Inside Benchmark's AI Bets | Eric Vishria

Key Takeaway

Audit one important plan this week as if AI has changed the underlying economics. List the assumptions it depends on—switching costs, implementation time, staffing, demand, and technical constraints—then ask which are now cheaper, faster, or obsolete. Do not merely add AI to the existing roadmap; redesign the roadmap around today’s capabilities and validate it with a customer. In fast-moving markets, preserving an old plan can destroy value faster than changing direction.

Episode Overview

Eric Vishria of Benchmark discusses what investments in Fireworks, Sierra, Cerebras, and Sunday Robotics reveal about the AI economy. He argues that AI will create multiple winners across the stack, but success requires technical depth, rapid iteration, high-quality data, and a willingness to discard operating assumptions inherited from the SaaS era.

Full Transcript

Transcript of Inside Benchmark's AI Bets | Eric Vishria from Invest Like The Best. Auto-generated from episode audio; may contain minor errors.

I am very skeptical. I agree with the idea that people will just " "vibe-code" your own [ __ ], that's not it at all. problem. The SaaS problem -companies in that every day when you carry out your plan, you are destroying the value capital. Only Think about it. All career we were taught: you are laying out a plan, you do it mercilessly and decisively. Are you following the plan or you overdo it, you continue to build —that's how it's created capital cost. And now it turns out that every time you you are carrying out the plan, you...

You're ruining everything. You You are destroying value. I like it every time we see you, ask you and your partners : okay, you have these unique investments . You are doing it wrong. a lot of investment in year, and those who " "shoot"—as, for example, Sierra and Fireworks —of course, they work . You will find out like this a lot about the world through through the prism of these companies. Yes. So I would with pleasure talked about both. Maybe we should start with Fireworks. What do you know?

or what you learned about how the world changes, watching him through the prism of Fireworks, which it could be strange Is it interesting? These are models for two, three, four trillion parameters. Well, it turns out, to launch such models are the [ __ ] hard. And launch them effectively—it is It's really complicated. AND it is obvious to everyone, The point is that...everyone: from AWS to Azure, GCP, neoclouds and companies like Fireworks or Together—they all launch these standard open source models available in their pools developers, etc.

AND it's a fact, the difference in productivity between Fireworks and clouds provider is about 5x. And that's just speed work. Add to this pass ability that is not visible from the outside. It visible only if you understand economics this business and you think: "Wait a minute, it's the same open-source model on the same Nvidia hardware, and the difference in productivity 5x and many times bigger capacity" . And you know, understand it's simple: these companies pay margin cloud providers, work on top of them and while earning money.

Yes. So how is it? perhaps? And therefore my main conclusion was like this: wow, run these things actually very difficult. It's simple. really hard implement. And, um, for this it takes a lot expert knowledge. It kind of very specific expertise, and that's it what, when you look like an investor from the side—and we should to talk about the first AWS days—looks like wide product consumption. Do you think that it's just a resale, a game on a scale that whatever, and then— wait, no, it turns out that this Not at all.

What do you think, is this just moment in time? I wanted would like to hear your thoughts on cloud technology. You are very keep a close eye on cloud; you talk a lot you know, about the curve implementation, about, like people use these things, and about nature these two businesses in comparable. Temptation arises to say that there will be one or two winners for on a scale like this usually happens on goods market. Where is the cost? service is everything, but the scale is decreasing her, and this is supposedly all history.

I just I think the AWS example very accurate. So, 2006 year, launching S3 and EC2, Yes? Their computational platform and platform for data storage. It the first two sentences AWS in 2006, and about this start talking at the end of 2006 or 2007. It seems that in annual report for 2007 Bezos year many talks about AWS, why this is important, interesting and this is different, but the reaction investors was not very much at all. I think if you gathered in room 30 the smartest investors of that time.

and asked them what the probability that this AWS business is good, has a stand long-term margin, is extremely interesting and not at all goods. I think the score would be zero out of 30. Among the truly smart people who we know what were around in 2007 year. Okay, let's move on from 2007 in 2014. I joined the venture business in 2014, and very a common thesis then was, "Oh my god, AWS "will eat everything." For corporate business is not left opportunities. They will absorb databases and infrastructure, and also all the apps.

AND they will offer this the cheapest and best. And you know, must remember that in we all had these great SaaS projects and software businesses, in which we were involved in did you invest? AND partly their people loved because they were annuities with gross with a margin of about 85%, as and everything else. So there was a feeling: oh oh my god, AWS and Amazon can to offer this with with a margin of 8%, and they just destroy this possibility. Yes. You- my opportunity. It everything, the whole story, if to think about the period from 2014 to date in corporate sector, you understand?

Of course you have Snowflake, direct competitor to Amazon Redshift, powered by Amazon . You are superior to Amazon on Amazon itself. Okay, but that's not it. only Snowflake. Were Confluent, Elastic, [ __ ], Databricks, all these companies. Okay, surprisingly. This is the level infrastructure. Further there is a whole level applications. At the level applications, remember the offers they had on the very beginning. Datadog— today it is hundred billion company. They had competitive proposal, right? AND they did it. AND, of course, there were many " victims on the road".

There were many victims. They really crushed many companies. But even then, in 2014 , the thesis that AWS will eat everything, was fundamentally wrong . Not because of all those examples that I just led. She was completely wrong, because Azure and GCP then were irrelevant. And if we move to The year is 2026, then this incredible business. So is AWS the biggest? now? I think it's somewhere. 40, 30, 20. As a result you received oligopoly. And even outside this the big three are Cloudflare, another type cloud provider, another hundred billion company.

So such appear smaller players who become hundred billion companies outside of them. What is the conclusion? The conclusion for me is because many people think in categories zero-sum games and does not realize, how big what if everything Did it work? What if everything Did it work? And it all worked. Of course, it is important. guess the relative the winner, were victims, there were companies , so all this has value. I'm not saying that it's simple, you know, " "random shooting". I don't have that at all.

attention. I'm just saying, that the market was like this huge, that one the vendor could not scale and to swallow it whole, and they just don't could absorb all industry. Now, again, everything otherwise, and all these things, but if you just take this concept that you and I see in AI , it really looks like history repeating itself, it's like Anthropic is going to do everything Yes, that's true, is this, is this really right for me? A view of what, saying, "one "the company will eat everything" , does not stand up to criticism.

And I will tell you what the clouds scaled very quickly, but not as fast as that happening now. So, for the company to really scaled and achieved success, scaling requires a huge development infrastructure, truth? From energy, capacities, you know, chips, memory, and, well, of course, algorithms and everything else on top. So it seems to me that we will get oligopoly winners, and I I really believe that these crazy people will be hundred billion smaller winners. So this, you know, it seems like what is happening exactly.

Can you tell about it with demand side and compare with how you watched cloud implementation in corporations and how currently being implemented AI? If you return until 2010, 2011, or even three, four years since the appearance AWS, corporations were very skeptical, simply incredible skeptical, traditional "blue" chips. You had digital companies; new ones appeared companies that used cloud. You know, seems to be famous the fact that Snapchat was built on GCP, and, it seems, in some way moment in this period, maybe 2012-2014, Snapchat occupied 40% of GCP.

Here such things and took place. It's like a Cursor that takes up 30% 100%. 100%. What happened? exactly. Corporations were very skeptical about clouds, until, you know , until 2014, 2015, 2016 they didn't say, "Oh, "Yes, exactly." You know, then the big banks, financial services, insurance companies and conservative corporations thought: "Oh, wait, that's actually different, and we will have to pay "Pay attention to this." And you know , this became a problem when hiring for them, because they couldn't attract the best developers, because they wanted to work for the most convenient platforms, and it happened many similar things.

So now the difference seems I really significant, because although it It is 100% true that AI in big business does not is digested and not is being implemented properly, it is not the same as going into Cursor. Compare atmosphere in Cursor with working in a big new -Yorkshire financial company regarding the use them AI. They want this. They want deal with this. They conduct experiments. They spend on it funds. They are trying this to comprehend. They They talk about it. They do not belong to This is disrespectful.

I think they are considering this, probably as a larger opportunity than yours time clouds technologies, considering on potential impact on their business. I think they probably see in this greater threat than clouds, and perhaps they just learned cloud lessons technologies regarding What is possible here? Therefore I I feel that they are in the future will try to figure this out and adapt. Despite all that has been said, I think one of the most interesting things, if you go beyond the limits Silicon Valley and talk to these corporate companies, there are understanding of what there is a pace of implementation, barriers and more—I I think there is huge opportunities .

Help enterprises to succeed—that's what what I advise companies with whom I work: gay, let's stand their AI conductor. Let's become their AI- conductor. If we we will find ourselves in the role such a conductor, combining both worlds, this will be very valuable. And I think this will continue to work. Ramp is the only platform, created to make your financial team more efficient , faster and better, saving business on average 5% annually so that you can to focus on growth. Ramp Clients increased revenue by 3.2 times faster than average American business .

Visa, Vercel, Cursor, Stripe, Notion, ElevenLabs, Shopify and 70,000 more companies work on RAMP. My company too, and yours has. Learn more at ramp.com/invest. The best AI companies and software software, from OpenAI to Cursor and Perplexity. Use Work OS, to get ready for corporate level in one night, not in months. Visit works.com to bypass boring infrastructure work and to focus on your product. Felix from Rogo—this personal financial agent, which transforms one request for ready-made client's work, using templates, context and your standards companies. Send Felix email to like: "Take these comments and work them for me".

Or: " Update my tracker, considering the context these letters". Or: " Calculate solvency this buyer", and Felix will send ready PowerPoint presentations, Excel models and selected research. Felix works the same as your team, performing tasks quickly and accurately 24/7. Learn more at rogo.ai/felix. AI/Felix, what? Sierra teaches you about implementation of such tools? It really interesting contrast with Fireworks, because Fireworks is its own kind of supplier infrastructure. Sierra trying to find really cool solutions for end users consumers, giving opportunities for companies to help customers, starting from service.

But now that we we are talking about Horizon, again, the same thing question, what about Fireworks: what do you know? something that the world doesn't fully appreciates, with considering how you Have you seen this business? Well, that's great. example. First of all, my partner Peter and Brett, it seems, are working together already in third company. So this is a 20-year relationship, what is simple incredible and excellent starting point point. One of the things, which, in my opinion, does Brett special, is that he and his team— technologists, but they actually lived a long time in the world corporate business, so good understand him features.

And I I think he embodied this idea: "Let's we will be their AI- conductor". Let's start from maintenance customers, where everything easy to automate , and now we have these long-term agents in Horizon, which can to do more and more more things. And one of the most interesting things for me is because, yes, on at first glance it is application company, but they do real work with artificial intelligence. They are experimenting. with models, build agents, being very close to the basics models, their capabilities and systems, and understand "ragged" the "edge" of AI possibilities, which is very differs from smooth human the arc that we we understand intuitively .

They understand this " torn edge" opportunities and building solutions around him. You saw the evolution of Cursor from IDE to autocomplete tabs and agent work, again and again. They were doing their thing. previous work outdated, like half a year ago, you know, this is what Brett Taylor calls sandy locks, right? Earlier we built castles. Now we are building sandcastles, which will be washed away with water. And that's if you don't perceive the program ensuring exactly Yes. Yes. I need this. accept, because if you craftsman and you say: " Hey, I built this.

perfect foundation a brick castle, it was perfect, I really am I take care of him, and he will stand for 100 years", then you just don't You will cope. So, because in the models new ones appear properties, you get new ones opportunities every four weeks. They understand this " serrated edge". They understand application of this " serrated edge" or gaps for your customer base, fill these gaps and adapt them. Not you can just superficially to apply things. I I think it's complete. inversion of how previously worked product development, compared to how it is happening now.

Product development, you know, it always was like this: "Hey, product- manager, you taught: "Oh, the product- The manager should not you know, you have to understand technology, but really shouldn't to think about implementation and not should determine implementation, not must do that and se". And the work of the product- manager is to so that really understand the customer and translate this language problem engineers so that they created a solution. It was traditional product management . Well, good luck to you. do it now. It terrible way work.

You can't to do it this way . You really need to understand nuances of possibilities models. What are they? wonderful, where are they they fail. And to you need to understand customer problem and combine this, to overcome these gaps, to create valuable decision. I think this is one of the things that Michael and the Kurser team did very well from the very beginning: they really are understood these " "geared" capabilities and created a product, which allowed to carry out this transition from developer to such opportunities, constantly improving it in as borders change.

It's ironic that all this sounds like return from technical knowledge is growing, even despite that the models supposedly leveling the technical advantage. Well, I, yes, I think, here there are two things. Yes, 100%. And actually I think, what's happening a strange thing when all these were happening discussions about traditional roles product manager, designer, engineer and something like that. I think that actually exist only models. people who understand the problems customers, people who have taste, and people, who understand boundaries AI capabilities and are interested in them.

And these are three things that they will do. If you have these three quality, you will achieve success. And it doesn't have meaning, are you engineer or manager product, or designer , but if you have taste, you know customer problems and the limits of AI capabilities, you will be successful. When we first they did it so much years ago, it's just madness. It would be interesting to watch some ideas we have spoke for the first time regarding SaaS. But you used the term, which I use from that time, you called it competitive border, that is, things, which will determine winners and those who loses.

My partner Bree has a great idea, which has settled in me to the head: now everything resolved in a fair manner fight. Yes. And I really interesting, except for the idea of sandcastles vs. real castles that you still see among people who become winners, are there any difference? Or different features winners now, with points of view personality, business strategy or business models, compared to what you learned in the SaaS era? I am very I am skeptical. to the idea that people will write code "on" "feelings" or something similar.

This is not at all This is the problem. It no problem at all for SaaS companies. The problem of SaaS companies is that competitive frontier completely shifted, and everything they relied, believing, What will it bring them? victory, no more works. Let's take the bases data, because they for a long time there were phenomenal niche for software software. Wonderful profits, because application developers built solutions on based on specific interfaces specific databases. Over time, you became more and more data. So migration applications from one database data on another was a huge project, which was very difficult realize.

Exactly so these were incredibly stable businesses that brought huge profits. Of course, in you were Oracle, SQL Server and a whole series of smaller ones players who have achieved very great successes in databases. Well, let's think about this in the context of AI. Now the developer has already does not base the decision on interface-based databases. Do you have cloud solutions or coding models that working with database interface data. First, beauty database interfaces data in that they very well defined. It turns out that AI is very copes well with things that are clearly defined.

Three agents are not get tired of monotonous work with translation of one specifications to another. So it turns out that now suddenly database migration, which was or was not before the main thing that should not be done in software provision, became quite trivial. Yes. Yes. Just invest in This is money. Move it. So what has changed? The criteria have changed. of what he does database company outstanding. This is not means that we don't databases needed or that everyone will build its own database or something similar.

This is not what will happen. It will happen what has changed criteria. Now you have will be much more applications that will be launched as we see it everywhere. People will be much more experiment, because experiment much cheaper. Much cheaper start a new one addition. So now you need bases data that scaled from zero use to full load if everything works out. This has a lot to do with more important. Your the cost has much more value. Do you want to be able to to launch these things, turn them off, take them apart again and again.

So speed iterations of what you require from the base data is growing. Ultimately, I believe that the cost becomes decisive factor. Therefore, cost, and also this scaling from zero to infinity, mobility and everything other things around this become criteria the one who wins, and who is not? This is really differs from when they say: " Hey, I chose this base data for our user-defined product, purchased license and launched her on such and such equipment". Of course, elements of this were always, but I think that just changed criteria.

And I think, which is one of the main messages for these SaaS- several companies years ago it was like this: "You have a choice: go to AI or cost three times as much more than income." And I think a lot of these public SaaS- companies now bargain with multiplier six or so that. But not forget that in 2021 year they traded with multiplier 30. Isn't that right? Almost everyone, truth? And here you have it. triple value. Oho. I have grown 4 times since then. time, right?

This is it. all this compression multipliers, this just horrible. I grew up in 4 times. The multiplier has fallen. six times. So I I cost less, although I grew 4 times in several years and went out on breakeven, and that's it like this. they said that this was the first message. But I I think that even for I felt completely It is obvious that every day, when you perform your plan, you destroy the joint-stock company cost. Just think about it. it. Our entire career taught: you compose plan, you are executing his mercilessly and hard.

You perform or you overachieve your plan, continue to develop, and that's it this is how you create shareholder value. This is what everyone was taught management teams . This is what happened before taught generals directors. This is what What was everyone taught? AND now you are getting in a situation where every day, By following the plan, you You're ruining everything. You all you are spoiling. You are destroying cost. And the point is, to tell them this, consisted in the fact that give them freedom. Other explanation of this from my friend Anne Lee Gates was that that general directors who passed through this transition period and had business with turnover in hundreds millions, thought that they are already working on with your business from 8 morning until 5 pm, and then they try to do AI from 5 to 8.

in the evenings. And what they had to do, this is the opposite, it was quite vice versa. And this is very difficult to do because all training, muscular memory, inertia and everything that we learned; we all of a kind we climb to hill. And you know, the model of success was like this: make a plan, to carry out the plan, increase the value, capitalize cost, and here: "Oh no, no, no, stop, you must completely everything change". And I think that still a very long way to go to answer your question about what is the profile or mentality today's winners.

But if we will look at Brendan from Mor Lynn, or Fireworks, or Max, or Brett, on any of them, they are so flexible regarding the assessment results. For example, over what they are the ones working, to optimize process? They so flexible in all this. And if look at each of these companies, business evolution is happening constantly. And they did a wonderful work in this regard. And I think that's very, very different from why me taught. What do you have? feelings about disorienting the nature of progress models and how we ...we talked about it before that every four weeks a new one appears " sharp edge" with which need to figure it out.

You are a little out of place. this as an investor, on unlike technical the founder, who, you know, it works directly "in" poly". How do you act? differently than, say, three years ago, with considering this pace? Every time I I am talking to founder of the problem in their companies, about their actions, steps or anything other—and this is what 80% goes to of my time. I am very I often say: "This is how we "They've done it before." M-hm. This is what we usually do did.

Here's how usually evaluated such a situation. It old approach to explanation why this candidate is better than that. Now let's let's reevaluate this in in the context of today . Let's reevaluate this is in the conditions unstable technological foundation. Let's let's reevaluate this in business context models that develops like this, and no different. And I started to question every assumption and lesson learned before: what of it what's relevant and what's not. AND that's a huge difference. I will give a specific an example that has become quite destructive for companies that scaling AI.

In these companies that scale AI, a lot came leaders previous generation with a big experience, and they completely failed —this is visible throughout industry. So there is Question: why? These are the only ones of the best leaders four or five years old antiquity. They learned all the lessons, they are wonderful, but for some reason it doesn't work, and there is a certain gap between needs AI founders, business requirements and what these bring people. I saw this. very sharply with one head of department sales, which we hired, and he said: "We can't to do these indicators, because the model software sales for decades was based on quotas ".

You have a model, based on quotas. Each representative does this. On initial stages company quotas make up, I don't know, 1.2 –1.5 million. Maybe internal representatives have 750 –850 thousand, and over time indicators are increasing, and for corporate sector is already 2.5 a million, and that's all; That's how it works. Exactly that's how all these are built financial models. You start with a model quota. Apply achievement discount results. That's it. everything we can make. Bam. Bam. Boom. You know, essentially, not even realizing that everyone implemented something that was based on stimulation demand, not its involvement.

And for many of these companies that work here, for these customers, appears new product on based on AI, and it's simple , tin, magic. So, these companies They sell magic. Well, real magic. It turns out that if you you sell magic too first on the market, you will sell a lot more than 2 million. You will definitely sell. much more than 2 millions. And therefore all this concept model quota potential and her work is not that so that she doesn't have value. She has a certain value, but this is definitely not primary factor or restrictions.

And here it is these are coming leaders with all this. Like, here are ours calculations, here division of territories, Here's what we'll do: first Western coast, then East, then center— and so on. And you are like this: "Oh, no, no, it's It doesn't work like that at all. ". And one of the main discoveries that I began to realize, communicating with these people—that's what you have to leave everything the old one is at the door. Just leave it all behind. . What is probably good practice in in any case, but throw all this away baggage.

Forget everything, what did you know, and learn it from scratch, from first principles. How is this? Does it work at all? Which actually exist bottlenecks in deliveries? How narrow are places exist in in demand? For it turns out that in many of these companies have sales reps , which make 10, 20, 30 millions. I recently saw 50. That is, you understand, and this means that, okay, it turns out that this a completely different story. What does the best do? the seller you seen which one works completely new?

Honestly, best seller in any of these companies—this founder. And what they do—this combine these "sharp "corners" of technologies from real client capabilities . That's all. And this It sounds so simple, but that's not true. But that's what they are do. And, you know, the market is simple huge, just like we talked about cloud technologies. I I think the biggest a mistake that everyone did—they just underestimated the size market. And it turns out, the market is simple incredibly large, and this market is even bigger . It seems that this is the moment, actually just today in the morning, we are with you in one big one group conversation, where discuss demand, a kind of demand for intelligence, which, not I know, it seems almost boundless.

AND it seems that what getting smarter technology, the higher demand, and perhaps the only bottleneck— this is capital. The world seems feels the need take a breath. Like the RSI concept, if you apply it to technology, they don't you need to breathe; agents are not get tired. But the world as if to say: "Oh my God, it would be nice to have at least three months to just a little bit of all of this to digest, to capital was formed and assessed his prospects", and the scale becomes so huge.

What do you think, whether can this go on forever , do we just have money will run out, which can be invest, because it seems that we could consume any quantity funds to build anything and to provide any What is the scope of the conclusions? What do you think about market? I I I will say this: me Energy is a concern. If we consider models as a tool transformation calculations on intelligence, well, simply, What do models do? They are very effective transform computational power on intelligence.

What is the demand? on intelligence? Well, It seems that very- very big, right? So it is logical that we we will continue to have an increasingly large demand for computing. I mean calculations in in a broad sense, not only chips, no storage or something else. But what do we need? for calculations? We need energy. A lot of energy. Now a lot are talking about distillation, Chinese software provision with open source and other things. But for What is more important to me is that, I think it's China.

next year will put into operation ten times more energy capacities than we have in USA. If energy is what is needed for calculations, and this bottleneck, and demand on intelligence unlimited, then It is quite logical that if we have significantly less energy, then we will have less intelligence, less tokens or significantly more expensive tokens. AND if we have much more expensive tokens, then by law supply and demand there will be fewer of them. And this seems very bad scenario. Therefore, in my opinion, problem energy deficit—whatever it is was not—will appear twenty different ways.

You know, gas prices the turbines are oscillating, on natural gas too, as in the sun energy, rare earth metals—for me it is, probably more anxiously. And if to look at from afar regulation or government, it seems to me that the administration does certain steps, you know, promoting investment and development of the whole energy. That is, I mean everything: solar, nuclear, gas energy, which whatever—do everything, we should develop all. Yes. And that's it. will get better, you know, but then again, this one of those things where, Of course, one thing will be a little better than the other, and I don't know, and I don't so smart, to predict that exactly, but everything will work.

By the way, about calculations, I would really wanted to hear the story Cerebras. I didn't hear you. told the whole story version. I think the reason why I I ask, is it because I am deeply I am interested calculations. I have large investments in computational power. And I'm passionate about it. It simply the most magical a thing that can be watch. It It's amazing when you see it. up close, what people were able to achieve on these chips and in these systems. Me it seems you invested around 2016.

So I think that's it. was your first an attempt to invest in super complex "hardware". This is incredibly difficult. And I'm glad that now The world is full of such people. opportunities, while it used to be real isolated case. Teach me everything, What did you learn about? investing in " iron" thanks to Cerebras. Mostly this It's really hard. For I think it's great. an example of what need a certain naivety. So, you know, the company appeared in 2016. It was five. founders and presentation. I I didn't really want to.

to go to this pitch. But it was my job because I thought: why should we invest in " iron"? That's it madness. It's been 10 years since our last investment in semiconductors, but the team was wonderful, and the first the slide was about what GPU actually bad for deep learning, they just accidentally turned out to be a hundred times better per CPU. And as soon as he does it said...yes, it is necessary remember what it was to the era transformers. Then OpenAI was simply strange research laboratory.

Nvidia cost about 40 billion, not 4 trillions. TPU not yet announced, nothing this has not happened yet. It it's still very early, but the idea itself—as soon as he voiced it, I immediately thought: "Damn take it, of course, Of course". Like, why? I have been for the last 18 months tried to find out areas of application deep learning, studied the issue security, medical visualizations and a bunch everything else, thinking that there is something here to be something that will change radically thanks to this. And here it is we are passing through this whole, well, journey.

We finally we invest, that was simple incredible. We are the first met on Wednesday, on Monday was meeting of partners, and there are still a lot between them conversations, and this helped shape confidence that this a great chance. I will say you that I understood, and I actually understood. very little, but There are generally three things, which we still know how to do so accelerate deep level training equipment. Increase quantity kernels, improve connection between nuclei, bring the memory closer to computational capacities. These are three things.

That's all. It the only three dimensions that we know in hardware part. So mine vision of what they are I was told—and honestly , all I understood, — Let's prove this. all these three things to their logical maximum. You you receive wafer chip on that time. It was on it b 450,000 cores. You would have about 20 gigabytes SRAM directly on the chip. So you never will have to to contact external memory. AND because they are all on one plate, connection between nuclei maximized. So this is the best that can be done was to do in that process.

AND I think the first one the chip was seven nanometers or something like this. And you are like: " Okay, that's all." "This is what what we do." It turns out that in software provision, if in You are so logical. diagram of why everything works, you are already 80% near the target. And it's just a question market entry and implementation. And in hardware you only provide 2% near the target. There are such things like physics, the whole this chain suppliers, and you you know, there is, of course, TSMC, which everyone knows about, but it's not just TSMC.

THERE ARE 30 others suppliers who matter, and all this needs to be collected together and so on. I didn't know all this and, honestly, not even guessed. So, let's move on from 2016 in 2019: they received your first details, and the process begins start, start by launch. Oh, yes, we are. received the detail, and again: launch, yes. Then you have to go through because of expectations, launch, there's somewhere around 14 launch stages and everything something else. And somewhere before In 2020 we already had your first work thing .

And then comes this one a long way to go to force it all work. One of the lessons I learned in this matter: you are passing through all these simulations in equipment, especially in semiconductors, and this becomes your "ceiling" ". This is the maximum productivity, It won't get any better than this. limit. And any software software, reality, compilers and kernels only take away part of this " "ceilings". You can start with 10% of this boundaries, just start system, and then the boys for months and years are working to to get closer to her close up.

This is completely other. This is very difficult. I actually surprisingly optimistically configured. If look back, then part of the reason our investments was what we saw the previous four generation computational technology in my life. There were processors, graphics, networks, mobile technologies -every time a new one appeared load. Were universal calculation that loaded central processor . There was a mass parallelization, which led to graphic creation processor. Graphic the processor provided massive parallelism , which led to development of graphics. Then for networks needed chips from ultra-low latency , and they got them.

AND for mobile devices needed energy-efficient chips . And in every case we were getting a new one company worth in one hundred billion dollars. And the first question is still In 2016 it was like this: there is artificial intelligence equally significant new load? Since there were many attempts to create specialized chips for another, that actually nothing influenced. Perhaps, there were some good ones results, but they didn't have much meaning, right? That's right. So we thought: "Well, okay, for this really huge is needed workload" . Okay, that's the first one, and here we had a big confidence.

And then the second was character working load: or it created new restrictions or problems? And what I have learned, was that AI -mass loading benefited from GPU parallelism, but GPU did not decide interaction problem between the nuclei, i.e. the problem of layers networks. So you are like this: "Okay, wait a minute, It turns out there is something new. limitation—problem communication speed". AND you think: "Okay, or is AI a huge new working load, for which are needed specialized chips? "You know, everything I just said, that was it everything I knew at that time time, and I really think that it is obvious confirmed.

AND, you know, in every previous we are generations received Intel, Nvidia, Broadcom, Qualcomm, ARM, in each of these generations had their own giant winners , independent players. Of course, TPU itself is winner, Trainium— winner, you know , there were Groq and Cerebras etc., this fight will continue in the future. And you know, I think that it will be massive, and I actually think that a new one appears, sixth generation. The generation that is... And I am very excited. , we did investment, about which not yet announced, but I think it's the first time for a long time there was a place for a new approach to CPU.

What is happening now—and this reflected in the value of all shares semiconductor companies—this is an LLM that working on accelerators and GPUs, generate code. This the code is executed on the CPU. And now he works on classic CPUs that have existed for eternity. But the CPU is loaded a whole series of restrictions, and they pull all this load of the past, which, maybe not anymore needed. So I, actually, very inspired by this opportunity. Next category. Next category. Does the experience make you with Cerebras to invest in companies much more?

Damn two. Why not? Well, let's say, 2019, we are on council meeting directors, and this The thing is melting. She simply, tin, melts. Okay. We attracted 500 million dollars or so that, and I am like this: Wait, what? Like, what? Is it melting? Is it burning or something? We we look, and I'm like: " "Oh, my god." You know, I understand that 500 million dollars —that's not true anymore a lot nowadays, but I just thought we we're going to lose all this money . That it's just not will work.

And, you know, we discussed this with Andrew, what you did this team. This is madness. It madness with from a technical point of view. Development teams iron arranged otherwise. They are arranged otherwise. They so different, and I I feel huge. respect and gratitude to them for what they done. Listen, Jokes aside, I I think there are things that Are you really proud of how venture capitalist, because you are financing what something matters changes. For me investing is only way to work with companies, not because that I fundamentally love the process itself investment.

Me like to work with companies. This is mine. favorite part work. And work with such teams and companies over such grand ones ambitious projects —this is something special. I I've said it before, even before 2018: regardless of whether did Cerebras succeed or not, it's an effort worth it venture capital, this is exactly what it is worth engage in. Worth it to try to create what people are talking about worked for 50 years and couldn't, but now we believe that we can, and for this there is reasons, application and everything other.

I like it. And that's why I like these things, jokes aside, and we have a company with robotics. We have there is also a defense the company that hired me very fascinating. And more this processor project, which we are talking about. I think that such things actually very interesting. This is good. using venture capital, but it's definitely not easy. I like that " productive naivety", about which you said when this even the advantage that you didn't know anymore because otherwise you wouldn't do this done. Yes.

This is definitely mine. experience with Etch: when we first started ask experts , in any of these spheres, they all say one thing: don't do this. Yes. This sucks. It too hard. Basic the rate is too much low. Young people can't do this make. Ahem. Are there areas where this is already too difficult? Or it could be bio- sphere or something like that where you just don't do you want to invest, if you are naive. Actually, mostly all of these stereotypical the statement is correct. They are not correct.

only three times out of four . They are correct. nineteen times out of twenty. Maybe ninety nine times out of a hundred. The thing is, Bruce, one of ours founders, always says that we have to ask yourself what can it go well, and whether Do we see this path? AND like, yeah, you know, young people cannot create chips, or you shouldn't open another one semiconductor company, shouldn't it to do this or that. All this is actually quite right, except in those cases, when it is not so.

AND, There is probably a saying, which someone said one of mine partners: if nothing will work out, then precisely for those reasons, what did yours say partners. If everything succeed, it is because that those reasons were not no meaning. Vantaa automates security and compliance for over 16,000 dynamic companies like Ramp, Cursor and Harvey, ensuring their audit readiness 24/7. It platform number one for agent trust, which helps such to companies like yours, track risks, arising between audits of yours suppliers, AI- tools and in the entire environment.

Every new instrument on which yours is signed team, everyone supplier that enables AI functions, — this is an opportunity for failures. And most security programs are not were created for at such a pace development of AI. Agent Vanta running in the background mode as 24/7 GRC- engineer, finding problems while cooking correction and reducing time assessments suppliers by 50%. Regardless of whether Are you a startup that growing rapidly, or global corporation , Vanta helps you to obtain and prove its reliability. Invest like the best listeners. Get a special $1000 discount at vanta.com/invest.

Ridgeline is first comprehensive accounting system with built-in AI for investment companies, which allows to manage a portfolio accounting, reconciliation, reporting, trade and compliance on one unified platform. Companies refuse to obsolete technologies on Ridgeline benefits through how far away Ridgeline AI Features ahead of everything other software provision for management investments. I I hear a lot from investment managers about AI, and they are divided about two camps: some not sure of what start, and others convinced that they can create your own management system orders for weekend. Reality is that management investment firm always requires control, management and a single source truths for your data.

And none enthusiasm for AI is not cancels this requirement. Ridgeline is built specifically on this foundation, therefore I believe that companies that will break forward in the AI ​​era, there will be use a single Ridgeline platform . If you are serious relate to AI strategies in your company, Ridgeline should become part of this discussion. You you can order a demo version at ridgeline.ai. You and I, I think, both we are interested robotics. Yes. This is not contradictory: if robotics will work, it can overshadow everything we we are experiencing now.

As do you think it has to happen, so that this Did it work? It, Of course, it's exciting. I I want it in my house. the robot was folding linen. This sounds great, but this is one of such classic themes. Before that, always 10 years, and it continues like this for a long time. What, on in your opinion, will be occur? What has to happen, to make it happen reality in the nearest or medium-term perspective? Classic robotics exists forever, she is everywhere: on assembly lines, productions and in places where are being carried out repetitive tasks in a controlled environment.

Recurring tasks in a controlled environment—it is more- less resolved problem, and that's will develop further . But for execution various tasks in real conditions for you AI is already needed. Here AI is needed here together with robots. AND the main difficulty in because for this appropriate is needed model. Of course, the first the problem is because there is no data internet scale for training the system, isn't that right? Languages models are trained on data from all over the internet. It huge volume human knowledge, and analogue for robots simply does not exist.

People are trying use different approaches: video, simulations and remote control. THERE ARE many different ways, but if you take remote control as example—how many need to manage manual labor to collect data Internet scale? The first step is obtaining quality dataset for model training, and I I think one of conclusions are as follows: in there are many on the internet "garbage" data. Even before AI started generating everything this, it was already there many unnecessary data, and some of them were more valuable than others. You can to consider certain things Reddit is more valuable than others ; you can appreciate Wikipedia is more than other forums; you you can appreciate GitHub more than others resources—and all companies that develop models, That's exactly what they did.

They prioritized data that was more or less valuable, and let them through model. So I think, that one of the most interesting things, who do these companies in the industry robotics and AI, this is the strategy: let's start of the most valuable data. If we focus on high-value data, we will be able to in a certain way run this model. And when you do that, thanks to the previous you will be able to study use very small auxiliary datasets during final training, and suddenly it works.

This is the magic of the great language models: you you have a huge previously trained base, and then add a little magic for through training with reinforcement. You train the model new, which was not in initial set data, and move further. And I think the same thing takes place in robotics. We are investors Sunday Robotics, company that operates exactly according to this scheme. This is a cool company, they are engaged in household chores, but the key point before going out household market technology is not about in the works themselves, and in so, can you fix this learning process.

One from the lessons I learned by looking at history, consists of because she is not repeated, but rhymes. If you take autonomous vehicles as an example—this, by essentially, robots, this is AI plus robotics. Look at Waymo and Tesla: both companies were designed as vertically integrated systems, each in their own way. You you have a Tesla, you had a whole park Tesla cars, which people owned. Of these cars data was collected that were used for training models managing them. That Same with Waymo. And I I think that one of lessons is that that they collected very high-quality data.

They spent previous training, and then worked on its basis. It was simplification, which allowed to create finished product or a decision that, of course, will be improve and will eventually become universal, so you you can install him on any car. Therefore I I think it's the same. takes place in robotics, where companies like Sunday and others, use vertical methods robot integration, models and data collection around him. Sunday uses gloves that developed especially for hands work. Therefore they They work perfectly. So, you get very good transferability data from one to another.

You spend this is previous education and have great data set for learning. You you have a good model, and then add these examples, URLs, and conduct additional training , and you get a cool behavior that occurs . Did you have any? most impressive moment? When we first invested in Sunday, my partner Peter organized demonstration. We went down to the basement laboratories Stanford, and they have was just like that a bad thing with cardboard in the form of gloves. And you see several stages of it development; and lastly, when we saw demonstration, we went down to the basement their current office, and there was about a dozen robots that just made up any linen.

It wasn't, it wasn't demonstration. It was just a trial and error method errors. And then people took clothes and measured it to make sure that he properly composed way, and they created a strict comparison base and evaluation criteria, and I thought, "Oh my God, this is really happening ". Do you care how choose the right one client for these companies? For example, each of these devices is underwear, but I don't think so what about someone I like to compose. linen. So this is seems good option use, but it looks like we not really we understand the demand for that that these things will be do.

Doesn't bother you that we create decisions that are then will look for problems? Me neither, and I I will explain why, and this, certainly, reinforced observation of the evolution of LLM, where, as in me, some people who were involved in early LLMs in OpenAI, really understood that the code will be important, and, of course, the Entropic team had this view that you could achieve AGI, if they fixed it code generation and automated AI research, etc. But if you think about it, first options use were very focused on language.

They were related to writing and essay editing, marketing, and I I think the first one an app that really took off, there was Jasper, who just wrote marketing texts, and I think he will develop a lot. I I think that washing is good job because it is arbitrary. It difficult. It requires skillful manipulations . And this is being done...it's not depends on time, so if this will last three times longer, then let it be. Whom Does this worry you? This does not have value. Simply let it work whole day.

So I think that he has some of these properties, but I I don't think the task is actually like this important. Hmm, what do I think? opinion, much more important: are you pre-teach a wonderful model, and then able to educate her and run this flywheel? And if you run this flywheel, you know, task possibilities will continue multiply. Maybe this is a question will be for you uncomfortable or a little unpleasant, but if to ask practically each founder, and , importantly, others investors of your type, almost all, if I ask who the best partner for board of directors, they will call you—you you remember a lot more often than anyone another one whom I met.

And me too I wonder why you are like this. Do you think it's you? do something like that others—although incentives are there not everyone does it, to be wonderful partner on the board directors. How are you? What do you think you are doing? on the boards of directors these companies, collaborating with founders, who actually distinguishes you from others very much talented investors who nominally perform the same job, but not are mentioned so often, when do they put this question? One of things that I understood, this is what each of we are attracted different types entrepreneurs and where We have "chemistry".

And I I think one of differences in that What I'm saying: I'm an investor. secondly, and firstly I try to be partner. And that's what I I mean. We we see how we are treated companies are coming. IN There was only one of us yesterday, and this is what I would call investment attractive opportunity. This is yours. kind of investment that can be done. She, It will probably work. You you make money, that's good. The investor would so he did, and partner—no, because for partner of this not enough.

Because, if you don't have real "chemistry" with this person when you do you feel that will you be able to work together very effectively, and I have a lot of things I will learn from them, and they —I have it, and together we we will nourish each other, right? ? If you don't feel, then you don't you can be partner, and therefore you refuse this. But for me it is of great importance compatibility element. AND it all starts with so-so mutual choice: strange, but they want to cooperate with us, and I want to cooperate with them, and I really am I am looking forward to our joint work.

And I will show you a great example of that, how does this manifest itself in in my case. If take, for example, Saji and Benchling, then Benchling is a SaaS company in the field of science life that has reached incredible successes and performed very well, very good And then, Of course, this happened. collapse biotechnological market, and the company it became difficult, it began a real routine. I I often joke: in the company's longest there was no time at all customer churn, so even in their reports, you know, like in every SaaS company: gross ARR growth, outflow line, clean ARR growth—all they do the same thing exactly.

They never have there was a line of outflow, they did not report him during the first six years, that I worked with by the company. And then they got seven-year level outflow in just 12 months. It turns out that life It becomes terrible when you get a seven-year-old outflow in 12 months, and because of this hard work , and through all this related to the fact that " a minute, a whole changed, we must do something different." Despite everything, they continued to think about how apply AI to these biotechnological and pharmaceutical clients to whom they are very close.

As we can do better for them? How are we? we can apply these models in their the world so that they Were you delighted? They continued improve, and it was hard, but I was nearby, and I just...it's one of those cases when you are you working with this a person with enthusiastically, because, first, of course, you Do you think that this special opportunity, and there is a way out, there is a way, and we we can find it; and secondly, through pleasure from process, from relationships and from everything else.

It part of this. And I don't I know what could be. bigger than that. This is very is different. I saw different models, many different work models venture capital, isn't that right? Meritz was writer, Dor— a salesman, you know, Gurley was an engineer, and Peter is professional venture capitalist . They are all different, truth? But for me, I think I like to work with them, I as if I'm calling this people, somehow I learn, they are mine deny, and then I ask them questions. . And I realized that a large part of me work consists in that they already know respond.

They know that want to do. They know the answer. AND my task, maybe , put them question to help strengthen their confidence or to formulate more clearly what they want to do and why. AND thanks to this process we, I hope we become 1% better, well, several times a year. We accept better solution for 1-2% several times a year. And if do it during decades, this gives real results. One of the questions that I set myself how to invest, this: there are all these people who I don't care.

throughout life, as well as your loved ones. Or could I persuade some of them to go to this company and honestly, intellectually honest in front of oneself, explain to them why this is can become a thing their entire lives. And if I can't do this I shouldn't do it. invest, because it's simple means that this The project is not for me. He just doesn't coincides in this plan. So while we have there is at least one of these things, not so for me it is important that this is exactly what like this.

It's simple. important, it can to make big contribution. And if it can to make big contribution, and this is a special man, I would pleased with this worked. Are there any others similar? questions you have set yourself before investing? It especially good question. Um, yes. Hmm, something else question: if this the person will call me at 9 pm in Saturday, shall I lift? the phone? I call it a test " "green button". Yes. Yes. Yes. Exactly. It's just one of those things when if you you don't feel it, then you know it's not.

You just know, for some reason it's simple chemistry questions, and they, obviously, must feel the same. AND one more question that I I bet this: if that is correct, or Does this matter? It differs from the first one, but there is so many things that we as a business can to be right I, I, I looked at one last week and said to the businessman: I I don't think you can. build here great company, it's just not worth it to attract venture capital capital; but there is many things that you may be right, but in the end they they just don't have value.

To no one no problem. It will be so. better to say. If Are we right, or is this Does anyone care? Yes. And if there is no one interesting, you just don't create enough cost of capital. Ahem , and I think that's still one useful point. One of the coolest things that are now is happening, is that everything that you just described, has higher rates and more levers influence. Yes. What is the simplest expressed in more dollars. Yes. Um, and higher prices. We talked to you.

about the concept of what means high coefficient invested capital, which he was and where was he moving. Yes. You recently took a step—for the first time for a long time raised a fund for investment in growth, and I think this is related to the concept of what to these companies need more capital. Prices are higher. Results larger scale. Perhaps, we can get the same coefficient at the entrance with a billion dollars, which is 10 years so when entering from 50 millions. Can you you tell about this evolution?

Tell me about your approach partnerships before this question, and why, on in your opinion, you adopted the following decision? Simply remember why LPs, starting with Swenson and others, started investing in venture capital? Fundamentally, no because they thought, that can bypass NASDAQ or index by three or five percent on a year or something else stupidity. This is because there were situations when venture capital could provide these crazy coefficients profitability on invested capital. WITH financial point sight, that's right. This is what what they wanted. AND for a long time time, almost all of it history of the industry, two things were synonyms.

Early stage investments stages and high cash return on invested capital. The path to heights monetary indicators The recoil lay through early stages. And that's all. . These two circles in Venn diagram almost ideally overlap. And I I think that has changed. recently, relatively recently, lately several years, because the results have become much more important, these markets grew up, and everything else, range of possibilities with high money the multiplier became more than just early stage. And it is not so much huge, so that there a lot appeared new companies, on which can be earned 100-fold profit.

This is not true. But, of course, there are many options outside the early stages where you can get very high profits. And here it is That's all. This is what what we want to go for. We can say that we were late for several years. And I I think I'll take this one. criticism. I believe that ...but I think that's it the opportunity exists in perspective. Therefore, we it's worth doing, and, you know, everything else that we represent— high conviction , high dedication, partnership, all of this must remain part of this discussion.

as part of this discussions. What were other parties discussions, for example, maybe we would say the same in 99 or 2020 when markets become fascinating. Yes. It seems that the possibilities —we all we assume this mistakes extrapolations. Yes. What were counterarguments against so that, despite everything this is still not... I mean, there are all counterarguments, on which you would expect. I I think the main one counterargument, which made this time correct comparatively with two years ago or at another time, it is what you need a team that can do this to do because it's easy a different mentality; there are the differences in that how do you evaluate and thinking about things.

All other things, or, you know, "Why not change and not to stay within the limits "within their competence", —all of this is true. But for me it was most important, and in there were several of us examples from the past a couple of years ago, when we was, I think, correct intuition about the company or opportunities, and we they didn't do this, because it was outside beyond the usual, because it was unconventional, and this, obviously, foolishly, and I think that this quite strong differs from of what he does branch; we really we are chasing these very rare special companies that have very high monetary opportunities returns, where, in our opinion, can be stunning successes , and we can in them invest.

Is there any lesson that can be taken out of many, let's say, 20-100 times success, who you personally are observed? I have I mean, this is like this crazy profit that early calculations are not possible converge. Impossible to calculate something ahead, because if everything it was so obvious, then and the price would be different. Um, yes. Why you? taught these cases growth of 20–100 times, if at all taught? Work with really special people. You know, I want to work very hard hard, I want to work.

smart and catch a little luck. All of this must come together. There are many things that depend on the choice moment, and on which you, as a company, you don't have no impact. Let's take an example. Cerebras is great case: we went out on public market this May, but tried to do this is still in 2024, and then the score the company would be much lower, and all it would be enough not easy, and not came out through CPHAS and other factors. So it's time matters, and those 18 months difference changed everything for the better thanks to things that, frankly speaking, there were beyond our control.

There were things that depended on us, as- here is the start of inference and so on, but many what was beyond our influence. It classic: yes to focus on because you can control. This is one of things that are essential distinguish us from software companies. In companies- software developers, if not counting AWS or like, you usually own everything technology stack. Therefore you are completely control your fate in this aspect. In companies that are engaged in equipment, it is not yes—there is a whole supply chain and much more.

HBM is important, DRAM is important, TSMC is important, and many of they are crossed geopolitical borders , so geopolitics intervenes, complicating everything, and this really huge difference. You have to to get lucky with time, macroeconomics and other factors. But I think that's all. starts with work with these crazy people people who have limitless possibilities. And if you work with so extraordinary people over boundless opportunities, you know, then from time to time you luck is smiling. This is inevitable. I am what I will say: when I was 20, I worked in investment bank .

Horowitz, Mark, Ben— they started LoudCloud, the company was still in the shadows, and, you know, Ben suggested I will become his. assistant, and I talked to this colleague, who then seemed almost to me as an elder (to him, probably there were 25, but, you know, for me he is was like that then), and he told me something that really me remembered. He asked: "Are you playing?" golf?" And this is classic. banking question: " Do you play golf? No, Damn, I don't play. golf. But he says, you know, in golf you you constantly train and trying to throw ball in three tricks, closer, closer, closer to the hole.

And he says you constantly you bring the ball close to the hole and continue exercise. And he says, this is it hard work. It how to work with mind. This is what you are you have to do it constantly. He says to do " "hall-in-van" is luck. Me like this wording, and I I even use him with his children , because it means that yes, luck is important, she really is plays a role, but there is way to increase your luck. And a way to increase luck—it's more common to bring the ball to holes, and then, finally , one of them will hit.

And I like this one. mental model, i I return to her, when I think about companies. Each of us invests in one- two companies per year, I think in 12 years I will invested in 18 companies in general. Madness. This is a relatively small number, so this requires a lot confidence and devotion. I invest own funds, I believe in these entrepreneurs and I believe in these companies, but if you you continue to work with these special people over such possibilities, maybe A miracle will happen. What did you learn about the best reasons and conditions for entering public market?

Benchmark suits these dinner on Monday, Have you ever been to one, and yesterday In the evening we had CEO private company worth hundreds billions of dollars, and we had this conversation. It still a little fresh impression. I think that after all, when you go out on public market, you you get a number new opportunities for development. You you get social trust, oddly enough, thanks to transparency, which comes together with publicity companies. You, of course, you receive currency, from which you can to perform operations. Ultimately, this is what remains.

You have appears incredible ability to raise capital, and that's why, I think, laboratories eventually will also follow this path . Although I think that a matter of trust—it actually very important element why they are worth it to go to the stock exchange, and that useful for the world and America, if they will become public— this, you know, the opportunity to see that is happening. Everyone can see it. I I think it's very useful situation. THERE ARE another aspect: what wants a student- athlete? Become professional. They want to play on higher level.

And is this more difficult? Yes, it's harder. Is it stronger? competition? Yes, competition is stronger . They are faster. They more resilient. They bigger. They are stronger. . The stakes are higher. Scene larger scale. WARNING chained tighter. All This is true. WITH companies approximately the same. There are a handful—and it's really a handful, three, four, however many they were not there—those who can achieve huge scale without going public thanks to the perfect completing tasks. Lots of free cash flow. They have a lot of free time.

cash flow, they are very good worked for a long time, and I think that this fantastically. Um, and you know, good for them. But you know, in general for everyone else—get out to the market. And what else would I told you: there are " "windows" for a specific type of companies. So yes, SaaS companies that went public in 2021 year, many of them encountered difficulties in public markets, because their shares sharply fell due to a problem compression multipliers. They were trading for 30 -multiple exponent. They grew 4 times, but now they are bargaining by 6 times.

It turns out , you are still in the red, Yes? This is not easy. situation. Ahem. However, I I will also say that there is about 500 private SaaS companies with turnover from 100 to 500 millions of dollars. What What happens to them? These employees never didn't have a chance to sell shares. In these no employees annual tenders. These employees are not have the opportunity exit, as well as investors do not have opportunities to get out. They are stuck. And, you know, I don't think that they will all disappear.

As I said, no I think they are all " "outdated" or something else , but in the end, AI- companies with their growth rate sucked the whole oxygen from the room and interest, and the moment was lost. And it's hard. What are the biggest ones now? internal disputes partnership? I always I love coming here. and communicate with you when it happens something interesting because you are having discussions, it's great, sometimes it's could be very interesting to watch, and I a lot of this I will find out.

What are these? discussions today? A lot is going on. controversy around AI ecosystems, infrastructure and applications, in particular basic models: where accumulates value, how exactly it accumulates, where are the "moats" "and how about this?" think? But also about innovation in business models. And I think, people not quite understand why SaaS achieved such success compared with traditional software provision—this it was not only because the best model delivery. There was true innovation business models, right? ? You really had this one subscription element, which eventually became fantastic for both companies, as well as for customers.

It was a situation where everyone in winnings. And I think that the same thing exists in AI, sale based on results and in other similar ones aspects. But then, in this discussion, arises the question is how much value is obtained laboratories. How much value producers receive semiconductors? This is really serious. discussion. I hold the opinion, that it all works. As I said, this is a very strange thing. It's like ... Will they be successful? cloud providers? Yes. Yes. Will there be successful some of these new cloud companies?

Yes. Or will be successful companies such as Fireworks? Yes. Will there be successful Nvidia? Yes. Or will these be successful startups with chip production, or at least some of them? Yes. Will we have peripheral inference on our phones? Yes. Or we will have an inference at the network edge in POP- nodes? Yes. Will there be we have big models in data centers? Yes. Too much thinking in the style of " zero-sum": " Okay, how are we? Shall we share this pie? And how much are they?

"Will they eat it?" Or: "Oh no, Anthropic or someone else will take 98% of the cost, and they will do it themselves all drug discoveries." And I say: "That's enough, "This will not happen." We have examples from so far away of the past, where everything it happened exactly like that, truth? When I say: " I think everything will work out." listing different things, very important understand: this is not means that each the company that is doing something with this, will succeed . Actually, everything is just right.

vice versa. Majority companies in each of these areas are experiencing failures. And now, as never, it is important to have real differentiation, to prove every point to the logical completion and understand: wait, you have to go to the end. Well, in these matters should be more real special. Is there anything else you're looking for? keep a close eye on whether on the investment market, or in the world of technology , anything at all? I'm actually surprised, that some people, so smart, are still there in this tech world where almost reach deterministic conclusions about mass unemployment and all that like this.

I will show you. very specific an example that I think simply wonderful. Let's take Jeff. Hinton and radiology. It seems that in 2016 he said: "We should stop cooking radiologists. AI will make everything better." Listen, Jeff. Hinton three times smarter than me. But he was wrong, as no one else. Although the very premise, which led him to this conclusion, was absolutely correct . Namely: if to look at radiological images, then we have to teach AI can read them better than people, and this is probably truth.

Actually research confirm this, and we invested in New Lantern company, which working on it, but the main obstacle, which they indicated, consists of the following: Aggregated set training data simply does not exist anywhere, company's reason focus only on for something specific, for example, on CT chest. AND ordinary radiologist deals with every day a lot of things: from X-rays and CT scans to MRI all parts of the body and everything else. Therefore, AI that specializes in in certain areas, such as- here is the computer chest tomography cells, there are only minimally useful, because he does just one thing from twenty or forty scans that the doctor reads per day.

So ok, problem number one: you don't have data, just like we talked about robotics and everything another to train AI. Problem number two: the entire industry health care focused on reimbursement of expenses to doctors for interpretation results. And how is that? will there be at all work? Who carries? responsibility for this, what are the consequences error or omission something, and also medical question negligence and everything another. So how do we we are going this to avoid and how to get around these obstacles? This is it.

problem number two concerning real life difficulties. I think, in the end, we let's get to this applications where AI will really help radiologists. It will help radiologists increase bandwidth ability, because AI will perform certain parts of the work, and radiologist—others, checking one one. You will find yourself in such a strange situations of "the second" "pilot" for a certain time , and then AI gradually will begin to analyze more and more scans, increasing its efficiency. But real way from now until then It will take a lot of time.

time. And during this time we need more radiologists, not less, because, to things everyone does more examinations than earlier, because the cost visualization falls because of the paradox Jevons. My opinion in because you have a very a smart person who really understands opportunities, understands, what is happening has correct data, but not taking into account application of these data in real time world, she comes to false conclusions. That's how I look at it. the issue of unemployment. I I think it's almost identical situation. Eric, me like to talk about the markets with you.

It real pleasure . Thank you. You know how small advantages accumulate from sometimes. That's fair. in investing, and so just in the way you manage your own by the company. Your the cost system is your strategy capital distribution. Ramp makes her smarter than default. The best data, better solutions, better economy with sometimes. Find out how at ramp.com/invest. In moderation your growth business Vantaa scales together with you, automating compliance and providing the only source of truth about safety and risks . Learn more at vanta.com/invest. Ridgeline rethinks technology asset management as a true partner, and not just provider software software.

They helped companies zoom in five times, providing faster growth, smarter operations and competitive advantage . Visit ridgelineapps.com to to see which ones they are possibilities can open for your company. The best AI companies and software powered by: OpenAI to Cursor and Perplexity. Use Work OS, to get ready for corporate level overnight, not months. Visit works.com to to skip a boring one infrastructure work and to focus on to his own.

Go deeper

• All Invest Like The Best episode guides