← This Week in AI
This Week in AI
Grads boo AI, Reese Witherspoon gets dunked + Karpathy joins Anthropic | TWiAI E14
01:38:43
Loading summary
Transcript190 lines
- [00:00]
A
Andrej Karpathy, topic number one, has joined Anthropic.
- [00:04]
B
Yeah, Andre doesn't really care about money, so it's because there's something there that is interesting to him.
- [00:09]
C
I wouldn't go and then say that therefore OpenAI is dead. Things have shifted so often over the
- [00:15]
A
last few months that credibility is a massive communication tool for Anthropic. Every time Dario speaks, he creates massive investigations in Washington D.C. massive hand wringing. He's got a pdum of like 90%. This has caused a lot of the hatred of AI. I believe he is like too honest about this scenario. I also think he's wrong. He's the leader now of the AI industry. He's the worst possible spokesman for AI in America. Whereas Andre Karpathy is a great spokesperson, I think. And that might take a little of the pressure off. Thanks to our friends at PayPal, the exclusive sponsor for this Week in AI. Try the payment and growth platform that's trusted by millions of customers worldwide. PayPal open source start growing today@paypalopen.com. all right, everybody, welcome back to this Week in AI. Yes, this is the new roundtable from executive producer Jason Calacanis. That's me. This is the new roundtable. Just like I do this Week in Venture Capital Sometimes or the all in podcast. I love to have three friends who are actively building the future and then have them on a roundtable to discuss the news of the week. That's what this Week in AI is. You can get more information at this week in AI AI and you'll get links to YouTube, Spotify, Apple, all those great places where you can subscribe and post a comment. The community is growing on X. We've got an incredible roundtable for you and a very full docket. Kanjun Q is back and she is the co founder and CEO of Imbue. She's making open source agents. She was last on this week in AI back on April 7th for episode 8. How are you doing, Kenjin?
- [02:00]
B
I'm good. Good to see you guys.
- [02:01]
A
Good to see you. You were a sequoia Scout for 10 years, helping place bets. I was the first Sequoia Scout. Guess who was in the first class and had the top two investments in the first Sequoia Scouts class.
- [02:13]
B
Do you know, I believe that was you. I believe they said it was you.
- [02:16]
A
Oh, they did, yes. It was actually me and Sam Altman. Oh. We placed two bets on two friends companies. One was this taxicab company, Uber, that nobody believed in. And the second was this very silly, simple Payments Rails company called Stripe. That first Sequoia Scouts fund I think was like a three or four million dollars fund and it wound up returning $250 million. So one of the greatest venture funds in the history of venture funds. Scoia Scouts. Great, great, amazing industry changing program. Tell me a little bit about what's going on in your world. Obviously we know you've raised ton of money, 200 million from some of the best people in the industry. Nvidia, Eric Schmidt and I see here in my notes you're running a 10,000 H100 GPU cluster in partnership with Dell, is that correct?
- [03:14]
B
We actually sold the cluster many years ago and make a lot of money off of that, which is great. But we rent out the cluster. We rent out the cluster, yeah.
- [03:22]
A
Take me behind that because my friend Elon, sorry to drop a name, he also was early into building a giant cluster and wound up having some extra compute. And our friends at Anthropic took over all of Colossus 1, I believe. So take me through that decision. How early were you to building out a giant cluster and then when did it become extremely, I assume, profitable for you to rent it out?
- [03:47]
B
Yeah, so when we first started Imbue, we were really interested in this issue of what we saw would be power concentration and incentive alignment. So at that time a lot of people were very interested in technical alignment. How do we align these technical systems? And from our perspective, one of the biggest issues is incentive alignment. How do we end up with these systems that actually serve us humans and not just the companies that build them when the shareholders profits? And so when we first set out, we were like, okay, well we're going to have to find, find a way to fund ourselves. And we were started as a research lab and explored these very small models. Actually, can we make models that are human inspired and figure out how to have one on everyone's computer? And that would change the incentives of the industry because that wouldn't be so centralized. When we saw the opportunity to buy a lot of compute, this was, I don't know, maybe in 2022 we were like, this is coming. We did. And so that was a great investment.
- [04:50]
A
We did. You're just like kind of blowing it off. You spent like nine figures doing this. This was not a small bet, this was a colossal bet. And if we were to look at your bet, is that throwing off tens of millions in profits a year for you or something in that range?
- [05:08]
B
I would guess I won't speak to it that much, but it does definitely fund A lot of things.
- [05:13]
A
Fantastic. Really great. When you make a good bet, it's
- [05:16]
B
great to make a good bet. And I think for me, I didn't want to raise venture money. All of our investors are actually either corporate arms or a nonprofit primarily. And so I wanted to figure out how can I fund the company in a way that's actually aligned with what we're trying to do, which is to eventually figure out this incentive alignment problem.
- [05:37]
A
Well, fantastic. Jeremy Frankel is back. He was last on this Week in AI March 31 Episode 7 Another great guest. He's the co founder and CEO of Fund Fundamental. They're doing large tabular models for enterprise structured data. Which is a fancy way of saying what Jeremy, say it in plain English.
- [05:58]
C
Basically what you see in AI. Like nowadays when people think about AI, they all F1 thinks about large language models. And you know, it makes sense because like ChatGPT or Cloud really changed the way, you know, we all interact with like text, images, coding, audio, every single modality. But what you realize is that most of companies or enterprises, data actually sits in structured form. Think about all the spreadsheets, databases, ERPs, transaction logs, all of those are structured forms of data. And LLMs really don't handle that type of data very well because LLMs work well with data that comes in sequential manner. So if you think about text, audio, video, it's all sequential. You're predicting one token at a time. There are next token predictors. But if you want to work with structured data and you want to use that data that comes in more than one dimension to make predictions, you need a very different type of architecture. And that's what we built. And so if you think about who
- [06:54]
A
uses it, who are the customers of this?
- [06:59]
C
We're working with a bunch of very large enterprises across different industries and use cases from energy to financial services to supply chain companies, shipping companies, healthcare.
- [07:11]
A
I'm sure some of them are listed at your website under your, you know, case studies, etc. Tell us about one of them if you can. Or are they all stealth?
- [07:18]
C
So we're actually, we're actually not, we're not naming companies yet. We came out of stealth three months ago. So working with a bunch of companies, bunch of Fortune 100 companies, but we're not yet naming those companies.
- [07:33]
A
Which category has had the most success? I can give it to you in an easy way to answer without disclosing a customer. So give us like broad strokes. You can, you can kind of meld together a couple different customers. But if we were to think of a vertical with a lot of transactions like marketplaces like Uber have tons of transactions in real time. Maybe doordash, maybe a financial services company. The stripes, the visas, the American expresses of the world. These people deal with very large databases that are real time and quite dynamic. Would one of those be in the top customer getting the most value?
- [08:08]
C
Yeah, and some of those may or may not be customers without me disclosing anything. But yes. But what's interesting to me is that when I first started a company, financial services, for example, was very much an obvious choice of a vertical where I knew there was a lot of tabular data. And I came from that industry, so I really understood that industry much better. What has been interesting to me has been that over the last few months I've just seen so many industries that I did not expect there to be a lot of use cases for structural data, but the country was actually the case. Energy is a perfect example. Predicting there are a bunch of use cases on the commercial side. Predicting the demand for natural gas can be one thing. Or shipping companies optimizing the best routes for ships. All of those are tabular data use cases. And I didn't have a specific knowledge in those industries to really know that ahead of time. But by talking to customers and large customers, I realized that all of them really care about that.
- [09:11]
A
Amazing. And you're also providing the service through Amazon Web Services, correct?
- [09:15]
C
Correct. So we have a first party partnership with Amazon which means the models are hosted natively on aws. And so like if you're a big customer, you can use us directly on AWS without having to go through any, you know, procurement cycle or you know, vendor onboarding like Amazon will take care of that. But we also, you know, we work with every cloud provider. We just have a special relationship with aws.
- [09:36]
A
All right, also with us. And now a first time guest, Carrie Saarinin. Saarinen. Did I get it right? Yes.
- [09:44]
D
Yeah, thanks for having me.
- [09:45]
A
Oh, great to have you. You are the co founder and CEO of linear. People have heard about that product development system for AI agents and teams. Translated into plain English. Who's buying the product, how much do they pay for the product and what does the product do for them?
- [09:59]
D
Yeah, I think like we kind of have an interesting collection of people here. I feel like I represent more of the application stack where we sell this product to companies like OpenAI, Coinbase, RAMP, Cursor, Cash App. Like many top product companies out there use LINEAR to organize their product work and then now they can also do it with the agents so you can create plans with the agents. You can kind of like ask the agent to research to customer feedback, like what are the products we should be building? Like, what are the features what these enterprise customers are asking for? Use the linear agent to actually write some kind of proposal or some kind of plan. Send that plan to the agent to look into the code base, like what does this actually take us to build? And then delegate it to specific people or to specific agents. So we started a long time ago around more like the issue tracking, project management, use cases, which is every software company needs some kind of way to coordinate work. But now I think now with the AI, we've been shifting more towards the actual execution of it as well. So very near in the future you will have a native coding agent built into linear. So you can just say, hey, I found this bug, go fix it. And then this is all shared in an organizational workspace. So it's not on your computer, but it's actually part of your team's workspace.
- [11:28]
A
And you and your co founders all came from Finland where design and development tools are all the Scandinavian countries tend to do really well in that space. And you started pre AI back in 2019? Yeah.
- [11:43]
D
Yes. So yeah, we are all finished. But we actually all been lived over a decade in San Francisco. So previously I had a YC company back in 2012 and then we were acquired by Coinbase in 2014. So I was there and then I was at Airbnb. So the linear really started from our frustrations working in these companies that why is the tooling of building products so bad or so frustrating to use? And it didn't feel like it's actually helping us. And yes, we started pre AI and then the last couple of years we've been shifting towards more of the AI. I think what we see our role really now is that a lot of our customers also have started pre AI. Like most companies out there in the world has started pre AI and now what they. They are in the process of like, we need to think through the. Which processes still make sense. Like if you are a large company like Coinbase or something, it's like you can't just throw everything out. Like, no, we just. Let's just forget every. Any kind of tracking, any kind of planning. Just let's let the agents run. I don't think that's like a sane way to run a company.
- [12:57]
A
What do you think of design in the age of AI and everybody being able to create what I would say is like a 7.5 out of 10, suddenly everybody, a product manager, maybe a developer, the CEO, co founders, they can all make a design that's maybe 7.5 and 8, maybe even an 8.5 out of the gate. What happens in design when everybody becomes what I'll call a serviceable, reasonably good designer? And you can disagree with me. Obvious if you think the framing of my question is wrong. But are designers offended? Inspired? What's your take on it?
- [13:35]
D
I think there's a lot of feelings, probably from different people. My feeling, I've written about this too, is that yes, the AI can produce an output which is like, here's the design, here's the design for the website or a product. But then a lot of times when you're actually designing things, you actually think through the problem. So if you just now not thinking through the problem, but generating designs, you might generate designs that actually, like, maybe they look nice, but they don't actually work well. And I've heard some comments from some, like early stage investors where they say that on average, the actual design of early stage products is worse now than before AI. Because people that knew that they don't know about design, like, this was also why I was hired at Coinbase was like, Brian Armstrong was like, he wasn't a designer, so. And like he asked me advice like, what should we do? Like this Coinbase was like using the Twitter bootstrap, kind of like basic kind of framework. And I think the product itself was actually pretty good or it was constructed well, but it wasn't just like designed in a good way. So what I always told you, just hire a designer, like, I don't think you can fix this like this way. And so what happens with the startups now is like, they don't hire those designers and so then they don't actually put someone to think about the design, they just delegate it. But then they don't themselves understand whether they're delegating. So I think the design actually, you might think that now everyone is a better designer, but actually companies might be sometimes worse off because they're not actually kind of investing into making the product design good.
- [15:17]
A
So to summarize, it's soulless. You haven't thought through what the design challenges are when you're using AI. So it's a bit of a crapshoot. It may esthetically look pleasing, but it may not be thoughtful. And then you have founders getting the impression that design is like writing a press release or it's just some task, you know, writing a term sheet or something that AI does perfectly. And there's no difference between a term sheet sheet written by AI or an attorney or an individual, generally speaking. So you lose that soul in it and then net net. It's fool's gold because the founders actually believe that they can get it done and they're not getting it done properly.
- [16:04]
D
Yeah, and I was like, at that. Like, obviously, like, if you know what you're doing, if you have like a, like, interest and like on some level understanding in the design, then I think the AI can be very helpful. It can help you with your design process, but you can't. I think it's a little bit with coding too. It's like, yeah, you can generate code, but eventually if you understand nothing about code, I think you will have a problem. Like, it's like there's some of those Reddit posts where people like, I've been wipe coding for six months. Like code base is basically falling over because garbage. Yeah, Like, I. I don't have any understanding of architecture or like security or any of these things. So it is kind of dangerous.
- [16:47]
A
I see you nodding. It's really interesting as a founder, because what we're seeing here is people know enough and they have this new copilot, so it's enough to get themselves in trouble or build something serviceable, but you still need human in the loop to hit maybe the elite level. Is that your thought, Kenjin? I see you nodding along.
- [17:10]
B
Oh, I just really agree with Kari here. But also I've. I'm nodding because I've recently building. Been building my own kind of like personal tooling and workflows. We have a product we haven't released yet that we will open source soon that allows people to work with their agents and build their own UIs that helps them communicate with agents. And I think, Kari, this is something that you think, which is that the UI is kind of a layer for. For humans to communicate well with agents. So I've been building my own UI for like, I made my own to do list so that the agents can do my to DOS for me. And I made, you know, my own email processing thing so that it's a bit much better, more tailored to me than superhuman and other people. And I can like batch process a bunch of stuff and batch compose things, but other people look at my UI and they're like, God, Kanjun, what is happening? What is going on? I hate that. And I'm like, yeah, you know, a lot of what we've traditionally thought about as design, I think in the tech industry is Kind of how do we design for a lot of people? How do we design in a way that many people can understand and that requires different ways of thinking and principles. Whereas if I'm just designing for myself, like, only I have to understand it.
- [18:19]
A
This is a good point. Bespoke design is now possible. So if you like the elegantness of Superhuman and, and you love the quick keys, you can have that. But if you prefer it to be written, you know, like a terminal window and you want to just have this like terminal of your email, which is how email used to work, by the way, you can have that version. You could have a version that looks like Tinder. You swipe left, right. It's really fascinating where we've gotten to. And I like this idea of adaptive design. It's kind of like adaptive learning. I just thought like, wow, I wonder if AI over time will just make a custom interface for every customer. Maybe that's the best design. Maybe as you're using Amazon, it's like, by the way, you don't like these never ending product boxes. You prefer lists that are ranked in a tabular style. I'm just going to give you that. And then my AI knows just everything. Tabular, please. Jeremy, I'm curious what you think here. Adaptive customized designs for each individual done by their browser or done by their local LLM. And you know, we're just going to live in a world where the website or the data source adapts to you, the individual, or is it going to be just this God designer in the clouds who just makes everything perfect? What is your take on design in the age of AI?
- [19:47]
C
I completely agree that it's moving towards a completely personalized way of working. Right. Like at the end of the day, you don't like. And that's, that was my point in during the last podcast of like, you know, we build our own CRM tool because we want everyone to essentially customize things to their own liking. Like I don't need to see something that works for a thousand people. I need to see something that works for me. I don't care what works for the other 999 people. And I think that that's where we're moving with everything in AI. But like the, the to carry point earlier and I think that was a, an even deeper point about the fact that at the end of the day, even if AI, like, even if you know whatever your design is, 7.5, but you don't think deeply through what it actually means, you're wasting time, you're wasting Time for everyone. And it's funny because we have someone at the company who we had a meeting and someone essentially wrote a seven page document completely produced by Claude. And I essentially told the person, listen, if it takes me more time to read what you're writing, then you're not saving time, you're just shifting the burden to someone else. And same thing with what we mentioned about coding. At the end of the day, if you want a human in the loop, I can write 500 lines of code in a minute. But then if someone else needs to review it, you just shifted the burden. You didn't actually create any productivity.
- [21:07]
A
That's fascinating. Using the AI tool to dump a bunch of slop on your coworker's plate is lame. Using AI tools to make you work faster and then putting the final chef's kiss on top of what AI helps you build, that saves everybody time. That's worth it, Correct?
- [21:31]
C
Yeah. And you should deeply think in the ways you use AI, Right. It's not just like creating slot, but really think through it. And I guess it goes to. I know it's that there's that old. That old quote that I think it's attributed to Mark Twain, but I think it was written by Bless Pascal. Like, if I had more time, I would have written a shorter letter, Right?
- [21:54]
A
Yes. Apologies for this long winded letter.
- [21:57]
C
Exactly.
- [21:57]
A
I didn't have the time to write a short one. Is exactly what people need to understand. The way I've been using this AI is people ask me questions, then I take Whisper Flow, which is awesome. We gotta have the Whisper Flow founder on our roundtable again soon. He's awesome. And do you guys use Whisperflow at all? Have you used it? Oh, my God. It's like super life changing. What it does is you dictate to it, like Siri might, but it's AI, so it understands the context. So if I told Whisperflow, hey, make me a bulleted list here, It's a shopping list. I want it to be check boxes. These are the things on the shopping list. It says, here's your shopping list, and boom, it does all the formatting for you. So you can kind of be having. It's almost like having a secretary in the old days. Hey, can you come in, Miss Moneypenny, and take a memo? Sorry if this is sexist, but, like. And, you know, they come in with the pen and you're like, okay, I need, like a salutation for, you know, the king of Saudi Arabia. And then I need an opening sentence. I want it to be something like this. And they kind of interact with, interpret what you want and then give it to you. That's what Whisper Flow does. So what I'll do is I just put it on. I do a real long winded prompt. Like the person who works for me, who's an associate at a venture firm, asked me how to do a deal memo. Can you please do research on the best deal memos ever written by venture firms? Which sections are important, and examples of crisp, clean writing that led to funding from a deal memo? Then make me a super prompt on how to make deal memos that I can save to memory. I just like, speak for 90 seconds.
- [23:32]
C
They'll end up taking your Sequoia Scout deal memo, I guess.
- [23:35]
A
Yeah, exactly. And so it then winds up going. And then I just share the prompt with them. I'm like, here's how I would have solved your question with AI before coming to me. And they're like, okay, boss, I'll do that next time. Yeah. You're referencing the famous Uber memo in which, talking about writing a short letter, it says, why should Uber exist? Why should this company exist? And I, I told Roelof, I don't want to write deal memos. It's a waste of time. It's like, can you just please do this so that my partners don't cancel the Scouts program? And I'm like, fine. So I write like three or four paragraphs on how, you know, Uber changes everything. And then I was like, this is. Anyway, this is our word. I delete it all and I say, because cabs suck. And I send it to Ruloff to troll him and he writes back, that's perfect. And then the way the original Scouts program was done is they took 5% of the gains and they distributed to everybody in the program who had done an investment. And then they split the rest between their firm. Back in the day, they split it like 50, 50 basically between the Scout and the thing. Now I think it's the 35% to the person who did the deal. Putting all that aside, I get like, Uber goes public. Roelof sends my deal memo 11 years later to the Scouts program and says, each Scout in this, like nine person original Scout class is dividing up. I don't know if it was $20 million worth of Uber shares. They each got one and a half million. I literally started getting presents from the other cohort. They're like, jake, Hal, you made me a million dollars for coming to two Scout meetings. And I never did a Scout investment. Thank you It's a funny story from the archives. All right, we gotta get to work here.
- [25:25]
D
Yeah, I think I was gonna add like, I think like the mistake probably like a lot of people do with AI is kind of like that, like one like write me a letter or write me this thing or write me this plan. Whereas I think like a better way is like ask me questions about I'm working on this thing, ask me a question what I should be thinking about. And the point of using the AI is gain understanding yourself, not just generate. Here's a plan for Jeremy. Go read it. But did you actually understand anything? What is in the plan? Is it a good plan? It sounds good, but is it actually good? So I think that I use it more like a reflective surface where it's helping me to think through problems instead of solving the problems for me or like generating the solutions.
- [26:12]
A
I got to, I got to, I got to remember to do that.
- [26:15]
B
I was just going to say actually, if you work this way, we actually open sourced something we call blueprint, which is a thing like an agent skill that asks you a bunch of questions. Really simple, but it's tuned to ask you really good questions because Claude by default asks you shitty questions. The questions are terrible. So if you search in View Blueprint, you'll be able to find it. But what we found actually is that if we first answer good questions, then we can have it one shot, much harder, much bigger things because now it has the context from my brain about what it needs and things like that. Yeah, exactly. We use it primarily for code, but I've been using it for email, I need to write or a document and it does a pretty good job searching for stuff.
- [26:56]
A
Amazing. Yeah. Knowing what questions to ask. This is actually kind of a sign of huge intelligence. And also entrepreneurs who succeed, also capital allocators who succeed is figuring out what questions you didn't know to ask and finding a tool for that. I think this all falls under. Jeremy, a new category we need to bring up and codify here on this week in AI. I'm going to call this AI etiquette. In corporations and in personal life, people need to understand AI etiquette. Yes, let's snap it up for AI etiquette. Do not just dump AI slop on your co workers. Figure out what the goal is and yeah, actually succeed at the goal instead of dumping slop. We need like a whole playbook for this. I don't know what else should be in the AI etiquette playbook.
- [27:49]
D
I mean, the code reviews.
- [27:51]
A
Go ahead.
- [27:51]
D
The Code review is definitely like the one where like if the agent writes the code for you, then you should be one reviewing it first. Like you can't like send it to review because you didn't even read it yourself. So like, why is someone else having to read it first? So I think first read it yourself, then like use the AI reviewer and then like maybe like get another person to review it. But like you can't send directly to other people.
- [28:15]
A
Your agent is your responsibility. Your code is your responsibility. Take responsibility for the output. That's the AI etiquette. AI output is your responsibility. Don't dump it on your co workers. Absolutely. And then take out the EM dashes. Okay. And always give the instruction, can I say this in less words with less fancy words? Always use less words, always cite your work, always double check your facts. I do this as like an ongoing prompt. At the end I'm always like, please double check your facts. Please do citations. And then I use this model council on perplexity. Have you guys used Perplexity Model Council before? Oh my God, it's so great. I wish I was. I was in bed with my wife and I used it last night. Sounds a little weird, but we're going to Japan in the summer. And I was like, she's like, tell me about this model council you keep talking about. I'm like, well, let me show you. We're going to Tokyo with the daughters for 10 days. We want to do five, six days in Tokyo. Then we want a three or four day road trip to a beach that is within a three hour train ride. The train ride can be part of the experience. Let me know which beaches I should consider within a three hour train ride of Tokyo. And it comes back with here's what Claude 4.7. Here's what Gemini, latest version. Here's what 5 point whatever of OpenAI. Then it says, here's where they agree, here's where they disagree, here's why they disagree, and here's what I think you should keep in mind. And here's a table with the three different results. It's like, whoa, dude, it's so awesome to have it just run that way. I don't know how many tokens this burns, but it's at least three. It's got to be five times the amount of tokens. I think highly, highly recommend checking it out. What, what do you got on Jeremy? Your AI. What's. What else is on your AI? What are we calling it?
- [30:11]
C
Etiquette list.
- [30:11]
A
Etiquette. What's on your AI. Anybody got anything else for their AI etiquette list? Things that are pet peeves.
- [30:16]
C
Coding would definitely be one of them. Where make sure to check your own code. Same thing with writing, to be honest. It's like, just make sure that, as you said, it's not different than the work you did before. Whatever work you produce with AI or the work you produce before AI, it's your responsibility to make sure that you're willing to present that work to your boss, to your coworker, to anyone else at the company. And that shouldn't change whether you're using AI or you don't use AI.
- [30:43]
A
Use AI. Yeah. Ownership over output or ownership of output. Ownership of output. Kenjin, you got anything else on your list? I don't think you should be responding to anybody who's a friend with AI content. Like interpersonal relationships, friendships, lovers, spouses, children. There should be a no AI response rule. If you respond to a friend with an AI response. That is super lame.
- [31:10]
C
I just had a funny story the other day. A candidate reaches out, out of the blue, and I look at his resume, LinkedIn, and I'm like, this guy seems pretty interesting. And so I schedule a call with the guy, get on a call, and I'm like, oh, thank you for, you know, whatever your thoughtful email. And he's like, oh, I don't even know where I. Who you are. Like, my agent essentially reached out and applied to. That's pretty amazing that you just, like, completely, like, had no idea who I was. And, like. And so he's probably done that to 200 people. But, like, he's done it. It actually looked really legit, so.
- [31:48]
A
So you get an outbound, you respond to it, and then you find out the person who sent it. It doesn't even know. This came up for me the other night. I was talking to this young couple who I met at a party, and they. The woman said, yeah, you know, you know my uncle, blah, blah, blah. And I was like, oh, yeah, I know Maybach Mike. And they said, yeah, I run his, I don't know, Hinge or whatever dating app he's on. I'm like, what do you mean you run it? It's like, I put in the filters, I picked his photos, I made the profile, and then I go in for Maybach Mike, and I do all this work, and then I set him up on dates. I'm like, do the women who are going on dates with Maybach Mike know your writing the responses in the chat, or are you doing the response to the chat? Or you hand it over to him to do the chat. Oh, no, I do all of it. He just picks up at the date. I'm like, okay, that's unbelievably dystopian or brilliant, I don't know which.
- [32:54]
D
That's like agent fishing. Like a cat fishing, but with agent. So.
- [32:58]
C
But I feel like that woman's job is going to soon be displaced. With an agent you can at that point just have an agent do that.
- [33:05]
A
AI phishing, then that's on the AI etiquette list. No AI phishing, no AI between humans in imessage or whatever it is. It's just. It's super lame.
- [33:15]
B
It's just lame. Yeah.
- [33:16]
A
Andre Karpathy, topic number one, has joined Anthropic. He did a post at 10am this is a breaking news story here on this Week in AI. Quote, personal update. I've joined Anthropic, period. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R and D. I remain deeply passionate about education and plan to resume my work on it in time. Obviously he worked at. He was one of the co founders of OpenAI. He worked on the Autopilot vision team at Tesla. He founded an AI education platform, Eureka Labs in 2024. He coined the term vibe coding. He introduced auto research on March 7th. That went crazy. That was the open source Python tool that let AI coding agents make large language models or machine learning models, experiment and get better over time. And civilians started doing it. Kanjan, how important is Carpathy to the future of AI in 2026? Obviously super important previously. And then why did he do this?
- [34:29]
B
Andre. Andre, in 2020, 2016, you. Your first question is how important is him?
- [34:38]
A
Is he in 2026? Obviously in 2016. But I like the way you went with it in 2016. I mean, invaluable person in 2026. There's a lot more people who are there. So I'm just objectively looking at talent in 2026. With the industry exploding to the extent it has. How important is this of a move or is this just, you know, grabbing a bag and getting a ridiculous offer from a company that is apparently trading at 800, $900 billion with 50 million in revenue or 40 million. 40 billion in revenue.
- [35:14]
B
Yeah. Andre doesn't really care about money. So I think if he joined Anthropic, it's because there's something there that is interesting to him where he feels like, you know, it's uniquely suited to his contribution or something like that. And so, yeah, I think, I think he's very, very talented. I think there are a lot of. A lot of other very talented researchers. But Andre has kind of like a nose for, you know, a certain shape of thing that, you know, he wrote this viral blog post a long time ago on recurrent neural networks and what was wrong with them. And I think he's really good at kind of like framing things in a way that hit, like educationally hit and help people kind of talk about things in a different way. And it's, I think, why he has such a following. And so for Anthropic, it. It's maybe even less about research and more about some other aspect of things for him, I'm not sure.
- [36:17]
A
Fascinating. Jeremy, what do you think?
- [36:19]
C
I mean, of course he's really talented, but I think to me the story is more than just him joining Entropic. It's the fact that over the last few weeks we've seen Entropic hire quite a few senior people and especially like City, I think, CTO of Workday or whatever became an IC at Anthropic or something like that. Right. Like, you've seen a lot of, you know, people from CTOs and people on top management levels just going back. Going back to technical staff positions. And I think that that is the more interesting story to me, the fact that Anthropic is hiring all of those. All of those senior people to join them. Now, I wouldn't go and then say that therefore OpenAI is dead. We were talking about that right before you joined, Jason. It's funny because things have shifted so often over the last few months. A few months ago, six months ago, whatever people were saying, Anthropic is a dead company. And then since the beginning of 2026, Anthropic has had an explosion in terms of growth. And then people shifted towards saying OpenAI is that company. But I don't know. I think that I would never underestimate a company that raised $120 billion in one round alone. So I think that both of those companies are currently dominating. The question is to see whether that will remain the case or whether that will change over time.
- [37:46]
A
Kari, why did Andre join anthropic? Not OpenAI, not thinking machines, not Google, not Meta, not SpaceX. Why would he go to Anthropic at this moment in time? And how important is any one AI executive in today's world with the industry expanding to the extent it has in your mind?
- [38:09]
D
Yeah, I Think continuing some of the other comments, I think to me, probably what I see the most exciting part is he has this quite original ideas of how to apply AI in different ways. He had this knowledge wiki kind of idea and then the white coding. And I feel like a lot of the AI is still hindered by the understanding how to use it or where to use it and what is the shape of it, how to use it. I do think maybe sometimes it's easier to go to a different company than return to a company that you know, because I think like, if you want to do something different or like differently, you probably can kind of carve out that more easily in a new company. Like in Anthropic, where he doesn't have a history of doing stuff. And then like, I think there's a danger. I don't know, like I don't know him, so I don't know any kind of like how he makes these decisions. But I could see like sometimes like going back like something like OpenAI, for example, is like he might get kind of stuck doing the same things he's known for or something. And so with Anthropic, maybe they found a really good place for him where he can have a lot of impact. And maybe it's not hardcore research, but some kind of applying or shaping that research to something more useful.
- [39:29]
A
There's three things to say about Andrej joining Anthropic. I think the first is he's an incredible communicator, as you were pointing out, Kari, and he has a way of, I think, building massive consensus around how to frame AI, how to apply AI. That credibility is a massive communication tool for Anthropic, which every time Dario speaks, he creates massive Investigations in Washington D.C. massive hand wringing. Because Dario, I think, can't shut up about his P doom. He's got a P doom of like 90%. He thinks like 90. This thing spirals out of control and everybody loses their job. And this is a really. This has caused a lot of the hatred of AI. I believe he is like too honest about this scenario, you know, And I don't have any reason to not believe him when he says, you know, he thinks all jobs are going away. I truly believe he thinks that. I also think he's wrong. But he's the leader now of the AI industry by, you know, the fact that his company is doing the best and people are listening to him and they're weaponizing what he says. The Democrats are weaponizing it to stop data centers, to regulate AI. He's the worst possible spokesman for AI in America. Whereas Andre Karpathy is a great spokesperson, I think. And that might take a little of the pressure off. He could, he could be doing that. Role number two, when you have a massive market cap and valuation, you put it to use and you take talent off the market. Google did this and Zuckerberg did this. It is a great playbook. Somebody's talented, get them off the market so they don't start a company, get them off the market so they don't join a competitor and leapfrog you. That is a tried and true playbook. So kudos to them. If they gave him $10 billion in equity, even if he's not motivated by money, it just locks him in. It becomes golden handcuffs. It becomes like the ultimate golden cage. Right? Like how do you say no to $10 billion in options? And how does Dario not offer him $10 billion if he thinks he could make the company 1% better? And he will make the company more than 1% more value? And 1% of a trillion dollars is $10 billion. Right. So it's. You have no choice but do that. I think the communication thing is an interesting juxtaposition, by the way, to OpenAI, which bought like a bro podcast, like a tech Bros Podcast. And I don't think that's worked. All it's done is made the competitors not want to go on the podcast. And the podcast now feels like, okay, it's just like a PR arm. So that's an interesting juxtaposition. And third, and finally, each of these companies is turning into a bit of a cult. And SpaceX is the tech libertarian monk cult. It's the engineering cult. We want to be tech monks and be in this town on the edge of America building a starship that goes to Mars. If that appeals to you. And it appeals to a lot of tech people to be multi planetary to answer questions like, you go work for Elon. If you're woke and you're left leaning, you're a Democrat, you have a P doom that's a little higher, you join anthropic. And if you're a cutthroat capitalist dealmaker, you join OpenAI. If Sam Altman's screw everybody and build the valuation and do deals and piss everybody off appeal. So you go work for him. And what has that done? I think OpenAI is losing the talent race to those other two. That's just my very surface level evaluation. Tell me what I got wrong.
- [43:33]
B
Kanjin, you always call on me first for these.
- [43:36]
A
You're expressive. So I have thinking machines hooked here. Mira gave me an early release of it. So I have thinking machines and it told me there's an 87% chance based on your facial expressions that you want to add. And the other two guys had like 40s and 50s.
- [43:52]
B
Oh, thinking machines.
- [43:54]
A
It's doing real time assessment of your propensity to drop a banger right now. So do it.
- [43:59]
B
I don't got a banger for you.
- [44:01]
A
No hot take.
- [44:02]
B
No, I think you're not wrong, but also that none of these people are doing this intentionally. I think like Sam Altman just is kind of like screw people over kind of person. And so the leaders will always attract the team and culture that they're building. And it turns out that screwing people over in business is not a good long term strategy. Which I think is why they, you know, ultimately it probably wouldn't work out for them. Whereas I think, you know, Dario is a very like, like he's very earnest. Yes, it's very hard to get him to be not earnest. And that's a, that's a, you know, he's kind of like a pillar. You can't move him. And in that sense, you know, if that pillar is pointed in the right direction, then the pillar is just going to keep going in that direction. And so I think that's what we see in the industry. Like Anthropic was like, okay, this might kill everyone. We should try to make them harmless and honest and helpful. And also we should invest in coding because that's how it's going to self improve and let's just keep doing that
- [45:07]
A
and we should hold back a release because it might not be good for humanity.
- [45:10]
B
We should hold back releases. But I think the thing that is weird about Anthropic is is that even though I want to believe that they're the good guys, they're also doing a lot of very anti competitive things that are like very capitalistic. And maybe that's not though, you know, that's not just the like, maybe it's not intentional, maybe it's just the incentives. But you know, they like locked out a bunch of third party providers like Open Code and Open Claw. They're like you're, you can only use the API. Like if you use first party Anthropic product, Anthropic products then you can use the Max plan which is like a 20x savings on your API token.
- [45:45]
A
So very subtle way to penalize open source and competitors. Most people would look at that. The FTC is not SAVVY enough.
- [45:56]
B
Exactly.
- [45:56]
A
To understand how cutthroat that is.
- [46:00]
B
Exactly.
- [46:01]
A
It took them 20 years to figure out what bundling a browser in Windows meant, right? It took them like 20 years to figure out what putting Google services at the top of search results in the first 300 pixels did to Yelp. Like if we're going to put local up there, you do know you no longer need to go to Yelp. We will take the oxygen out of it. They're just not savvy enough to know how anti competitive that is. Is a really good point.
- [46:29]
B
No, they're not. And so like how's the market going to play out? You know, like Anthropic is saying, okay, use Anthropic first party products. Like, you know, try to try to keep the data inside Anthropic first party products. You build your whole workflow on Anthropic first party products. You can't swap out even when OpenAI is winning. Has better models going to be expensive to swap out.
- [46:50]
D
But to be clear, that's only for the individual plan. So companies have to play for the API plans anyway. But I do agree that this pricing is kind of messed up there out there in the market. And my take on the what they're doing is that I don't think they realized what they were doing. So creating this plan is like, oh well, people code some stuff on eight hours a day. They didn't expect someone's going to hook this up to some agenda system that will run 100 agents every minute of the day. And now they're realizing, yeah, that doesn't work. So we have to put some limits. But that's the mistake. The classic pricing mistake is like, it is always annoying. You shouldn't give out things that you will take away later because people always hate you. You should have started from like, no, don't use anything. And then you can only buy this plan if you use it with the first party products. And people would be, maybe people still don't like it, but at least they knew what they were buying but now they bought something and then now it's taken away from them. And so obviously it's not fair.
- [48:00]
B
The other thing though is that what's really interesting is that Anthropic is like, oh, you can't use Claude Pro Max subscriptions programmatically using Claude P. But now with Vibe coding, anyone can actually just hack this terminal agent to do Claude P without using Claude P. So I think there's this really interesting battle that is happening or about to start. I was calling this punk software. People who are building software to tear down the walled gardens and try to keep data their own and try to make it so that we're not captured. You know, we don't get stuck in these walled gardens of the next generation of AI SaaS tools. I think this is. You know, you were saying, Jason, that everyone's booing AI at commencement speeches.
- [48:45]
A
Yeah, it's our next story.
- [48:46]
B
It's our next story.
- [48:47]
A
Okay, well, we're about to go get Jeremy.
- [48:48]
C
You want to take me?
- [48:50]
A
So, like, the all you can eat buffet is an incredible draw in Vegas. It's an incredible draw, you know, in your small town, to get people to come for lunch until a competitive eater shows up and eats every one of the shrimp. And then when the shrimp plate comes out for the second round, they're waiting, and they just take the whole plate. Like, it's all you can eat with a star on it.
- [49:13]
C
Yeah, but. But, like, to be honest, I'm. I'm a bit puzzled by why anyone is surprised by that. Like, I mean, that has been like, it's not new to AI, right? Like, Amazon is doing that with resell, with sellers. Apple, as you mentioned, did it with the App Store. Google did it with search distribution. I mean, that same thing happens over and over again. Right. I guess the difference now is that the platform itself is programmable intelligence, which is different than. There is a difference between the boundary, what you had before, which was the platform, and the application layer. Now, that gets blurry. And so that's why, personally, I would be worried about building on the application layer, because you know that those platforms are going to try and verticalize everything
- [50:01]
A
they're going to steal. Your idea is basically what every developer needs to understand, whether you're dealing with. You know, if you think that Sam is inherently, you know, chaotic, evil, you know, in terms of Dungeons and Dragons. And he's the scorpion on the frog who's like. Frog's like, why did you sting me? And he's like, well, I'm a scorpion. I mean, duh. It's like, it's in the packaging. There's like a giant stinger. Like, did you not see it? And then Anthropic is like, we're the good guys. We're looking out for humanity. But we do need to figure out how to pay for Colossus 1. We have $100 billion bill due to Elon and Oracle and whoever else is building capacity Google for us. They're going to need to do the application level and verticalize Everything here and if you're training it, you know, then maybe you're the frog with the snake on its, with the scorpion on its back. Whether the scorpion is, you know, upfront about being a scorpion in the case of OpenAI or they're camouflaging it in some way and it doesn't look like it.
- [51:08]
C
I mean and yeah, and I, and that's why I think that you know, if you're a developer or company like you need to find way unique ways to keep your, to keep your customers. Right. Whether it's, I mean like think about you know, Apple, right? Like Apple did it, didn't they do that with Apple Music? After Spotify launched they saw there was so much popularity they built Apple Music and what did Spotify do? They found creative ways to keep their users whether it was at a certain point I believe they paid to have Joe Rogan only be able to have his podcast on Spotify. But what Cursor is doing in terms of creating a really great experience for some developers or you just need to make sure that whatever you build is not, doesn't depend exclusively on the Frontier Labs in order to maintain an edge over those Frontier Labs building in those verticals.
- [52:01]
A
Kenjin and Kari, you both I believe are building on top of Frontier Labs. Does it make you nervous that they could rug pull you, they could study you and copy you and they probably are doing that already. And then what's the solution there? Do you eventually have to take an open source model and train some models for yourself and stop sharing your data with the anthropics of the world? With the OpenAI's of the world?
- [52:28]
D
Yeah, I mean I think like no one really knows like where this will lead and like I think that's still the question is like how much they're going to verticalize and like to what degree. Like we haven't quite seen that yet. Like yeah, like you have a cloth legal or design but it's kind of like a prompt you run like it's not like a full fledged product in a way. And so I don't know if they will have the appetite to solve a lot of problems and for example linear A lot of times I think about it, it's not really a technical problem, it's more like a human problem. Like how do you organize things in organization or how is the organization productive and gains the right understanding about how things are going. So I don't know if they will have these labs will have the motivation to solve people problems this way.
- [53:16]
A
Like do you worry about it and do you have a plan? Like if you are the founder, you must be keeping it front and center in your mind and what's your plan if they do suddenly say, hey, here's a straight up competitor to linear.
- [53:28]
D
I think like it's that the product is the plan and I think the, the, the more we can, like I, I think like we can build like we, we are building one product, they're building maybe hundreds or thousands of products if we assume that they will verticalize everything. So I think in the end it's like who will have the higher motivation and what is the exact advantage the Frontier lab can have? Because in some ways we have the ability to do our stuff and then gain the AI LLMs to use in our product too. So we can use their technology, they can really directly use our product or like the product thinking in a way. So I think like what we, our advantage is the product thinking and the product like the team we have and the relationships or whatever understanding we have.
- [54:19]
A
You'll always be focused on the customer. But I guess the question Kanjun is at what point do people on the application layer say yolo, I'm going to build, I'm going to build on an open source model. I'm going to host it locally. I'm going to have my customers hosting, you know, a verticalized small language model in slm. When, when do people start, you know, protecting their kingdom, putting walls up to stop those frontier models from absorbing everything? Or am I being paranoid about.
- [54:58]
B
Not paranoid, not paranoid. No, I think responding to your point, Kari. Like if we look far enough into the future, like this thing is going to be incredibly like all of these models are going to be super capable, much more intelligent than we are as humans. And so, you know, they can come after every application, every industry and anthropic and OpenAI and the model labs will verticalize in the industries that are profitable. So if you have, you know, if you have customers that they want and they think that this is a good industry to go into, they will go into it. So I think that risk is real, like I said, a huge and real risk. And that's why I'm concerned about power, concentration and the centralization because these models are incredibly powerful as a technology and they're incredibly intimate. You know, they know everything about us. They can control us, they can influence us. Intimate in the way that, you know, our software even today is very intimate. But the models are even more intimate. Claude knows more about me than my partner knows about me. And so and that does not say anything about my partner. I love him very much. We're very close.
- [56:03]
A
But. Well, you can get a key if you get a keystroke logger on your devices and then you could just send that to your partner and then their LLM can.
- [56:11]
B
Yeah, he can digest, you know, digest
- [56:14]
A
it and be like, hey, honey, I understand you got your password wrong four times. That must have been incredibly frustrating for you.
- [56:23]
B
Exactly. If Claude doesn't get there first.
- [56:25]
A
Yeah, exactly.
- [56:26]
B
But to your point.
- [56:28]
A
Yeah, continue.
- [56:28]
B
I was just going to say, to your point, I think actually that it's very important how we architect the industry as we're going forward. So one of the things that we released recently is an agent orchestrator where you can swap out the underlying model or harness really easily. So whether you use Claude or OpenAI or an Open source model, and everything that we're building is on top of that. And so that means that everything that we're building, you can easily swap out the model underneath. And also everything we're building is like, meant to keep user data yours, like local or in your own servers, and, like, meant to not verticalize. And so just trying to figure out, like, how can we create an alternate ecosystem, an alternate path where it's not just like everything getting swallowed into this giant juggernaut machine. I think actually as an industry we need to decide, like, each of us needs to decide, like, no, I'm going to choose a path that does allow me to commoditize the underlying models and I'm going to use these orchestration systems that do allow me to swap out the underlying models really easily. Like, I think that's actually really important in order to stay competitive and keep it a competitive industry, because the FCC is not going to help us.
- [57:40]
A
This is the key, Jeremy, in my mind, which is if you make your product headless and it can swap out, it has an orchestration level. When I use Perplexity, I don't know what model it's going to use. I just ask my question. It routes it intelligently or I use the model council and I'll pick ones just so I can stay up on them. And then I did the same with my Open Claw. It had Kimi 2.5 in a virtual machine that we started referencing, or it was using CLAUDE and it would just pick itself, or I would tell it in my prompt, use CLAUDE for this. Use this for this. That seems to me to be the way to defend against everything going to one player, as is open source and these very powerful Macs. I have the M5 right here. I unplugged my Mac Mini and I started plugging my system into the M5. I have 48 gigs of RAM. I'm thinking of just giving this to somebody on my team and getting the 128. I don't know if there's a 256 gig one, but having all that extra memory is just so amazing. And I want to start running local models. And people are now daisy chaining them. They're getting like five Mac M5s, putting them into like a stand, daisy chaining them together. There's a piece of software to have it addressed as one cluster, and that's actually the most affordable cluster. Now you can Exolabs is the company that does this. We got to have him back on the program. But somebody do a quick picture of it and pull it up on the screen here. Alex Chima is the founder over there. And until the studio comes out like this, to me, Jeremy feels like the future. Yeah, here's Exo Labs. Those are pictures of like stacked studios.
- [59:23]
C
Jason. You know, one place where I've actually seen that with my own eyes has been with, you know, enterprises. A lot of enterprises, you know, have like data is their moat and they want to keep total control and sovereignty over their data. And so what we've, we've actually embraced that. So what we have done is I believe we're currently the first and only company to have built something called Confidential Compute with Amazon, which means that we can deploy our models fully within a customer's own environment, own VPC and encrypt both the model at the architectural layer and at the weights layer. So the enterprise knows that the data stays with them and we know that they cannot decrypt our models. And so that essentially gives both sides a win win. And that's as opposed to what we've seen with the LLM world, where if you use OpenAI1Tropic, as you mentioned, you move your data to their infrastructure, which a lot of large enterprises are not willing to do, or you use open source models, but then you have to host everything yourself.
- [60:26]
A
Yeah, we have a company, Abacus, that is doing proprietary language models for organizations and we incubated them and they pivoted their way to this.
- [60:36]
D
But
- [60:39]
A
it's very interesting. Companies now are saying, I want to every week take our data and train a model and then essentially run a job every week to do reinforcement learning on it just based on our data and never share that model with anybody. Like, this is clearly the future. Here's a picture by the way Kenjin of 5 MacBook 128 gigs and I think these cost 4k each. So 5 of them would be 20k. I believe that this. I'm stalling so my guys can put it on the screen. Wake up guys. This is the cluster of the future. So here is I guess three or four. That looks like four of them. And yeah, you just use exo labs and put them together. The fans must be blowing really hard on those because they don't have great heat sinks. But I believe that this is the most affordable stack today. Wow, here's abacus. So it's Goabicus Goabacus co. But they're making this box and you put it on and you put it in your organization. I think the box costs 75k or maybe it's 75k for the year and they incorporate building your language model. And this is for regulated industries. This is Jeremy, a good partner for you. I'll introduce you to the founder Offline or Oliver. Can you do that? You might actually have a good partnership here of putting your software on theirs as well. But yeah, on prem I think is coming back.
- [62:21]
D
But I'm wondering about the training. Like does it like isn't it one of the like the bitter lessons that like the training always kind of like ends up kind of useless eventually because the model capability keeps increasing. So like yeah, you will kind of eat this bitter lesson every, all the time and maybe it still makes sense. You just keep training and training but it's like kind of the capabilities are getting maybe like greater faster than your training produces results. So I'm just kind of questioning like how much actually will help any experience.
- [62:52]
A
Kenjin on this specifically of like hey, I'm going to roll my own model. I'll fork Deepseek or Kimmy and then use all of my own personal training data to. What do you call that when you use your own training data and put it into a model?
- [63:07]
B
Fine tuning or like. Yeah, kind of like continual learning on your own data. Yeah, I think there are a few ways to think about this. One is kind of the way I think about open source versus closed source models is the models will keep getting better. The open source models and the closed source models will keep getting better. It seems like open source is about a year behind closed source and has roughly stayed that way. And so the frontier, you know, frontier tasks in terms of capabilities, like the hardest things, they'll always require the frontier models. But there are some things that even now are Starting to only require the capabilities of an open source model. You know, not all tasks are like infinitely hard and so probably what you want is something that routes things to cheaper models. And I think cost is going to be start being a concern next year is you, you want something that routes things to cheaper models and open source models are cheap. And so how do you do think about fine tuning or training on these open source models? Well, you do want to be careful. Like we don't train our own open source models. We just wait for the next ones to come out. But we do do you can do things like there's a, there's something called Lora L O R A which is a little adapter layer that you can train that you can kind of like put on like a band aid or a patch on top of an open source model or any open weights model and then that will kind of fine tune it to a specific task or specific purpose. And so if you have a training corpus and you train you like make a little Lora patch, then you can kind of like in theory patch it onto any open source model and get your own data and context into that model. Now this is all very early. Like I think a lot of this research is still very early and so it doesn't work. Lora works really well for certain applications but not for others. But I think that's the kind of thing that we have to develop and we'll see more of these little patches you can tag on.
- [64:58]
A
Well, this story silently went through this week, Kari. Chrome didn't tell anybody and they installed a 4 gigabyte LLM. I think it's Gemini Nano and it's doing interesting things like I'm assuming you know, text based things and they are now I think being a little more upfront about it. But your browser has it and I guess websites can call it and it will do. Proofreader, writer, rewriter, all this kind of easy layup. So this to me feels Kari like a step in that direction is everybody will have some Microsoft Tiny LLMs in their products.
- [65:43]
D
Well I was also curious about the harness. So do you create those? Because I think we've been talking a lot about models but model is just, it doesn't do anything really. But then the harness kind of what it makes useful which I think then you could have cloud code is a code specific harness. So could you create more specific type of harnesses for specific tasks and then run whatever open source models with them.
- [66:10]
B
We made this open source harness that this is for coding agents kind of but it doesn't have anything specific to code. And so you can run any agent and the agents will run, will write their own code and things like that. So there's a harness harness that we made that's open source that lets you swap out, it supports, you know, PI and you know, codex and all of these other agents in addition to Claude. And I think that the harness layer, this is actually one layer above the harness layer. The harness is kind of like the thing that orchestrates the model itself and then this is the thing that orchestrates the thing that orchestrates the model. And I think the harness is actually probably going to co evolve with the model. And so you, you know, what we see with anthropic is that the cloud code harness, you know, they're probably training the underlying models along like with the cloud code harness. If you use a different harness with anthropic models, they won't, it won't work as well. And so I think we're going to see that kind of like adapter where the model harness is like a unit. And probably someone will figure out how to make it a little bit easier to tune your harness to a model or tune your model to a harness. But we haven't seen that yet.
- [67:21]
A
I don't want to end before hitting on graduation season. Everybody loves all these commencement speeches. They're so inspiring. And tech leaders have been doing this forever. Seminal, amazing, inspiring speeches for graduates. Steve Jobs obviously did one, but we've had a trio. Eric Schmidt, friend of the pod, Google CEO, Big Machine Record CEO Scott Borachetta, and Gloria Caufield, Tavistock development Vice President. I think these are all different layers of commencement speeches that are commensurate with the school. Eric Schmidt, seemingly a great get. I'm gonna play just maybe 30 seconds of Eric getting booed every time he speaks, and then I'll have you respond and then we'll do Scott and Gloria. So let's start with Eric Schmidt. Here's the high res version of his keynote. It will touch every profession, every classroom, every hospital, every laboratory, every person and every relationship you have. I know what many of you are feeling about that. I can hear you. There is a fear. We do not know. We do not know the precise contours of what this transformation will look like. Choose a diversity of perspectives, including, let me add, and if you, if you'd let me make this point, please. All right, we can, I think stop there. Jeremy. Why are these students booing Eric Schmidt? That's the question, I think, to be
- [69:20]
C
honest, I Think the boots are ridiculous. First of all, I think that those students are hypocrites. I guess they profess to hate AI, but I can guarantee that every single one of them has used either chatgpt or cloth for every first draft of every essay they've written during college. So that's the first thing. And second thing, I think it's also, there's never been a better time to be graduating from college in history. I think that AI is a really powerful tool for your creative expression, and I think creativity and critical thinking will still be very important. But on the other end, I do understand that there is anxiety about job displacement, but I think that that ignores the bigger picture. And I think that Marc Andreessen made the point of fear of AI is more dangerous than AI itself. And I do believe that that's the case. You know, even though, like, today, you know, like, you, when you look at AI, you look at the fact that it's a technology that essentially does what humans do much, but much better and faster. And, you know, maybe AI, whatever, tomorrow, next week, or next month will do things beyond what humans can do, but that won't happen overnight. And I think that, you know, a lot of those fears are overblown.
- [70:35]
A
All right, so, Jeremy, your take. They're entitled hypocrites who shouldn't be booing. They should be embracing. I like the hot take. Let's play our next candidate on who wants to get booed at a commencement speech for referencing AI Here is Scott Borchetta. And then, Kari, you're going to get a chance to comment on the first two, and then I'll play one more. And Ken June, you'll be batting cleanup here. So here's our second commencement speech. Go, Scott. AI Is rewriting production as we sit here. I know it. Deal with it. Like I said, it's a tool. Hey, like I said, you can hear me now or you can pay me later. Hey, then do something about it, okay? It's a tool. Make it work for you. All right? Okay. So, Kari, there we go. You've seen, too Eric Schmidt, who was like, please, let me finish. A little condescending. Then this next person, he's just like, hey, straight up, it's a tool. Deal with it. What's your take on our first two commencement speeches and the boos that they got for invoking the AI super God?
- [71:53]
D
I mean, I think this connects a little bit on the whole problem with the AI Comms that, like, it seems like people in the field don't know who they're speaking to. And I think these commencement speeches, I think the failure really is understanding the audience or reading the room. You don't go to an oil workers union to talk about, hey, do you know what's cool? Electrical cars. And even if you believe that, but what you do go there is like, hey, yeah, there's electrical cars. But you know what oil does? It does a lot of other things, and we should be thinking about more ways to do that. And so I think that they. They. I think these people think they kind of, like, give the speech. They would give, like, I don't know, some, like, tech conference or like some, like, boardroom or something. They're not thinking, like, who are they giving the speech to? And maybe they think that they're giving some kind of lesson here or some kind of. They're trying to help these students, but they, like, if you make a speech, you need to, like, address the emotions first. Like, if you, like, if. If you. Whatever you say, it generates this kind of response. There's, like, an emotional reaction. And like, you. If you are a great speaker, you should be able to, like, work around that.
- [73:05]
A
Like, you should be able to respond properly. I think the person who said, hey, tools. They're just tools. Deal with it. He was a little too hot. Eric Schmidt was a little too cold and condescending. Let's see if our third speaker gets it right. Here's our third commencement speaker. And then I'm going to give you a secret fourth. I'm going to surprise you guys with a fourth hot take. That, I think, is the best one. Here is. Our third commencement speech is Gloria. Go, Gloria. Change is exciting. Very exciting. And let's face it, change can be daunting. The rise of artificial intelligence is the next industrial revolution.
- [73:57]
D
Wow,
- [73:59]
A
this is pretty interesting. Okay, I struck a chord. Yes, you did, Kanjun. You get to bat cleanup here. What is happening? Why are students so vocal when the word artificial intelligence is spoken at their commencement addresses?
- [74:22]
B
I mean, I think it's real. Their fear is real. This is a technology we're building that is threatening and doing what we as humans conceive ourselves to be, which is intelligent species. We are not animals because we're intelligent. And it threatens our very identities as people. You know, where we. We believe as humans, we work to provide. You know, we're here to work and do work, and this technology threatens and says, no, no, no, it's going to do work better than we do, and it's going to do everything that we're doing today better than we do, except maybe have relationships. You know, there are things it can't do. And so I think there's a real truth to what people, what students are feeling. And there's no narrative that replaces it. No one is saying in the tech, people are like, this is a great industrial revolution. It's like, okay, that doesn't really resonate with people. You know, the real truth is like, this is telling us where we fit as humans in an ecosystem of intelligence, species, animals, life and other things. And perhaps the thing we should not be doing is cranking out emails every single day. And that's not something that we're very excited to do all the time. But I think the other thing that it points at is also a real sense of disempowerment that people feel that, you know, students feel the, that the younger generation, a lot of younger friends I have, feel that they feel like kind of the powers that be are controlling them or they don't feel like they have very much say. In American democracy, we kind of see this and so it reflects a moment in time. And I think that disempowerment is real in a sense, yes. Kind of. Why? You know, I, I, I speak to it as like, oh, we're building technology and that technology controls people. And that, that's real. The addictive feeds the phone addiction, you know, all of that profit seeking. Like, we have teams that optimize for engagement, engagement of people's attention. Like, we have internal teams at tech companies that do that. I think there's a real kind of like moral question as to whether that's a good thing. I think, Kari, I doubt that Linear has a team that optimizes for engagement. I love Linear. I think it's a really good product. And I don't think that you think about product building that way. And I think there is a morality to building products that we aren't respecting and people feel that.
- [76:41]
A
Yeah, it's very interesting. There is a moment in the TV show the Sopranos where the grandmother, Tony Soprano's mom, who is like kind of mentally ill, you know, Alzheimer, whatever, and she just says to Tony's son, Liv Soprano, she says it's all a big nothing, like incredibly nihilistic, incredibly like demoralizing. And she says it to like a 12 year old kid and he's kind of crushed by it. He's like, life is just nothing. And that's, I think, how these kids are feeling now about their future. They've been told, they hear Dario saying all jobs are going away. They hear See these clips on social media. Then they see Mandami in New York saying, these billionaires and Bernie Sanders saying, ronna, the whole cohort saying, hey, we are these. This is the largest wealth creation ever and you're not part of it. And even if. And then they see the layoffs happening in a very large way. And you know, Jack says, hey, listen, I'm just going to cut half a block because the writing's on the wall. And Coinbase does the same thing and they get rid of 15, 20% and they see Meta not hiring. And Zuckerberg cuts 20,000 people here, 10,000 people there, and says, hey, it's all because of AI. They're listening, they're going into the job market, they're talking to career counselors, and all they're hearing is all the entry level jobs. There's no training program, there's no professional development. All those first rung of the ladder, first, second, third rung, they're all just being taken off the ladder. And these companies are going to become wildly profitable. So if they hear that, they go, well, this kind of sucks. And then you have the audacity as Eric Schmidt to come up there in their minds. I'm not saying this is my belief. And he says, listen, you're scared. I understand. It's like, well, fuck you, we're not scared. You're telling us that you don't need us, the world doesn't need us, and that it's all a big nothing. Okay, if it's all a big nothing, then fuck off. And we're going to boo you off the stage because you're an out of touch Gen X or boomer who made their money and now is torching the entire fucking world with your AI. That's their feeling. Now, are they right or wrong? Are they, Luda, it's, you know, okay, yeah, I think they're misinformed. It's the greatest opportunity to start a company ever. But it's the worst opportunity probably to go work in big tech. All we are thinking about is how do we make companies that have 10 or 25 or 100 million per employee. It's the best opportunity to do a startup because two or three people can get to a million in revenue faster than ever and that's $333,000 each. You're a golden and you own your company. It's the worst opportunity to work in big tech because they don't want to train people anymore. They're just taking their 10x employees and making them 11x 12x 13x. And that's a better strategy if you're Zuckerberg, if you're Brian Armstrong, if you're Jack at Block. It is better to just give more tokens to your top performers, give better hardware, give them more H1 hundreds, H2 hundreds, whatever it is, that's a better strategy. So it's like a best of times, worst of times.
- [79:49]
C
Yeah, but thank you.
- [79:50]
A
I mean, go ahead. You're. You and I might be on either side of this. I don't know if we have common ground or space between our two different views.
- [79:56]
C
I don't know. I just think that, you know,
- [80:00]
A
a
- [80:00]
C
lot of those young people are also, in a way lazy. Now is the time to reinvent yourself. And the difference won't be education, in my opinion. It will be whether you're leveraging AI or you're not leveraging AI. You say all of those companies are firing people. That's true, but at the same time, OpenAI I was talking to someone in the co dev department a few weeks ago and he said they were at roughly 4,500 people. They are going to double by the end of the year. You know who's getting those jobs? The top, the top people. People who are using AI and people who are reinventing themselves and people who are, you know, performing better with the new skill sets at the same time. Like, you know, Jensen Huang, I believe, said something along those lines at gtc. You know, if you're a CEO and you fire half of your people, that means that you're not a creative CEO because, you know, you should essentially have those people and leverage them with tools to perform better. And so I think that all of those things are true. While it is true that there is real fear about job displacement, but I think it is overblown. And when you look at the Industrial revolution, I'm sure that the same things you just mentioned, Jason, were also felt by people then. It's different now. Yes, we're automating cognition. It's different, I agree. But still that same fear on an individual basis was felt then. And I don't think that, you know, anyone looks now and says, you know what, I want to go to a pre industrial revolution world. That's just not true.
- [81:25]
A
And so, and then, well, you know, and to Eric Schmidt's credit, he does say, you have the choice, you have the agency to do this, go use the tools. And they didn't like that either. I think he has a bit of off putting. And I know Eric, but he's so Hyper intelligent and he thinks he's giving them the best advice. But the way he packaged it was a. Was a bit kind of.
- [81:50]
C
It's not, it's not just Eric Schmidt, it's all of that, like the three people you showed, right? Like, the reaction seems to be the same from those students. And the point here is, like, just make sure that you know you nothing will be handed to you. Just work hard and make sure that, you know you do something with it.
- [82:04]
A
Okay, here we go. You each get to give your five sentence commencement speech that you would have done better. Okay, so I give you all a chance to think about that. This is your assignment. Here we go. I'll go first to give you guys
- [82:20]
D
a little bit of time.
- [82:22]
A
Graduates, parents, administrators and teachers, thank you for coming out today. It is an honor to address you here at Harvard University. We stand at the crossroads of the creation of superintelligence. And this is an incredible milestone for humanity. And it's a huge technological innovation. But all of these innovations mean nothing without humans. The creative spirit, the intentionality. So as you go into the workforce and maybe you're hearing about job layoffs, or maybe you're hearing your career is going to be automated, understand that nothing, nothing is more powerful than the power of intent and individuals working together towards a common goal. And these will be looked at. All of these AI tools and all of the nonsense you're hearing from these AI billionaires will mean nothing in the future. They are but paintbrushes and canvases and you get to decide and define the future. Thank you, everybody. I wish you great success at roaring cheers. Okay, Kenjin, it's your chance to give your commencement speech now to the stage. Our top graduate from the class of Stanford. When did you graduate Stanford?
- [83:45]
B
Stanford students. I went to MIT.
- [83:49]
A
And now, the incredible entrepreneur Kenjin Kyu, graduate of MIT, giving MIT's commencement speech here for 2026. Rounding applause.
- [84:02]
B
Five sentences. So at MIT. At MIT, we talk a lot about building things and progress and how cool technology is. And I think going into this world, the important thing to think about is not just, oh, what can we build? How cool is it? How much progress can we make? How much can we automate things? It's to think about, what is this all for? Why are we doing this? Why are we building AI? It's not just to drive some capitalist machine. It's not just to make money. The purpose of it is because we all want to live better as humans. And it's very easy to lose track of that when we're just building technology because it's cool and we have a bunch of enterprise customers. And so I would challenge you all students to think about what kind of future you want to live in and to build toward that future. Don't build things that you don't want to see in the world and don't build things even that you would want to see, but that are not the most important thing. Instead, I think you should build with AI the things that you want for yourself going into the future and find some way to make money. Unfortunately, I think that's the, I think that's the hard part. Like if people had enough money then they would go and be creative. But I think a lot of people are in the game of trying to make money.
- [85:19]
A
Kari, go ahead. Your commencement speech. Where did you graduate from? What's the great university in Finland? What's the great university of Finland? What's the top five?
- [85:28]
D
There's the Aalto University in Helsinki.
- [85:31]
A
All right, welcome to the Aalto Helsinki graduation. Today we have somebody who never had a degree, but we're giving him an honorary dual PhD doctorate for his incredible work on behalf of the 6 million people who live in Finland. What's the population of Finland right now?
- [85:48]
D
Five and a half, I think.
- [85:49]
A
Okay, I rounded up. Kari, you have the stage. Inspire the people of Finland.
- [85:55]
D
All right, so I didn't graduate from school, but you all are today graduating. And we are in this now, in this world of artificial intelligence. And for many people, including me, it creates a lot of uncertainty. It raises questions about jobs, what is the structure of the organization, the government, so the structure of the societies, what it means to be a human, how do we be spending our time? Those questions are real and we shouldn't ignore them. But you should also know that the news, what you read can be one sided. Today I would want to more speak about hope. The hope for AI is not really that the machines replace the human potential. The hope is that the machines help more people reach it. And AI can help students learn new things. It can help you to do things you never could done by yourself. And it can help you gain more understanding and hopefully the whole society can learn from with it. Not to replace the society with machines or with AI thinking. So I think technology never creates alone a better future that people who use it do. So all of this starts from you. It's like you can take the technology, direct it in a way that makes a better world.
- [87:11]
A
Amazing. Standing ovation. We'll put some post production applause there and we'll put a Bunch of students roaring after each of your speeches, standing up, giving a massive stand ovation. Jeremy, we'll close with you.
- [87:27]
C
Thank you.
- [87:28]
A
Where did you graduate from?
- [87:29]
C
Johns Hopkins.
- [87:30]
A
All right, today, Jeremy is giving the commencement speech for John Hopkins. And did you get a PhD there, Jeremy?
- [87:38]
C
No, that was my undergrad.
- [87:40]
A
Undergrad. And we're giving an honorary PhD in Philosophy and computer science, a dual degree to Jeremy, do you have the stage?
- [87:50]
C
So there has never been a better time in history to be graduating. AI will make it easier for anyone to create businesses, art, ideas, solutions at not just a local scale, but at the global scale. And while I know that many of you may be fearing AI replacing your jobs, every major technological revolution was doubted before it actually transformed society for the better. And the future will not belong to people who compete against AI, but it will belong to the people who learn how to use it creatively and responsibly. Because at the end of the day, no matter how powerful technology becomes, the world will always need human creativity, judgment and courage. And so I am very excited about the future and possibilities that you will unlock with AI.
- [88:45]
A
Amazing. Here it is. Oh, my God, everybody's going crazy. Incredible. Applause. You know, I'll just leave you with this. The New York Times dunked on Reese Witherspoon this weekend because she's a girl boss and she was giving people good advice for AI. Here's what they said. This is like the most incomprehensible, woke nonsense victimhood I've ever read. But all right, quoting from this New York Times article that barbecued a bunch of successful female entrepreneurs, including Reese Witherspoon. Quote from the editorial page. Nobody wants a parasocial bestie who shills for the plutocrats who are nullifying their votes, degrading their educations, jacking up their our bills, stealing their wages and rigging the system. There is no feminist case for scaring people into adopting AI. Why would anyone even try? And then they get into all of this nonsense of like attacking these women who are entrepreneurs. This is in the New York Times, so I don't think anybody read it. But I'm bringing it up here because this is part of the problem is people are just saying, like, you know, this is all bad and your electricity bill is going up. You're not going to get a job. And here's Reese Withers Witherspoon, who got dunked on it, and she has a pretty simple take time to learn about AI.
- [90:03]
E
I was with 10 women at a book club yesterday and I said to the 10 of them, how many of you guys use AI? And only three of them used AI. And then I said, how many of the three of you feel like you really know what you're doing or they're using it the right way? And that was only one person. So if three of them of 10 women are the only ones using AI, that means 70% of that group is not keeping up. The thing I've learned about technology is if you don't get a little bit of understanding from the very beginning, it just speeds past you. So you have to have little bits of learning just to keep up. And let's get real. Our kids are all using this every single day. I think we should learn the basics together and learn some really good, good tools that are going to make our everyday lives easier and better.
- [90:51]
A
Kenjin, you're a woman, so I'll have you speak on behalf of all women. And the feminist move in. Second, first, third, negative, retrograde waves. I can't keep up with all the different waves of feminism, but I have three daughters and a wife, so I'm soaking in it. What's your take here on this condescending bullshit from the New York Times? And then Reese Witherspoon's completely. I don't want to lead the witness here. Reasonable advice for people speaking on behalf
- [91:20]
B
of all women everywhere, as is your right and want. As is my right. Yes. And want.
- [91:27]
A
Yes.
- [91:30]
B
I don't know, man. This is kind of bullshit. I think it's great to adopt AI and use it to make things better. I do think that the New York Times is trying to point at like, hey, like, our system has problems. But I think a lot of the issue with that is just saying, hey, our system has problems, doesn't point at the underlying mechanisms of the problems or what's actually going on. And so, like, these two are completely separate, divorced things. Like, one is use AI. Yes, obviously you should use it. There's a really good tool. And the other one is like, the system has problems. And it's concentrating power. You know, not to go back to concentrating power, but it's real.
- [92:10]
A
Sure.
- [92:11]
B
And so the. And they're like, ah, if you support AI, then you support power, concentration. Like, I don't know if that's true. You know, if you use AI, you're gonna support giving yourself more power. So I think this is a, Is a maybe a puzzlement, a confusion.
- [92:27]
A
Wonderful.
- [92:28]
B
Speaking on behalf of all women everywhere.
- [92:29]
A
Yes. And since you are speaking on behalf of all women and for all women and future women, I think we'll we'll end on that since you did it so perfectly. I'm sure all women feel heard and you held space for them and for the New York Times. I mean how does Reese Witherspoon take a taking strays? She's wonderful. She's delightful, she's inspiring, she's talented. She's making a smoothie. Like let her cook. Literally let her cook and let her give advice. She's a great entrepreneur and a success story. Nobody hates women who are successful women as much as the New York Times is what I've come to the conclusion here. And yeah, the balance of power is not going to be swayed by Reese Witherspoon's use or opinion on AI of like yeah. The top models. Okay, you all get a plug. Jeremy, give people a plug of how they can engage with your company if you're hiring anything you want to give up a plug too. Now the plugs.
- [93:32]
C
Yes, we are hiring very quickly. You can go to fundamental tech careers or Follow us on LinkedIn and on X and we would love to talk to you.
- [93:44]
A
If somebody is incredibly talented and they wanted to just show you that and reach out to you directly, they can guess your email.
- [93:51]
C
Jeremyundamental Tech.
- [93:53]
A
Okay, there you go. That's for the super high performers. That's for the top 7%.
- [93:57]
C
Kari, top 1%.
- [94:00]
A
Top 1%. That's what I said. Yes. 0.7. I'm sorry. Point 70 basis points. The top 70 basis points. Okay Kari, now the plugs. Plug it all.
- [94:09]
D
Yeah, we are also hiring. We, we have quite kind of like more like a talent density idea where we don't hire a lot but we try to hire well and I think like for anyone who is interested in some of our new new stuff like building this agenda product building system for lot of the leading companies. So this includes working on the agent itself, working on coding agents, coding harnesses also like different kind of automation systems. So I think there's a lot of like AI and agent related work where you actually get to apply these things into real kind of product like a user customer problems. And it's not just like building like one tool for, for the model or something. So. And also we've been profitable the last five years. Oh and the growth has accelerated quarter after quarter. So we are also in very good standing on a business wise. So if you want to want to join then we have some open roles on engineering, product and design and others.
- [95:07]
A
So yeah it will be. It's a profitable company. Your venture capitalists right now are banging their Heads against the wall, please. Start losing money again and grow faster. Yeah, that's always the, that's always the battle cry. Grow faster, burn more money. We have money. You have the ability to burn it. Kendra, now the plugs.
- [95:28]
B
Yeah. So at Embeat, we work on building open source agents and making the open agent ecosystem win over closed so that we all have power.
- [95:36]
A
Let's go.
- [95:37]
B
Let's go. Let's fucking go.
- [95:39]
A
Let's fucking go. Let's fucking go for the win. Power to the people. Yes. And hiring anybody or where can they try the product if they want to engage it or, you know, their organization with it.
- [95:53]
B
So we have released a variety of open source tools in the last three months, but some of the products that we're building kind of. I can give you a sneak peek maybe. I currently use, I mean this is internally codenamed Minds, but I currently use this kind of like agent UI mix. In order to do all of my email, do a lot of my to dos, I put my to DOS in and ask the agent, hey, can you, you know, can you draft a message to everyone who said X? And it'll do it for me? And so it's a really cool kind of. It will be fully open source and it's really cool UI plus way of, kind of, you know, I just think that agents, they will allow us to operate computers in a different way and interact with computers in a different way. And we want to build a version of that where all of your is
- [96:48]
A
yours and where can we download it? Where can we sign up?
- [96:50]
B
If you email me@canjuna.com you can get early access. It's currently only in private testing with a bunch of friends.
- [97:01]
A
When do you think it'll be released? When will people be able to sign up for it?
- [97:04]
B
It's a good question. Probably, you know, a month or so.
- [97:07]
A
Okay. And you're going to charge for the hosted version 100 bucks a month. What is it?
- [97:12]
B
The hosted version like so if you run it locally, it's going to be free forever. And the hosted version we will charge for great. We don't know how much yet. As Kari said, pricing, you know, pricing is a tricky thing to figure out.
- [97:23]
A
So I always, my best advice is to charge a lot at the beginning so that you challenge people is always my advice to founders. Just, you know, make it 500 bucks a month and say this is like, you know, reasonably unlimited and you get to be in this discord chat room with the developers and we have room for 500 people and just start there. Everybody starts with the low end and this low end pricing and it's like, then you have to like retrain people, pay what it's worth and you get the looky lose. Just start at an absurdly high price and say it's incredibly, you know, it's like the Amman Hotel. Take the Amman Hotel approach to it. Hey, we have a limited number of slots in this beta. It's 100 bucks a month. Just challenge people to do it. That's what superhuman did you reference them before? It was a dollar a day. I was the first investor in that company and it was a dollar a day for email. And I said so you want to charge $365 a year for a product that Gmail does for free? They're like yep. Like who would pay that? I was like, you? Why would I pay for it? They're like we can save you five hours a week. I was like if you can save me five hours a year, I'll pay you $365. And can you find five more hours I can get back? Yes. There you go. Luxury software. All right everybody. I love that Luxury softwares the future. All right everybody, another amazing episode of this week in AI. We'll see you all soon. Bye bye.