Skip to content

Podcast

Balaji & Benedict Evans- When Tech Breaks Industries

The a16z Show

Source ↗ ← All highlights
  • Attention Peaks During Transition
    • The moment you finally understand a technology is often the moment you should stop paying attention to it.
    • Conversation peaks during the transition, so focus on rate-of-change not absolute adoption. Transcript: Balaji Srinivasan You finally understand a technology is often the moment you should stop paying attention to it. What matters isn’t the absolute level of adoption, but the rate of change. We talk about a technology most during the transition, then forget it exists. Today, that transition is happening simultaneously in AI, crypto, smart glasses, and robotics. The conversation about each is at maximum volume, which means the interesting question isn’t whether they matter, but what each one actually disrupts and what it leaves standing. (Time 0:00:40)
  • Smartphone Dividend Enables Hardware
    • The billion‑smartphone supply chain made many device components cheap and widely available.
    • That commoditization enabled drones, VR headsets, and many edge devices to scale quickly. Transcript: Benedict Evans Now, like it happened. Time to look for different questions. Balaji Srinivasan Well, it’s funny because I think when we were overlapping, it was right in the middle of the smartphone dividend, the smartphone explosion. And just to, you know, actually there’s a few things. One is the smartphone dividend. That’s a useful concept, right? Like that the rise of a billion smartphones meant that everything that went into them became cheaper and that enabled VR headsets, that enabled drones, right? All this stuff. Benedict Evans Yeah, all the components that came out of it. Yeah, so sales are now from memory like one and a quarter one and a half billion units a year and all the supply chain from that all of those components is then available off the shelf if You want to buy 5,000 of them or 10,000 of them all the Wi-Fi chips and the batteries and the cameras and all the other bits and before if you wanted to put computer into something, you basically Need to use PC components. So ATMs and so on are all basically PCs. Elevators are basically PCs. And that has size and power and cost constraints. And then smartphones become the thing and then all those components are available. (Time 0:02:56)
  • Consumers Lead, Military Follows
    • Consumer adoption now often leads military and institutional use rather than trailing it.
    • Cheap consumer supply chains and scale drive downstream adoption in security and enterprise. Transcript: Balaji Srinivasan Of supply chain. And often the military supply chain is often just a subset of the consumer supply chain because you sell a billion units of this and maybe you have 100,000 or a million units of a military Thing. Benedict Evans Actually, it’s almost kind of the reverse now in that it used to be, so the way I think about this is like in the past, like before we were born, the intelligence agencies would get the cool New stuff first and then the military would get it and then big corporations would get it and eventually consumers would get it like 30 years afterwards. So this is like the canonical thing. Nazi inversion, yes. It’s like microwaves were invented for NASA. Right. And eventually consumers get them. Balaji Srinivasan Yeah, or like GPS was invented to guide missiles. Exactly. And now it’s used for tagging cat photos. Benedict Evans And the shift is like a combination of the stuff getting cheap enough that it can be for consumers instead of you needing a billion dollars to have one. And then the scale of consumers once it gets cheap enough. And so now the way it works is the consumers get the new stuff and the military gets it 10 years later. Because that’s how long it takes to the bureaucracy to assume that a the bureaucracy be the to harden it and productize it and turn it into what you need if you’re going to get shot at or It’s going to be cold or hot or warm or whatever it is. (Time 0:04:18)
  • Prompt Precisely And Verify Outputs
    • Treat prompting as higher‑level programming and craft clear prompts with domain vocabulary.
    • Always verify outputs, since generated text and code can be plausible but incorrect. Transcript: Balaji Srinivasan With AI, there’s, you know, one way of thinking about it is like now we’re two and a half years in. Let’s say, let’s call it the chat GPT moment, right? And it’s interesting because it, I think what people really overestimated was how much it’s agentic intelligence versus amplified intelligence. Like that to say, you still have to prompt it. So prompting is like higher level programming, number one. You still have to verify the output. And that means you kind of need to know what it is you’re looking for. For example, if it spits out a bunch of mathematical symbols in an area of math that you don’t know, then you have to be Terence Tao to verify it. It might be gibberish, it might be real, who knows, right? And so the prompting and verifying are actually the bottlenecks in many areas. Now, Karpathy and I, Andrei Karpathy, we were just having a discussion on this like a week or so ago. And the thing about verifying is if you’re using the GPUs that we have built in and you’re looking at images or video or front-end code, right, like a user interface, your eye can just Instantly pick out and you can verify pretty quickly. So for that side of things, AI is quite good. Anything that’s images, video, your ear can also pick out audio, right? And front end. But when it’s back end stuff, right? When it’s like database code, when it’s like crypto, when it’s mathematical equations, that you don’t have like GPUs, you can’t just like hit it with your eyes and quickly detect it, Right? Whether it’s correct or not. You have to deep read it carefully, right? So it can generate reams of text, but then you have to verify it. (Time 0:15:53)
  • Where LLMs Sit In The Stack
    • Traditional software is deterministic while machine learning solved tasks hard to explain to computers.
    • LLMs sit between these: they can follow short human instructions but struggle with complex kickoff planning. Transcript: Benedict Evans And maybe there’s sort of a, I’ll talk about the slide and there’s an observation around it. I think a lot of discussion of LLMs is sort of hunting for the light, what’s the right, what would the right way to conceptualize this? So with machine learning, the right way to conceptualize it was this is pattern recognition. And we’re sort of hunting for the right way to conceptualize LLMs. The slide is that traditional software is deterministic and does things that are easy to explain to machines in fact automation machine tools selling machines typewriters adding Machines right things that are easy to explain to a computer there may be things that are very hard for people to to do but they’re easy to explain so it’s hard for you to drill a hole a hundred Times or to calculate a mortgage in your head but it’s easy for you to write down the logical steps to explain how you do this. So that’s traditional software, like databases, data processing, the whole 60s, 70s mainframe thing. Machine learning is stuff that’s hard to explain to a computer. So it’s hard to explain why that credit card transaction is weird. Or how to move your hand or something like that. Yeah, it’s hard to explain why that’s a picture of a dog and not a cat. You think it’s easy until you try and do it. And then it’s like you try to make a mechanical horse. It always falls over until robotics comes along. So that was machine learning. I also think that as a quiz for you, do you think machine learning is still AI? Or is that now just software? I think there’s a process. Once it’s been around for a while, there’s no (Time 0:17:47)
  • Double Descent Upends Intuition
    • Deep learning defies classical overfitting intuition via phenomena like double descent.
    • Massive parameter models can generalize better despite seeming to overfit the training data. Transcript: Balaji Srinivasan And so your error goes down and then your error starts going up on the holdout set. So you train your model in machine learning and you want the minimum number of parameters to be able to explain the training data and predict the test data. And if you overfit, then you’re no longer predicting out of sampled stuff. But double descent is when you do AI, you get actually a second wind when you start going to a very highly parameterized model and the error actually drops again, right? And which is just a really weird phenomenon that there’s papers on this and so on. And it’s one of the most counterintuitive things about the whole thing that just having these gigantically parameterized models would generalize well, right? (Time 0:19:59)
  • AI Needs Workflow Context (AIOS)
    • GUIs convey developer decisions and constrain choices; prompts do not provide that workflow guidance.
    • There’s room for AIOS that embed workflow context and suggest actions like a super‑smart Clippy. Transcript: Benedict Evans Yes, that’s right. And so if you’re in a workflow as opposed to just a blank screen, you know, it’s one thing if you’re in like Photoshop or Excel. Yeah, it can prompt you on the prompt. But if you’re in a workflow in Salesforce, then there’s a decision taken. This is I’m going to offer the user these five options here and not 750 options. And with a prompt, you don’t have any of that. So you’ve got to shut your eyes and think for a minute of like, well, what would I do here? Balaji Srinivasan And you don’t have that help. This is, you know, Karpathy has talked about this also, but I do think there’s room for AIOS, right? Like, in a sense, and we can talk about crypto in a second, but I think AI and crypto are both actually operating system level innovations. And, for example, it may be someone who just does it as an app or like a downloadable thing and just does it as a layer on top of the Mac. But if you have the full context of all the actions that are happening on your Mac, you can suggest which apps to use, suggest which apps to download, suggest, hey, you probably want to Change these keyboard settings. And so like there’s, you know, it’s funny to put it this way, but Clippy is finally vindicated. Clippy, but for everything, right? And because Clippy can now be really, really, really, really smart, right? Like, you know, it was Anderson’s line. It’s like everything in tech works. It’s just when, right? (Time 0:25:15)
  • Design For Recommendation, Not Just Price
    • Track whether users search for ‘cheap’ versus ‘best’ over time to see internet move from price comparison to curation.
    • Design products for a discovery-first experience as recommendations gain importance. Transcript: Balaji Srinivasan That’s so interesting. So let me see if I can understand the psychology. So it starts from 2004. Benedict Evans In 2004, you go on the internet, and you already know what you want, and you look for the cheap, what is the cheap X, and then you put in a scoop or you put in a product or something. Whereas over time, that goes down. And best goes up. And best goes up and crosses it. It’s a perfect X on the chart, unfortunately. But the thesis is you’re going further up the funnel. You’re looking more and more for, I want someone on the internet to tell me the best X or Y. (Time 0:28:33)
  • LLMs Amplify Domain Experts
    • LLMs excel when the user already understands the domain because they can prompt and verify better.
    • For unfamiliar domains, outputs can look polished but be factually wrong like an overconfident intern. Transcript: Benedict Evans Last long thing I wrote about this was about looking at deep research, which OpenAI launched. And one of the kind of traps in looking at the news thing is to test it based on what was important to the old thing. So, you know, to look at the Apple II and say, does this match the uptime of a mainframe? No, so it’s useless. Well, no, but that’s not the right question. Can you write and build an Excel model on an iPhone? No, but that’s not the point. It can still replace PCs. And the reason I mentioned this is, so deep research, open air launch this thing, and it’s whatever it was, $100 a month or whatever. But then you look at the marketing page and the marketing page shows it doing a research project about mobile, which as we said I know a lot about. And it got the answers wrong. That’s verifying. See, you could tell that it was wrong. But it looked positive. Exactly. So this is the thing and it got stuff wrong in several levels. People are remembering now what I wrote like two months ago. And so there was a specific, it was make a table which shows mobile smartphone adoption in a bunch of countries and then the operating system market share. And then this is like an intern teaching moment. Because first of all, what does adoption mean? Does that mean unit sales share, installed app store sales? Like what metrics specifically are you asking me for? Then it had given a source for the number it had come up with, which was Statista. And Statista is an aggregator that steals other people’s data and re-complishes it. And when you jump through a bunch of registration hoops, you discover that the actual source was, I think, Cantar. Cantar? It’s an ad agency. It’s part of Group M. I thought it was part of one of that. It’s consumer survey data. So it is a proper company. Yeah. So it was actual proper consumer survey data. But the two things, so then when you go to the Cantar chart page, you discover that deep research had got the numbers the opposite. So it had flipped percentages. I see. And then it had also said… Right, because it was an ABAC. It had copied them from the website wrong. I see. And then the other source it gave was StatCounter. And StatCounter was… Was just using the same wrong data. Which is a traffic measure. Right. So that’s not going to tell you adoption. Because high-end phones get used more and iPhones get used more. And there’s a bunch of things in here where you’d like, this is what I’d expect from an intern. I would go back and say, no, this is what I mean by adoption and this is a good data source and that isn’t. And it’s like a great first version. The problem is A had to copy the number out wrong, which is not what I would expect for an intern, or at least not a good intern. (Time 0:33:29)
  • Make Hard Problems Visually Verifiable
    • Where possible, convert non-visual tasks into visual representations to speed verification.
    • Use spectrograms, charts, or diffs to make backend outputs immediately checkable by eye. Transcript: Balaji Srinivasan So give me, I’ll give you a small, simple example. Let’s say it generated an audio file, right? You know, you know, like a spectrogram of an audio file, right? You could maybe immediately see if there’s some artifact there, right? That’s a trivial example. Benedict Evans So I think it’s a fascinating concept. I would wonder whether that’s the right split. Balaji Srinivasan Okay, it’s at least one split I found useful for now, but (Time 0:37:19)
  • Language As A High‑Dimensional World Model
    • Language proved unexpectedly powerful because text encodes a very high‑dimensional map of human experience.
    • Billions of typed words bootstrapped models that generalize across many tasks without direct spatial input. Transcript: Balaji Srinivasan We were talking about this question before. I was surprised you could get so much mileage out of pure text. So much what, sorry? So much mileage out of pure text, right? And the reason I was surprised by that is you’d think… You mean like reasoning and stuff that looks like reasoning? Reasoning and also spatial manipulation, like having cameras, having eyes, seeing the world, reasoning about it, like a baby and so on and so forth. It is amazing how much of that world humans have assigned machine-readable labels to with text and the way that, know it’s just it’s just very surprising how well that worked like language What i’m trying to say is in a few in like 40 words you can describe it’s like code you can describe many many many different kinds of things in like 40 words right and and it’s just more general Um one of those things where sometimes when you’re really close to a space, you’re actually more surprised by a breakthrough than if you’re farther away. And I should say, like, you know, even seeing all the style transfer stuff in the mid-2010s and seeing ImageNet and seeing the benchmarks and so on and so forth, I was surprised that it Got beyond – you know what markov chain is it well um if you saw the stuff before gpt3 right uh it was like semi-coherent but it didn’t look like it was converging on something you know it Just looked like you know it repeated itself many times and what have you and the fact that it broke through to what it did just based on language, was so counterintuitive. And it’s I think it’s because it’s such a high dimensional thing, it captures so many different aspects of the world, like anything you can perceive in the world, there’s a word for it, There’s many words for it. And then we also have billions of people who’ve been typing those words for two decades, right? So in a sense, like the entire internet, the video games and social media were like this bootstrapper for AI. (Time 0:40:47)
  • Conversation Tracks Derivative, Not Adoption
    • Public attention measures the derivative of adoption, not absolute usage.
    • Technologies are loudest during rapid growth and quiet when ubiquitous or obsolete. Transcript: Balaji Srinivasan Proportional to derivative rather than absolute value. So let’s say you have a sigmoid that’s going like this, and then it flattens out, right? So when it’s like a nullity or ubiquity, you know, when it doesn’t exist or when it’s everywhere, when 0% or 100%, it’s just not notable. It’s not worth talking about, right? People use Uber or Dropbox a lot more today than when they were talking about Dropbox and Uber a lot, right? So the conversation is maximum at the time of maximum growth, and then it’s just much less because now it’s like not notable. It’s just a feature of the environment, right? Benedict Evans So you can do Google Ngrams that show exactly this. I think that’d be a (Time 0:51:40)
  • Elevator Attendants’ Bell Curve
    • Benedict tracked elevator attendants’ employment and found a perfect bell curve tied to automation.
    • As elevators automated, attendants peaked then declined, illustrating tech’s job‑pattern lifecycle. Transcript: Benedict Evans I was fascinated by elevators. I get these kind of autistic, autism spectrum fascinations about things. And I, there’s a chart I did of the number of people employed in the US as elevator attendants, which is a perfect bell curve. Balaji Srinivasan Oh, interesting. Benedict Evans Yeah, it’s all curves up and down. And this is because first half of the 20th century, you deploy a lot of elevators. Second half of the 20th century, they become automatic. You have a button. And you can go and find all this advertising. Why were they at the beginning? Was it just like switchboard operators? That was how it worked. Was it technical enough? There was no button. Well, if you think about what it actually takes to have an automatic elevator system in a building, you’ve got to have all the dispatching. Uh-huh. You’ve got to have the dispatching and the queuing. I see. There’s an interim stage where you have an elevator attendant who would just stand in the elevator and you would say, I want floor five, please, and they’d press the button for five. Right. But if you get in, you know, an originally elevator. What was it originally before the buttons? There was a lever that’s an accelerator and a brake. Oh, so it was like a car almost. It’s a streetcar. Balaji Srinivasan It was a vertical streetcar. I didn’t know that. Benedict Evans So there’s a fantastic book I have called Cultural History of Elevators, which is all about how weird this was. So it was a vertical train. Yes, it’s a vertical train. Wow. And that’s how people thought about it. Yeah. And so an elevator attendant, you can kill people. And there’s this wonderful story I tell everybody, which is that you press the buzzer to summon the elevator, but it’s literally you’re just ringing a bell and a light goes on in the elevator Car. And there’s this story from the war department. It’s like hailing a taxi. Yeah. Balaji Srinivasan There’s a story from a war department or ringing for a servant. Benedict Evans There’s a story from the war department in D.C., which is that you would buzz more based on how senior you were. So imagine you’re like a lieutenant and you get into the elevator on the second floor and you want to go to the tenth floor. But on the way, the buzz rang, it rings four times. And it’s a general. So you have to stop on the sixth floor and go down to the first floor, and then a major gets in. So theoretically, this call of 10, it could spend the entire day in the elevator going up and down. So interesting. And we don’t see any of this now, which is your point about conversation. (Time 0:52:50)
  • Elephant Graph Explains Political Backlash
    • Globalization produced winners in low/mid global percentiles and top elites, while many Western middle‑class jobs stagnated.
    • That income redistribution helps explain political backlash to trade and tech. Transcript: Balaji Srinivasan Probably gesturing at something that’s right it shows um percentiles or deciles of the world in terms of income. And it shows over the last 20 something years, I think from 91 to 2008, or something like that, where the growth went, like whose incomes rose. And basically, most of the world, so the lower 10% in Africa didn’t gain that much. But like maybe from the 10 to 20% through the 70% to 80% had huge growth then it drops off in the 80 to 90 percent almost zero and then it picks up again at the very top right and so that means Is the like global you know elite in every country did great and so did china india vietnam eastern europe all these countries are no longer socialist communist etc, etc. Right. But the Western middle class didn’t. And that is a big part of, I think, the silenced really. Now, when we’re looking at it is, you know, in America, they have, you know, obviously red versus blue. But one way of thinking about it is starting in, you know, certainly in 2008, there’s a ramp where China flips US manufacturing. And so China puts all this pressure on Red America, and that leads to Trump and trade war. And you’ve seen that graph of print media disruption, right? That’s the internet suddenly rising after 2008 to flip Blue America, and it takes all the ad revenue away. And it’s not just ad revenue, it’s also Craigslist’s classified ads, a bunch of other things. So the internet disrupts Blue America, and that leads to wokeness in the 2010s, I think. And also tech clash, right, which is the anti-tech movement. So we look at it as red and blue, but there’s also China and the internet over here where the internet is disrupting blue and China is disrupting red. So the thing I think that’s coming next is AI disrupts blue America and robots disrupt red America. And so Chinese robots internet AI. And so that artisan movement kind of thing is going to accelerate where people are going to be mad about that happening. I think on balance, there’s going to be a lot more productivity in the rest of the world. (Time 0:59:16)
  • Billionaires, State Power, And Decentralization
    • Over the 20th century the world moved socially left and economically right after communism waned.
    • Today’s billionaires reflect changing state power and new decentralizing forces, not just inequality. Transcript: Benedict Evans Tension in looking at progressive ideas and saying, because if you look at the last hundred years, the social progressive ideas have always won. Like nobody today says like being gay should be illegal. Like, so, you know, a little bit like what we were saying about AI a while ago today, you could deterministically say that what is woke today in 30 years time will be what every far right Conservative agrees with. Yeah, people have said that kind of thing. Theoretically, you know, maybe, maybe not. Yeah, it’s interesting. You also have these kind of overreaches around this. It does strike me that one of the differences between the US and UK politics is that what happened in my lifetime is that the right, for want of a better term, won the economic argument That state ownership and government control of the economy is bad. Right. And the left won the social arguments that, like, gay marriage is okay. Well, it’s… And so on. And what happened in the UK was the right embraced that, and the Conservative Party is the party that brought in gay marriage in the UK. All right. Whereas in the left, in the US, it’s kind of the other way around. The Republicans kind of… And Tony Blair sort of brought in kind of… Yeah, and he brought in the and middle economics whereas what happened in the u.s is that the republican party in the u.s never kind of accepted that it had lost the social arguments well Balaji Srinivasan It’s interesting i think um from 1950 like the moment of 1950 you do have something where because communism fell, basically because Nazis was defeated, the world moved socially to The left. And then as communism was defeated, it moved economically to the right. And so thus, for example, like the immigrant billionaire or gay billionaire is like in a sense— Can be white wing. Well, they’re far to the left of 1950 socially, and they’re far to the right in an economic right in a sense of 1950 economically. (Time 1:05:30)
  • Crypto Matures Past The Hype Tourists
    • Crypto today is driven by engineers and financiers building infrastructure after the tourist/speculator phase.
    • Its strongest near-term use cases sit in finance rails, stablecoins, and developer tooling rather than mass consumer apps. Transcript: Benedict Evans Which I hope you won’t tell me I’m wrong, is like there’s a bunch of clever people working away, building, like all the tourists left. Like the whole NFT thing was all nonsense. Not all there. All the tourists left. The tourists and the grifters basically all moved on to AI. Yeah, a lot of them, yes. And all the kind of people trying to build content brands, saying this is all wonderful or this is all bullshit. They all moved off to AI. There’s a bunch of people sitting and doing like abstruse, very clever, very technical stuff. There’s a bunch of stuff working or being built that may work around a financial finance industry around finance rails around stable coins various kinds of financial instruments Most of which is storing money or speculating in money or moving money around Yeah. There is a thesis that you could build Instagram on this, that this is sort of an open source computer in which you could write software that consumers would use. And I have a bunch of questions about how that would work, whether that would work, whether you would need to abstract the crypto stuff away so that the consumers didn’t see it. And if you did that, then why would they care? Totally. But none of that’s kind of there yet. Like there aren’t billion scale consumer apps built on blockchain yet. So there’s a sort of watch this space around that. And then there’s the finance side, which I think is sort of theoretically very interesting, but I struggle to get very interested in it just personally. It’s not what I’m interested in. And I struggle to see ways that I could add value in talking about it. So I kind of pay attention to it. And every now and then I point out like, my newsletter on Sunday, I pointed to the Shopify and Stripe and said, like, there’s stuff happening here. Yeah. And you should pay attention to this. And there’s people still interested in trying to build things. So if you’ve just written this off as all bullshit, you’re kind of wrong. (Time 1:29:20)
  • Block Space Is Crypto’s Bandwidth
    • Block space is crypto’s bandwidth: applications arrive as blockspace and throughput increase.
    • Think of blockchain capacity gating consumer use cases just like internet bandwidth once gated media types. Transcript: Balaji Srinivasan Have you heard that parameter before? Okay, that is the most important parameter in crypto that people outside crypto don’t realize governs crypto. Block space is to crypto what bandwidth is to the web. So if you think about the early internet or the early web, I should be more precise in the 90s, like it was very bandwidth constraints is 28, 857, 6 modems. And so that’s why like Google was 10 blue links. And I think Amazon even had many images at all. And in fact, you remember Six Degrees? It was a social, right? So that was a text-based social network. It didn’t take off because without images, people didn’t really- Yeah, you got nothing to share. You got nothing to share, exactly, right? ICQ was a chat app that did work. AOL and some messenger worked because that was just text that could be sent and that low bandwidth thing. It was only in the 2000s that you started to get more graphical things when bandwidth increased. Like Facebook, the reason it took off at Harvard, everybody had a T1 connection being at Harvard. And they finally had digital cameras so you could have photos. And as digital cameras propagated out, so did Facebook, right? And you go further and further and like, you know, the internet only or internet explorer only got disrupted by Firefox in like the late 2000s, right? It was only really by the early 2010s that you had the full JavaScript stack of like jQuery and then only later for React and what have you. So this concept that we have today of like a mobile web app where you can download JavaScript and run an app in the browser on a phone was a vision in the 90s. But it took a long time together because bandwidth had to increase for that, right? So what’s the analogy here? Block space. Basically, block space is the amount of storage that you have on a blockchain. Like think of a blockchain is like an armored car for data, right? Because this is data that people want to corrupt, right? In a sense, if it’s a file on disk, it’s important to you. If it’s a file online, it’s important to others. And if it’s a file on chain, it’s really important to others. And it’s so important that they might try to screw with it. And so Bitcoin came up with like an armored car for data, where you could guard the minus one or plus one of who had what Bitcoin. And over time, that block space increased so that you could do some basic smart contracts on Ethereum. And now it’s increased enough that you can blast millions of stablecoin transactions a day on like base and Solana and so on and so forth. And so you should conceptualize it as, oh, why hasn’t this happened yet? And instead think of, okay, these applications are gated by the amount of block space. And so they’re coming online similar to the amount of bandwidth. You had like text-only apps, then you had images, then you had videos. And like Netflix only did streaming video in like the early 2010s, right? I mean, we think about all that as reset. That’s the way I’m thinking about it. I don’t have a problem with the idea that you couldn’t build Instagram on this because the infrastructure isn’t fast enough. Benedict Evans Blockspace wasn’t there, yes. (Time 1:47:25)
  • Build For Blockchain Tradeoffs
    • Treat blockchains as new OS layers: slower than bare metal but enabling new capabilities.
    • Design apps to trade off latency for composability when you need global, tamper‑resistant state. Transcript: Benedict Evans Except that it allows you to do a bunch of stuff that you can’t do if you want on the bare metal. Balaji Srinivasan That’s exactly right. That’s exactly right. And the thing is, blockchains are, in a sense, one of the frontiers of operating systems research. Like in the same way, like there’s an operating system like Windows, there’s a browser, which is itself an operating system because you can run apps in it, it’s got a full programming Language, like that’s how Chrome layered it. Benedict Evans Were you at A16Z when Martin Casado was there? Balaji Srinivasan Yeah, we overlapped just a bit. We invested a bunch of things together. Benedict Evans Martin had this great observation. You remember when YC said that for a quarter of their companies, 90% of the code was written with AI? Yeah. And he responded to this by saying, yes, but if you write an iPhone out, 90% of your code is written by Apple. Yes. And so there were all those levels of abstraction. Prompting is just a higher level of programming. That’s right. Yeah, exactly. And so there’s a, I suppose the, you know, another way of answering your question is like, the finance stuff is there. I can see it. I get it. I’m not sure I can add any value to that. It’s interesting. And I will tell people it’s kind of interesting, but you pay attention to this. I think you’ll be a leader. Go ahead, sorry. The running, the building more generalized consumer applications on it is conceptually more interesting to me as something that I could make money telling other people about. Yes. Except that it isn’t happening yet. And it probably will at a certain point the curve will curve up, the block space will expand, the stuff will get faster and cheaper and can store more stuff, and people will be able to build Stuff on this. Deterministically, it won’t be exactly Instagram. I think that’s just kind of a useful mental model for thinking that you could build something like that. You could build consumer network apps like that on this. (Time 1:51:01)
  • Anchor Provenance On Chain
    • Use crypto for verifiable provenance and tamper‑proof custody when authenticity matters.
    • Anchor camera hashes or data digests on‑chain to create immutable chains of custody for evidence or science. Transcript: Balaji Srinivasan AI makes everything fake, crypto makes it real again. Because AI can fake all kinds of stuff and give you this very convincing thing on like the deep research thing where it said 40% of the phones or whatever you’re saying. But it cannot fake the private key, so it cannot show a non-zero Bitcoin balance or non-zero Ethereum balance without actually having the cryptographic solution there. Benedict Evans Yeah, but it could probably just tell you that the balance is zero because it might be. Balaji Srinivasan Yeah, sure, sure. It could make it up. But what I mean about that is, like for example, all kinds of, let me give you a, you know, CAP right, websites? So AI can bust a lot of CAPTCHAs now. It can get through, it can, am I a robot? It can figure it out, get through. But if you had to log in with a crypto wallet that had $1 in it or $10 or $100, AI can’t fake that. It cannot fake the possession of that cryptography, right? Like, to give you one, here’s one motivating example for why crypto will get, maybe this argument will convince you. Maybe not, but it’s fine, you know. Google login, you agree is at billions of users, right? But Google login, when you log into a website, you only can log in basically with your email address and the permissions to your Google account. There’s something very obvious that somehow even Google with all of its strength has not been able to implement which is an international balance a spendable balance right google Login could not have for whatever reason a spendable balance across different countries they’ve solved that for google itself where everybody can pay google and subscribe to google With a zillion credit cards in all these different countries but somehow they couldn’t make it work so you could log into a third-party site with a spendable balance. Crypto did solve that. Just that alone means that every Google and Facebook login will eventually be either augmented or replaced by a crypto login. So (Time 1:55:59)
  • Oracles Align Incentives For Truth
    • Prediction‑market oracles and on‑chain feeds reduce partisan noise where money bets encourage verification.
    • Markets and cryptographic oracles can improve truth discovery when stakes exist. Transcript: Balaji Srinivasan Because if you’re making a financial decision, I don’t know if you’ve seen that stuff, Alex Tabarrok has talked about this. When people have money on the line, their partisanship reduces and they actually get a different chip in their head where they’re like, is this true or not? They’re trying to dispassionately figure it out, right? They’re not just cheering my tribe, your tribe, whatever. And is this true chip basically means, okay, I’m going to double click into this. I’m going to verify this. I’m going to look at this. And that’s where like oracles come in. They’re like feeds of data that have some degree of verification. And right now they’re like mostly price data, but people use it for weather data. They use it for this, that, and the other, right? All these different feeds of information that people trade on. And over time, I think those feeds, once you can guard price data, weather data, you know, health data, et cetera, eventually you can guard any kind of data. (Time 2:01:06)