Very True by Verissimo

Alex (00:53)
Welcome back to another episode of Very True by Verissimo. Today I am super excited to have a very close friend of mine, Michael Stoppelman, joining me. We were hanging out together less than a month ago in LA, but we decided to do this 7,000 miles away from each other on Riverside, virtually. I'm just really excited to.

Open up this conversation. Here's some of his story. Our stories have been kind of intertwined for the last 12 years, 13 years, 15 years, something like that. And through mutual friends, through working together, investing together, and just having a lot of great defining, meaningful conversations together over many years.

Michael (01:31)
Well, first of all, I want to say hi to the audience and I'm excited to be here. And Alex and I have been friends for something on the order of fifteen years and super excited to to be on this on the podcast.

Alex (01:42)
Yeah, absolutely. So Michael grew up in Virginia, right outside CAA headquarters, made his way to become a boilermaker and study computer science at the illustrious Purdue University in Indiana. And then finagled his way into this little company in Silicon Valley called Google.

Michael (01:51)
Yeah.

Well, before that, I actually tried to move to San Diego and get a biotech job because I wanted to do systems biology. So in college in 2003 was the first year that they offered a genomics program in the computer science program at Purdue. A couple of the professors, Chris Bailey Kellogg and and a couple other professors got together at Purdue. Chris was doing protein folding.

as one of the things that he was working on. And so they don't did an undergraduate course that covered a lot of the stuff that they do in grad school around biology. And I got very passionate about biology, actually worked at MIT over a summer at that time was called the Whitehead Institute. And then ended up moving my stuff to San Diego to stay in a like extended stay hotel. And then couldn't find a job there.

this was 2003. the Iraq war, the first part of it was winding down. and the economy at the time was kind of crappy for biotech, and then got an offer from two companies, one that was working with the NSA on intrusion detection systems, and then the other at Google as a post-sales support person. and I took the Google job because I called all the VCs that I had networked with and figured out.

And they were like, you'd be stupid not to go to Google. So I was like, I guess I should go to Google. So I moved up to the Bay Area.

Alex (03:17)
Sounds obvious now, but this was before Google was a public company. And but and was it all it was also before Google acquired double click and launched AdWords?

Michael (03:21)
Correct.

Yeah, yeah, it was it was way before that. on my floor, I was in building pi. so they had like three buildings at the time. One was called E for the constant, and then one was called Pi. And I forg I think one was building zero.

And it was it was a special time. It was like Larry and Sergey were just kind of postgraduate school. And then they bought segues and they'd ride around on the segues like Job and in arrested development. And and it was it was wild. so Marissa was a friend and she would have these part elaborate parties all the time for like Halloween and all these things. I just remember being at Google and it was

Such a wild time because like there was so much money flowing into the company and it was so profitable by the time I joined the machine of Google was really kicking off and it was it was kind of in this moment that we see with anthropic,

everyone at the company was was brilliant and they were all just working super hard. And there was Larry before I got there had like cleared out all the all the engineering management. So he had decided at some point to get rid of all middle management. So all the middle management was gone in engineering. I had started in post sales support and then over that that year I figured out quickly that I should get into engineering. So I architected my way into engineering.

and so so I built a bunch of tools for click fraud. And so my contribution that got me into engineering was building this tool set that the ops team used to figure out whether a publisher was faking their clicks, clicking on their own ads, that type of stuff. And so I built that out. and before, you know, before that I was like, Disney would send in a ticket for like, hey, my web search is down, and then

You'd have to go look and try to figure out and email some engineers and try to figure out why web search is down for Disney. and so there were all these syndication partners that were across the world that were using Google search. And so I was on the team supporting that. and then moved into engineering and then became the tech lead of the of the like this click fraud traffic quality team with another with another colleague.

Alex (05:20)
Yeah.

Michael (05:31)
And so that was that was kind of the Google trajectory so Wesley Chan was my first PM who's who's now a famous VC and then Rob Niaz

who's a VC that you know was at Hoxton. he was my second PM. and so and then Gogol Ramajan he was like the PM boss of all of them who's another famous investor was at Doordash and all and then Elad Gill was another PM that was on the mobile mobile team. So it was kind of like a a crazy time. Alex Roter was another colleague that was he ended up becoming VP of engineering at Twitter later. And then WISC, I think.

the Larry project to build the EV toll type type planes. So just a crazy set. David Friedberg's a close he was a close friend that was on the BD team. I think he led an initial attempt to acquire double click that was that was rebuffed at the time.

I remember being at at like the Google Cafe and Larry at the time was, you know, there and we sat down with him and he was like, How would we like help Linux? Like, what what are some ideas for like how do we how do we make Linux better? And so like people were

spitballing ideas of like, you know, he just was trying to get feedback of like, well what should we do, what should we do about Linux? And just wild.

Alex (06:40)
Do you do you

ever did you ever feel like it sounded like you had some presence while you were there or the presence of mind in those moments to be like, wow, these are special people and or on a rocket ship and taking a moment to just sit and appreciate that? Sounds like you were able to do that, which I think a lot of people struggle with. But there's another question I'd like to ask, which is, did you ever have moments where you're like, this whole thing's gonna fall apart?

Like the wheels are just gonna fall off. Like, here's how this thing breaks down. Cause obviously you know, you hear everyone hears the rosy stories of like, yeah, it was a rocket ship, and everything just continued. And like, no one talks about the challengers, right? Like there, there have been plenty of those stories. but did you ever have those moments where you're like, This is just like everyone's dumb, this isn't gonna work? here's the 10 ways this can fail.

Michael (07:19)
Yeah.

Well, you know, the click fraud part of the problem to me felt like a existential issue for Google, where it was like, hey, if we can't protect advertiser ROI on clicks, like if everything's a fake click, then I mean it's a crazy idea to be like, hey, every time you click on an ad, we're gonna charge you a certain amount of money. It it just that that original business model just sounds so

To an engineering brain, it's like, really? We're we're doing that? And then it becomes a problem of like, okay, how do we protect this thing that's like generating all this revenue and defend defend these properties? there's been a huge business model built, many, many businesses built up over clicks, impressions, all this stuff, and later conversions. and so my arc was working on clicks, working on protecting publisher, you know, keeping keeping good publishers in, getting bad publishers out.

And then working on CPA ads because I felt the end result is, if clicks can't be defended, then CPA is kind of the defensible signal, right? Where it's like, hey, if someone actually bought the toothbrush, then it happened, right? And so you can, no matter how many clicks or impressions you got, you got a toothbrush thing, and then you can use that as your signal because that's that's kind of a defensible signal. It's like you gotta use a real credit card, you gotta like

actually get through the transaction flow. that was my goal was like defend Google against that. and then through that period I published a paper with another Googler the time. He's gone on to do incredible stuff. we wrote a paper about finding this like botnet then it went when I was at Google. And so we built we built this whole like malware lab.

Neils Neils Provost was part of that. and so finding malware that was clicking on ads. it was a very like crazy, four years. I I had three and a half, three and a half years there. And while I was there, my brother happened to start Yelp. so then I was like, I've always wanna do a startup, always want to do a startup. My brother's starting this cool company that like has real utility and and can help,

At the time, the offerings on the web were like City Search and Yahoo local. And like you type in pizza and it would be like random pizza place that had paid them money. and then you go to Yelp and you type in pizza and it's like little star pizza, and you're like, that's like a better result. That's more what I expected, you know?

Alex (09:43)
one of the things that pops out to me is like it sounds like when you when you join Jeremy with Yelp, like the business model was very clear.

Like here's what Yelp delivers as a product. Here's what its unit is. It's like a recommendation. You're building a marketplace of recommenders

Michael (10:00)
by the time I joined, it was like three and two and a half to three years in. they had raised like series B. they'd already gotten an acquisition offer for a like a large enough sum. And so my thought was like

when I saw Yelp's results, it was kind of the similar feeling that you got when you saw Google results, but it was in local. It was like, okay, these are the these are the businesses. And what I knew because I had been part of the Yelp growth story when Yelp launched originally three years before that, I took a week off from Google on a vacation. I took vacation time and spent it with the Yelp team to help on launch of Yelp.

So I'd been intertwined with like like Yelp Engineering and like Russ Simmons, who is the CTO and co-founder of Yelp. Like we were, kind of like already connected. And I was kind of always swirling around, like, is this the right time to join? Like, are things coming together? Is Jeremy on some like goose chase? and then what I saw Yelp solve.

the there's this emotional intelligence piece where it's like, okay, local content is all in people's heads. How do we get the in the head onto the computer? And Google was not a how do we get it off the brain and like get people motivated to type. Like Google was, hey, we can build the infrastructure to crawl everything and we'll do BD deals to get all this content and

we'll build machines that cut off the spines of books and scan every page, and then we'll pay people to scan book pages, and then we'll put all that content on the web and we'll do all these crazy deals. that was what I saw. I saw there was a disconnect between Google was like infrastructure and like hardcore engineering. And then Yelp was solving this like fuzzier problem of emotional motivation around giving local content.

And so I saw that Jeremy had figured out and the the Yelp team, Nish, and and a bunch of other folks they like cracked this with like the Yelp Elite and all these different pieces. And so that got the marketplace moving, where it was like, okay, we have once you have local content, you get traffic. And then through conversations that Jeremy and I had, Jeremy figured out that there was this new distribution mechanism at the time, which was.

Get content, get it crawled by Google, and then grow your business. And then that ended up, you know, as Yelp was exponentially growing content-wise, then Google, they were like, we don't necessarily want this to grow, right? And I could feel that strain, you know, like at Google, like they were watching, they were watching this Yelp thing, you know, exponentially growing.

And saw it in their traffic and got a little bit scared. So that's what led to all of the antitrust stuff that you see that's played out over the last, I don't know, probably fifteen years or something. you know, Google trying to buy Yelp and then not work that not working out

Alex (12:45)
if we take it back to Google, because I'm I was feeling a parallel to I think what we're feeling now with a couple of key differences, which, you know, one part of it, like we talked about, was all these smart people working super hard, things moving really fast. we have that again now in 2026, which is really cool. The there is another thing that I think probably doesn't get talked about as much, which is that there was a lot of

open questions about like what is actually our product and what is actually our unit that we sell. Like what are our SKUs? What are we selling? So back then you were talking about, you know, there was CPA, there was clicks, right? There were all these things. And then there was all these risks associated with them. I feel like now in AI land, people are talking about token pricing and seat pricing and usage pricing and results based pricing. And like they s we still haven't figured out what that is.

Right. And it might be some combination of them, because that seems to be what Google landed on. And that's another kind of similarity, it feels like, with the that's just nebulous when like a new technology manifests itself in the world and start people start using it because there's clear utility. It seems like everyone's got to kind of grab at, like, okay, now how do we actually make money on this? And that's where the stories seem to converge. Because what you said about Google early was that money was flowing and we were making money.

Right. Like people are talking about, you know, how much r run rate revenue like anthropic and open AI have, but you only talk about run rate revenue when you don't have profit. If you have profit, you talk about profit. That's like that's a rule in life. And so it seems to be that again, these two things, these two events, we'll call them two eras.

They didn't happen in isolation from each other. People are reading books about and or lived that Google experience, that Facebook experience, whatever it may have been, that changed the way the world interacts with each other, and are mapping that and what the learnings were now, which was probably the main takeaway from the Google days, was like, we should have gone faster, hired more, raised money to sell at a loss and taken over the world. Ironically, they did take over the world.

like name one other search company that people still use. Like no one can do it. Right. They indexed the whole internet. Google owns the internet. that's the product, or like the core technology. But then it was, okay, what do we need to acquire and what do we need to do to turn this into a massively scalable, profitable business that takes advantage of all this? And I've, you know, we've talked about this. I just had a really interesting conversation right before this call.

With an AI company trying to figure out like, what does that look like? And are people focused on it in the right way? and again, you tell me, was it just that money was growing on trees back then? And even if Google raised money, they'd still have been profitable. and they couldn't have even moved any faster. And today this whole hide behind your growth numbers and justify lighting money on fire. And

If that's what you're doing, then you're really, you're not getting to the brass tacks of how are we making money, like actually making money versus money in, money out, which feels like what's going on in AI. I I don't know. I'm interested in your thoughts on this because this is exactly what I'm wrestling with and trying to figure out right now. And obviously it's many layers deep.

Michael (15:55)
so I see a lot of analogies and and they seem to like map on to what we see with anthropic and open AI. so I'll set the stage with this,

When I was at Google, one of the first core acquisitions they made was Android. Okay. So they saw around the corner we need to have a position in mobile because we have this like, you know, if Apple controls where the users are interacting with, they can then redirect to their search engine or they they they then have control in the marketplace send our users to a different spot.

And so then you get disintermediated for a weaker product. and Google when I was there, had this like very visceral knowledge of like all of the wars of Microsoft versus versus every you know, Netscape and and and where Eric Schmidt had been before,

Eric, who joined as CEO right around the same time I joined Google, had all this knowledge from how Microsoft had put in bugs to like make Netscape's browser crash more on Microsoft products. And Microsoft controlled the user. So if Netscape's crashing, then guess what? You have to use Microsoft Explorer because that's your only option. And then they did this bundling stuff and all this.

Google knew the value of having a user with a Google search bar they had done these original deals to have the Google toolbar in browsers. I don't know if you remember those days. that was the first acquisition. I think the second, the second great acquisition was YouTube. that looked pretty crazy at the time. It was like,

one point seven or two billion dollars and like Google was taking on all this risk of all the, you know, Warner Brothers and everyone suing YouTube and it looked like this litigation disaster. And then Google through its capital and also through its legal department and relationships was able to like get this thing straightened out. and and the biggest the biggest risk to YouTube actually, I think you know, at the time was like it becoming friendster where they couldn't scale the infrastructure.

you know, Friendster was a social network computer, you know, same space as like Facebook, but like died because its infrastructure collapsed. they built on, I think, Oracle databases at the time. I'd heard rumors that it was like an infrastructure disaster. but I don't have the facts, you'd have to ask Jonathan Abrams.

I think that when we look at anthropic and its strategy so far, it's packaged up its product. So so A, they discovered that their models work really well for coding. And then they've they've now put all of their effort on ignore all the video stuff, ignore all the like chat, most of the chat use cases, focus solely on coding, and then

Let's do all of the work to get into the enterprise. And we know that that's a lot of work, right? You got to do SOC2, you've got to do the HIPAA compliance, you have to do a forward-deployed engineering team. Like, how are you gonna set this up as a BPC or a whatever? Are you gonna work with Palantir for government people? Are you gonna put it in FedRAMP? all the infrastructure, all those engineers that understand that stuff and know how to deploy into these secure environments,

That is a thing and that takes time. And so you've seen OpenAI on their heels being like they didn't really have like an enterprise go-to-market motion. And then you had you had anthropic embedded in like government clients and all this stuff, working with enterprises in a way, and then also focused on this narrow use case of coding that's ended up being this amazing product market fit.

It's a product market fit for the agentic workflow. the chat interface is less about like it's important that it's like an agentic workflow, but it's not like critical. But like in the coding case, agentic workflow is killer. And then because Anthroaptic has this like myopic view on coding, all of the data set acquisition of like their BD team that's going out and acquiring data is only acquiring coding data.

So they're getting security data, they're getting, they're getting better, you know, like workflows for DevOps, they're spending the most on that stuff to like make their models better for coding, right? And that's giving them a big edge. And that edge is leading to more and more customers dedicating enterprise resources to integrating their models and putting it into their harness. And then you have another layer of the cloud desktop. So cloud desktop is kind of like that Android.

like interface with the user. for an enterprise user that's probably on their desktop all day, that claud desktop is really the like connection with the user where anthropic's hoping that doesn't get disintermediated. But you could totally see like some disintermediation happening. Like when you're on Chrome, I have a big button for Ask Gemini in it.

when Apple comes out, it's gonna be very it's gonna be everywhere, right? The the Apple, AI, and that's gonna be powered by Google. Cause they did a huge deal with Google. So there's I think there's two fights. There's this consumer fight for AI, and then there's the enterprise fight for AI. And right now, the one that's monetizing the best and going ultra exponential with this like 10x growth on anthropic.

Alex (20:44)
No.

Michael (20:57)
Is being driven by this use case of coding agents. And that's why you see everyone reorienting around coding agents. you hear rumors that Gemini, they lost a few people. I think that's because probably Sergei is trying to be myopically focused about coding because coding it's like the most product market fit in enterprise. Everyone always needs more engineers.

And it's like becoming better than a lot of humans, right? you can kind of see, if you're doing agentic coding now, you can see the value, right? instead of higher, you know.

Alex (21:27)
For sure. We we talked about this last month

where you're working on a project which we can talk about a bit later, but how many people would you have needed just to get it done? And then I have my own personal experience. Like I had to take CS 106A at Stanford as a mechanical engineer, and I really struggled with it because I am not a syntax guy. I'm not good at languages, but I'm good at thinking like a machine.

And so this has kind of opened up a whole world to me. Truthfully, no code, low code, like Airtable, Notion, Excel also opened up a whole world to me. And this could light it on fire. Truthfully, I haven't had a specific thing that I've really wanted to do with it. I'm still kind of looking for one, amongst other things. But that coding use case to me, it feels like the biggest single use case where

a single type of expertise flow and and like workflow and edge cases get unified because it's it's kind of this universal translation problem solving methodology that all happens in a really closed, like digital kind of 2D world. One of the things that keeps coming up now is like other models, right? The small models, The ones that's like.

Hey, like you don't need all these different things, right? there should be a model. I shouldn't be using Opus 4.8 for my fitness optimization. there should be like a fitness model that's just trained on fitness data that like Strava is helping develop. because they have tons and tons of data. So there's all of these other cases outside of that coding.

Part of me wants to say, why would Google go after the same thing? But it's if you want to have one massive company with one massive model, you got to pick the biggest problem out there that's approachable right now. But I think there's millions of

Michael (23:11)
No, but I think what anthropic is what anthropic and everyone else is chasing is they they crack the nut on the fastest, OpenAI was trying to boil the ocean. they're building a hardware device, they were doing video stuff, they were pouring tons of compute on that. they were doing this chat interface to compete with Google, because they had product market fit in the chat, like, okay, we're gonna take over search, And then

Anthropic comes out of nowhere, 10x growth. every year they're growing at 10x. And like they're gonna do a hundred billion or more this year. That's the like rumor, right? And it's like yeah, run rate, whatever. But it's like, I know for my use, I've got five tabs here open in Claude. It's like I'm not using, open AI's model. I'm using Claude, right? And I've got my max plan of $200 a month.

Alex (23:43)
Run run rate. Point in time. Yeah.

Michael (23:59)
Like that's serious money, you know.

Alex (24:01)
For sure. I mean, it sucks up a lot of resources from other things. I think a lot of people, what they're doing right now is, you know, the unroll me or whatever's of the world are probably having a field day. I can't wait till anthropic and the others roll out a product that will just look through your email and like, cancel all these subscriptions for you that I don't need anymore.

Michael (24:17)
But but I think it

but I think it's an important moment to pause and say software engineering has just been completely dissolved and recreated in this agentic development world. I'm sure there's still engineers that are coding by hand, but it's gonna look

Like the carpenter that doesn't use a like a power nail gun or a or like a for a good

Alex (24:41)
Use a they

don't use an electric saw or a drill, they're using screwdrivers and handsaws. Yeah, yeah, yeah. And and and again, it's like it's almost like AI is just that interface, right? Like once we understand like precision manufacturing and you could build motors and things like that, you can take this power of electricity, which is just like compute, and you can start turning it much more quickly into something valuable, and coding.

Michael (24:45)
Yeah, they're not using any electricity. They're like building like an Amish you know, they're building like the Amish and like

Alex (25:09)
Like software engineering is the thing. that

Michael (25:11)
No, but but

I I just think it's like this type of transition is likely going to hit all different professions. like if you like and I think we're I think for for doctors that are like in the field that are like coming up with you know, trying to figure out what's going on wrong with the patient, sixty percent of them are using open evidence. Right? That's a new world. Yeah, as they should be. And like that number's probably higher now. And that's a US number that I

Alex (25:19)
Yeah, so that's that's what I wanna get to.

Yeah. As they should be. So so but that's a whole other world that like

so here's the this is the the question that keeps in my head. Like the biggest, fastest growing opportunity right now is this human computer interface layer, which used to be called a coder and is now named Claude, right? Like you could argue that that isn't gross oversimplification of what's gone on, but it used to be there was these people that were called coders, which the whole learn to code thing never made sense to me because I was like, it's not about learning how to code, it's about learning how to think.

Like an engineer and solve problems creatively, which when you study computer science, as I mean, you would know better than me, like that's actually what you get. You get the creative problem solving around numbers and optimizing for what a computer's capabilities are. And a chunk of that has been digitized and automated with AI, which is fantastic. But there's all these other opportunities out there that I'm trying to wrap my mind around, right? Like law is an easy one because it's like all text-based.

it's just it's perfect for this. I my favorite they some stupid article is like, this AI made a mistake and made up a briefing. I'm like, you don't think a legal associate has ever made up a briefing before? Like, are you kidding me? Right? Like, it's it's a whole different level. It's like every time an autonomous vehicle runs somebody over, they're like, my gosh, front page news. It's like, you know how many people ran people over today? Right? so

That will start to level ize. But when I look at like the capabilities of an anthropic, of an open AI, of Gemini, like whatever they're doing, it feels like right now that is kind of the end all be all, and that's what everyone's talking about. Is this software engineering layer and how that's all changing? But the medical example, like food optimization.

It seems like the margin profile of software engineering and building digital tools is just bound to collapse. Whereas the ability to deliver medical advice, medical care, to deliver again, food, transportation, much things lower, frankly, on Maslow's hierarchy of needs, which historically were viewed as a lower margin, may actually build steadier margins based on the specific models in those areas.

Michael (27:36)
those

Alex (27:36)
So

the first question is are margins actually durable for coding agents? That's that's probably the shortest way to say that. Because whoever's using coding agents, if you just look one line down the food chain or a couple down, are they actually solving important problems? Right. if the whole world of enterprise software is the biggest customer of enterprise software?

And enterprise software collapse, then selling to enterprise software is not so interesting anymore. And it collapses in a good way because it becomes just massively efficient. And then the next question is: okay, so Anthropic this is an aggressive way of saying it might not be accurate, but you get the idea, has bet the farm on software engineering being the core driver of growth, which it unarguably has been over the last two years. But if that collapses and will that collapse,

Are there other companies primed like open evidence? is Waymo gonna just like wipe the floor with everybody because that matters so much more and is so much more differentiated because the stakes are so much higher? And you'll always be able to charge a certain amount for a ride to get a person to a place, but you may not be able to charge that amount of money to get a coding thing built for a use case that is fading in its own utility.

Michael (28:46)
So let's say that the improvements in the models continue for the next 20 years. it's kind of like you suspend reality on semiconductors and you're like, we can never get better than seven nanometers. And then two years later, they're like, you know, we got to three. And then they're like, a year later, they're like, you know, we got to two nanometers. You know, and everyone's just

How'd they do that? They like stack stuff. So

Alex (29:10)
It does become asymptotic. So for just in that example, with with semiconductors, it becomes asymptotic because you start having electron interference at a certain like atom scale, right? Once you start getting into angstroms, it's like the things stop working.

Michael (29:20)
Yeah.

So

there's going to be a bifurcation of the coding space. And so we're gonna have CRUD apps. So by CRUD apps, I mean simple apps similar to like what I'm building with with CC Marvin, like things that aren't mission critical things. You're not building a system for nuclear weapons, or you're not building like the the this the system that runs a power plant.

if you're just building an app for managing you know a database with interactions with UI, you're gonna be able to use a GLM six. It's at five point two right now, but GLM six is gonna be great. And you can get away with doing nine ninety-nine point nine percent of things with that model. And then every level up, you're working on a satellite system.

Is your company not going to pay for the most intelligent model that is going to get every edge case and think through every little fine detail of that model? would you pay for the omniscient, the omniscient intelligence? And my supposition is that there's always going to be a company that's willing to pay for the the most epic model, just in everything. the bike behind you, special is that a specialized or a track?

Alex (30:32)
This

is a pivot.

Michael (30:33)
Yeah, so like is Pivot gonna pay for the omniscient like model to design their next bike and spend like you know $100,000 on compute, or are they gonna go with a GLM GLM 5.2? Like they're gonna use the best model that's out there to get the best bike so that when they do their production run. So I think there's gonna be a bifurcation for like the hobbyist use case, you're gonna see like, in the next Apple release, we'll all be using a lot of models on our computer.

They're run, know, some stripped down Quen version or GLM. And like that will work for a lot of these use cases, but the spread of all of these like things that you're gonna want to be able to do. we haven't even seen like the progression of these models into like mechanical engineering and to like electrical engineering and to like we're just starting to see it with chips.

Alex (31:18)
But my question

Michael (31:20)
And then we're gonna get to like humanoid robots and factories and things. And like those are gonna need to have like very sophisticated cognitive systems, and that's gonna use a lot of this edge computing. So, when you think about a company being able to go public, have cash flows, DCF, how many years do you need before it makes sense to say that that thing is valuable? I don't know like what the edge of intelligence in Prado, Prado curve or something is that

Is that the like word that they use? But the frontier intelligence is going to have a big chunk of the market that's gonna shrink over time. I think the question is how much how fast do the commodity markets eat into the mass intelligence? Right now, I want to use the most intelligent model.

Alex (31:57)
Yeah, it all moves though together, right? Like

Michael (32:01)
I wanna have the least security bugs. right now, I have to use a secondary tool to do code reviews of my Claude created code. So I use this thing called CodeRabbit. And CodeRabbit goes through the models aren't good enough to like find all the errors. And so the the Code Rabbit thing's finding real bugs in like the code that's being generated. And if you had like a security expert looking at the code, you'd find a lot of stuff that's like.

Alex (32:08)
Yeah.

Michael (32:25)
not perfect.

Alex (32:26)
that brings me to the question of is bigger actually better? Or is there a thing such as specificity? Like if you just map that to human beings, Like you don't want an astrophysicist from Harvard doing heart surgery. Even though your heart surgeon went to some no-name university that you've never heard of, you'd still rather have the no-name university heart surgeon.

Doing heart surgery on you than an astrophysicist who has a PhD from Harvard. So, is there just this global, it's the smartest and it's the best at everything? Or is there room for specificity on those leading cases? Not I'm not talking about the stuff that gets kind of trickled out behind. It's almost like, all right, this is good enough because we get the right answer every time. So we don't need to move beyond, let's say, Sonnet.

I always kind of laugh at, I'm using the Claude like web interface. And it says, Fable for your toughest challenges, opus for your complex tasks, Sonnet, most efficient for everyday tasks. Haiku, fastest for quick answers. And I'm like, but I'll just play with it and I'll ask a question to Haiku. Then I'll ask a question to Fable. And I don't know which one.

You don't even if you're not aware of how hard your question actually is for the model, then you really don't know which one to ask it to. So, it gets back to this like forget the cost efficiency and all that and the fact that I have the luxury of just being like, I'm only gonna ask two questions to AI today because I'm going on bike rides. So I'm just gonna ask them to fable because why not? Right, like at a certain point that needs to end and we need some more discretion,

on how we actually use this thing versus just bigger and more is always better and there's an absolute intelligence that we want to expose to our heart.

Michael (33:59)
mean, I mean we are

we already have this happening. there's well funded companies in material science. There's a company called Periodic, there's a company called PSI that that I'm an ad I'm an advisor to, PSI.inc. and PSI is trying to build a super physicist. So using data from from all sorts of physics experiments and all of that stuff,

put it in on top of like an anthropic and then in improve that model beyond to have super capabilities. I'm also an investor in Chai Discovery. So Chai Discovery is doing specific stuff in antibody design. they have their own models for protein folding and all of these different things, similar to like Alpha Fold. And so those are those are examples,

The PSI thing is like a little bit different. It's like training on top of one of the models already existing, or at least that was that was my understanding of it. in the case of Chai, like they have their own foundational model, right? They have their own, that's my understanding. So they're training on this biological data. And I'm watching this space of earthquakes right now. So like I really am interested. I think the perfect use case for transformers.

Is like taking a bunch of unstructured data, all the seismic information from around the world, and then tell me where the next earthquake's gonna be. it's perfect use case. And so you're seeing lots of different teams from all around the world, because this is a very important thing. And we just have the earthquake in Venezuela. it's very useful to be able to have a model that's specific to seismic data. so I think there's gonna be different use cases where you either build on top of an open weight model that you can extend.

Or if you're trying to do the seismic stuff or you're trying to do this biological stuff, it probably doesn't make sense to put that on top of an open weight model. But I don't know. I mean s researchers

Alex (35:40)
Well then it's like how

do how do you find that out? Like you just ask the open weight model, like, Hey, who's better at this? You or me doing this other thing?

Michael (35:48)
out. No, this is research. you know, some people will add the functionality to the to the base models, like a Quinn or a GLM five, and see if it does better than them trying to train a model from the beginning. But like you have to understand the amount of compute, GPUs, time, expertise to train a foundational model is way harder than trying to fine-tune one of these open weight things. And so

Alex (36:09)
For sure.

Michael (36:11)
if you were in graduate school, Alex, and you're trying to train a model to like be better at figuring out like what the design for a wing is, you could shove a bunch of data in a day into an open weight model and have it fine-tuned on it, or you could spend the next six months of your life trying to train with PyTorch your own model. You choose, you know.

Alex (36:11)
cost question.

Yeah.

Think that we're gonna live in that back and forth for sure for the next, I don't know, a couple years. But one of the things that I always come back to is my nerdy kind of calculus-based view of the world, which is really trying to understand the shape of a curve and the equation that's driving it, so that you don't get caught in a local maximum, a local minimum, or

part of the line that looks like it's really straight and is gonna continue forever. I was talking to someone recently about like S curves, right? An S curve you could describe with a mathematical formula. I think it's a third derivative, right? It's like X cubed,

Michael (37:05)
You're talking about like a sigmoid?

Alex (37:07)
Just like a yeah, just an S curve. Like those used to be how we talked about company trajectories, where like it would inflect once and go faster, and then inflect again and go slower. it would keep growing, let's say, the whole time. And like obviously real companies, you know, it's a little bit more complicated than that, but there's always that curve shape. And I wrote about this recently as well, is like connecting dots and drawing lines is a very different thing than trying to understand the equation behind the curve and then extrapolating forward.

How's that going change? And what that kind of maps to is first derivative, second derivative, third derivative. And I think in our case, in the world we're living in right now, what that kind of comes down to is like the who. who is going to do these things? Who is trained to do them? How quickly do they move and change over time? Are enough people actually, frankly, smart enough to like learn fast enough?

To be able to grab onto that. those secondary and tertiary effects is what I'm trying to think about. And then feed those back into the conversation of if software engineers are not a thing anymore and we're not training them, like what happens? Right. Like people have said the same thing for like legal associations. I don't either. But I'm saying, if there's way fewer, are we just picking out the elite tier?

Michael (38:13)
I don't think I don't think we're getting rid of software engineers.

I don't I think there's way more software. it's like same number of software engineers, just way more software. the question now is moving to like, what do you what do you decide to build? Like every day I come to my computer for CC Marvin and I'm like, what do I want to build today? The time that I then spend that day coaxing the AI to build the thing in the right way, figure out all the bugs, smoke test the thing, do all the unit tests.

Alex (38:21)
And like we trim the fat, like

Yeah.

Michael (38:46)
It takes time. I start a project and then I'm spending a day or two polishing it before it's shown to the user. That takes real legit time. And so you think about it on a bigger scale with like a company. engineers now can go off and like spend you know, have 10 tabs running running cursor or running clot or whatever or groc,

You end up two days later with like tons of code, but like you went completely orthogonal to like where you should have been building. Right. the skill that's going the skill that's really valuable is product management, is what to build. What to build, how to build it, how should it show up to the user?

Alex (39:12)
Yeah.

I would argue that's product design, right?

Michael (39:23)
Pro yeah, product design, yeah. I mean it's really like that's the that's the killer skill set for for this AI era is like, hey, you you need to decide where do you want your tokens and your engineering horsepower that's managing the tokens to like which direction to build in. I don't know if there's a mechanical engineering analogy, but it's like, if you have a 3D printer,

You don't want to just go printing random stuff. You wanna like print stuff that's gonna matter, you know? If you're building if you're building a rocket ship.

Alex (39:51)
But I guess the

but if we think about it, over the last year, the coding agents have gotten better and better, which presumably means that they can bite off more and more of what people did before. And then if you have this overarching layer of like, what should I build? What will people like? What should it look like? How should it be used? That's still that human intelligence layer. At what point does like that tide rise and like the kind of the

The water just fills the room. And what's left if we extrapolate a couple more years of improvements?

Michael (40:22)
I have not seen a like a team at like Carnegie Mellon or Purdue or anywhere else in an infrastructure division be like, hey, we just developed a better Linux kernel that's 10 times better than the current Linux kernel, and it was developed by AI. I don't think that we're we're doing that with like CUDA kernels to like improve the CUDA kernel so it like runs the GPU computation a little better. There's

Companies that do that and try to optimize that with AI, but they're not depending on that for AI. So why is the AI not good enough to do that? Like, why don't we have an auto research loop that's making all the algorithms around us that are critical to our lives, like fast-forward rate transforms or these types of things? Like, why isn't it making it better yet? And so I say, like, there is a big delta between what the model companies are saying is possible and what they can do.

Alex (41:06)
Yeah.

Michael (41:10)
if we had this omniscient model, do you think we would have a better C compiler that could like recompile everything and make it five times faster? Like we don't. no one's to my knowledge, no one's written a paper like a 2x speed up on a C C or C compiler. or like a faster renderer for Chrome, like the V8 system that runs JavaScript on Chrome. Like to me, those are the obvious things that you'd apply AI to if it was really like

omnisciently better. It's not there yet.

Alex (41:35)
So it

it seems like now like the name of the game is there's like some there's some line that's not a straight line. And you've got to figure out by pushing on that line, the AI is better at this and faster at this. you still need a human working on this, they're better at this. And if we keep pushing on that and testing the limits of

What people can do, what AI can do, how they can work together, and how we can lift the whole thing. presumably that line will only move in one direction. No matter what the shape of that line is right now, of what people are focusing on and where there are still holdouts, presumably, that that whole thing will move in one direction overall. And we'll just continue to eat into it. So it's not good enough right now. But if you think about two years ago versus now, about what you were able to do.

Right, which was very basic, versus how much faster that line is moving, you would I would argue now than it was two years ago. It's hard to even know what this is gonna look like in two years. I'm a business guy. I'm like a physics of business financial hardcore fundamentals thinker. You know, I like reading Munger and Buffett and Graham. what's durable?

Michael (42:39)
Yeah.

that's the crux of it is what's durable. And so going back to your original question where we were talking about Anthropic and Claude and I was talking about the analogy to like Google building out like Android and then grabbing YouTube so they could have a connection with the user, right? So you had Android, Chrome, YouTube, YouTube ended up being an important set of data that Google needed, but

Chrome and Android represent the connection with the user and they could control the endpoint and the choke point. so you have 50% of phone users or I don't know what percentage of the world uses Android versus iPhone at this point. But Anthropics trying to do that with enterprise. And then if the cost of tokens goes down, they still control the choke point. So now today, I would posit that

Alex (43:08)
Yep. And Gmail.

Michael (43:27)
A bing web search or like an equivalent, like all these upstart web search companies, can get to approximately the same quality of web search that Google can get to. So Google doesn't really have a moat that it li that it had long time ago with search.

Alex (43:43)
this is all in

once something's figured out and people know what it looks like, it's not that hard to represent.

Michael (43:46)
No, no, no, like search has become

kind of commodified. But st still, but still people go to Google. And why, Alex, why do people still go to Google? The choke points. They have position. This is why they got the monopoly designation by the EU and all this stuff and in the US. No, so so they have they have Chrome. So every time you type in the Omnibar, it goes to Google.

Alex (43:50)
As I've mostly look at the quality

Met Muffle you you buy

Yeah. So it's but

Michael (44:08)
any sort of search it's going to Google, like they have all of the choke points.

Alex (44:11)
Broaden this,

this is not unique to Google, right? I was talking to a friend about this yesterday. Look at the automotive industry, right? I live in Israel. There's a 90% tax on cars here. So every car costs roughly double here what it would cost in America. It used to be that in America, if you bought a $15,000 car versus a $30,000 car, you are living in a completely different existence of driving that car. Think about the 90s.

Right. just I'll just pick on them because I'm a fan. But like if you look at the Korean cars, Kia and Hyundai, 20 years ago versus now, and you look at the gap between those cars and BMW and Mercedes, the driving experience, the infotainment experience, the safety experience, that gap has closed massively. And part of it is reverse engineering. And part of again, I'm a sucker for the physical world, but the beauty of it is it.

tends to move a little bit slower. So you can kind of note these benchmarks with a little bit, more on solid ground. But like the gap has closed. Not completely closed, but what they say about like for example, a Genesis, made by the same, you know, Hyundai Kia company out of Korea it's 75% as good as a Mercedes for 40 to 50% of the price.

Twenty years ago, that was impossible. So you have the you could say the same thing about Google, but what did Mercedes and BMW invest in? And why do people still buy those cars when it's clearly not economical? They invested in distribution and brand, which is the broader customer-facing portion of it. So they got into everyone's again, it's the distribution and brand. So it's the same thing with Google.

Michael (45:37)
What what is anth what is anthropic doing?

Okay, and what and how how

how is Dario getting into the heads of everyone, government and users?

Alex (45:48)
Yeah, it's it's through this coding angle.

Michael (45:50)
And fear.

Alex (45:51)
Well, yeah, that's I mean I've been listening to a lot of Cal Newport lately and he's like beating that drum hard. Yeah, it's fear tactics. Right now, I again I don't want to get into like

Michael (45:52)
Yeah.

No. No, no, no. But like he might,

I believe he's sincere. He does believe it, but like also it's a great sales pitch. every security company knows that they sell with fear, right? It's like, like, do you want to get hacked? Alex, do you do you want your do you want your state secrets to be on the web? Yeah. so so but like

Alex (46:14)
Well, that this something just came up

in an article that my dad sent me. I just have to throw this out there. This is a mutual friend of ours who shall remain nameless, was quoted saying something like, the government should have no part in intervening with AI technology. And meanwhile, Cal Newport is on the other side being like, How are you going to run around on every news outlet saying this is more dangerous than nuclear weapons and not expect the government to want to get involved? It's their job to protect people.

That's another whole can of worms that I don't think we can get into another time.

Michael (46:45)
think about the analogy of what Anthropoc's doing. They're getting embedded into the enterprise. They're building out their Ford deployed engineering team. They're marketing the hell out of all of their stuff, Fable and Project was what was their I for plot Methos, which was for their security stuff. They're getting all the headlines that NSA is worried about how good their model is. They're getting all these people to validate it and say, shit, like if

Alex (46:58)
Mythos, whatever or whatever.

Michael (47:10)
So they're they're winning the war of brand around who has the most intelligent model. And then behind the scenes, they're buying all the data to like make their models stronger and better and faster at like those specific things. Way more money than I think at least six months ago than OpenAI was throwing at the problem. right now their product strategy is get the brand, be the definitive place, pound.

Alex (47:16)
Yeah.

Michael (47:34)
Pound everyone that clawed is Claude Code is the place to that real developers build. And it's hard for everyone else to catch up. it's hard to catch up to that branding because the government's doing its they're doing the best job of marketing for anthropic.

Alex (47:48)
Sure.

So, but my whole point is that give credit where credit's due, which is that this is a playbook that's been played for centuries. But what I'll ask you, and I will wrap up on this question, is so what should we do about it?

Michael (47:59)
How should we invest or what like?

Alex (48:01)
should we invest

our time our money what skills do we need like how do we hedge all of this

Michael (48:07)
it's the equivalent of like being like, hey, you you're a candle maker you were handed a light bulb and electricity, and do you want to make a new business with this? it's the equivalent of that, right? maybe not as extreme, but it's hey, you need to adapt as a software engineer. I have to adapt to becoming an agentic software engineer.

it takes a mind shift. It's kind of like a weird feeling to like hand off all of this responsibility to coding. But but also I think it's very freeing. you know, as you get older, your sleep gets more finicky. So some days you wake up and you're tired or whatever. And like as an as a software engineer, when that happened, you were really wouldn't be that productive as a coder because you'd be like, couldn't keep variables in your head, couldn't keep state in your head.

But now it kind of normalizes the field where you're just working with the model, giving prompts, like guiding it, debugging it, figuring out what the next thing to to build is. I think it's very freeing. I think it's amazing. I the question of like employment displacement and worries around that,

it's something to monitor and and and governments like California are building systems to like measure like how many jobs are being lost and things like that. But like we it's ridiculous. there's so much anti-AI political political backlash happening. It's very reasonable to measure it. I would say that the government systems for measuring it are probably pretty arcane.

Alex (49:29)
That's that's what I mean. That's ridiculous.

Michael (49:29)
That we had that we had

before. you can see how we responded in COVID to EDD in California. We we ended up being defrauded out of $30 billion here. So the systems are not great. and then also people are in massive cues to get to get checks out during COVID. So I don't think our systems are super well oiled. And so for sure a plausible argument to like make those make those systems better.

Alex (49:39)
Yeah.

Michael (49:53)
But you can't put your head in the sand. You can't be like, it's gonna have no effect. And like it's not gonna happen. you measure it, you see like when we're gonna have those effects. I'm like just waiting for the first announcement of a 10,000 deployment of humanoid robots. when is that gonna happen? when are we gonna see the first announcement of a thousand robots being deployed by Figure or Tesla?

Alex (49:55)
Yeah.

Michael (50:15)
In a factory. when is that announcement day going to happen? And you're going to go into a factory and it's a bunch of robots moving around doing stuff. that's an interesting moment for society. we demographically have a challenge in the US. We're not having as many babies. Like the West in general is not having as many children. Like we're going to need extra workers if we want to keep our GDP up.

And if we're keeping our spend up, we want to keep our GDP up and GDP growth up. So I think that like we we're gonna end up in this is a great solution to like keeping the country moving, at least in the US.

Alex (50:47)
I love how you ended off with that. It keeps the country moving. I think the complacency is probably what's driving a lot of the fear. Cal Newport's talked about this as well. And I tend to agree with it is like people saying, this is the jobs lost due to AI. And then you look at like the actual job descriptions, and none of them were like software engineers. They were like HR people that got fired, Like

It's actually just because they overhired during COVID when there was like this weird competitive hiring bubble and everyone was remote and then productivity was low. And then AI was kind of like the catalyst or the excuse to shut it all down. But overall, outside kind of that specific example of like, the data and where does it come from and how do you do the root cause analysis properly, which is a skill that's lacking in most of the world, the overarching piece here is like this is called progress.

Michael (51:21)
Yeah.

Alex (51:33)
And be scary, but it creates some amazing opportunities for people.

Michael (51:34)
I'll leave you yeah.

I also think that the agentic world increases the number of potential entrepreneurs. it's just like AWS decreased the need for more DevOps engineers and then it ended up leading to more DevOps engineers being needed. when Yelp was moving to to AWS, all like a bunch of the DevOps folks that we had at the time were very fearful that it was going to get rid of their jobs.

It only increased the need for DevOps engineers because we built so much more stuff. So I see the software engineer agentic workflow increasing the amount of software. Then we need more DevOps engineers, more security people, more IT people, more everything. It's like it's gonna be an entrepreneurship boom, in my view. And that's why I've started to work on this CC Marvin thing. you know, you can check it out at ccmarvin.com. It's made for investors and

And I'd love your feedback as well.

Alex (52:28)
All right,

you heard it here. Check out ccmarvin.com. All my investor listeners, give it a try. And now you know the guy to complain to whose name is not Marvin, but

Michael (52:37)
Yeah, so

yeah. But my as my pitch for Marvin, Marvin's a chief of staff for you, accessible via email, and you can create newsletters that scan over X, YouTube, and other news sources and get up to date news without Doom scrolling. That's my pitch.

Alex (52:51)
Amazing. Michael, this was a pleasure and extremely intellectually stimulating. And I think that our listeners will enjoy it. And I appreciate your time. we'll have to do this again soon. But it's always good to see you. Thank you. And talk to you soon.

Michael (53:05)
Awesome. Thanks, Alex. Appreciate it.