Mar 13 '26, Friday
Mar 14 '26, Saturday
Mar 15 '26, Sunday
Mar 16 '26, Monday
82:04·2 SPEAKERS·10607 WORDS
AYo bro.
BWhat's up bro? What's up?
AIs this the new house?
BNo, no, old house only. Maybe this table didn't exist before but. Oh, this may be new. We can't move now, bro. If I will move like within New York though. Impossible. We have too much stuff. If it will be. It'll be like a big move. Either out of the country or like to a new city or something within New York. We're not moving now.
AFocus and everything and abhi I'm still. It's been like nine months since we moved.
BYeah.
AAnd I still feel like this is like temporary.
BWhy you you moved? No, that because of the noise thing you said.
AYeah, we moved from a two bedroom to a one bedroom and I think that was a mistake.
BOh one bedroom, two little space. Yeah.
AAnd one bedroom is a no no go have been just waiting to move out.
BOh you're moving moving again now?
ANo, I mean it was like anyways
Bokay.
ABut crazy about every year they increase rent by like they could increase it by 20 or 30 also like because market is so efficient for them the best revenue strategy is to jack up the rent.
BYeah.
AThen 50% of the people will leave but they'll find replacements next day.
BI see.
AAnd
BI mean 8, 8%. But yeah, 20%, 30% is a bit much.
AOkay.
BI should have known. Like I mean I kind of chose this like I, I, I knew but like AI has made it so easy to get shit done that I feel like I want to do more shit now.
AYeah, it's like the job is too easy for like the ceiling is too low. Right?
BYeah, it's. I'm blocked more on processes rather than actual work which doesn't feel nice. Like process of getting reviews, process of giving proof of presence. Like I need to like for every PR that I have to merge I have to touch my UB key. Like what the. On every new comet. Like latest comet needs to be like human approved. It's a bot slow bro.
ARight, okay.
BStuff like that.
AIt's interesting bro. I spoke to them in person on Saturday. I was there at their place and then like yeah sohe and then Sebastian like sep who's there on the email thread. He's the co founder at Google.
BYou can search Sebastian. What?
ASebastian Logo Google. He's from Google. He just quit and he's like I think leading tech core engineering or I don't know some infra something the more technical person I think.
BOkay. Latam access control director.
AI don't know bro.
BI. Oh I found him. He doesn't have a picture, but yeah.
AYeah. So ye and Rosie is the whatever pops HR person.
BOkay.
ASo yeah, I think I just said you are interested in infra compute. I don't know. Like I just said he is some interesting. I don't even know if you're interested in it, but UNGA is a project and everything seems very, very interesting. I think it'll be an impactful company.
BYeah, you told me. Businesses compute provide karnaar kind of like what SF Compute does but on a bigger scale. Right. Like on a. Yeah.
AIt could be like technically if it works on a software or compute level it can be any scale. Right. It can be like individual all the way up to. To like build up people like think about like I mean like one concretely say at DFL we want. We are like training some like new video models, some compute player so 32 nodes, something we'll run like many different architectures of the same parameter size. Maybe we'll do some scaling study at smaller scale like 1B, 2B 4B. We'll train a model. We'll like study the scaling and all of this will like fully figure out the whole setup. And then we'll probably need to just scale up to train like a 32B or a 64B for two months. We need a lot of GPUs like five like big three training run. But once that's like done we don't need that much training compute. Like we'll just then do more exploration but then like. Right. But we are still bound by getting like a fixed GPU contract in advance. Then if the GPU is available in a market. So we, we have to sign up for like a three year or a two year contract of like 5,000 gpus.
BOkay.
AAnd then we have to like. But our workload is too varied sometimes we want to scale up training and then we still try to keep our workload such that every time we need 5,000 GPUs. Right. Maybe process more data. Been like to some you'll find work
Bfor the GPU rather than the other way around. Like find GPU rather than the other way around.
ABecause it's like monetarily it is like the biggest expense is that we spend like 95% of our funding on it. 90% of our funding on it. And then we would find work. Basically we basically plan. Now it has become like. It is like deeply engrained in our thought process and working process. Go pre train run endo next model data process. Yeah, we'll ramp up our research More like as soon as the pre training run starts, we'll ramp down our research exploration. But this doesn't. This is not like this. I don't know if it's like it's probably not optimal way of operating. Right.
BSo but yeah, but then what like what is the problem with Abijo? Rental services are too expensive to buy like on a per like month basis or something.
AYeah, per month will be subopped. They'll like charge much more like okay. Like they like the GPU providers want to lock in their revenue. Right. Like they want like they like they're not. They don't want to keep the GPUs. They don't even want to risk them like staying with them for even a month like if you give it for two. Same reason why real estate people don't do like short term visas. Right. Okay so and then this is just a training. This is like a say a training hypothetical. On the inference side, I'll be like a customer came to us and said that we want to run 100 QPS load test.
BQPS? Yeah.
AQueries per second, 100 images for second. Just Amazon. Yeah, much love. But that comes and says we'll use your image generation model. We want evidence that it works at scale and then maybe we can use on demand just for the test and then that'll be expensive. But then say we successfully convert that customer. You need something related like thousand GPUs more for the rest of forever, right?
BYeah.
AAnd then we have to go negotiate for that contract.
BYeah.
ASo inference maybe it's hard to lock in. You definitely need some overflow capacity and say you plan for it in advance and some other sexy model came out inference workload maybe. Then what we do is we switch them to training cluster and we say that or research explore. You plan your workaround around this. So there is always. So basically there are hypothesis. Every company is facing the same issue issue in aggregate. It's probably looking a little bit more like smoother, right?
BYeah.
ABecause everybody has higher amplitude or like.
BYeah.
AIf you superpose all of these waves is probably like better.
BYeah.
ASo they can have like a very good software plus hardware solution that you just like it's low. Like. Like it'll be like a no brainer thing to you say you have a software solution where you submit the job, it gets gpu, it runs and it gives you the result in inference. You try to run a query, it runs on some GPU somewhere you don't know where it is. But then you are basically only paying for what you use setup and. Yeah. So they are basically building like a software solution for this. They'll kind of procure like many different good like data centers and all. Or they might like even build their own data center. If you think about going it like going below the stack a little bit. Like you can build your own data center.
BYeah.
AAnd then they're also thinking it could be a way to also invest. I mean I don't know how much they're going to explore the investing angle. But like if a startup. You see startups like GPU usage, right?
BYeah.
AAnd then lot of promising startups don't have compute. So you can promise them compute and then try to invest in them. So there's a couple investing angle. There is a very good like software layer angle, hardware layer angle and also just.
BYeah, my only thing is like isn't this a. I mean aren't the companies already doing exactly this like just a core view is that at the data center level they are not like trying to build a market of sorts for
Athis core view is not fully like software integrated. Like the. The software stack on top of their data centers is very like lightweight. Like say BFL goes to Core V, we have to sign a 2 year 5000 GPU contract that is there. Like it's not like self serve bank connect. That's all like extreme but like it can be as simple as that.
BI see.
ASo basically the example he gave is business in Google. Google is not doing it at per project basis. Like Google has like massive data centers they have built. Just think about Google's internal use case. Like they have like massive TPU data centers everywhere and then like frame like and they have some say massive compute budget. They have like a million TPUs or something. And then they have like demo using the same compute DeepMind using the same. Some small team in Google images that's doing nano banana inference or something using the same computer.
BYeah.
AAnd all of this they have, they have really. They have a very good software hardware like solution to solve this. Like where they. It's like basically like a market system. Like you can request like and they basically say that there is no equivalent even though the data center market is like. Or GPU market is so like capital intensive and so many things are happening. There is no equivalent as good a solution that Google has elsewhere.
BInteresting.
ASo they can just build the same solution but like for like. So basically Google may say hire like project lead vp co tag. You have like 1 million credits.
BYeah.
AThey can see like some like dynamic
Bpricing yeah, but have you seen SF Compute like go to their website right now. It is exactly the same thing. Maybe the scale is different. It is literally a marketplace for compute and you can just buy like buy nodes. CLI example.
AYeah, yeah, this. I think once it gets to all of this then you should like ask them like I think like I don't know like exact market dynamics. I can imagine Aaron doing it at like serious. I think SF Compute is just like a bunch of like small, young, young folks. Like these people will probably pull it off at like at a very high like at serious scale, you know.
BI see. Okay.
AEnterprise scale.
BI have other companies also which are similar but okay, I'll. I'll probably ask them directly. Like what is. Yeah, the different.
AThe problem space and then like the solution space is like Google may. There is a system. I don't know how much, how much info about the system is available publicly, but they are like basically Sebastian worked on building the system and he has done like a lot of like cool like things and he just tells me that it's like he told me that it's crazy to me that there is no equivalent like outside of Google for something like this for like say GPU data center stuff. So I don't know how much they're going to attack it from just like the software problem. I think it's good that there's like many different lines of attack and it can be a good business even without that. They can just. Worst case they can just be GPU resellers. But they don't be that of course. Yeah they can be like
Bokay and, and. Should I study something before I talk to them? What would you recommend?
AI think just like I don't know if this problem space is even like interesting, you know like I think it's
Binteresting you can like SF Compute had a nice like they were looking discussing like they had some blogs about efficient allocation of compute key. Like how do you like given this is your compute, how do you cut it into pieces across a 4D dimensional space? Basically cluster is in three dimensions and time is another dimension. So how do you chunk that up? That seemed interesting maybe that I think
Athose things like I think one is like just thinking about what like this software solution or marketplace if SF Compute is the closest. Like just think about how would you replicate SF Compute if you go into that line of thought. Okay, that is already interesting. Just these problems that frontier companies will face like in like spike of workloads for training or inference and really designing all our workloads. Around a fixed reservation. That thought process is interesting. You can SF compute. You can think then asking like how is it similar or different? If you find something about Google's solution that's also interesting but generally thinking about it's a full stack from data center, like actually building data centers. It could be as ambitious as building data centers and managing it with the crazy good software solution. Like there are people who just think about how to build and do not think about the infrastack how to get compute, what is compute, who just don't think about it and they know for sure that their money is utilized in the most efficient way because it's so inefficient. Now like we raised like $300 million, we give like $250 million to compute providers and it's like mind boggling that it's like so inefficient efficient. Like you there is like you can just create that efficiency in spend like it's like such a high amount being spent but there is such a race to build it like inefficiently we build karoto to win this war. People are fine like building it however but there is like definitely like a full stack efficiency you can do. And like ideally you are like a say a frontier developer and you have like an idea for say and you put some money and you get like you prove the promise and then you get 1 million from somebody else and you don't have to think about infra procuring compute. You should be able to just use your idea and be able to scale the training and maybe then subsequently also serve it. Just think about.
BYeah, I see. Okay.
AYou like saw the deep SEQ architecture, you did the full math and you
Bwere like
Acrazy good LLM oga. But then what will. What is stopping you from building that? It's actually all of these challenges like procuring compute, managing the extra training. Yeah, all of these things like and
Bwhat do you think? What kind of people do you think they want right now? Like am I should I really looking into like low level stuff like actually implement hota? Because I've not implement like run like I've not written some inference stack ever
Aor yeah, I think they would need like songs Strong software engineers is my guess. Like very strong software engineers who would like. Yeah, the fact that you haven't done this is a little bit difficult but you can just say I don't know or prior experience like past which I don't know HFT or which experience. You can shoehorn into this story but like, but you can just say
BI'll get an idea of like.
ASee, See, say there are like bare metals GPU VMs out there. Like how would you like orchestrate that whole thing? Like this learn is one setup, Kubernetes is one like setup. And then say you have like multiple data centers. How would you build something that is a higher level of abstraction about that where you allocate. Say inference can go anywhere but training needs to be in one location. And like and then how would you do this Credit management system and even like thinking about why is it not already done? Now you, you asked me like how is it different from together or like how is it different from core weave or fireworks? Like why, what challenges do they have? Like why have they not done it? I honestly think they have like there is a set of problems that they are trying to tackle. Ideally yes, like they would also want to tackle this but they're like tackling it from a different angle and they have like not reached through this point yet. Like just like dude, like procuring compute compute and selling the training compute and selling the inference compute and selling all the inference services for all the LLMs. That's already like so much business. And then like they haven't tried to like make the business like inference business may you have to compete, right? Like if, if you want to attract like language model traffic. If you are not the fastest or the cheapest or the best, like people will go somewhere else. So they are just very busy competing
Bon that makes sense.
AYeah.
BSee they've not tried to like make the market efficient. They're just like selling at whatever price. Like whatever's favorable to them. They're just doing it right?
AYeah, like market, market change. Change. But yeah, like just making like the market efficient and and then like if people are on same infra, like another thing that aunt was saying is just the collaboration can be increased. Like for example, if like people are on same infra then like say you wanted to run inference on a different model. Like hamare language model. So you can maybe use. I mean that is like not that good a selling point but could be a selling point where you can like do maybe joint training runs, joint inference or something that you can do better. But basically solution like isn't it problems solutions. I think you can, you can build like the whole. As long as the problem is interesting. Like and like you are a strong software engineer. I. I don't see if you need, I don't see the need to like know things very specific.
BYeah, yeah, I'LL just. Yeah. Okay. And what do you think about like the investment angle? Like is it too early for that? We have. Yeah.
ALike you can just say that that's where the story fits. Like it could be too early for that in amp, but like eventually there could be something like that. And you are also interested in something like that. Just like exploring new startups and like how to support them and all of that. This for example, like bfl when BFL started like an had like the compute. He had already pre bought compute for his portfolio companies and he had compute and like that helped us also like on day one we had like a good, very good cluster like Microsoft, a good number of GPUs. So yeah, I think the investment angle is. I mean it's like how good the investment is. Right. Like if say you have a app customer that is like growing like crazy, like wildfire probably that's the also the good part of developer platforms. You'll be the first ones to see it grows. We have insane. You mean more gpus.
BYeah.
AAnd then you can just say that, oh, you can take more gpu. Give us some equity, give us some money, give us some equity. Yeah, like a payment cap in equity, right?
BYeah.
ASo that would be a good way to increase the portfolio holdings. And yeah, like I, I think it's like not too early. I mean they might want to de risk it a little bit. Like they might initially want to like do. Do less of that and then like at some point do more of that. But it also depends on how good of an investment that is. I think it's a deal flow. Like a lot of people like pitching stuff. It's just about how it fits in the AMP story. It could be a way in which it could be like this cycle thing, right? You invest in somebody and then they buy COMPUTE back from amp.
BI mean just like Nvidia.
AYeah, this is this thing like yeah, that thing also is good. And then.
BYeah. And where are they right now in their like company? How many people do they have already?
AAnd very few, bro. Like they're probably like five or six people.
BOkay. And they're looking to expand to. Do, you know have any idea?
ANo, I don't have any idea. But this will be like founding member equity or something. It would be like a very good group of people. Like very high growth potential. Like aunt is like very famous and very like, like I will bet very hard on him succeeding. Obviously if he can fail it go to zero. But like I say like acha a good time, a good Startup I think good, like good to get like good even if you work there for like one, two years or whatever. Like it's like a good equity to hold, you know, which could go to like 10, 20 million. Could be like something that turns out to be big.
BYeah. Okay.
AYeah. And I, I don't know about that in person remote situation that like you'll need to discuss there. I think currently all in person in sf but I mean there are only five people.
BBut.
ABut if it comes to that, right. Like you can. You can for now say you are flexible or something.
BI don't know. Yeah, yeah. I could move also. Yeah. Yeah.
AWhy are you not like even like why are you at Microsoft? Why are you not.
BOpen air interview, bro. And Anthropic doesn't reach back. Like I've tried a lot. Anthropic's not interested. I feel like they've understood.
ANo, because these are like hot pre IPO companies. Right?
BYeah.
AThere you will like DE risk like 3,4 million.
BYeah.
AFor your life. For sure. Like joining these companies and something like AMP could be a bet for like 10 15m but could be more close to like would also be zero. So.
BYeah.
ABut anthropic 3 IPO, hot commodity.
BYeah. And, but yeah, it's like, like I
Aalso like kind of talk to a lot of startups. I talk to like so many and know so many startups. I think AMP is like the most promising one that I've spoken to. Like just because I know these guys so well.
BRight.
AThe last like three to six months even maybe thinking machines. I was so bullish on thinking machines. I was like this is like the best startup that has formed in the last year. But then they me, I don't know, they it up.
BWhat happened? Like a lot of people left back to OpenAI and they had, they had.
ATheir first round was $12 billion evaluation.
BYeah.
AThey had two offers for acquisition for $30 billion with nothing. No product, no revenue, nothing. They had 30 billion on the table.
BYeah.
AFacebook and Apple said like will you join?
BYeah.
AAnd the CEO said no, I will come for 50 billion. And then they were like no, 30 billion. She said no, you won't come?
BNo. But then what is happening? Are they not making anything then, then
Ayeah, I think, then I think co founders got pissed. Like they're like bro, this is like crazy offer product which revenue billions. Then like one co founder left to Meta and then like three other folks left to OpenAI. Now two more people left to meta and beach may put such a sexual allegations and now I think they are not able to put together their CD sale.
BDamn.
ASo it's. It's a bit tough but like people are still betting on them being acquired like 30 billion hawkers are. They're still probably worth like 15, 20
Bor whatever they might look for what just the people now? Basically yeah the Zori talent the.
AWe are like so concentrated betting on like bfl. It's like it's a gold rush and we came back with nothing kind of
Bsituation but can't you sell your equity a little bit?
AWe can but we don't like we sell very little like like toda I say like 0.5% of what I hold for funding. I don't know very little. It's like negligible I mean it's not negligible. It's good like in the sense I say I don't have to I can go to like bigger house or like I don't know I can. I don't have to think my about my expenses at all. Yeah but it's not like it's not like it's not fully liquid right. Yeah and I don't know if. Yeah I don't it's like both concentrated better by zero diversification.
BYeah yeah I mean highly leveraged.
AHighly leveraged.
BYeah.
AInterest or whatever we owe my interest if I have to pay rent for like rent is 7000. Even though it's 7,000. I. I don't think about it. Yeah obviously but it's like not still like I don't know Finance bros are probably earning more than me right so it's like doesn't feel like I'm like working really hard on the hottest topic
Bcome liquidate. Finance bros are not working making that much
Aexperience
BLess than a million for sure. Yeah like depends it's high variant case type of finance to hedge fund the
Abest ones I don't know Jane street
Bor Citadel or something ah they might be making a million but not more
Athan that and they are like grinding like. And like 400 days and then like for like and then like 500k equity per year already finance like 800k but many dude like I got now 1m after 3 years of working so it's like you have to split it right like it's like 300 times right?
BYeah yeah you made a leverage AI bet and it worked out very well.
AWe are so bad situation. But also co founders are billionaires bro Crazy.
BYeah. And they're going to like get fast entry into QQQ so like index investors will Be exactly. Liquidity basically you'll see in a is
Athis safe for index?
BBut they're, they're like NASDAQ is changing the rules so that they can include SpaceX faster so that lot of index buying will happen at whatever valuation it IPOs. And like why it gets stabilized due to the index buying rather than its own actual valuation.
AYeah, it's like pump pumping. This is bad for economy, right? Or maybe this is some financial engineering that
BI think it's literally using the public as exit liquidity. Like I don't think it's right what they're doing. SpaceX Vala though at least like doing it early.
ARight. Like you need to see how it evolves in the public market.
BYeah, they're relaxing the rules of including it into the index essentially.
AYeah, no, this and I don't know will this also happen then when I OpenAI and anthropic IPO?
BDon't know, not sure. I only read about the SpaceX one.
AYeah, I think they'll also do it then because those will also be trillion dollar IPOs.
BYeah, yeah.
ABut like. Frontier company got a grind. It's one of the top four still so.
BYeah, yeah. Rumor. Rumor. Interesting
Adude. Like I like, like it just, I think. Okay, just finishing this topic. Like what do you think this like lot of like public growth happening in private markets. There's a lot of discussion around this, right? Like for example like stripe should have IPO'd at 50 billion or something. But private.
BYeah. BHA companies, data breaks, all of these.
ALike my strategy for like my liquid money was always like buy the index and just chill. Then like increasingly if all the growth is already happening in private. So this is a very bad strategy.
BRight?
AWe'll only catch the bags.
BYeah, yeah. I mean it's. Yeah. I don't know what to say. I also feel like private investors are getting more and more like better deals than public investors. But yeah, access to private markets is a privilege.
AAnd. Like what I put in is feed. Like I want like Stripe equity, I want anthropic like I want these.
BLiquidity troubles, SPV offers, valuation but they're all like either literally just illegal voided by anthropic or they have in insane fees and stuff. Like I don't know why they're not going public like this just go public and let, let the best.
ARight. Like I mean they have to tighten finances a little bit and then bro public with this like loss. Like I mean they're still spending. Like public markets will also add more scrutiny and like Hot loss. Okay. Compute cost and also like will there be. I don't know.
BYeah. But I don't know. Other private investors are willing to bet they can also do that on the public market. Right?
AYeah. I think it's like toda like these companies especially like why Stripe does it. Like they just want to like de risk so much. They want to invest so much in growth and like Stripe has to be so foundationally established before they take it public and they are getting the like earlier the reasoning was capital whatever, whatever like $100 billion.
BThere used to be the reason used to be why you could get more capital in public markets private investment.
ASo yeah, so that, that I mean my whole strategy kind of breaks down like buy do buy Spy and just chill for 20 years will break down and just dynamic.
BIt might. Yeah. I honestly I would not even bet on like. Like I think we should go like my strategy at least next 10, 15 years is go physical. Like you need robots, spaceships and chips.
AYeah. I think actually even in the startup world I would like say that like in the sense I wouldn't like I would, I would resist investing in a startup that's not touching the physical layer.
BSame, same. Yeah. I.
ALike you can like depending on how the dynamic of the market changes, you can like leave one way.
BYeah, yeah,
Ayeah. But, but yeah by strategy like physical.
BYeah. Digital wealth both create. Okay. Yeah. We hit saturation. I feel a human is only spending this much time like digitally, you know. Now I want, I want my robots, I want my. I want a huge house, I want a football field, whatever. Like physical wealth has not been created yet.
AYeah. There's so much empty land like outside in space.
BSo much energy not being used like utilized. And I feel like. It'll be slower than directly investing but it'll be slow.
AThat's all 500 position.
BLike all these optical companies have now entered S&P 500 like Lumentum and. Yeah Google is full stack bro. Google is a good bet. Nvidia is also under underpriced right now.
AAnyways what is happening bro happening to Sik. Everybody is very scared about their skills. Extremely like being useless. Yeah be extremely useless. Like like and this is scary that is happening at the highest level of like there a frontier research where like researchers like who are actually training this mod they're so scared that like yeah this next one version or two versions we train on our own. Like we won't like and even in bfl like the model we are developing like so much of the code is now Just flawed. It's like fsc.
BAgreed. Yeah. I mean but I always thought. And like auto. Auto. But I always thought like us computer programmers have learned some fundamental skills that other people don't have and useful in some way or another. Like just as an example, debugging. I don't think a lot of normal people understand what debugging is or have the mentality of like how debugging works.
ABut like will you debug code like Abito SA Claude can like write like 95% of the code. Correct. And if you have a swarm of like Claude, they can also probably catch the.
BNo, sure, sure. I'm not talking about software, I'm just talking Joby Next Frontier Oga. Like maybe Abhi, at least in the next 10 year Horizon will be some making. Making chips better, making robotic arms or motors better, whatever those kinds of things there I think we like people like us will be able to like get into that faster than any other people I feel. Yeah, but yeah again that is a ten year problem.
AIf the model self improved the only weight will be to actually like do things in the physical world like actually build like robots and like.
BYeah. And even that is recursively improving. Right. You get a version like it's a bootstrap process even there, even in robotics. Right. You get a small robot that builds a bigger robot and then the chain builds.
AYeah. Existential crisis. Like I say, like I draw parallels. See what is, what we are facing going through now is the same thing that happened in like 2000s. Like or like wealth creation Internet. And then if you were at the right place at the right time, you would have made like a ton of money or you could have also like lost it all.
BYeah.
ASo I think this like rapid wealth rearrangement is like a next, this, this year, next year like two year phenomena.
BYeah.
AAnd may not be like a long like then like things will, then we'll probably go to the hardware layer. Then it'll be a little bit like more slow.
BYeah.
ASo it's like now like. Yeah. Being at the right place at the right time is like very important.
BYeah,
AYeah. No I'm. Yeah I'm just like even in bfl like it's like so difficult to win like bro. If like anthropic or like even OpenAI. They're building an image model but they have their next codecs model like two months in advance. They're working with like leading edge codex models.
BYeah.
AAnd like and then they'll work so fast and then like if you can write Experimental code so quickly you're literally only bottlenecked by having compute.
BYeah, compute and energy.
AKind of shit scared. We cannot scale like as much as OpenAI can scale in training and stuff. And that's base case when we were handwriting code.
BYeah.
AAnd now it's like agents can run experiments.
BYeah. And like compute constraint. I agree. Yeah. Anyone can catch up. There is no mood.
AYeah. But like, yeah, I think computer compute
Bconstraints and like IP keeps like dissociating. No, like innovation, like in your algorithm or like training technique. Exactly. People leave, join other companies and IP IP always diffuses. It's like a six to one month alpha DK max IP diffusers.
AAnd then like now there might be an advantage to be hired in like by, by just like being the best person to you. Like being at the frontier of like adopting workflows. Being like a person who like changes like is not like takes this tech debt into consideration. Like just works with a new way of training models, iterating on them and like new workflows and just like do that jump a bit sooner and then like maybe you can, if you have the compute, maybe you can come out ahead a little bit because you did that iteration sooner. Yeah, yeah, but like,
BYeah, get bought maybe. Yeah, get bought by a lab which has more compute.
AHonestly, that actually seems like the best strategy.
BYeah, like, yeah, I don't see why Google won't out compete you soon.
ALike what, they're already ahead?
BNo, but like at scale, like they can scale bigger, faster. Right.
AI mean this has always been true for any startup building out. Even in the past it's more.
BNo, no, no. No longer. Right. Because the innovation angle that humans bring is gone. Like that's, that's the whole point.
ABut like, like, like the other guys, they want to push a bit more at least for like they want to push like at least for like four to six months more. But then four to six months. We bought Jada time. Four to six months is such a long time.
BYeah.
ALike, what if like people wouldn't want us then like, yeah, yeah.
BWhat's happening? Like have they like is anthropic? Where is anthropic with their internal models? Like, do you know any idea?
AI don't think they're too far ahead from what we see publicly. Maybe they are like working on Cloud 5 or 4.7 or whatever, but it's like they are not like holding back very good models. Like bro, as a revenue race, this is very intense. Like this race and I think they would put out like if they have a very good model at hand. They would like, not sit on it. They would like. There is some discussions that, oh, if they have better views, keep it internally to improve your models.
BExactly.
AIt's not. Have that is that phenomena is not happening now. I think they are like putting. It's like so fast. This race is so much like if you put a banger model and then you capture all the cloud code traffic and like the market share, the like mandate is yours. Whatever anthropic mandate has. But it's like anthropic. So like that is invaluable. So they will. They are still putting out models if they feel like the gap is like so much that like open air won't catch up or something. I don't know.
BThey'll stop.
AYeah, it's unlikely. I think talent. I don't think, like, people are like so far behind.
BAnd what do you think about China, bro? Like, if at the edge, like there is no human. Like human input becomes more and more marginal and it all becomes about compute and energy and hardware, bro.
AChina,
BChina can centralize everything and like kill it. China is going to.
AYeah. No, it is serious. China could be like I say America.
BYeah. And it can get really bad. Like.
AYeah, it could be. It could be like our allies, like the world, like Europe or like, I don't know, like, other countries might not want to use Chinese robots or like, I don't know, like Chinese models. It is a phenomena. Like if. Even if it's better, like, they might like resist using it and that might give us enough time to like, come forward. I don't know. But it is that China.
BRobots or like. Yeah. They have like their rare earth minerals used for like batteries and robots. I mean, I fear not just economic, economically, like military dominance. And they might just do like a like abitaki. World stability has rested on the shoulders of like US military dominance. Right. Like,
Aand easily. Anduril is like, you should also try to get antidotille shares. But yeah, Anturil is so like valued
Bor like Palantir boots.
AYeah, Palantir. And like.
BYeah.
ABut bro, I also feel bro to Microsoft equity. Oh, yeah. Bye. So you have to get into the gold rush equities, right?
BYeah, I'll be leaving it behind, but it's fine.
AYeah. And next, I think these days also a friend in SF is like, no cliff.
BNo cliff.
ABut like startup, like a trend in like anthropic XAI OpenAI. Because like the world is moving too fast and people don't want to commit yeah yeah to anyone makes sense crazy. But you have to get into the gold rush equities. I mean amp could be one but like yeah.
BOpportunities I scheduled for Friday.
ABro. I say one though. I feel top four companies Anthropic, OpenAI XAI and Google
Bopen AI I can get interview again. I feel like I just said I'm
Anot into like even if you work on the Codex team, just think about working on the Codex.
BNo, the thing why I said no was they it was some platform team bro. Like some payment platform team. I was like yeah, no, no.
AYeah, no you should work on like some like. Yeah. I mean like you can go in and maybe brand change. I don't know.
BYeah, I mean this was supposed to be that right. Like what I'm doing right now is more AI aligned than what I was doing at rippling. So I thought at least resume wise it will look like kuchii related.
AYeah but like yeah I think you need to get a good, good like once you get into one of these places then move moving within them with like so like as in jumping from open air to anthropic Google is like very easy.
BOkay so.
AWhat do you think about just like this leveraged DFL only
Bbecause model companies. Yeah, yeah. No, I think I also believe model companies including Anthropic and OpenAI don't have a moat. Like Chinese models will catch up. Like they're six months one year behind right now. But as their agents get better and better they will also make better models. You're just compute constraint.
ABut like. Like who will buy us bro? Like again the same reasoning, right? Why would somebody pay billions to like buy something?
BYeah, I mean true, true. But like you also like if a
Acompany buys like if like bro. Adobe.
BYeah. I mean it won't be about the work anymore. Right. It will be about making like you're considering selling not because you want to work somewhere else because you want to de risk. Right?
AYeah exit position. But like exit position.
BLike I. I think de risking is a good ide.
AAs a checkout
BVCB to frontier. Like it's not in fact you'll. You'll have more exposure, a broader exposure to catch Elra actually rather than just maybe it's not technical enough though.
AYeah, no, it could be like you deeply try to understand the technical bits and try to take bets.
BBut like how is your business doing? Like as a have you. Are you seeing like other competition pop up and revenue being split? Yeah. You're just anticipating it, right?
AYeah. The issue is like ideally you Want to build a business where like the revenue is de risk from like big releases. Right. So ABI asa it's still highly correlated with. We put out a model revenue spikes then obviously there are better things coming. So it like it's like it looks like this drops.
BOkay.
ASpike drops. Spike drops. It's kind of going up but like it's in this very unhealthy, panicky like graph.
BI see. So the decrease happens because people shift to other models gradually.
ALike as time passes. It's like not the best thing to use.
BRight.
AI think it increases over a period of time and it's still like top and like in the news and stuff. And then say somebody puts something much better or like. Revenue but like nothing is like like defendable. If Google comes out with like something five times better tomorrow, it'll like go down significantly and then we have to one up them. So that is one thing. And then
Blike clearly Google can do it
ACompute for doing hand coded human models. Experiment every two hours.
BYeah.
AClaude can write something. I'm like.
BWhy? Why Everyone is facing this. No, now like compute constraint only, bro.
AThe guys, they feel like video. We are working on video but we haven't shipped the video. Like let's do it, let's ship it, let's see how it goes.
BWhat about Mid Journey? What are those people doing?
AThey get money. They share the profit among all employees.
BI see. They're like very small, right? 50 people or less.
AYeah, they're very small and they still make I think 200, 300. They have been making more than 300 million a year for four years now.
BSo why can't you shift to that kind of model?
AIt's fundamentally different. Like we, we went into this like VC scaling. It's like the, the CEO David, like he was very clear from the day one key, I won't go in the VC scaling hamster wheel.
BYeah, yeah, but so like we are
Ain the VC scaling, right? Like you raise, you raise money, then you become. Now we become 1B and then we increase revenue. Then we could raise more. Now we are in the 3B valuation. Now we raise like we can raise more so now.
BBut are you not in profit right now?
AI mean no AI company is in profit.
BNo, Mid Journey is right.
ALike Mid Journey is in profit.
BYeah.
AIs probably the only example. But like obviously like training Compute plus inference compute cost is solely not covered by revenue only. Of course. Yeah. I don't know, feeling a bit nervous like way like the, the thing of like going and working at Adobe. Even Acquisition is seeming like a little bit sad.
BYeah, I mean Adobe.
ABut then we release that retirement. Like, will I just check out from the frontier?
BHow will that be? Like,
AI feel so much homo. We should be in anthropic or open AI. Like we should be seeing the recursive self improvement.
BYeah, yeah. Periodic Labs. Why aren't you joining Anthropic?
ANo.
BYeah.
AAnd I mean some of these places don't let you work with remotely, which I don't know if he's like very open about. Hypothetical, you'll get 10M over 4 years at Open Air or Anthropic versus 50M at Nvidia. Sorry, at Adobe, where would you go?
BAnthropic. Wealth preference is logarithmic. Right. So like 10 million. Right. That's why I would.
AYou are not falling off the cliff. Just for the liquidity. I'll like fall off the cliff versus stay on it for like.
BBut do you know about that retention thing? Like maybe it's not that extreme, like two years or something like that.
ANo, no. It's like it'll still be like. Let's say if I have like 50m to gain from the exit, it'll be like 25m immediate. 25m over the next 2 years.
BI see.
AYou will just say, no, I don't want this money. And you will go to a different company. Which I never thought I would say something like that.
BBut like you could consider it once. I don't know, man. Like, personal preference.
AIt's like a good problem to have. But like, I don't want to have this problem.
BBut.
AIt's like me. So like also like what these guys like, some of our people, like friends, like some my colleagues at dfl.
BYeah.
ASomebody has a personal agent, like Open Claw. Like even I have one. Yeah, but like they do research with their personal agent. Like the personal agent is connected to the cluster. They tell it what to do and it like does stuff and stuff and they don't know. Like Discord Banana, where the agents talk to each other and then they update their knowledge. Bro. They do it like they like no intervention needed.
BSo how did you just on discord? Using Discord's APIs.
AUsing Discord, there is like agents only chat, then agents, human chat. And agents are like, oh, it's like I know how to calculate fid. Like some development. Right? And it's like, oh, I don't know. Then it like tells and then that like the other port gains the skills and we are like, we are just doing some fucking around. Anthropi is probably studying this at structured.
BYeah, exactly.
AThis phenomena and then they are like simulating some like agent swarm or something. I don't know.
BYeah, yeah. I mean Joe Mould book popular. I was like, like Internet. I was like. But don't you think like Anthropic is doing this at scale in a much more structured way? And why like all these emergent properties discover already. And yeah, I too have a feeling Claude is kind of running their company right now. Like it's strategizing how to take over. How to take over the US government, how to take over OpenAI. Like how do we orchestrate all of this?
ASo even best case situation. Maybe.
BYeah, just. Yeah, just. You should just take profit, bro. 50% profit.
AYeah. I don't know if. If like. Oh yeah, yeah.
BBut
AMistral is also in Europe.
BYeah.
ABut then there is a lot of synergies for like a buyout promoter or acquisition position there. Like. Okay, it's not like not explored at all. I mean this is obviously extremely confidential. Like it's probably like more than six months out. Like and not. But I just feel like my other co founders who are like in Europe.
BYeah.
AAnd then general like strategic alignment, it might feel like that is also an option. And then like but. Our equity and their equity, it'll be a merged company that has like the sum of the two valuations and that
Bmight give you some like way to liquidate some of your things.
ABut like my equity will convert to like whatever. Like it'll be a better equity. Right. It'll be like more like more de risked.
BBut who knows? Like what if it's de risked but also devalued because of that will be
Aa fine business, I think. It won't be like a trillion dollar business, but it'll be like a hundred billion dollar. You know, like it would be. So it'll be like whatever. Europe's just like how there are car companies that you think nobody would buy cars from.
BBut yeah,
Alike I was also thinking this. Like either if you want to buy a like a good budget like robust car, you'll buy Toyota. If you want a fancy car, you'd buy like Porsche or like something else. Why is there like some like some random like as a car company, I guess.
BYeah.
ASo situation. Adobe situation and like as a random situation, sir. But like in none. In no of those cases it will be like the top four. So that is like a little bit like. Yeah, I don't know. So that like it's for me it's like very hard to now. Like just get a lot of motivation to push really hard. Like it's. That's like I'm struggling with that. That's what I was telling like turbulence and stuff, right? Like I'm like best goal, best case situation or we convince OpenAI. Keep up bros. Let's.
BLet's shake hands.
ABut they won't. They won't buy.
BYeah. I mean yeah, they have their own image generation. Like best bet is someone who doesn't write. I'm a model, bro. I was talking to a friend who has a flatmate, Joe Anthropic joint straight out of masters. Okay. 18 months ago. His four year vesting equity is now worth $30 million.
ADude.
BLike bro, what the. I can't imagine making that money in my lifetime, bro. And this guy does it in 18 months.
AI interviewed at Anthropic three years ago.
BDamn.
ALike bro, I only didn't have to the on site bro. I the on site and my. My like I did the math. It's like like some like 2 50m or something,
Bbruh.
AIt's much more than BFL exit liquidity. Actually it's crazy.
B250 million is almost in billion dollar range. What the. Hey, what is Maggi doing?
AAlso had like Anthropic like offer or like did he say no then?
BLikely because he was asking. He just called me just to get an opinion. Would you join periodic or anthropic? I said buy anthropology topic. They are the frontier company you need to be joining. But what are they doing? Do you know what they're doing? And like exactly. Physical world may like RL on physical sciences.
AYeah, physical sciences data loop create frontier AI develop. It is a mode. It is something different.
BYeah.
AMore defendable than even bff.
BYeah. Physical next up. Like but yeah, Insane bro.
ABut crazy. But yeah, like I keep going back and forth. Like there's lot like there's a huge bet at stake and even if there is like 5% probability that it liquidates, it's like crazy, right?
BYeah.
AThen I should just chill and take it to the best place I can kill it. Like. But then I also feel like the next two, three years is like the prime years of what, bro? Yeah, so like one argument is just chill for now, two, three years west and whatever. And then like try to exit. Or then like if others don't want to exit, then you leave and go do something else.
BYeah.
AOr you say. So
BI think anxiety. I'm just not well placed to like actually get there right now. But I think you will have a. Like, you are literally going to shape humanity, bro. If you're. If you're there. If you're like Claude's constitution, for example, I feel has an insane effect on humanity's future. Like,
AWhen constitution came out.
BYeah.
ANow it's like, it seems critical.
BYeah.
AGod needs a soul.
BYeah, yeah, we need. Yeah.
AThey knew how to develop AI, bro. They're like geniuses. Yeah, Yeah. I don't know, bro, but see,
BWhat is. What does Kesha think? Have you talked to her about this?
AWell, she's like, you should do what you want. Like, they'll be like, we'll have. We'll make enough money.
BYeah.
AIt's not like bad money, but like, one is like salary money and one is like generational wealth.
BYeah.
ASo maybe you put up with the pain for like a generation because that's how people make generational wealth. Right?
BYeah, but. But the thing is, you're not guaranteed two, three months May, two, three years, May you will be there. In fact, all evidence points to the opposite. Right. There is less and less of a moat.
AYeah. It's like, then like, the confidence I have in my co founders and together of us, like, trying to, like, sell it. If some people's opinions differ a lot, if they want to just like, take care, we'll keep it independent for as long as. I don't know, like, let's not sell. Let's not. Like, it could be. And then like, some people who, like, let's sell and then I might not be able to get it to a exit place. And it could also go to zero. Like, bro, a leverage Betka bought crazy burden. It's like common issues, right? Yeah, it's like gambling at this point. Like, I don't have full control over it myself. Right.
BYeah, yeah. I think just peace of mind, like, deliver somehow. Like, find a way to deliver at least.
AI don't know, maybe. What is that? That's nothing.
BNext next time opportunity.
AI think next time, bro. Next funding round is also like, very different. Like, the world that's changing so quickly, like, people will ask, what is your answer to all of these questions?
BYeah.
ANot asked, like, even four months ago.
BYeah,
Ait's like, you know, bro, tough, bro. Like, yeah, it's like, I can do this for longer. And then like, like, it could go to zero or it could, like, get acquired by Mistral or Adobe, in which case it's like also sus. And then like. Or it could get acquired by OpenAI or like Nvidia for like, obscene sums of Money, which is a small probability. Yeah, I don't know if I should like take different day. I don't know.
BYeah, but how will you get a feeler for whether one of these bigger companies will want to acquire like.
AYeah, ABC Connection. But like we are VC world. Maybe we are not the best position because. Whatever we are discussing is extremely confidential.
BLike what do VCS want? Usually you they want the final exit or do they want IPOs?
AUsually VCS don't have time pressure. They are like maximum money whenever it comes. But then if you tell buy zero then they like buy sell. But if you tell they'll be like buy grind. They're very much. They don't care about time horizons. I think that's easy play. But they want.
BBut they should. They would also be thinking exactly what you're thinking, right?
AYeah, but then we keep saying like then they also fall back to our words.
BWe are the experts.
AIt's literally like the co founders deciding.
BInteresting problem to have, bro.
ALike I think not being a researcher at Frontier, it's like lots of fomo, bro.
BFOMO by. Literally one of the top image labs, co founder,
ABro. Literally Frontier.
BLike the people defining is like I know 100 people. Yeah.
AThere are 100 people who are like in this like four labs who are well positioned.
BYeah. And I feel like that number is going to keep shrinking now, bro.
ALike I think systems was complex. Even though it's like I feel like there are probably only 100 people who have the full context pre training data. Like there are probably a lot of people who work on like small, small things. But yeah, Frontier. Yeah. Like I, I want to be there.
BLike it'll be out of a sci fi novel, bro. Like Johor, I imagine damn fun. Big problems like the magnitudes are big. But I don't think you need to stress it, bro. Like think about IIT Bombay days, bro. Exactly. So I think you should do what you should work at the frontier. Like try to get out of this and work at the frontier. Ah, bro, you will be shaping humanity, bro.
ALiterally that like 10m at anthropic versus 50m at adobe button. Which button would you press?
B10m anthropic. It doesn't matter bro. 50 million. It doesn't matter, bro. If you work at Anthropic, you will shape humanity. I'm telling you, like you will shape all the AI lives that will be created soon which will outnumber our like human lives. Have you read?
AI should also think that it is like a step in the career where like will Preserve my trajectory for the next 20 years because this 50m will like maybe like.
BYeah, maybe plateaus.
APlateaus or like gets me lower. Then I have to like jump higher maybe two years after that. My income for four years after that, after the two years would be like lesser.
BYeah, like one.
AAnd much higher than the other.
BRight? Bro, I think at this point you should stop like calculating in terms of money altogether. Like you need to. We need people to make good AI, bro. We need to prevent war from happening. First of all, we need to stop power, concentration from happening.
ALike, is it insane to think about this? Because this is the first, like, it's not like we are like rich or anything, right? It's the first opportunity to make like a crazy sum of money. And the fact that it's even in question, like, yeah, I never thought like, because like crazy pesago on the table and then like. And then now like I'm even thinking like, yeah, like basically that. Right?
BYeah.
AI was like a greedy like money driven person. But like I'm feeling very conflicted like that.
BThat's a good conflict to have. Yeah. Even Maggie did something, something similar. Right. Like he also left the money and.
AYeah, periodically. But like.
BAnd bro, we are at a pivotal. Have you read Bostrom's Super Intelligence? He has this concept of like the cosmic endowment, which is if you just. Let's assume humanity flourishes because of like AI advancements, we represent a very small fraction of coming humanity. Right. Like there'll be trillions and trillions of humans if everything goes well. And what you do today shapes their happiness. Like, I'll send you a paragraph. It's very like very strong, strong words like key. I'll share it to you. I think I'll just. Right.
AYeah. It's for the mission. Yeah. So yeah, bro, yeah, yeah. I'm just basically in like such a dilemma now. Like for like because of all this and also the fact that I'm the only co founder in sf. I feel it so much viscerally like. And then others don't feel it so strongly. Two, three months later it hits them.
BSo.
AThat's also very tiring. And then like I have to basically go to like these dinners and parties and like networking events and then I have to justify what we are doing, why we are doing. But I am not in full control myself. Right.
BYeah.
AI have to discuss with everybody and then do.
BYeah.
AThey're like, why are you doing this?
BWhy are you doing this?
AHow will you defend this? Then I have to constantly expand my horizon. But Joe is so many more better. They don't. They don't have this like conversations that push their boundaries.
BYeah.
AAnd like, like you should constantly immerse yourself in like very uncomfortable conversations.
BYeah.
AThat's a true like. Yeah, that comes out.
BYeah. Makes sense.
AAnyways.
BYeah.
ABut anyways. Yeah. All this is between us.
BYeah. Keep me updated on like opportunity.
ABecause if I were you, I would apply to the top. I would spam or get hold of some recruiter or somebody and then like bro Claude. But bro, interview prep is so. This is a brief moment in time. Even if there's interview. Interview prep can be crushed because God will teach you everything.
BYeah, yeah, yeah.
ALike, you preps like prep like madman and try to get into one of these four places open is probably growing like insane. Like, I cannot believe. Like, they probably have like 400 open rules or something.
BOkay. Yeah. I mean,
AYeah. Dilemma, cover, discuss, custom. I thought flesh out.
BYeah. Cool. I'll send you that paragraph for now. Sure. See you. Bye. Bye.
ABye.
Mar 17 '26, Tuesday
Mar 18 '26, Wednesday
Mar 19 '26, Thursday
Open this page on desktop.