a little bit. Also, you know, with review.com, like we're now making large amounts of revenue, we're growing very well, we're growing our sales team. So it's a little bit less like, oh, this will go completely to zero. I think those days are kind of over. But you know, you can still be very excited and there's certainly still deals that you're pushing for and so on. And sometimes you win some deals and we don't lose too many deals, but sometimes we lose deals and that's really frustrating too. They're still ups and downs, but they're, they're slightly smaller for sure. And, and yeah, you do get used to it. But so come back to your question. Like, so the founders, then the founding team is what you look at and make sure the dynamics between the team work out well. Then we look at the overall technical risk. For AI companies. We have kind of a big selection bias. Like people who pretend they can solve the world with AI don't come to us because we know we can look through those and know what's actually realistic right now and what isn't. But you want to also not be too obvious on the technology and on the risk side and upside, you look at the risk of the industry, you know, especially in healthcare and biotech. And then we look at kind of first principles of like, where is the world going? I love predicting the future. Had a surprisingly decent hit rate on predicting various things. And I love kind of enabling founders to build that a future in sort of an optimistic and constructive way.

Interviewer (13:14)

How would you describe your ability to predict the future?

Richard Socher (13:16)

Often it's first principles like what can be done, what people would benefit from, what are goods and services again, that only a few people have access to, but more people would love to have access to. That's one. And then when it comes to AI is like, which ones of these are bottlenecked on intelligence. And then having a deep understanding of what the technology can actually do, and in some cases doing the actual research to push it forward certainly helps.

Interviewer (13:45)

I remember a famous management consultant went to state telling them to prepare for 4 million third graders and they said it's absolutely impossible. But there was 4 million kindergarteners. All you had to believe was that these kindergartners would age and become third graders. Sometimes first principles thinking is just kind of the obvious. If you just ignore all the noise around what people's preconceptions are about a certain topic.

Richard Socher (14:06)

Yeah.

Sponsor/Advertisement Voice (14:07)

When you want more, you start your business with Northwest Registered Agent. They give you access to thousands of free guides, tools and legal forms to help you launch and protect your business all in one place. With Northwest, you're not just forming an llc. You're building your complete business identity from what customers see to what they don't see, like operating agreements, meeting minutes and compliance paperwork. You get more privacy, more guidance, and more resources to grow the right way. Northwest has been helping founders and entrepreneurs for nearly 30 years. They're the largest registered agent and LLC service in the US with over 1500 corporate guides. Real people who know your local laws and can help you every step of the way. What I love is how fast you could build your business identity. With their free resources, you can access thousands of forms step by step guides and even lawyer drafted operating agreements and bylaws without even creating an account. Northwest makes life easy for business owners. They don't just help you form your company, they give you the tools you need after you form it. And with Northwest, privacy is automatic. They never sell your data because privacy by default is their pledge. Don't wait. Protect your privacy, build your brand and get your complete business identity in just 10 clicks in 10 minutes. Visit www.northwestregisteredagent.com investfree and and start building something amazing. Get more with Northwest registered agent@www.northwest registered agent.com invest.

Richard Socher (15:29)

Bre is certainly a lot of noise. Again, there's so much hype on both sides. Like some people think, oh, it could kill all of humanity. And no one can ever give me a realistic scenario where that actually is the case. And eventually they go down to like, well, it could hurt some people. And we're like, yeah, we should definitely regulate it when it actually is applied to real people in like the fda, you know, for food and drug issues and we have FDA trials and we have FDA approvals or in self driving cars. Right. Certainly they should be regulated and, and so on. So the more real and impactful AI applications become, the more we can and should regulate them. But some people get ahead of their skis and just say, oh, we need to regulate intelligence. And I'm like, oh, that sounds pretty dangerous. You know, you don't want to regulate math, regulate intelligence, regulate. Like, European Union has so many, like, really unfortunate regulations and laws and taxation ideas that destroy their entire AI sort of economy before it could even start.

Interviewer (16:19)

If you look at crypto as analogy to AI in terms of the evolution of a space, first you had obviously the currencies, then you had the infrastructure, then you have the apps. Is that how you look at investing in AI and that there are certain layers of the value chain that have to develop first? Or do you just invest into kind of best ideas with the most ambitious team?

Richard Socher (16:39)

The best ideas and ambitious teams look exactly at what is the right time. I often say, as a researcher, if you're right and ahead of your time, you're called a visionary. As a startup founder, if you're right, but ahead of your time, your company is just dead because people don't know or don't want your product yet, or something is not quite ready yet for really mass disruption. And so I think we are looking at what data, for instance, is available. It's a very easy way to predict where I will have a lot of impact is looking at where is there a lot of data? Or can data be very cheaply and efficiently collected in a space? And then you'll know that that space can be more likely disrupted than those spaces that aren't. So I'll give you a silly example to some degree, like plumbing, no one really collects a lot of data on how a plumber is crawling in some, you know, below a house space to fix a pipe. And so there won't be any AI plumbers anytime soon. And you know, like, at some point that physical labor will be so much more expensive than digital labor that then it makes sense to really get humanoids to really work that can crawl into spaces that previously humans were also able to get into and do physical labor and roofing and tiling and plumbing and whatnot. But there's a weird future where the prices of that might go up, up, up until it makes economic sense to automate it.

Interviewer (17:53)

You're competing not only on the US scale, but also worldwide. And you see this geopolitical rivalry between China and US in AI. Where do you think that plays out? And what are the upstream battles that are being held to determine who wins the AI race?

Richard Socher (18:09)

Anyone who actually participates in the IRA will benefit from it. I don't know if there will be like a single winner. I don't think it's a winner take all market, but it's certainly a non player lose a lot market. Right. And world where if you don't even engage with AI then you'll certainly fall behind. Yeah, it's not a zero sum game overall in human productivity and progress that is mostly pushed forward by AI. Now I do think China has some structural advantages of just hyper competitiveness. It's so brutal and so competitive in China as a business. A lot of folks aren't able to do regulatory capture the way we see in the US in more places, you see a lot more competition. China is and has been for the last three decades very, very good at taking ideas from the west and then making them cheaper, producing them at scale. And we're seeing the same thing first with trains and cars and now also with AI models. Once these ideas are out there, China is very, very good at making things more efficient. We'll continue to see that. We haven't seen many super exciting, extremely novel ideas that really change the field out of China. And maybe that is part of what they're not as well set up for like complete failure. That is likely if you try something extremely novel and out there, kind of out of the box thinking is, is more acceptable in Silicon Valley for instance, but also not everywhere else in the western world. Countries and places that allow for some failure to be, you know, recoverable in your career are more conducive like Silicon Valley are more conducive to people trying out different things. China has less of that. And so that is the fact that the US pulls in a lot of amazing people from all over the world that want to build that future and that want to come to Silicon Valley and sort of have this constructive optimism I call is unparalleled in the world. And that will continue to be a big driving factor for the US.

Interviewer (20:08)

I mentioned earlier in the podcast, you're one of the most cited NLP researcher, I think 230,000 citations. When do you think that we will achieve AGI and ASI, which is advanced General intelligence, Advanced superintelligence.

Richard Socher (20:22)

I sometimes don't even call artificial superintelligence artificial because there is no superintelligence natural or artificial. So we can kind of drop, drop the A from. From super intelligence. I used to be more cautious. I actually now think that depending on how you define AGI, we're already there. Like these models are quite general they can write you a poem and then they can tell you about your medical, you know, results and then they can talk to you about the Macedonian empire and Alexander the Great. And like it's, it's pretty general and it's pretty incredible. To get to superintelligence you need a couple of different things. You need to have either domains that you can simulate or domains where you can verify all the outputs. I'll give you some examples. Any game you can simulate quite easily is obviously solvable to a superhuman capability by AI, right? So we'll see these pockets of super intelligence already emerging. I was never that surprised that an AI will eventually play a game that it can infinitely sample from and infinitely play better than a human.

Interviewer (21:21)

But it certainly, because it's kind of locked constraints and it's able to constantly simulate without needing even outside data.

Richard Socher (21:27)

Then you just collect training data. You're like, I tried this. Was this good? Yes, no. And so if you can get into any state where like you try something and then you get feedback on was this good? Yes or no? Then you can the. I can now infinitely try things like, not infinitely, but you know, millions or billions of times. And so in a game like chess or go where you can like simulate everything perfectly, you have full information of everything, you can now try billions and billions of moves until you gain that intuition of like what is a good move or not. And so that is an example where AI is already super intelligent. Now where it gets interesting is the real world isn't. You cannot simulate it. And if you go out and you try to make some money or have a job and get a salary or something, no one will tell you, like this particular action just now made you more money in some places you can, in finance, you can, in math, you actually can, like proving things in math you can get the feedback of like, yes, every step here was provably correct. And so you got to the right place. So math will get majorly disrupted. And another really beautiful domain is programming. And a lot of people like to say software is eating the world, like to say AI is eating software. So as we can build verifiable programs with AI that will disrupt the entire digital economy in very exciting ways.

Interviewer (22:37)

You're known as an AI optimist. What's the down case?

Richard Socher (22:40)

AI is only as good as the people, the policies, the infrastructure and the data that influence it. AI is kind of like an omni use technology. It's like a hammer or the Internet. The Internet can be used for wonderful things and communicate with the world. And be connected to everyone and learn about things. But it can also be used to share horrific content of people getting tortured, right? Like, there's all kinds of horrible things you can do on the Internet, and we need to regulate those things, right? Like, we need to regulate really bad things from happening. And there are other industries, again, that need to be regulated with and without AI. Edison is another one. We don't want some crazy drugs or an AI neurosurgeon to just, like, practice and get its reinforcement signal in my brain, you know, while it's figuring things out. Like, you need to regulate those industries because there's a huge downside risk, personally. Also, military applications are kind of scary, right? Like, you don't want to have super intelligence kind of given the objectives and the goals of murdering people. I think, you know, that's a really terrible way to think about efficiency. And so those are areas that we definitely need to regulate because they have real downside risks. I think one of the most realistic negative scenarios is probably in biology where we don't want to create like some super virus or something, and instead we should work on creating some super vaccines that actually work the way, like, traditional vaccines work. You know, don't get people sick and keep them healthy.

Interviewer (24:03)

And why do you discount pdoom or existential risk of AI? There's many different versions of this. There's a paperclip problem, which is if you tell AI to create paperclips, it'll turn the entire world into a paperclip. But there's other edge cases. Why do you discount that? And how do you look at that risk?

Richard Socher (24:17)

It's really interesting. The paperclip example is a great one for a failure of prompt engineering and reward engineering. Ultimately, I think that will be a new kind of job, right? And you can already see this now, like we'll have in the next few years. AI agents for your enterprise, where you say, get my CSET score to be higher in my service department, just go make it higher, right? In the AI, if it has access to a huge amount of actions and different kinds of things, it can do just be like, okay, easy. I'll just create a million bots. They all call, and then they Give me a 5 out of 5 rating on my CSET score. And then, boom, I just improved your CSET score. But you're like, that's not what I meant. That was. I guess I have the wrong reward, right? So as I'll give you, I'll give you a better reward. It has to be done with real people. Then the AI will say, okay, real people maximize CS. Boom. The easiest way, I just give a $10,000 gift certificate to everyone after the call, and then boom, you get five out of five perfect CSET score. That is another example of a very poorly thought out reward. And so if you're this stupid to give an AI that is super intelligent the reward to only maximize paperclips, then you know, you're just really, really dumb. And like, you probably wouldn't be given access to billions of dollars worth of compute to actually accomplish something. And so you just have to like, be realistic. And as the technology gets better and better and could eventually have these like real life ramifications, people will also get better and better at defining their rewards properly. And that will be one of the many new kinds of jobs that are going to come into existence in the world. That's the first problem of the paperclips. The second one is at some point the AI also gets smarter itself enough so to question rewards and to question a context. And so you, if you really think an AI is somehow smart enough to be able to destroy all of humanity, get access to all the physical resources to build paperclips, but then at the same time think she's dumb enough or it is dumb enough to like not realize that if no one's there to buy a paperclip, you don't need to buy a build paperclips in the first place. You kind of assume this like very weird type of intelligence that is both ultra brilliant and ultra stupid at the same time. And I think that's also a very unlikely scenario or like basically zero.

Interviewer (26:24)

People ask me almost on a weekly basis, I have a son or daughter in college. What should they be learning? Used to be computer science was the answer. Now maybe it's the most subject to disruption. What's the best way to think about what the next generation of students and people early in their profession should be learning?

Richard Socher (26:42)

Computer science is still one of the best things to study. I disagree there with some other people that I otherwise respect a lot. I think if you understand the basics of computer science, that means you understand the basics of logic and math. We know that training AI with programming improves its reasoning capabilities. Why? Because same thing happens for people. Like when we learn how to program, it improves our reasoning capabilities. And then this whole technology becomes less magic and more like a program, a piece of code, something that you have control over, you have agency over, you can actually modify and make better and improve in different ways that you think is valuable to Humanity. And so I'm still a big fan of computer science. I think computer science should be kind of like math and physics in high school. Every high school should learn, should teach skills to program people.

Interviewer (27:27)

Take it literally. It's a way of thinking versus writing this code. Yes, the code might become obsolete, but how do you think, how you construct this reward prompting these new jobs that will come out? You still need the same skill set.

Richard Socher (27:38)

That's exactly right. And then I would probably recommend people to combine computer science with another passion, another applied field of application where AI can have an impact. That can be biology, it can be chemistry, it can be physics, can be economics. Right now we're making economic policy based on like oversimplified linear models that obviously are wrong and incredibly so. And so, you know, we published this paper called An AI Economist where you can build sophisticated simulations where you actually deal with AI actors that adapt to your policy, that try to circumvent your policy, that then you know like are intelligent themselves the way people are intelligent when they drag the economy. So many more areas, medicine, but even you know, history and like philosophy, like all of these areas will be impacted by computer science and AI.

Interviewer (28:32)

$250 million venture fund. You're running one of the hottest AI startups. You're doing AI econometrics. What do you do for fun?

Richard Socher (28:40)

I just finished a book on AI. I was right. It's called the Eureka Machine. AI for Science. I think that is really incredible. It would be a good superintelligence if all it does is create some memes and answer our emails. You know, our average intelligence can already do that for the most part. But in scientific sort of frontier, there's still so much more to go. And so I'm thinking right now a lot about ways and actually starting another organization to really think about recursive self improving superintelligence that can eventually not just improve itself, but then also improve our understanding of science in the world. So that's, that's really it. I don't, I don't have that many more other hobbies. I used to paramotor a lot. I love paramotoring, but I just don't get quite enough time anymore for it. I did get a couple of days this year where I got in the air in between work. But yeah, paramotoring is a beautiful hobby. It's surprisingly not as popular as it could be despite like enabling you to see the world from the most incredible vantage points.

Interviewer (29:36)

If you could go back to 2014 before you started MetaMind, what's one piece of timeless advice you would have given a younger Richard that would have either helped you accelerate your career, helped you.

Richard Socher (29:44)

Avoid costly mistakes, work as hard as you can until your health and your both mental and physical health kind of cannot take it anymore, and then you have to tone it down a little bit. So I've been doing that for a long time and certainly during the phases where I was most productive, that's, that's how I operated. And I think that's generally good advice. I'm pretty happy where I am, so I don't know if I have any like, don't do things you've done, but maybe I could have been have even more constructive optimism even earlier. You know, we invented prompt engineering but didn't scale it. So now is the time to, you know, have exciting ideas and really scale it. And I think the biggest thing we need to teach our kids outside of being intelligent, is to develop a certain amount of agency that they can, quote, unquote, just do things. Like there are a lot of things where you're like, that seems impossible, but actually it can be done. And sometimes you have to be in the right place at the right time. I was very fortunate. Eventually gotten into Stanford after multiple rejections and eventually got through Stanford into Silicon Valley and then be surrounded also by other people who have this constructive optimism that is still, in a positive way, infectious and allow you to think bigger.

Interviewer (30:51)

Richard, thanks so much for jumping on the podcast. Looking forward to sitting down soon.

Richard Socher (30:54)

Thanks for having me.

Interviewer (30:56)

That's it for today's episode of how to Invest.

Sponsor/Advertisement Voice (30:58)

If you're a GP with over 1.

Interviewer (30:59)

Billion in AUM and thinking about long.

Sponsor/Advertisement Voice (31:01)

Term strategic partners to support your growth, we'd love to connect. Please email me@davidisburgcapital.com.