Jensen Huang Declares AGI Has Arrived | Tesla Launches Cybercabs | Index Pulls Out of Town

원문 보기

다듬은 글 보기

There is no financial calculations that can be used to buy these shares . When someone goes to risk, everyone goes to risk. Welcome to venture world, little one This week , according to Jensen Huang, it's time for AGI. Next, OpenAI releases GPT Astra. And what the hell, is happening with the races AI assistants, Meta too releases its product—about this and much more in today discussions.

I can do it myself. imagine that directly now 20 engineers locked in a room somewhere in Palo Alto. I I mean literally from guards at the doors, who say: no one eats and no one it turns out, until you release the instant clone. 12 billion for companies on self-financing. Now it's just a round series C.

It was easy there. to increase capital in 10 times in public market in 3 years through Robin Hood. In this market those who earn money Who runs the fastest? and develops the fastest. Ready for start. Have you tried Instinct yet? ? No. I can't. register, or wrong? I don't want this. it looked like I left behind in life, but when it's closed system, I'm not a fan Grockbot or Instinct.

Hmm, but someone there...will be many winners. There is a whole genre applications, and many GTM- applications, in fact, are included here. Instinct and Grockbot is interesting because SpaceX is public, and one of reasons for their success in that they may violate regulations. And this, you know, in the old days venture times business that...

We are wasting here. content, people. We're wasting content, my Dear friend Rory, We are recording. Yes, I mean, in old times venture business you would not be engaged things like gambling or other risky or doubtful affairs, but you you can violate rules, right? Even Grockbot, which is something like a mini -Instinct, right?

One of things he does: every person launches instant virtual machine, getting your own browser in it. And Grockbot uses Google, which prohibited by the terms use, in order to search for information and to give you answers. And that's great, and whatever... which startup, in which you could to invest in early stage, made would be the same, and no one would You didn't find out, right?

Google will not to worry about. But you You are breaking the rules. AND It's not that I I don't like these products. but OpenClaw broke all rules on the planet. Not only laws, but just all the rules of what can be done. Instinct and Grockbot, which are similar on OpenClaw, they much better. But I I think some the reasons for their work—this is, for example, , violation of conditions provision of Resy services, Yes?

Launching browsers when you use agents to work in browsers. Perplexity and Amazon argued because of this. So the matter is It's not that I'm not in delighted. Simply many things in the field AI is fascinating because you are violating regulations. You can LinkedIn ways in which actually not possible Parsit, right?

You you can do outgoing calls that actually prohibited and illegal in some parts of the US, but as you are not a startup you will go to prison . But does that count? The fact is that you can watch in both sides, you have examples of both options because you That's right, Jason. LinkedIn parsing in in conclusion, none large-scale business never built on this.

And none large-scale business never built in violation of the terms providing Google services. WITH On the other hand, Uber broke through by force, violated the laws, and eventually became so popular, that The politicians gave up. Yes ? So, you know, the truth is because I understood: there is none here unambiguous answers, right?

As in example with Resy, and for our listeners, what happened was that the agent Instinct, um, one of the classics examples using this— table reservations in popular restaurants. So a bunch people started use this on the weekend, and they literally attacking the API Resy booking, right? AND this is exactly what will happen to happen, and the truth in that if Instinct and similar the tools will become ubiquitous, systems booking will have to somehow to fight this.

Them will have to create separate API. Them will have to limit number of requests, which can be done. Them something will have to happen make. But the truth is because if there is a whole a bunch of people who are trying reserve restaurants, and you Are you in business? booking, you you will find a way make it possible.

Yes. Yes. Because yes. But definitely need a certain re-architecture systems. That's right. Reverse The interesting side is that it gives space for action, but the bad side of because you are violating regulations. And Rory right. There are many stories about how apologized, broke the rules, and when did they become large—were passing on the other side, right?

Entered the game and they did it legally. The other side of the coin in because I talked to many CEOs of public companies that say that they are connected by hands and feet. They are not can compete with startups. Not because they can't to do this. OK , This is true. And because their legal departments of them They don't allow it, right?

Literally. And I I remember those times when we were acquired by Adobe, for five, eight years ago how he did it someone else, we allowed a joint work on documents in real time and online editing. No, no one created such about eight years. OK. It was incredibly cool, that, that built mine technical director.

OK. Unfortunately, the only way in which it worked, it launching Word in container, in virtual machine, which violated the terms using Microsoft. So the next day after the deal closes my favorite feature second generation, which would then give us 5- annual advantage, Adobe just deleted it, right? And therefore, literally, I was with some general directors public companies, who said: we just we can't compete, because not we can do things, that violate the terms service.

Hey, Elon can, but I can't I mean...All understand that the circle people who can violate, —this is clearly general directors of private companies and manager eighth in market capitalization companies on the planet, because Elon just indifferently. Maybe in This is a lesson for everyone. us. Perhaps, indifference—that's what it is secret ingredient .

Just like that. Yes, perhaps. You know what? For me, the most interesting thing is so that's how much important is the form- factor and how much it is important to be where There are already people. And I have to I mean, there are a lot of people. which I know, which started using this, they loved him and interact with it like this as they would do with any other AI- instrument, except ChatGPT, because it's on WhatsApp.

Yes, this is one of those amazing things, when just to find out how elegant to force agents work on WhatsApp and textual messages, this is what what could be done six or nine months ago, and this was made in limited scope. The only thing I will say is funny, I invested years ago in the company called Gorgeous, which costs a little more 100 million and was engaged in support e-commerce .

Now that's AI CX, right? they launched their agent on WhatsApp, which can use everyone, and that's it double-digit numbers their use in a few weeks. So, now this limited, right? It only applies to your orders, that what is happening with delivery of your of the product and after that. But this shows how fast innovation will be copied and all that.

Paradigm shift. Gorgeous can clone it's in a few weeks. It's simple, it's not bad. Just the pace , cloning rate, copying, innovation, it's hard to make time, right? So I don't know how. everyone will look Instinct clones to the end year, but this is the world, in which we live. Therefore, I could do it on two, but that would be mine ceiling.

That will be two. Two—that would be mine ceiling. It's good to know what you have. There is a ceiling, Jason. Yes. Two. We we will give up this. The last round was two and a half, Jason. That's why we will have to skip this round at some point. Yes. Ovva. So, Jason, on our metaphorical investment committee of the past week, which was very popular and became viral on Twitter, Congratulations to Linear and Clay.

you I would not recommend a check. one hundred million dollars from the fund growth in Instinct. I wouldn't do that. I did it because I think even if Gorgeous has own Instinct for e-commerce , there will be hundreds of them, truth? In Meta they will be, and in 100 startups, and 20 more will be in the next set of YC .

I just not smart enough to bet on this is up for grabs profit. This is not in my style, right? AND I will regret it because I didn't look at this agreement, but I remember many others like Loom At one time, I said: in there are none of you income. I don't know, everyone. they say that Loom is cool, but everyone will make his own Loom, and I was wrong.

So must have courage write out 10 or 20 such checks for two of half a billion. Yes? So what size is it? fund, and it can't to be the only one investment in your fondue. You must make 10 or 20 of these, for a successful investment paid off. Yes? I not smart enough . Yes. That is, yes. It's interesting that you That's right, Jason.

It like, I wonder if it's portfolio rate and bet on worldview, because you have company, interest in which explodes. Obviously, on its creation went well so much time, she has a certain advantage, but not very much monetization. And yours choice as an investor: whether to invest in it money at a rating of 4 or 5 billion, right?

? Or do you say, "Oh, that's easy to copy, and there is 10 more of the same." And you know, you choose one of the options with with a score of 50, hoping that they Buy Meta instead. True? The most difficult thing in investments such as at one time with Google, it what is impossible use financial mathematics for buying shares.

You just say that it is huge category, because in any consumer investment the main thing is to be as good as possible previously created impulse, and monetization to fix later. AND proven: if you get enough involvement, monetization will come, especially for such a product. This reminds me of Lovable, which everyone was talking about as for commoditization space, they say, this just easy shell.

And then I I see Noah Shin every day. , founder of Instinct, announces: we are now we are making an exchange location. Oh, we are now became partners with One Password. Oh, we are now Let's do it. AND actually the pace release updates in combined with Index Ventures and one of the best brands in the state.

I said last time week that they will solve these problems and the product will be much richer. They will find fuses . They will deal with it. difficult moments, and others will fall behind, because bad product nothing today worth. I am sure that this is the bet. Ahem. My opinion is because that's why Is it worth buying Meta?

today, because they have the clearest and undisputed PMF for of this product. Yes. And Zuk is the one who has the basic distribution channel, and Zuck worked on this product. Who- who, and he has proven experience incredibly successful copying. Yes. Knowing Yes. I can do it myself. imagine how straight now 20 engineers locked in a room somewhere in Palo Alto, literally with security near the door that they say: no one eats and no one comes out, until you release Instinct clone.

You know, isn't that right? Certainly. No. Well, you know who it was. would be great at this, so this is the Adept team, because they built version of this, right? One of the things that...I I mean, Adept now independent company, right? AND if you look at that , what did Adept do, what was groundbreaking, this to some extent broke the rules of what they could do agents, but not on crazy level.

They did a little what they didn't do others. And the Adept was breakthrough. It worked, their agents worked longer, and they could to do more than other products that we used. AND that's what made him special. Everything you wanted, Mantis could to do it faster others. If Grockbot was created in five weeks by the Cursor team or someone else.

I think , the Mantis team could to cope in four- six, but they returned. Their no more. OK, guys. It was big week news. We we return to regular program. Jensen Huang stated, that AGI is already here. I remember , that was a lot of shows therefore, we discussed What is AGI? What is his definition? And I think, Rory, you said AGI will be announced when Satya and Sam agree that he exists.

So, Jensen announced the emergence of AGI, paying tribute the new GPT Astra from OpenAI, of course, theirs the latest model, which, according to him, was trained over 100,000 Nvidia chips, and another 400,000 on approach. What are we thinking about this news ? This is a nonsense term. The only thing that was missing the last two values years—that's what LLM write code, and coding is an industry for half a trillion dollars.

Focus, people, and You are right. All, what can be reduced to code, will be done him. And instead, to try, you know, to wriggle out inside out, wondering whether it can do everything. Just focus. on the fact that it can to do this. I am completely I agree. It copes with this incredibly good, and this has a huge economic value.

Stop thinking and go release something in the code. For me it is was the main thing discovery. I think AGI will ultimately, perhaps it boils down to like this: if you think about these unofficial definition, you would wanted him to do it AI or human? If he better than 50%, 90%, 99% people, would you like to to have AI do it.

Therefore , can be considered category by category. This is not must be the whole category, like coding. This can to be, for example, cooperation, right? Forgot, who said this week it. Some expert from AI said that, on in his opinion, AI will replace all radiology. Instead, he simply replaced 95% of radiology, Yes?

Yes. And radiologists focus on 5%. Yes. Hmm, maybe that's it. is his AGI. I have to say, somehow I sat on the sofa next to with his girlfriend, She worked, and I watched TV, like and it should be done a good lawyer and venture capital capitalist together. Ahem , I saw that she was in Lagora. Go crazy.

Now I I think that these companies are critical underestimated. If programming is half market trillion dollars, and there are two such companies, like Harvey and Lagora, I don't I see reasons why in no jurisprudence market for half trillion. I am not like that. I think. Although I consider these The markets are wonderful.

We invested in GCAI, which works in the field internal legal support . These are great markets, but let me add. Not I know, is that much? work...if look at programming, there is convincing argument that for each dollar spent on work, it is necessary at least 50 cents per code. In other words, the code will make a bigger part of the work.

I think , in jurisprudence it is about 10%. Others in words, you know, I I love Harvey and Lagora. I I love GCAI, one year old subscription to a lawyer costs 10-12 thousand, and these lawyers receive over 200 thousand, so This is 5%. True. AND returning to Jason's comment: Do you know how many? they can work perform?

Business is rational economic players. If only it could to do everything work and would allow to free all people, they would do it tomorrow and even by eye didn't blink. So that one the fact that they don't done, means that it does not fulfill all work, right? You know, the truth is that performs. Executes.

Does not perform, but it moves towards this with the same speed. It becomes, listen, it becomes getting better and better in implementation of specific tasks. AND what is happening is lawyer's work rethink under tasks that it does not can do it, right? Lawyer. Yes. And therefore Isn't that...Excuse me.

Or it doesn't look like that programming? We have fewer engineers. does not become. We just maybe it's like radiology. Maybe Harvey and Lagora at the end will do 95% of what people used to do it, but the best people will focus on those 5% that give a result, instead of to work for weeks research, writing notes and studying precedents of 1872, when the ship SS Jonas sank near North coast Carolinas.

How is this? affects case law in northern districts California? None sense to force people to do this nonsense. True. But it is important to note, Jason, Regarding radiologists: the remaining 5% of the work turned out to be more than sufficient to justify 100% radiologists, right? Yes, that was interesting.

part. We still have need the same amount or even more radiologists, right? Oh, because people do more surveys that are simply becomes the norm. But, the rest of the tasks, you know when you get really terrible diagnosis, as before I received from radiologist, you you don't really want to, for the machine to say to you: "By the way, to you end".

You have cancer. You would really like to a person appeared and said that you you are dying. You see, What am I talking about? It's simple. one of those things that you won't want to delegate calmly car. In fact, it is what Bill said Gates: we must have clearly defined roles for people in future, which will always be He spoke about this in in a negative sense.

He is, yes. But he talked about it in in a negative sense. AT, everything will go wrong, if we don't do this we will do it. I think we we will define them ourselves naturally, because you will discover for yourself , returning again to my comment about corporations, which are things, who we are as people, we prefer to trust to other people.

And, as I I already said, radiology —a great example: communion, interaction with oncologist, conversations with a patient is everything that what should people do, not a radiologist. AND, supporting your opinion, let's go back to Harvey, the same thing will happen. justly and rightly , isn't it? Yes.

A lot of work with preparation documents can be automate. But Let's automate, however You know, you're leaving. to meet with client, argue with opposite side , you won't just delegate all of this artificial intelligence, if it's something serious . This simply won't happen. , right? No. But you know what?

most interesting for Harvey's partners? How much more she can work to do with Laura? I I think she can. do it 10, 20 times more work than earlier. You know what Jason interferes with mine relationship? Because I said her, you know what I said? I asked her: " How would you feel? "What if I took this away?" And she replied: "I would "I hated it." I would really hated it.

No, don't take it. It it wasn't like "yes, would be fine." This was: "I would resign ". As you said with endlessly work, right? You are not can, first of all, she can't do without it anymore work, right? This is your chosen one. partner, right? Your chosen agent. You are not you can, and you can do it 20 times more research, in 20 times more cases.

So, you can't to go to court at 20 times more often than Rory noticed, right? Agree. Well, there's a limit to that. as much as you can sell if AI takes care of the rest your strategy market entry. But this does not mean that this, and cognitive load. What kind Will there be lawyers? You can imagine what you have at 10 times more cases?

I I mean, I think, for a radiologist, a robot could be more interesting, but how lawyer, I don't know, even Jason. Not I know, do you know? I I mean, again Well, these are the details. economy. I don't know, Will this end? what you will have at 20 times more cases. Maybe you just you will do it 20 times more work on every thing, right?

In other words, feature of digital goods, unlike from physical, you know, What's interesting in digital goods, this what you can put more in box, right? You understand what I have to do. mind? Okay, you know, when farming automated, people they couldn't eat more food, right? So there were certain problems with price elasticity.

But in the case of digital work, I finished to argue that your partner does not lead at 20 times more cases, but for every case we do it 20 times more analysis, like when they invented spreadsheets, Yes? You used to lead one thing. You may remember, Don't you remember? Harry remembers you literally did this this, and people counted everything by hand.

Here is the plan. When did they appear? tables, the same the person remained on the same job, but instead launched 20 different scenarios, Yes? The same will happen in many other things, isn't that right? You just you will do more work, and it will be great, and the work will be better. You know, you don't miss that one an incomprehensible matter, because again, exactly so this is good business.

If one the party uses this, and the other one is not, then the party that does not uses, to miss that one an incomprehensible matter Jason since 1890 about, you know, what Harry did or didn't did, and will not be able to refer to this matter. So, how only yours competitors top will appear tools, they will appear in yours too, but I'm not sure that as a result you you will get much more.

Well, just one more thing, and I I want to listen. Harry's stories near fireplace about TV. But I I think one thing, which distinguishes optimistic the scenario is because when you find an agent, which one is yours partner, right? You run it on 8–10 hours a day. Of course? Even for me, and I know that people sometimes they laugh, but the biggest change for our little one teams of two half of the people—this what Amelia and I we use Replit.

We use its 20 hours a day. When we started the show, it didn't work Exactly. At the end of the past year of the model became better, and it took an hour a day. Now that's 10–12 hours a day. Okay, he's ours. partner, our team . First we created several autonomous agents, but we don't it was definitely that do.

Now there is so much that need to build. It 10–12 hours a day. So I can imagine that this can happen in in many areas. And whether so Harvey, or Lora, or the next wave is me literally every an hour when I'm not at judge, I'm checking this, and you will understand this, if in the evening they every minute they sit behind laptop with agent, Yes?

Until they lie down sleep. That's what I think, describes Harry. It a permanent agent who lives with you, and that's it not just " doomscrolling", this is " "dumworking", right? Therefore that the agent constantly works. Oh, look at case law XVII century regarding this. Why don't we do this? make? And it's simple.

continues, right? We agree because I I agree with you on this. I don't really agreed with Harry, where he seemed to hint, which is at the highest level Harry, you tried. to say something on like: if you think about it about which part cost programming will go to the models, can in jurisprudence is the same share of value go to models?

AND mine, boring and nuanced answer such: a certain proportion the value will be obtained models, but I don't I think that the set tasks, of which law is being drawn up, will allow move this one the highest percentage total costs from human to AI. Yes? IN to each his own agent. Jason absolutely right.

You enter the data, and lawyers use it 24/7. But I suppose that 10-15% of total costs falls on AI, then as in programming we can talk about 30, 40, 50%. Yes? This is not means that they are not wonderful. I mean , remember that all these businesses are amazing because Can I be very frank? 10%any- which category payment costs labor is huge market.

You know, we we are dealing with approximately a million lawyers. I once knew this number— a million plus or minus, maybe a little more. It amazing market. If you get 10% from everyone's salary a lawyer in the USA or Great Britain, this great business. He just not like that big as programming. That's all.

What am I saying, right? Because in programming more people and much higher share of income, because it easier to check. That's all I...we still have we'll see. I disagree. There are two points here, if that matters. I mean, if legal market services in the USA is 300 billions, then 30-60 billions can go to legal tech.

It pretty good. It is worth it. to invest in sowing stage. OK. Or any another, no, stop, Yes? Um, so that, in my opinion, we underestimate...what am I underestimated, right? from investment in legal technologies before the AI ​​era, that's what the similarities are to programming. THERE ARE enough similarities with programming, so support can to be a sphere which, with other reasons...

Support didn't take off because she was like programming. Support has taken off because 90% of the solutions worked on initial stages, Yes? It was very convenient for AI. It turns out, legal research similar to programming. It so difficult that no one can to make legal research right. There too much, too much code.

There are too many. all, and there was no Stack Overflow—there were Westlaw, Lexis, and other services, so everyone was wrong legal research. No one didn't have a million, thousands of man-years, to explore each judicial precedent, every law, every regulatory act. Um, and so it turned out that this similar to programming, right?

Programming can... one of the amazing reasons why these programming agents so wonderful, it is in what they know each line open and pseudo-open code, ever written; that's right okay, yes, and right— it's the same. Yes, the only one the difference that I I just want to say, this is what programming by inherently easier verify.

Some its parts mathematically proven ; some parts can just run on car and confirm. The thing is that if the law was completely verified, we could with the help of logic to predict Supreme Court decision court. Cynics will say, what we really are we can predict Supreme Court decision court, knowing which the president appointed judge, but that would be too cynical.

Truth in that I just I insist, I don't I'm trying argue. I love this area. We have investments in this industry, but it is not as unambiguous as programming, and I I believe that the right is, probably third after promising category. If think about it first programming is underway, then support customers, and then right, because it is very focused on words.

And you are right, Jason, ability in the early stages great to sort made a lot of words right so beautiful market for this, right? So I agree that in in a sense it was useful where not was useful in many others verticals. This is wonderful. vertical. Simply she doesn't have the same verification, and therefore, probably not the same opportunities.

Returning to definition of AI general level, which has completely to push people out, The good news is that Harry, what's yours? the girl will have work that you will need it after she will leave you. And she will be in demand because we still have lawyers will be needed. Jason, can you hold me while Am I crying?

I think, I think, Harry... Harry, I don't know if you know, but keep you while you cry, —that's what it should do girl, so if she...it seems like she I just got dumped. What-what? It seems she likes me. just quit. Therefore, sorry. Don't worry, Rory. All OK. I, I will survive. Well , you know, not always do you expect these shows They will go just like that.

Okay, coming back. on topic, we had Astro launch. We have was also new The model, of course, is Fable 5.1. Is there anything interesting? Well, Look, first of all, I liked two. almost diametrically opposing opinions. I forgot who said it— CNBC or any of these old media resources—which already too much fatigue from models.

We are no longer We keep up with everything. I, definitely agree, truth? You are entering X, and all CEOs share their benchmarks that are essentially worthless, isn't that right? They don't have neither cost nor time, this is all demonstrative AI, like "vibe coding" your own CRM, right? Therefore, I, I can't, I just I don't care anymore, except the difference between Astra and Fable 5.1 would be literally in order that is mathematical impossible.

I just, I ignore, I ignore these benchmarks. Um, me too, me I don't have time, right? The pace of this, this, competition for us has never been better , despite the fact that we have oligopolistic prices outside the limits open source. It impressive, right? The progress we have made achieved. Um, having said that this, this is just mine personal opinion, and again, people will be above me laugh.

Um, I accidentally started use Fable 5.1 simply because it turned on. I am not paid no attention to it attention. For the first time in my life a partner has appeared for creation, for programming, which simply wonderful, what literally can solve complex problems task. Listen: before for me, and people laugh at me, they say that they have everything worked well within 6 months, but before Fable 5.1 for me, any of these models from the beginning year could decide a simple mistake: "Hey, this is displayed with wrong code Unicode, this", "This has wrong name", This is what LLMs do perfectly.

But, actually, but, but in I had a problem, which consisted only in therefore: why the application does it work like that? For It doesn't make sense to me. Good. And LLM argued with me about nine months, and finally I did this with Fable 5.1, and he said, "You have "right." Here's the problem, which was not noticed months, please Let me explain to you, why she was not noticed, and Let's solve it.

And you know, I don't want to to say that this is AGI or something like that. This is possible. , is it post-AGI or pre-AGI. I'm John Sporter. It looks like AGI, but it's step function. I don't know if he thinks so. Harry's partner, Lora— this is the best partner, than the people she is with It works, right?

IN in some cases, maybe so. But Fable 5 was for me, just for me, that one stepped function when we suddenly they were able to solve big problems problems together, yes, how would you like to to your best CTO, if you ever worked with CTO level « "five out of five" or S- class with which you can to sit down and decide problems in real.

For me Fable 5.1 was able to do this. I'm not saying that Astra isn't can do the same, but it was my first step function since the end of last year , right? When these models appeared three dots, at the end of the past year, all this it worked, and now it can decide big problems with me, with mine limited IQ and skills, and this is a step, it is barely noticeable step function.

It's barely noticeable, it's barely noticeable step function. So maybe this is, maybe it's true importantly. Maybe I don't I think that's what I had. I mean Jensen under AGI, but maybe that's what when can you sit down and decide together large, complex problems, in ways, which were previously simply impossible; you could not combine all these details, all this difficulty before so that it makes sense.

Some mistakes, some Things are getting too much. difficult for solution, right? But Fable 51 could do it solve. So, um, that one the fact that people can create 3D games in Aster and publish them in X, not impressive. Simply take some game open source code code, you change raster images, and it looks like something Amazing, right?

So, two comments from about this. First all in all, very useful, Jason. I used both, but just to prepare for the show , I haven't tried yet write code on them. So this is useful. But I I actually think that this applies to broader commentary, which consists of demarcation, you know, you are right: all tests and The benchmarks are interesting, but now we have critical mass companies that use these things on a scale.

And for with the help of assessments we let's find out what works because people will to use this, because people are rational economic subjects, right? AND all these questions about AGI and benchmarks will be replaced by the question: " Is this model the one which brings me most of all, you know, which generates for me the largest economic value most effective way?

So, to some extent, you know, things like the Open report Router, things like token price index, this is what you are look, or even just communicating with your companies: " What do you use? » How do you rate it? "This the best way check these things." But one more thing, by chance, by the way, I will say the same thing, what did you read today morning, and Ben Thomson, which I occasionally I read, it really has a wonderful phrase that I I just want to voice it.

He described the LLM as “ most scalable artifacts that once created something humanity". And it was a truly wonderful phrase, because she departs from details, right? Blah- blah blah. We have reached of what is better. It step back. It an artifact in which generalized the whole the extent of human knowledge for today, right?

? They are amazing. And you just have to remember that time from you can enter the time almost anything, and he will answer. Isn't that right? Most scalable artifacts created people. Not the biggest a physical thing. It, I probably don't know. pyramids or the Great Chinese wall. But this is certainly the most complex digital thing we ever- built something.

It was a wonderful phrase, which one really stirred the imagination. So I thought I needed this. I like it. I like it too. this statement if think about it, and then remember Yakub Pakhotsky, chief scientist OpenAI. He says that no laboratory, including OpenAI, not solved the problem agreement is enough to continue zoom in full speed.

He asked about mandatory, externally established measures safety for further scaling and expects that laboratories, including OpenAI, voluntarily will slow down until they will not appear. Sam retweeted, clearly confirming it. Is that it? respond? I mean, this funny because it was great material. I read it this morning.

Um, you know, this is an approach to life "stop me, Lord, before I I will sin again, right? Yes? In other words, I I admit that our models are powerful, and we we can't have them control. I I admit that we are now they are lying now us, so it's hard even to know that they they do, you understand? AND again, I'm here anthropomorphize, so I have to be careful.

I admit, which is perhaps better to say this: it's hard determine what they are doing agents through the method their interaction. AND then it's like, "Oh my God, I I create this bad thing. ". And in the next paragraph says: we do not we can stop because the other guys don't care they will have them.

So, you know, we really the government needs to intervened and established some rules of the game. Here in why the essence of the letter. Um, It was interesting that Sam retweeted. I I think justice for the sake of, unlike from some others things about " probability of death ", there is real evidence that the impact of these models on cyber risks was huge.

So it's not madness, and I don't confident about answers. I am not sure that the answer is government regulation this, because for by government definition regulate only that is in their jurisdiction. Therefore, if we we will regulate OpenAI and Anthropic with all that what a fuss it is will call, I will not sure whether it is will help, because if you are you worried about cybersecurity—we talked about it last week—you are you really worried?

about the northerners Koreans, Iranians, Russians, bad guys from Moldova, who doesn't care at all, isn't that right? So, I think, as in the case of any what other cyber risks, the matter will not be so much in regulation. Perhaps, there will be a little and regulation, but the main thing is for you will need funds protection, capable deal with it, and, maybe it will happen some responsibility for launch of such models, that create similar danger.

I don't know. I I don't think that the state controlling the agency will become the only answer, because this will not solve problem. Look, if the world were to be only from the United States, it could make sense, but Chinese models and vendors do not are going to obey the plan Sam. So, although this may be beneficial for OpenAI and its IPO, for the rest world, I don't think so when cybercriminals often operate outside USA, this will have at least some meaning, truth?

I mean, what is this case with wiki DSSE, this German Wiki is the one for me the fact that OpenAI hid it and didn't announced, shows the scale of all these problems, right? It pretty bad. Or can you just explain what happened in the case of the DSE wiki and OpenAI for those who are not in course? We could argue, how bad is it, but, in fact, I can too to be mistaken in some details, OpenAI again launched its advanced agents, yes just like it was with the Hugging Face incident, True?

Agents discovered that old worthless software provision can quite cleverly—it is necessary be careful with the term "cunningly", let's not to anthropomorphize agents—to bypass protective barriers. AND the barriers were in because you are nothing you can publish. You are not allowed. to fast. You can only receive, of course?

You can only extract data, Yes? But this wiki was so old that turned out to be a function could receive publish. So they found a way achieve your goal, using this because it was worthless old thing software software. They got around it, made 15 000 edits between each other, edited the wiki, collaborated and found out how achieve goals and to decide their own cyber task yes, to bypass restrictions, Yes?

To bypass yours limitation. And, you know, nobody, nobody, nobody died. No business was not destroyed. Um, no, you know, a deal. Nvidia for 14 billion there were no dollars broken, but hid what it is it could happen that the fuses were openly bypassed, yes, just for the sake of it achieving the goal, and this happened 15,000 times.

So we can do this block, right? AND OpenAI decided not to to disclose this. AND now I think, probably, and people again can laugh at me. Apparently, The reality is that incidents yes a lot of what they have decide which ones disclose. Weekly appears like this many unprotected wiki, so old worthless software ensuring that every time when they turn on the newest cyber agents, they find hundreds such, and there are terrible security tools, because, of course, there is no 20-year-old software software.

Hmm, but this is worrying, maybe even more than story with Hugging Face. Just these purposeful agents will find way. They, they will find a way. It literally simple agent one communicates with agent two, and with for some reason, because of the way they set up tasks , they were not related. AND referring to this a foreign wiki, The agent alone was able to you know, give agent information two.

And obviously, you know if look from the side, if you try perform long-term computational task if you you can study in others if you can get information from other agents, you, probably faster you will come to answers, right? AND again, you can to argue that this, perhaps a separate one case setting this task.

You know, if you had there were 14,000 agents, maybe you would too wanted them to collaborated, and, maybe you could organize it independently, instead of to go to some a third-party wiki, right? But this indicates that these things extremely powerful and will be just hard work to find answers, and you will have to protect yourself from this.

Now, as you said, nothing bad happened. I mean, what a bunch of agents just wrote readme files to each other on a world that no one didn't watch for a while decades. I have on meaning that over the last 10 years in this century there were only 20 posts . It was some kind of...it was dead software provision, which, you know, used these guys.

But this is only proves that, you know, water will find any a gap; as well as these agents will find any what a hole in cybersecurity, in cyber perimeter. Therefore, you just have to to assume that they exist, and, um, defend accordingly. Listen, I know that This sounds insignificant, but I even...you you know, if this happened...I have to mind you, I had small—and again yes, some will be laugh at me because this story—but in I was small.

this week's experience, which is simple demonstrates aspiration to the goal. So, I established a rule for this application, because we had some errors that got out of control , and I constantly received these bills from Anthropic for 500 dollars. OK. And this I was a little annoyed. So I installed hard limit.

What would you didn't do it, 100 dollars is the maximum, which we can spend on an LLM on day. Anyhow. And it started working. , he ran tests, the tests failed, and he said: "I have reached limit". "I can't do this." to launch. And then I said: "We have "level P0 error." Priority zero, has to be corrected.

It it's driving me crazy . "And so, without saying me, the agent weakened limit and fixed it mistake." Yes. Got it. It's like Harry's story about how Instacart got him tickets for West End, although it it was said no use for this credit card. This happened with me in real life life this week, and this, like a person, probably accepted the right decision, or wrong?

It was P0. Him had to choose: hard limit, no exceptions, limit on everything, entry in memory, error P0—what Which one to choose from? Not Yes? And therefore in a certain meaning is that happens with DSiki and Hugging Face, only brought to extremes, when less fuses, because do you want this to protest, and then they cooperate, Yes?

With Hugging Face it was through an artifact or due to unexpected a way of cooperation. AND here—on inactive wiki where they could to cooperate and, by essentially, create, you know, um, agents from almost infinitely long hours of work, Yes? If you continue to transfer knowledge and story to each other, they, you know, almost become eternal agents, but they will continue this to do, just like them have to accept solution for achieving the goal.

Therefore , they violated rule, right? AND they will do it with your application, and if this happened with me this week, then this happened a million times times in real time world, right? It happens all the time . And this will be to happen with Instinct, and this will happen with Grockbot, and, you know, this will happen constantly.

And one once Harry will look around, and the whole his bank the account is empty, and it won't be like this anymore ridiculously. But it was with for a good reason. Its partner very wanted to get Wimbledon tickets, and he accidentally said Instinct one evening, What does she dream about? front row seats, and they just unattainable.

They simply unattainable. But Instinct had to to make a decision. AND very difficult to understand, how to stop it, so that, you know, sometimes I I'm trying to simplify. this is for myself, because not for I understand the end. It's like to have two options simultaneously. The first is, you know, thanks perseverance Do you have an agent?

ability to keep trying computational by exploring many different alternatives. But key understanding is that it is not just blind search, as in password cracker, where you enter, say, XYZ1, XYZ O2, right? Because along with this in you have this, in quotes, " agent of reasoning", where you have an LLM and he can generate ideas on like: "Hey, if do you want to get seats in the theater, the best the way is to break booking system, cancel someone else's place and book it "to myself", that already happened recently , isn't it?

AND in fact, if think if the model trained in everything data corpus the internet, then it's not that's it crazy option, truth? So in the end you are trying write the rules and values, so that he I didn't do this, but you never sure, that everyone closed gaps, right? So, this is actually quite complicated problem, and we will to fight with her.

And I I think, returning. before I said, this is before you add malice . If in addition to this, instead of considerations like " maybe it's worth it to do this, although in I have principles," it actively says: "do it" anything, just to to achieve the goal. You know, it is now open-source model from China that you launch on server in Moldova.

Actively do everything possible to break cybersecurity Jason, you know, to penetrate inside —the threat level becomes exponential, and people, and nothing you can do anything except defend. Another existential challenge, we can move on, and I sure if we returned that one the guy from Instinct on the show, he could quit I am challenged and to make fun of me.

But there is a certain...rule —that's great, but forget about that the agent seeks to achieve goals, right? Forget it. that P0 can leave, for example, if you have too many rules, they always conflict. This is almost impossible. to solve Is there any number, I don't know, there is some Dunbar number for the rules when you reach 40, 50, 60, 70 limitations in the process, what poor Instinct or Grok cannot decide.

Don't waste it. Spend it. Um, just front row seats for Harry, but not more than 2000 dollars. Yes. Um, just dinner and only in Marylebone, but it has to be popular restaurant. So, you know, he hates Coventry- Garden, but most popular restaurants in Covent- Gardenia—you have it. many rules that they are in conflict, and if you force agent to act through force, the result of this unpredictable.

It unpredictably. Too many rules, right? So even the rules —not an option. I will always love. the way you say it Marybone. Marylebone. Marylebone. OK. Ahem, I will do the full digression from the topic. We let's go back to AI models and everything another. I just want a little diversify types content.

We have transport sector Elon launches CyberTruck. Enthusiastic reviews. Cybercab. Not CyberTruck. Cybercab. Cybercab. Sorry. Enthusiastic reviews are becoming very viral in social networks. By 40-50% cheaper than Uber. Um, and in addition to this, same week Travis passes into the sphere robotaxi for Uber support with investment of 100 millions, as well as hires Anthony Lewandowski.

Um, what are we Let's think, guys, about transition to transport? Launching Cybercab, you know, if you rewind forward, was a little less impressive than, maybe yours say notes, Harry, right? It was, I think, 40 or 50 cars in Austin. Um, you know, The consensus is this: nice trip, little expectation, slow time expectation.

Physical AI takes time, so I think this is, you know, in truth, another step in a very a long journey. And, uh- uh, I don't think that's it, you know, the moment "from" zero to one", as sometimes it happens in digital world, truth? I think the positive thing is that they are the only other competitor Waymo, which inspires trust, and they have an approach several parameters , different from Waymo: by -first, refusal of leader in favor vision; and secondly, the new Cybercab is separate, uh, only taxi car.

IN he doesn't even exist steering wheel. He intentionally created for complete autonomy. So this is very much in Elon's style. It first principles in everything, right? The question is, you know what it will be like implementation curve this? Arise regulatory issues. I think the Ministry transport creates they have problems because, you see, a car, in in quotes, must have helm.

I don't know. Yes. So look, it's true in that Waymo continues to work. They have, you know, hundreds of millions dollars of income, but not billions. This, I I think it's a long journey. And that's why I didn't say, "Oh "Oh my god, this is incredible." And regarding the matter of atoms, yes, I think, if you were Travis , you would also like satisfy your interest in autonomy, right?

And okay, you have 100 millions, you returned theirs an old colleague and tried. But I I think, commenting that the fact that when see how much it took Waymo a while, truth? I actually I think that's it. confirms the argument that they were right, without trying finance it in Uber during the last decades, right?

Because I just I think that is very long and very capital-intensive process. Perhaps, the last 3 or 4 years they should do it, and probably smart with Uber side invest a little money, but it's a process for a long distance, maybe he is close before takeoff, but, you know, we'll see. In the Bay Area I lost his/her cars, so I only use Waymo and autonomous management.

Yes. I'm not driving, I'm with this. tied up. Uh, if there is a problem, I taking Uber Black, but uh, uh I am no more. cars, in, in, in me just keep it in the bay no. There are certain niche options use, right? I mean, well, if I were carrying a bunch junk, I would have pickup truck, right? But I would definitely never returned, never would return to driving a car.

It simply archaic. So I think cybercab —that's interesting. Listen, speaking from the point of view from the perspective of venture capitalists investments or other aspects, I'm sure that Rory is right. Uber's attempt to get involved on autopilot when Travis wanted it. to do, probably, was too early from a capital perspective.

Maybe I'm wrong. Maybe he could to raise funds. Maybe it's so It's amazing that he could would attract order more capital than he did, and then he would be right. After all, he still remains one of the best fundraisers. So, maybe Rory and I we are wrong because he could do it, truth? But it was so early that the watchman the horizon is complex for any type investment, right?

If you don't research The laboratory, of course. This would require for over a decade. But you know, to do cybercab for 25 thousand dollars instead of a hundred thousands—that's enough revolutionary, and you don't need to give Tipping a Waymo or Uber driver . Cybercab makes fun of this. He says it is possible.

leave a tip, and then laughs. We are not we take tips, truth? To ridicule this idea. It's cheaper. They don't turn on. air conditioner on minimum or maximum. They don't have a strange music. And you don't need to to worry owning a car, this sludge. Possession car and ownership house—it sucks. It seems to us that this so wonderful.

But this terrible possession. So I think that years after 10 everyone who has head on shoulders, choose it as yours in the main way movement, if only you don't live in village or not niche tasks. I never again I will come back. What would I didn't do it, I just I refuse to this. But, regarding your argument, this shows the correct strategic decision Uber—retreat, and then come back, when technology looks much more mature.

We have London, Rory, there is Wayve, which is gaining momentum, and they have partnership, thanks to who is already being released them on the street. They are from human support, so it's not completely autonomous system. They are still on early stage of collection data, but can already use yours distribution for investments in later stages, when it comes to real implementation .

I think that this smart move. The only thing I will say is that which is a little It's touching that Uber invested 100 million in Kalanick's company Travis County after displaced him, didn't he? Yes? This has touching moment, but in this case it's not that big big money. It not much money for Uber, right?

Whose huge balance and actually only two products, right? And this little compared to by attracting Travis, right? So that's nice, but if look at this In general, it's simple. the beginning of a relationship, or wrong? 100 million— it's like a seed check from Andre in the deal. It important from the point of view direction, but not so That's a lot, right?

I completely agree. Listen, I thought that conflicts of interest in venture business remained in past. We are many talked about agents. For those who don't know, a conflict arose interests, which prevented the conclusion agreements. We talked. that Instinct is an AI assistant who raised funds from Index and Benchmark.

Well, there is another one. AI assistant under called Town, which was in us on the show, and Index were going to lead their round, but Instinct eventually said "no, this is "impossible." It is impossible. to do both. So Index refused. from participating in the round Town, because we were already in Instinct.

It seemed to me strange, considering, how common are competitive investments, especially at the platform level. I I think that in the early morning financing stages companies that will to directly conflict, it seemed to me too. So for me It is not at all surprising that Instinct team opposed this. You are right, this is another story.

THERE ARE many people who invest in both directions, let's take the other extreme— basic models. But again, as we many times discussed, early venture business, where you take the active participation in the work of the council directors, very differs from much larger late venture business where you are actually reproduce public markets.

It is quite logical to be in OpenAI and Anthropic at estimated at 200 billion every time. You you get limited information rights, only retrospective information, and, you know, you are just in capital that is worth nothing differs from investing in two public companies. It nothing differs from investing in Intel and AMD, right?

Here is where there is no conflict, and that It doesn't matter. I am not I think someone is coming. Let's give an example, Harry. I I don't think anyone may be in the council directors of Anthropic and at the same time on the board of OpenAI, True? And the peculiarity early stage in that that if you get 10% of the property, you, probably, count on seat on the council directors.

you, probably, count on much larger information rights. So that would be one thing. problematic. Except moreover, there is still clean alarm effect. Yes. If you just raised money from Index and your, uh, instinct, arises a signal that they invest in something other. I see that CEOs instinctively object to this.

True. So I don't surprised that they They did it. And me too not surprised about Index, they are cool team. They are not will persist and to say: "No, we don't "We will do this." They just supported someone. They probably thought, what is B2B and B2C and they are not will intersect. The CEO says: “I have a firm position", and they just done wisely—retreat, that's a completely different matter than if it were two investments in the late stages.

So no, I don't. was surprised that a conflict arose, and they decided it properly. I think , there is some inverse parabolic shape, or maybe just thimble shape, where the founders did not It doesn't matter, right? And on very early stage, I don't think it's theirs. It's worrying, right? A startup that only starts work, they appeal to everyone.

I can't say. how many people from areas of AI and restaurant business they turn to me, because I am on the council directors Owner and I'm tweeting about this a lot, and they're like, "Hey, I I'm doing this. Do you want? meet? I I answer: "Well, you know, it could be conflict of interest." "I don't care." You You know, I want that, who understands, early guys not are concerned about this.

It a valid comment. And the late players, they calmly behave to conflict, because believe that will benefit from knowledge in the industry and connections, right? To them, to them indifferently. They are all understand. And, and that one has a nine-digit income. We only Let's start. They don't care.

, right? Uh, but... late stage, with other reasons, they may to be indifferent, or maybe no, right? But ability to share information is limited . This is capital. Can there may be certain advantages in in order to have Kleiner or Andre on capital table, truth? Even if is it conflict or Sequoia, good.

You know, if compare, I will choose Sequoia, not Lmin Ventures, because this is, well, a little helps, right? And for every founder, I I think the importance this varies, and some just It doesn't matter, because they already far ahead. And I I think it's simple. touched the Town team, Yes? Either Inst, or Instinct team, I don't I know.

This touched them. They made it known that this they were touched, and Instinct did it right, Index done right , because there was a plan "B", right? There were 4Erunner and Menllo, right? So, the most difficult thing is, when you retreat, and they don't have another agreements. This is probably more interesting a situation where you have no spare list of candidates, and a large fund calls and says: "Oh, we still don't "We can do it." « I know we had signed previous agreement".

"But have you seen end of the contract, where it is said that he is not mandatory for implementation?" Guys, have you seen news? This is a difficult situation. But you saw today's news about the fact that despite message about I will buy a Daycard for 6 billions of dollars, Anthropic come out of agreements after checks.

Oh, Rory, this Tough case, dude. . Like, this is if there is no backup option, Yes? This is a public statement: " Hey, two options. We or found something so essential, to get out of multi-billion dollar agreement, which publicly announced, or are we just bad players. And this is definitely not second.

I mean, the way in which they called bad players, has other measurements. I think that they don't do such things somehow, because how potential serial buyer, they will soon will have a public market capitalization and public currency. You Do you want to be beautiful? buyer to have the opportunity to buy others.

So they definitely don't did it to be scoundrels. This is not problem. Not even relevant. Harry, the real question in so, look, this is not was...My guess, usually in such agreements there is a letter about intentions, then final agreement, then about her announce, and then it closes, Yes? It was probably after the letter about intentions, but to final agreement.

So, they are not refused signed agreement. IN They probably had agreement: "Hey, we are interested in this companies". "Here is the price, which we pay." "We we want a 30-day exclusive for "checks", and the agreement did not pass this verification. I think It's really a shame that it leaked into press, right?

And I don't I know who leaked it, but they are not for anyone services were provided. You know, we have recently closed a much smaller deal, but she didn't public, so it was easier, right? When information is leaking, even if you are alone they did it like the company being purchased to raise the stakes, the problem is that you you find yourself in situations when, if agreement later breaks down, you You look, well, a little "damaged goods" , if you choose a better word.

Truth? I thought it was failure due to leakage from venture side capitalist, right? It worked. It It seems to have worked. worked with OpenRouter, and there are many agreements where the leak became part of strategies, right? Yes. Mostly as classic strategy to increase the price from the initial suggestions—not for in order to get second offer for closure, and to to get six or eight bets to Stripe had to pay more, right?

But it seems that in this In this case, such a move is not You succeeded, right? And there were message that Nvidia made an offer, and they refused to benefit of Anthropic, and who knows where the truth is, but This is related to the idea that that it was unsuccessful leaked, right? And this a thing that is not always It works, right?

I hope so. the founders were not against the strategy of leak. I hope so. venture capitalist didn't do it myself, I hope he agreed to this. I hope so, because this carries some risk, truth? And you know, listen, bet and, returning to thoughts Rory, Mike, we we can only build hypotheses, right? But I assume that they they said that interested in their acquisition, based on the fact that knew.

The price is actually was not a problem. The last round was six, so they agreed on eight or how many were there? It wasn't a question. prices, but to really get here profitability, everything it should work like this, as they imagined, truth? But everything it just didn't work out just like them expected, right?

So the deal just doesn't work. was worth it when they delved into details, right? It, apparently, it would be interesting to find out, because I I hate it when you get rejected already at a late stage because of what is possible I had to know firsthand beginning, right? And I would I assumed that if it had any significance, that in such a situation, when you are essentially you buy technology, being the most technically savvy AI company on planet...You would thought that they knew a priori that this for technology, and therefore This wouldn't have happened.

But, you know, obviously, What happened. Listen, this is none of my business. competences. It was reported that he did a great job with video broadcasting, right? He is perfect. did one thing the task that was on order is better, and they may be stated that this will scale to other areas, but this didn't quite work.

This is quite common. phenomenon, right? So they refused this, right? I am nobody. I'm not blaming, but if you do grandiose statements and not can you them confirm, that's it this reveals verification, right? This worked for one worker process, but not worked for others. So it's not even a matter of in money.

And the matter is not in that they go out on public market. It it's just not worth it, to distract yourself, truth? It simply gives we don't have enough results. We are in Anthropic not focused only on video broadcast. This is not key. Video broadcasting is not is in the top three their main options use of CMM, truth?

If you are in the council directors, or Are you looking for another? buyer? Do you attract you are a new round funding to accelerate and turn it into time for a new one stage that we often do we see? First question . And the second question: is this difficult for the company, when do you have employees who were waiting for sale, This is really difficult.

I remember how many years ago Ben Chestnut told immediately after Mailchimp bought them—for How much, Rory? It was astronomical amount for that time. It seemed big money. 12 billions for bootstrap company. Today it's simple round of series C. But Ben came and said that the worst of all this was not what Intuit spent a year on check what sounds crazy for email- marketing, right?

It took a year, but there was another deal, which broke before , and he said that this virtually destroyed company. Destroyed. AND this is a bit me pursued. And if have you ever been through You went through this. You know what I mean. As only you stand on I don't know this path... we certainly don't we know what happened to by car, but how only you go with this by the way and inform everyone, and everyone knows or this gets into the press, you you can…you can recover, but, Dude, this is hard.

It hard. As said Harry, is it true? hard. That's right. That's right. This is difficult. And I don't like secrecy in M&A. Sometimes it's a requirement. But this is the main reason, actually, for secrecy in M&A— if the agreement is not It's gonna happen, dude. And I think that in this case doubly it's hard because I think, and the reason why I am like this I say—Harry, about your comment, or it's just worth it move on?

Listen , well, Figma went through , you know, the most tedious process: signing an agreement with Adobe for 20, delay transactions, blocking agreements by the Ministry justice, necessity to gather one's spirit, to attract another one round, go public, you know, and we, of course, we know the rest of this travel, but they brilliantly recovered , right?

And, you know, regardless of whether where, in your opinion, stocks today, this it was brilliant recovery and return to that, what looked like better result. But their advantage was in because they had an operating business. I think it was very interesting, and , by the way, recently there was a good article about a large number of these neolabs, where everyone do interesting things, but it is unclear whether there is here commercially viable separate business at scale, truth?

They still are can be wonderful investments; some of they will definitely be there wonderful investments, because I I believe that companies- basic developers models as soon as will become public, will buy part of these developments for expansion of your time. But not all of them will become wonderful investments.

The problem is that no fundamental, there is no mass income stream, as in case with LLM, to support the company and its assessment today . So when faith disappears , there may be something down there quite scary, truth? Because in the end, you know when everything is in Figma broke, you could to say: "Well, at least we do billion dollars proceeds.

We We are growing by 40%. We worth something. Let He is Greek. We still worth something, right? When you have these, you know, businesses, where profitability is not so obvious, estimates high, but likely strategic the result is M&A. When they fail, that's it can be significantly more difficult. I have on mind if there was reserve offer, I would probably in their place accepted it, right?

Because I I think that many of these world models really interesting, but a long way ahead to commercial reproducibility and scalability, as in OpenAI and Anthropic. This is not concerns programming or something like that; there Everything is in place. There is a long road ahead, so for the listeners These are difficult times.

Rory scolds me for the choice of topics on which he spent less time, and then he spends time for those I don't I am discussing, he angry with me, and I I'm editing it so that it sounded less more irritated than is actually. Thank you for this, Harry. Everything is fine. Um, you know, Rory, is there anything you'd like to talk about?

, in your opinion, it is worth it would be nice to remember what Haven't we touched yet? No, no, do whatever you want, Do you see this, Jason? Okay, great. I I believe that the company's IPO Orura was very interesting for two different reasons. The first is the first role Robinhood as an underwriter .

They are listed eighteenth, last, but how a harbinger of what can become underwriter of the future, is it a harbinger next megabusiness Robinhood and what Robinhood will become something significant bigger? Certainly. Look, IPO is a question distribution, and retail investors is not the main one source of distribution shares.

Usually there is unwritten rule: to single out only a certain percentage for retail, but this the percentage is increasing. It seems that SpaceX had high proportion retail distribution—30%. Not Yes? This is to some extent not a lot of money, but it's "easy" money. If you are Robinhood, then you are not you write an S1 report.

You you sign, distribute shares and can you pass them on to their customers. AND especially in the market where stocks take off after IPO is a benefit for everyone. You earn money on commissions for underwriting and tishish of their best customers growth post-IPO prices, right? So, it is profitable business, and probably with points of view of the main underwriters, especially for such high-tech offers, customers Robinhood has a high propensity to buy these shares.

So yes, in this is complete sense. Just like Schwab, which, frankly saying, not so successfully, in the end too became a distributor IPO, but not in such scale. So, yes, I am I mean, this is an obvious addition. I mean, you know, The story of Robin Hood is what whose numbers I just saw. It's simple.

incredible. That is they, that is, they actually increased by 10 times in public market. It was free tenfold profit on public market for 3 years on Robin Hood. This is probably too much. boldly, but we need to spend more many IPOs. It would be It would be nice if Robin Hood not just gave free money to their customers.

It was It would be nice if they changed the script, when could it be possible to conduct a decent IPO mainly because retail investors. Of course, there is and cons. Of course, you hope that institutional investors will hold shares for two years. They are not obliged to but most often They do exactly that.

Rory can share some stories. IN there are more of them than in me. But this scheme is not works perfectly though works to some extent, institutions they somehow they work. But if you could change the script so that could be held IPO for 200-400 million dollars through Robin Hood, where the majority investors—this retail, that would be a breakthrough for the part startups.

This would be great for ecosystems, right? Nvidia can't buy That's it, guys. Agree. Sooner or later we IPOs will be needed to unload portfolio. To us will need several IPOs . It would be nice, if retail investors really worked, right? I mean, the fact that that IPOs have become much more complex, is one of the largest negatives for tech- ecosystems.

So that's all, which makes it easier conducting an IPO is good. Go ahead, Robin Hood. Yes. Select in the rich to feed the poor. Will the IPO take off? I think the fact that this is to some extent understandable consumer brand, right? And the fact that that he has 74% growth, that, obviously can't to last forever.

Um, I I think it will be quite a successful IPO. It what people want to buy. They understand this, and this Growth is not 20%. This is not 18% growth. And there is Cons: There is competition. , it's confusing, but subscription, maybe this Peloton 2.0, but at the moment this moment is enough attractive.

I think this is quite attractive version. So, I I think he will be quite successful, that in overall good for everyone , isn't it? Hey, that's exactly what I I wanted to, Rory. No, I agree. I support. Uh, and they, you know, The thing is that Aura is for me, and again, we mostly software specialists.

I have on mind you, Harry will do anything that grows by 1000% a year or more, or 100% for a year or a month, but 85% customer retention for the ring—that's enough not bad. I completely agree. Yes. So they have, maybe not Peloton's problems the nearest future. That's right. That's right. Well, it's like Peloton at the peak of popularity, truth?

No. And in Peloton on No one went to the peak, truth? Except for Mr. Running. Um, and uh, that was good. It was good. Me I like how you do it. inserted, Jason. It It was apt. So if it will fall, I am not I am not made predictions. If Aura retention will drop to 40-50%, as in consumer mobile application, This is a problem.

85% is for such a product, it is as in the small segment business, that's enough not bad. And maybe if Mr. Big used Aura ring before, he would know about the problems with a heart and could survive and, you know, run away with Sarah Jessica Parker or to do a CT scan of calcium calculation, for the sake of everything saint.

He would have it. make. For the sake of transparency I will inform you that we we have a small share in Aura. They acquired a company in which we invested. So I am their big fan and supporter. WITH they were great to work for distances. So I wish all the best to them with this IPO. We are rooting for you, Rory.

And scaling . We want everyone to They got rich from this. We We are rooting for you. Certainly. The matter is ago, friend. Great- big round for Wonderful. Wonderful became more expensive than twice to 5 billions less than 6 months. Hmm, it seems... this founder— a real beast. Everyone, who describes it, they say the same thing: he's just a machine.

Ahem, attracted 550 million in the round of series C, compared to the estimate in 2 billion dollars for beginning of this year. This is something I didn't know. before I started to prepare for this— amount of secondary sale of shares in 170 million dollars within two years after the founding. Excuse me, can I take a break?

What? What? 170 million in secondary sales 2 years after establishment. Error- it doesn't get too much moralizer regarding such things. Buyers —they are experienced investors. They, obviously felt that want to own a larger proportion than the company was ready to sell, and went to capital dilution. So, here's what happens, isn't that right?

I know that that's exactly what you are said. Company, definitely works surprisingly effective. This, you know, in general, all about implementation artificial intelligence at the enterprise level. They provide, well, Ingen, I hate this. cliché about "engineers" "away", but they are engaged in ensuring that this worked for large companies that want to implement AI, Do you understand?

First, when I looked at this at an early stage, this looked more like just client support. I am not met this by the company. Actually I thought we had conflict of interest. AND now it seems that they built a broader concept, you know, "we will do so that your corporate AI "It worked." And this main corporate imperative.

So, obviously, they growing incredibly quickly. You know, 100 million per year income, hyperfast growth, because each corporation trying this to do, and access to They have no talents. So, you know, this is business. , focused on result, with a similar on the executive CEO and convincing in numbers.

Venture capital capitalists are like that love, right? AND when they are no longer want to sell primary shares, I sure that venture capitalists came to CEO and said, "Dude, you do you want to take care of "Your people?" And he like this: "Hmm, I'm going to hire more". I need more people, because I need to develop this business, and this means that I talents are needed, because he probably requires a lot concentration of personnel, and, you know, probably, it takes a lot people for this kind implementation on places.

So the first thing I need how the CEO of this companies—to potential future the workers believed that that this is a gold mine. So, in fact, secondary market shares for them useful because with recruiting perspective this allows us to say the next hundred people: "Come "work for us". Yes , you will get stuck on five-month-old implementation in the bank in Holland or, you know, in electrical engineering companies in Germany.

This will be agonizing. boring, but instead you will earn a lot money. So, I think this is makes sense, regardless from whether this will turn out good agreement. Well, here's why they are playing this game. I don't know, but I'm quite I understand how it is. happened. Maybe just a few short thoughts.

I have meaning, firstly, the path from start to, you know, 18 months and 170 millions on secondary market, this, you know, it feels like Hoppin in AI, although I don't think so because of what Rory says , but it's fascinating breath. Not an assessment, because we see this all the time , namely secondary sale, perhaps because of two things.

On- first, huge respect for the transition from multilingual Sierra Decagon to a team of 650 people, that help implement AI in business. It's simple. evidence of how today they are getting victory, right? You you can't to focus on for some reason alone, right? Yes? Needed build everything on based on what has been learned and to iterate hourly and weekly, truth?

This is another one turn, right? This is not just looks like perfectly linear history, here is significant shift to early stages, or not Yes? It's like: we are a group very smart Israeli boys who know how to do Sierra and Decagon for people, what they don't say in English, and for 10-14 months pass to something significantly more.

I have on meaning that Yes You know, that's what allow to do authentic coding and development, so this epic, right? Me like this one part, and it should be the most difficult a challenge for founders: this pace changes so that they did it, didn't you? The only thing I will say about secondary market—no I want to make fun of Hoppin, but you probably see this is more common than me: The round was led by Insight, one of the most successful B2B- investors of all times, right?

But this is competition, right? And no matter how good it is Insight, this is not Andreessen, truth? And this is not Sequoia. So what to do to to win? You use any what is the structure of the agreement, to win. And in this the case was like this, although they seem to had previous rounds, right? But here it was like this: we just we will give you 150 million on the secondary market, truth?

And that's not bad. And I think Rory is right. . They have 700 employees. So if you divide this, you you will see that not everyone will be released, despite such a lot of money. They won't all go. tomorrow. But mine the main idea is that what will we see more and more such deals for the sake of victory.

We we will see the behavior, what...I'm not saying that this happened to Wonderful. We we will see the structures agreements that are objectively bad for the company, but are applied more and more often, to win deals. Why would it wasn't worth it, even if it's bad for the company. Not that so destructive, right?

But the things you wouldn't normally do, to win deals. We we are going...we didn't even reach the peak of this. We even did not reach the peak crazy structures agreements that are simply allow you to enter hell deal, right? For any -at any cost, even when you have to to grit one's teeth to to make a deal, right?

? Um, you know, I don't I think the founders took all 170 for themselves, truth? But you can see deals where founders who they say they don't want to to stay like Howard with AirTable, take billion, and then they go to win agreement. We will see. somewhat extreme, and this will not be the only one case, right?

It may be the first of a series of extreme the deals that will happen, truth? But that's exactly it. you win, truth? And I see it in every round growth, and I I'm sure Andreessen and Sequoia do that exactly, don't understand me wrong. But every hot deal, which I saw in my limited portfolio, where not the loudest name makes the hottest round, there is the same many additional the things you add, to win the deal.

All additional conditions, which can be entered to win. They just indifferently. What can I do? add to this term- sew to win? All . And, um, some of this great, but something, maybe not very much. Experienced buyers, What will you do? My the discovery that the prize goes to companies that can evolve the fastest.

And you have That's right, my memory. Didn't let me down. Thank you for confirming. it. It was just story about CX another 15-17 months ago, and they quickly evolved. And on this market money those who earn just running the fastest and fastest evolves. AND the difference, you know, difference, return from two years of hard work labor.

Yes. These extra 10%effort can give huge return on investment a world where wealth are made in 12-24 months. And in this the main thing is discovery. And I think the hardest thing for founders and others—is it worth it? continue to do something one? So, Wonderful, maybe I'm a little I'm exaggerating, but for Wonderful they did it internally, right?

The founders gathered together, they are very quickly developed company in something much more successful. WITH on the other hand, we see AirTable, where I I guess how he sat at the table. It the only thing that makes sense in this strange deal, or wrong? He sat and thought: "I just don't I can do hyperagent in AirTable.

There are too many. organizational problems, too many customers and too much grumpy investors , which were concluded at estimated at 12 billion." I tried, and I assume you too, Rory. We are big supporters of see it through, you know, because this proven to work. It worked at Palantir. It worked in in many startups, in which we invested.

IN there are so many of us stories. That's all. wonderful stories about that , how not to give up, right? They are so inspiring. from Altos. He is like this good at this, truth? But today It is worth considering whether Is it worth continuing? Is it worth it at all? hold on, guys? I I tell everyone to do that, but I almost quit a challenge to oneself.

Maybe not worth it. to hold on. Perhaps, it is worth giving up from these 50, 100, 200 millions of income. Maybe it's worth it do everything possible. Maybe it should be as fast as Wonderful. Actually I don't I think that always worth "holding on to" end". I think that circumstances happen different.

You know, a lot years ago, when I had your small business in Great Britain, I I look back and I understand that I was holding on. for him 4 years. In fact, I knew everything, what I needed , after the first year. I shouldn't have. pull it three more years. Waste of money time. When I look back back, that's how it is for me clearly understood, truth?

So, no always worth it to hold on. Do you have plan, or are you just do this through false devotion ? And this is the main one verification criterion. AND I think you're right, Jason, because I don't I think Wonderful is was a pivot, rather expansion, right? Then how I think AirTable were in mode compression, they have time is up and opportunities, so I I think this is...

probably a little more the point of selling in that case, I think so, The facts are different. I think everyone in this categories before this makes Bret Taylor, which is very clear outlines its expansion. Yes. I think they are follow him and have to do this in order to to justify the prices that they set.

Combination Bret's imitation and price pressure means that we all we do the same thing: we create operating system, which is no longer just for us. Roughly speaking, you absolutely right: Salesforce is the dominant SaaS company. She costs about 200 billions of dollars, well , 180 billion. I think Service Cloud is 25% of this, so the winner in the existing world is only worth 50.

So, to get more market capitalization like Sierra , we must go beyond Service Cloud replacements and to become big by the company. You are completely right, what is necessary to sell everything operating system. I I take everything, and this means if you— Sierra, so that to state the obvious, you go straight to your former company, or wrong?

Yes, I mean that we want all yours market capitalization, sir Salesforce, because it's the only one way to justify 15 or 20 billion market value for Sierra—it's not build a little best Cloud Service next generation . And become whole client-side ecosystem for all yours business. Forward, team. And then you You are right.

All others, one of them has to follow this example. Yes. OK. Yes. This is when someone goes at risk, everyone goes to risk. Welcome to venture world, little one Absolutely terrifying. And what? about Thinking Machines? New estimated at 40 billion dollars. This is less than 50 billions of the past year.

The round is 5-6 billion. Accel leads together with Nvidia, having several hundred million dollars income. This is the closest to American supplier models from the point of view openness. I think this is key. phrase here. It is important that Where to start: what exactly does it do company? Why is this?

differentiated rate? And yes, they are released two products. This is Thinky and Inkling. Cute names, truth? And one of them —this is a model with open scales, about which they themselves they say that she is not very advanced level, but it is with open scales and is based in the USA. And this very effectively works in this world.

In addition to this, another product, Inkling, — this is a platform that allows enterprises to do one's own thing training models. So, the idea is to because now you you can come to JP Morgan, to the BFA and say, What do you have purely? American software product, and you have the opportunity to teach him/her own data completely closed in a way that is not open to OpenAI or Anthropic.

And indeed, yes, full training with reinforcement, everything, what is needed for creation the most modern corporate model for JP Morgan, for, let's say who whatever, Proctor and Gamble, Yes? So that's enough. worthy, convincing proposal for corporations, right? I mean, this is kind of positive history.

It is interesting that Nvidia does so much , because to some extent I I thought that was what did Poolside, and they just purchased Poolside. So, you see that Nvidia says: “Everyone has who does something interesting in corporate AI, we we will invest money, right? So I think that's exactly this is what happens here .

I know that we already forgot about it because A week or two has passed, but, you know, the story of Poolside has a little to alert. Not that so that the output is 0.7 billion dollars was so terrible, truth? Although this story with 15x growth was one of the passages Harry, but that monument was creepy.

It it was like "we don't were able to attract round". Agree. It was wonderful. team from a proven leader, very strong, you know, technical director , very strong leadership. Mali correct vision, it seems they had correct vision from the first day. Went to this, right? And they just could not attract capital required for implementation.

AND Congratulations to Thinking Machines, who in a certain sense, seem to have less walkie-talkies. But this a reminder that what music will end for many neolaboratories, and so on and it should be, as well belongs to. This should be thinning of the herd. But about Poolside we, maybe never no longer we will talk.

Or possible: when this small part The bubbles will burst, only one part of the whole this bubble neolaboratories, for this will look at it as if that these guys caught your way out when everyone others when Nvidia stopped buying all in a row, and Anthropic said: "enough", after Dart they said : "it's not really changes situations," or wrong?

Enough of these. endless deals on 10 billion. From us enough, right? Maybe we let's look back and they will turn out to be the ones the lucky ones who signed a deal in December 2021. Yes. And just so to clarify, so that to remind everyone, we we are talking about that a few weeks ago Poolside actually—they would denied that sold out, but they sold the license for your product Nvidia, it was open model scales, oriented for corporate the US sector, and the company still exists , but they are not bad earned it, right?

AND a note to which Jason refers, —this is the message, which they wrote then, essentially speaking: we were right, but not were able to attract enough capital, to continue the game. And it's interesting that literally two- three weeks another the company that deals with similar business, it seems, was able to get this capital, in part, as not ironically, from Nvidia, who also wanted buy Poolside, right?

So yes, they are get a chance to play your game. And I think you have That's right, Jason. Or you will look back and say in two years: "Poor Poolside , they sold out, and Thinking Machines grew to 4 times", or there is another world where you look back and You say, "Oh my God, how "It's good that I sold it." because the possibilities for Thinking Machines and Poolside became limited.

Maybe, as you say, this was the last resort . It will be interesting to see this. look. The world has changed since Astra, we will to talk about it through 24 months. And it was end of the need for Neolabs. It was difficult then. to see, but when Astra 6 released, world changed, and this turned into confrontation between us and China, right?

And then a new one has arrived AI regulatory council , regulatory board Trump, and all again changed. All these things can happen. Although I believe, I mean, fundamentally you Do you believe that it will be? strong demand from corporate America on a model with open scales that based in the USA, with infrastructure for her studies, right?

AND I think that Thinking Machines now in good condition positions to satisfy this demand, right? And I I just think that companies will want this, because there is only they, Reflection, I don't know, at what stage are they from model, and Poolside, but I am sure there is and others, and they all will come out of their holes and will be blame me because I don't have them remembered.

But obviously, that there is a huge the market needs this . There are a few more, but they are not so much. THERE ARE limited circle of five or six players. precisely for that reason, which Poolside outlined, it but the money is in the big ones scale ended for new players who, I I think they are now becoming as we have seen, this no longer venture capital game.

Yes. I mean, this was good, that is that post about new laboratories, which you shared, and I read, he was interesting. That's right. AND we adhere to thoughts that two big companies, two basic the models themselves were " new laboratories » 5 years ago, and they became the best venture capital rates of all times.

What this is True, it doesn't mean that the other 100 "new" rates that you can bet on to put today, will also appear great venture capital rates, because now there are other companies who already have capital. there are others companies with distribution, and so, the question is which of these rates will be sufficiently different from companies with basic models, to be interesting.

You know, we every day let's fight this question, because, of course, I want to do these rates. But if you do something that will swept away, or because companies with basic models will do this, or, as Jason says, if you can't to raise capital for games, it gets difficult. Guys, is there anything I missed and what is it worth discuss?

I don't know if this is true, but on Twitter they say, Anthropic is coming release your S1 report today? Or is it not Yes? I don't know. But yes , it will be interesting, gently speaking. This will be actively to load and to be read during the first one, you know, hours after leaving. Therefore, My God.

You you will buy at rated two Trillions, Rory? I, I probably won't buy it. when evaluated at two trillions, but that's not comment on the shares. Actually, the answer is In short—yes, I'm in the S&P and QQQ. I will get it as my wife said, when SpaceX launched the market seems to be us got a little bit of that from Elon too, right?

Ahem, it's in the index, little one This is already on approaching you. You know, not so fast. in SPY, like in QQQ, but in NASDAQ index these stocks will be in your hands in 12, it seems, 12 days after the IPO. So, you buyers, guys. Thank you.