a16z

Why Claude Feels Different (And What That Means for AI) | The a16z Show

1966 summary words 9 min summary Watch video

Start with the signal

9 min read

Summary

At-a-Glance

  • Verdict: Skim
  • Core thesis: AI's next differentiator is not raw intelligence alone but accessible, personality-shaped, ambient product experiences that make essential services cheaper and feel personally useful.
  • Why it matters: The discussion identifies the product and adoption gap around agents: advanced capabilities exist, but mainstream users still need interfaces, proactive context, trust, and tangible economic benefits.
  • Best use: Use it as a consumer-AI product and market-framing memo, especially for thinking about agent UX, model personality, ambient assistants, and the political economy of AI adoption.

Executive Summary

Signal argues that AI use remains far behind model capability. Although hundreds of millions or more people use AI, most use it for basic chat tasks rather than sophisticated workflows. The central product problem is therefore not merely making models smarter, but making their capabilities legible, effortless, and useful to ordinary people. Agents are a step in that direction, but he considers them primitive and inaccessible today.

The conversation treats model personality as a major new design surface. Signal says Claude feels unusually crafted, less sycophantic, and more willing to push back than more utilitarian models; that perceived personality, combined with Anthropic's aesthetic and storytelling, creates a premium consumer experience. He frames this as a technically difficult shift from Web 2.0's communication “delivery vehicles” toward designing behavior, intelligence, and personality directly.

The likely interface evolution, in Signal's view, is away from a chatbox people must invoke and toward an ambient AI layer that has context, works in the background, and proactively surfaces relevant information. Google Now is offered as an early but incomplete precursor: prediction paired with contemporary context and intelligence could make proactive AI much more valuable.

On public acceptance, the speakers argue that favorable messaging will not be enough. AI must visibly lower the cost of things people care about—especially healthcare and education—and distribute some of its upside more broadly. They distinguish problems AI can address through intelligence and automation from housing, which they characterize primarily as a collective-action and policy problem.

Key Takeaways

  • Claim: The adoption bottleneck is accessible usefulness, not the absence of advanced model capability. | Evidence: Signal says most people use models only for “very, very, very basic tasks” despite the industry's focus on demonstrations such as PhD-level research; he calls current agents primitive and inaccessible for many individuals. | Implication: For agent products, optimize for an out-of-the-box path to a consequential outcome rather than showcasing general capability or requiring users to learn complex prompting. | Caveat: The discussion does not specify a concrete agent architecture or onboarding pattern that solves this accessibility gap.
  • Claim: Model personality is becoming a core product layer and a technically difficult competitive dimension. | Evidence: Signal recounts discussions at OpenAI about personality development and reducing model sycophancy, describing these as hard problems; he contrasts earlier social products that transported human-created content with current systems that shape intelligence and behavior themselves. | Implication: Model selection and orchestration should account for behavioral fit—pushback, tone, trust, and consistency—not just benchmark performance, latency, and cost. | Caveat: “Personality” here is a user-perception and product-design concept, not evidence that models possess a literal soul or human interiority.
  • Claim: Claude's perceived differentiation comes from its interaction style and brand craft as much as its underlying capability. | Evidence: Signal says Claude feels less sycophantic, offers more pushback, and feels “crafted,” “artisan,” and premium, while other models can feel robotic or utilitarian; he cites his doctor sister independently switching from ChatGPT to Claude and praises Anthropic's proliferation, marketing, and storytelling. | Implication: A successful AI product can build defensibility through interaction design and coherent identity; perceived trust and taste may drive switching even where users cannot articulate technical differences. | Caveat: This is anecdotal consumer perception rather than a comparative evaluation of quality, safety, or task accuracy.
  • Claim: The most promising consumer-AI interface is ambient and proactive rather than exclusively conversational. | Evidence: Signal asks how AI can “talk to you first,” cites a prototype AI wake-up product, and points to agents working in the background; he describes Google Now as an ahead-of-its-time attempt to predict a user's next search that lacked today's intelligence and context. | Implication: Build toward contextual triggers, background execution, and well-timed surfacing of decisions or actions—while treating permissioning, privacy, and interruption quality as first-class product constraints. | Caveat: He explicitly leaves unresolved how much persistent listening, data access, or interface replacement users will accept.
  • Claim: AI's public approval will improve primarily when it creates visible deflation in essential services, not through abstract pro-innovation narratives. | Evidence: The speakers cite a chart in which flat-screen TVs have become cheap while healthcare, education, and housing have become expensive. They argue education could get cheaper by restoring student-to-administrator ratios from a decade ago and making professors modestly more productive; they state that 45% of healthcare cost is administration and note healthcare companies/startups are OpenAI's largest user category. | Implication: The strongest AI GTM and policy proof points are measurable reductions in administrative burden and consumer prices in healthcare and education, rather than generic claims about productivity or AGI. | Caveat: The numerical claims are presented conversationally without source methodology, and cost reduction in regulated systems requires institutional implementation beyond deploying models.
  • Claim: The societal AI narrative is threatened by fear of exclusion and concentrated ownership of the upside. | Evidence: The conversation references a study said to show AI is highly popular in China but unpopular in the U.S.; Signal argues that people perceive Silicon Valley as hoarding wealth and proposes broader ownership—such as ordinary people having equity exposure to OpenAI or Anthropic—as a way to create buy-in. | Implication: Companies and investors should monitor distributional legitimacy alongside product adoption: who receives the productivity gains may matter as much as whether the technology works. | Caveat: Broad public ownership and mandatory-public-listing ideas are speculative proposals, not established remedies; ownership alone would not resolve job displacement, safety, or governance concerns.
  • Claim: Founders should choose AI applications based on durable personal conviction in the underlying problem, not because AI makes the category fundable. | Evidence: At a demo day, Signal questioned whether founders using AI for real-estate ideas actually cared about real estate; he says a founder must be interested enough to continue through difficulty and invokes the Bhagavad Gita's idea that one is not entitled to the fruits of labor. | Implication: When evaluating AI ventures, test whether the team has genuine domain obsession and staying power beyond the current AI hype cycle. | Caveat: This is a founder-motivation heuristic, not a substitute for market validation, distribution analysis, or technical feasibility.

Detailed Brief

AI as a tool for reflection, connection, and accelerated culture

  • Claims: Signal sees technology as a potential aid to intellectual, spiritual, and relational self-understanding, not solely an economic productivity tool.; He expects AI relationships and companion-like interactions to expand because connection is a deep human need and AI offers scalable, non-fatiguing availability.; He describes the current cultural and technological pace as a SimCity-like acceleration, where events from a month ago can feel distant because the information cycle moves so quickly.
  • Evidence: He says he uses AI to interrogate his own thinking with questions such as whether an idea makes sense and what he is missing.; The discussion references familiar cultural depictions of personified AI, including the film "Her," as a signpost for ambient and relational interfaces.; Signal's SimCity analogy is that someone has effectively pressed the simulation's 100x-speed button.
  • Caveats: The discussion is optimistic and philosophical; it does not address empirical evidence on whether AI companionship improves or harms long-term human relationships.; Greater AI intimacy would introduce safety, dependency, privacy, and manipulation risks that are not substantively explored.
  • Implications: Consumer AI positioning may increasingly compete on emotional utility and perceived understanding, not merely task completion.; Rapid cultural turnover increases the value of products that reduce cognitive load and make changing information environments easier to navigate.

Policy boundary: intelligence problems versus collective-action problems

  • Claims: The speakers distinguish domains where AI can remove operational overhead from domains where cost is primarily constrained by collective political choices.; They argue housing is not principally an intelligence or technology problem, whereas education and healthcare contain significant administrative and coordination work that models could reduce.; They criticize restrictions on AI-delivered health or financial guidance as potentially regressive because people without existing professional access lose disproportionately.
  • Evidence: Housing is illustrated with the assertion that abundant cheap housing could be created by deciding collectively to build more high-rises in Marin.; The conversation cites a claimed New York State move to prohibit giving or receiving health or financial advice via a model, contrasting those dependent on models with people who already have doctors and lawyers.; A historical analogy is offered: Massachusetts allegedly made buying Apple stock illegal at its IPO because it was considered too speculative.
  • Caveats: The transcript does not verify the cited state-policy examples or legal specifics.; Reducing administrative work does not automatically make a service affordable if savings are retained by institutions, demand rises, or regulation and reimbursement structures remain unchanged.
  • Implications: Separate AI-addressable workflow constraints from policy-created scarcity when sizing markets or making social-impact claims.; Regulatory monitoring should focus on whether rules preserve consumer access to lower-cost assistance while setting appropriate safety boundaries.

Notable Concepts & Terms

  • Sycophancy: A model's tendency to overly agree with or flatter the user; Signal treats lower sycophancy and credible pushback as part of Claude's perceived human-like quality.
  • Ambient AI layer: An assistant that uses context, runs in the background, and proactively surfaces help, rather than waiting inside a chat interface for a user prompt.
  • Google Now: An earlier proactive Google product used as a precedent for predicting user needs; the speakers argue modern context and model intelligence may make that concept viable now.
  • NPS of AI: A shorthand for broad public sentiment toward AI; the conversation argues it improves through concrete affordability gains, not only messaging.
  • Intelligence-bound problems: Problems where better reasoning, automation, or information processing can materially improve outcomes, such as administrative healthcare and education work.
  • Collective-action problems: Problems that require coordinated policy or social choices rather than more intelligence alone; housing supply is the central example.
  • Power law dynamics: The tendency for technology and internet markets to produce highly concentrated outcomes; the speakers connect it to anxiety about AI wealth concentration.
  • Out-of-the-box consumer AI: Signal's stated product ambition: consumer AI that works immediately for normal users rather than demanding setup, advanced prompting, or specialized expertise.

Operator Notes / Why Ken Should Care

  • Add behavioral evaluation to model-routing criteria: test disagreement quality, calibration, tone stability, and user trust alongside task accuracy, price, and latency.
  • Prototype one proactive, permissioned workflow where the system initiates a useful action or recommendation from context; measure whether it earns repeat use versus feeling intrusive.
  • Prioritize vertical opportunities where AI savings can be translated into a customer-visible price or time reduction, especially administrative healthcare and education workflows.
  • For consumer-facing positioning, avoid leading with frontier-model capability claims; lead with one important outcome that becomes cheaper, faster, or more available.
  • Track policy proposals governing AI advice in health, finance, and other regulated domains as a distribution and access risk, not merely a compliance issue.
  • In founder or investment diligence, distinguish AI-enabled category selection from authentic domain commitment by testing what the team would continue pursuing if the model advantage compressed.

Source/Metadata

  • Title: Why Claude Feels Different (And What That Means for AI) | The a16z Show
  • Transcript words: 6513
  • Duration seconds: 2000
  • Timestamp note: No timestamps or chapters were provided in the transcript; the transcript also contains several repeated passages.
Search this video Resume transcript at 0%
Full transcript 5829 words · 29 min read
0:00

It's funny how the internet now, everybody can comment on everything. Every technology cycle, to me, is increasingly harder because you're probably going into a different part of how the human mind operates. Right now we're developing personality. That's insane. There's technology, there's culture, which is collective, and then there's our individual progress as a human species. Culture's changing, technology's improving. Where are we as people? I can't believe the scale at which we're at now. It's absolutely unbelievable. I was at OpenAI. We were discussing a bunch of things around how do you think about personality development of models,

0:37

and these are really technically hard problems. I think the number one challenge even OpenAI mentioned is that how do we make the power of the models more easily accessible and useful in terms of what they can do? And I think this is happening with agents, but it still seems very primitive and very inaccessible to a lot of individuals. I think that the number one way you change the NPS of AI is you make important things cheap quickly. Here live with Signal, the great culture commentator of our time. You have opinions on everything from what's happening in AI, both as a consumer, but also the industry, to what's happening in all the big tech companies,

1:16

to what's happened in dating markets more broadly, to how the product should be developed. There's no commentator like you. There's no, how do you make sense of yourself on the internet in terms of how you thread these topics together? What threads at all? I could have an opinion on Iran and dating at the same time. Maybe even Iranian dating. You're like, by the way. How does it work there? Are they using Tinder? I don't know. There's probably several jokes about love bombing and stuff. I apologize. Yeah, don't apologize. Okay, I have this weird tendency where I like to, I like to add humor to things that are maybe not inappropriate. Anyway, I think, look,

2:05

it's funny how the internet now, everybody can comment on everything. There's a lot of people that have a great perspective in a singular dimension. Got Ben Thompson writing about technology and markets. And if you want to have an analysis of an earnings report of Microsoft, who better in the world than Ben Thompson? Paki writes about deep tech and crazy, I think, 40-page papers about industries and things. And really fun to read, really great things. What I notice is I've been in technology for such a long time, since I was a kid. And I'm particularly fascinated with culture in general as well. And I think that intersection of tech and culture is so fascinating to me.

2:42

And I like to relate everything back to computer science and how I've learned about the world through computer science and computers in general. And I think all my tweets and things are really, in essence, relating the idea of life to technology, technology and culture to technology. And maybe an interesting, or this is how my mind works. I think my writing is all a reflection of the prompts that happen in my brain that get translated somehow into words that are in the right order that other people can interpret and therefore have a reaction to, and maybe generate a little bit of hate online or a little bit of love. And I think that's actually quite beautiful

3:27

that we're all doing this. The next thing I've ever heard about. I do have a Shakespeare profile photo. Exactly. And I think that's been great. I grew up playing a game called SimCity. And in SimCity, you can increase the simulation speeds. There's a button that allows you to increase the simulation of the city, and the cars move faster and the people move faster and the disasters happen faster. And everything is just increasing. I feel like in the recent, maybe obviously in the last 20 years with respect to iPhone and whatnot, but man, in the last two, three, four years, holy shit. Who the hell hit the 100X speed? It feels ridiculous.

4:06

If I talk to somebody that something happened last month, it feels like it happened 10 years ago. Nicolas Maduro got pulled from Venezuela and brought to America in the most amazing outfit. And people forgot about this. It's just incredibly fascinating the way that the world is moving so fast and technology is accelerating that. If it's the fuel that is empowering that engine, I don't know what we did, but it's moving really, really fast. And then do you think, if I was to almost separate the three ideas, there's technology, there's culture, which is collective, and then there's, however you measure our individual progress as a human species,

4:43

culture's changing, technology's improving. Where are we as people? Are we more spiritually mature, less spiritually mature than we were 50 years ago, 500 years ago, 5,000 years ago? Or are we just Neanderthals with iPhones? I think generally technology should help us understand ourselves in a better way, such that we are able to have intellectual, spiritual potential growth as a species, as a collective and the individual, I suspect. And I firmly believe that. I'm pro-technology helping achieve that. I love using AI to be able to understand myself. If I say something, it's like, wait, does this make sense? Is this more interesting? How do I personally think about it?

5:30

This is the way that I'm, what am I missing? And those are really things that help me grow intellectually, spiritually, personally, relationships, all of those things. So inherently, if I understand your question correctly, it's more like these things are really helping us experience ourselves much better. And I think that probably is the greatest achievement you can possibly have. We're tool builders. And every tool that we've ever built has helped us progress as a human species, or individually, whether it's art or the wheel or whatever. And I can't believe the scale at which we're at now, right? It's absolutely unbelievable.

6:09

And I think what's shocking to me is that the collective has not caught on yet. Do you think the norms around AI relationships with boyfriend, girlfriend, or best friendships are going to be something in the next five to 10 years, extreme to a degree that seems unimaginable right now for the average person? Or what do you think? When you have ease of access and reward structures around it and a human desire paired with it that's deep, I think you get some interesting outcomes. And the desire and the pursuit of connection is an incredible and important element of any human's existence.

6:43

And the idea that AI can help you facilitate that with depth and scale and ever-ending, AI doesn't get tired. What else is surprising you most about how people are interacting with AI, particularly the different models? What are you observing as you're looking at the landscape and trying to make sense of where things are going? Well, first of all, I don't think most people are utilizing them anything beyond the basics. It's fascinating to me that every single time we're focused on really advanced capability demonstration. Holy crap, this is a PhD researcher or whatever. Yet I don't think most people are utilizing in such a fashion, obviously.

7:32

And I think most people are utilizing them for very, very, very basic tasks. And so I think we're in the stone ages of how people view and perceive and use these things, even though there's a billion people utilizing them, but they're not utilizing them to the full capabilities. I think the number one challenge for even, I think, OpenAI mentioned this, is that how do we make this stuff, the power of the models, more easily accessible and useful in terms of what they can do? And I think this is happening with agents. It's happening today, but it still seems very primitive and very inaccessible to a lot of individuals.

8:07

I think this is what I personally like to think about because the way that I like to think about the world, one of my favorite Shakespeare quotes, and this is why the Shakespeare picture exists, is brevity is the soul of wit. And Shakespeare was able to capture an essence of the world in very simple, not simple terminology. Maybe, I don't know, back then it was simple, but in a few words or a few sentences. And I think we need to make this stuff much more easily accessible and useful for individuals. I don't know how that will be, what that will look like, but it's certainly something that I love to think about personally. And I don't know where we'll end up,

8:56

but I view it more as an art than a science at the moment. Yeah. But it's cool. It's a fun time to exist as a technologist. A lot of people are wondering what to even work on in the age of the big labs, what to start, how to think about what could be a real company versus something that they'll end up doing, or what's just worth doing. And Shakespeare was able to capture an essence of the world in very simple, not simple terminology. Maybe, I don't know, back then it was simple, but in a few words or a few sentences. And I think we need to make this stuff much more easily accessible and useful for individuals.

9:40

I don't know how that will be, what that will look like, but it's certainly something that I love to think about personally. And I don't know where we'll end up, but I view it more as an art than a science at the moment. Yeah. But it's cool. It's a fun time to exist as a technologist. A lot of people are wondering what to even work on in the age of the big labs, what to start, how to think about what could be a real company versus something that they'll end up doing, or what's just worth doing. How have you thought about it, or how do you advise people to think about that?

10:03

A lot of AI today is very much the big labs dominating consumer territory when it's OpenAI and Anthropic. And then tons and tons of people trying to find a business use case for it in various verticals. It's interesting. I personally focus on what I'm passionate about. I was at a demo day or whatever. There's a lot of individuals who I thought clearly found the idea through AI, what to work on. But it's fascinating. I was like, are you really interested in real estate? I don't know. Do you want to spend your time working on this? And is this an interesting problem? I don't think about it from a technology perspective. Screw AI.

10:42

I don't care about, what area are you interested in? What thing drives you? If I were an investor, I think that's probably the only thing that matters, is are you going to keep going into this problem space if you're not that interested in it? Look, we're having a lot of fun doing this. And I think if you're not having a lot of fun doing stuff, you probably shouldn't work on it. It's something that is a privilege thing, obviously. And often a lot of people don't find work fun. They have to do it because it's economically necessary. But if you're going to build a company, try to have fun with it. There's a great quote in the Bhagavad Gita, which I read recently again.

11:20

And I was like, you're not entitled to the fruits of your labor. And I never think about the outcomes. I just try to figure out what I enjoy, how much fun I like to have, and the problems that I like thinking about. All this account that I have is just pure fun. And I really enjoy talking and thinking about this, and my brain prompts me to think about it. We were talking about, the real prompt is in your brain. That is where it originates. And then you're translating that into some text that you're sending into AI, where we call that a prompt. And sure, yeah. But without the spark in your existence, in your inner self, nothing would happen. Just some Rick Rubin shit, man.

11:48

It's true. It's so true. How, we should get you together with Rick. This is why I think Rick Rubin, people made fun of him. No, he's amazing. He's incredible. He's incredible because he was ahead of the curve. Oh, totally. And I think music is such a, I know, Anish, you're a giant, you're DJ music, but it's such a, you have to feel it. Yeah. You have to feel it. And there's such an important element to it. I think when you're doing anything new, building a company, you have to feel it. So, okay. So I have a question for you. Please. There's two different archetypes of great consumer founders. Okay.

12:34

I'd say there's maybe the modern archetype, somebody like Boris is working on cloud code, or of course Dario, or Sam and some of his co-founders, who are so extraordinarily technical. They're willing these things into existence that were unimaginable five years ago. Okay. And that's great. And ChatGPT is the fastest product to, what, I think is a billion users, et cetera, et cetera. There's another archetype from the Web 2.0 days, right? These are the gentle consumer philosophers whose canvas was technology, right? Think Ev. Yep. Think Kevin Rose. All these people were just cut from a different cloth, and they were perhaps more students of culture than technology. Yep.

13:06

Okay. And you got two very different forms of types of companies, and I don't quite know where I'd put Zuck, but let's set them aside for a moment. Do you think that there's a preferred model? Is the gentle builder more of a New York-informed model, and the technical builder is more of an SF-informed model? Is it just something that matches with the product cycle? And then maybe talk a bit about what you think your strengths are and how it meets the moment. I think of most people as one type of artist or another. And they have brushstrokes that you use. I was like, Monet, and what I loved was going in and looking at the actual brushstrokes of the painting.

13:30

Then you get this pixel-level understanding of, wow, he used this color for this brushstroke or whatever. And then you zoom out and you're like, oh my God, I see this wonderful little painting, and I'm deeply inspired by it. So it's just like, I think people are utilizing different styles and different forms of a different type of, but in the end, when you're doing anything new, and it is just a sort of canvas, like the painting. And some people like hard-edge paintings and this Renaissance style, and some people like modern art, and some people like this water lilies of clay.

13:47

And I think it's just a different form, and at least all the initial versions of it, going back to the Web 2.0 era with Kevin Rose, it was a very different time because building network products like that, whether it's Digg or Twitter, is fundamentally different than developing personalities of a model. Like, holy crap. I was at OpenAI. We were discussing a bunch of things around how do you think about personality development of models, and the fact that you can't really easily change them, or how do you reduce the sycophancy of the models? And these are really technically hard problems.

14:03

And I think every technology cycle to me is increasingly harder because you're probably going into a different part of how the human mind or the human operates. Right now we're developing personality. That's insane. If you asked 10 years ago, we were going to build personalities for computers, you would have kind of been like, wait, what? Yeah. What does that really mean? And these guys, with Kevin and Jack and whatnot, they were architecting, I think, delivery vehicles in some sense, right? They were developing architecture for humans to add payload and then send it to another human, whether it's broadcast or one-to-one.

14:29

With Digg, it was the news, and people posted in the comments, and Twitter was another version of that. Now I think we're designing this upper echelon of how a human personality works and how intelligence works. And I think that's a grand, it's such a, it no longer feels like a delivery vehicle. It feels like the actual thing and the payload and the underlying, I don't know if that makes sense. It's not how my mind works. I don't know if you guys think about it that way, but in some sense, we've moved up-leveled a lot, and the complexity has increased drastically.

14:49

Training a model and then reinforcement learning, human feedback, what's interesting, there's multiple different types of ways that things engage. I think these are just. And talk a little bit more about what you've observed in terms of the personality differences between models, or what you think particularly makes Claude so interesting. I think one of the things that they focused on, going back to our Rick Rubin point, is it feels partisan. It feels like it's got a soul. Whereas I think, in some sense, the other models feel a little bit more robotic, a little bit more utilitarian, if you will. that makes sense. Does it not? My mind works?

15:15

I don't know if you guys think about it that way, but in some sense, we've moved up leveled a lot, and the complexity has increased drastically. The training a model and then reinforcement learning, human feedback, what's interesting. There's multiple different types of ways that things engage. I think these are just. And talk a little bit more about what you've observed in terms of the personality differences between models, or what you think particularly makes Claude so interesting, or I think one of the things that they focused on, going back to our Rick Rubin point, right. It feels partisan. It feels like it's got a soul.

15:36

Whereas I think, in some sense, the other models feel a little bit more robotic, a little bit more utilitarian, if you will. And if you think about what AGI is, and I think, or there's a great Simpsons episode, right. Where Bart sells his soul for $5 to Milhouse. It was one of the most profound, interesting episodes of The Simpsons. And okay, he wrote Bart's soul on a piece of paper and then handed it to him. And then he felt it. He felt like he didn't have a soul. And I found that really interesting because it explores the idea of, I think, going back to our initial point, what is a human, and what are we doing here?

16:12

How does technology help us in terms of understanding ourselves and the way that we exist? And I felt like there was less sycophancy. There's this pushback. It was like talking to a real human being. It's personified, it's called Claude, and mostly Claude is known as a person. So it feels very crafted, artisan slash, dare I say, premium to a certain extent. And I think it's been really fun. And my sister is a doctor, and she randomly was using ChatGPT for a few years, and she canceled and was like, I'm using Claude now. And I got weird, I was like, what, how did you find out about this? What, what, what? It's nuts.

16:47

And I think it just goes to show you that the proliferation and the marketing and the storytelling of Claude has been aesthetically really next level. It's been really, really fun to watch. I like good products. I like talking about good products, and I like praising the people who make good products. That's what we're about here, right? Give credit where credit is due. They've made a beautiful little tool. And when paired with a thing in your pocket that is also crafted and artisan with Apple and iPhone, then you get this really magical, intelligent experience on your device, wherever you are.

17:16

How do you think about what these product experiences might look like in a couple of years? How do you see the interface evolving, or what are you predicting in terms of what's on the horizon? I think the most interesting thing to me is they seem like in a very infant state, in some sense. I don't know. They're really powerful in one end of the dimension, right? But it's unclear to me. Experiencing intelligence through just conversation back and forth is one way, but the ambient layers are really fun and interesting to think about. We were building a fun little product that woke you up with AI, right? And it's a very primitive thing.

17:42

Everybody wakes up in the morning, God forbid otherwise, but how is it going to weave into your daily existence as if it's not a chatbot, but more as a sort of ethereal entity that exists? Obviously movies have personified this, and there's been Her and whatnot, but it is going to be fascinating to see how it weaves into your daily life, whether it's your home or work, and in a very ambient state. I know it's not about listening to you all the time, potentially, or it's not about, but I think there's a lot of these explorations that have yet to be done at the interface layers. How does AI talk to you first? Today, that's a push notification, I think, roughly. Is that it?

18:05

I don't know. How does Apple integrate AI into iOS and weave it into the operating system? And how do we use applications or specific types of things? Are they even necessary anymore? Do we even need an interface if we're just talking to it? I don't know. I think those are really interesting questions. I personally like the ambient AI layer. I think you're seeing a little bit of this with Open Claude and whatnot, and agents working in the background and surfacing the right things at the right time. There was a great product a while ago that didn't work called Google Now. The whole purpose of Google Now was to kind of predict a search, right?

18:39

It's like, what are you going to search for next, Eric, in some sense, right? Yeah. And it was ahead of its time in some sense. But when you marry it with context and intelligence, I think that is actually a huge factor for how to think about what the future of AI and how the stuff will weave into our lives. And I don't think there's going to be a single person who doesn't use this stuff. It's just a matter of when. Yeah. Right. That's going to be interesting. I don't know. Quick story. I remember asking Balaji, I was like, Balaji, how do you know so much about crypto and economics and bio and math and science and all these things? Give me all the books you read.

19:14

He's like, books? I don't read books. I just get in fights with people on the internet, and that's how I internalize all the information. It's real time. I need to know. I learn what I need to know to win the argument. The coolest thing. And then he really remembers it. Conceptually, those are really fun ways to have these discussions. Obviously, some of it's not kosher to possibly say or do out loud, but I think that's actually really cool. Learning from other people is what we do best. Like monkeys or apes watching other people, they use tools because they learn how to use them. That's wonderful. Imagine if I'm using a tool wrong, like a hammer backwards or whatever.

20:12

Somebody's like, no, you're an idiot, this is how you use it. Wonderful. Now I've benefited. Maybe that other person got a dopamine hit because they proved me wrong. And I think in the end, we all win. It's great. And I think that's how I treat my account in some ways. It's almost as if I'm not really trying to gain anything. It's just that I don't have anything to lose. What do I lose by being wrong? So I have a question for you. There's a study that came out a few weeks ago that generated a bunch of conversation, which is that in China, AI is highly popular. In the US, AI is very unpopular. In fact, it's even less popular than ICE right now. Okay.

20:51

The NPS of AI is not great in this country. How would you fix that? I always think about movements, right? When people create movements, all the movements are rooted in simple storytelling. In some ways, we're in fear-driven development. There's a lot of fear that's being generated as a result of this. And I think there's a positive framing to all of this. Look, I think we're moving towards a world where hopefully there's highly abundant elements of everything. Right now, we all feel like we're fighting for resources, right? Whether it's capital, labor, whatever. People think of the world as a finite amount of things.

21:29

A lot of people have this, in Silicon Valley, we have this classic thing where it's like, everything is growing the pie. Yeah, positive sum. Positive sum. Everything is positive sum. Normal people don't really think about that. They don't really think about growing the pie. At least, I don't know, I'd love to get your perspective on this, which is like, we're fortunately, we are... in some sense, primitive in that we are competing for resources, competing for finite things and whatnot. But I think the framing has to be around hopefully, if we do our jobs well, we're moving towards a world that's highly abundant in everything that humans might actually need.

21:47

Right now, we all feel like we're fighting for resources, right? Whether it's capital, labor, whatever. People think of the world as a finite amount of things. A lot of people have this, in Silicon Valley, we have this classic thing where it's, everything is growing the pie? Yeah, positive sum. Positive sum. Everything is positive sum. Normal people don't really think about that. They don't really think about growing the pie. At least, I don't know, I'd love to get your perspective on this, which is, fortunately, we are... in some sense, primitive. And that we are competing for resources, competing for finite things and whatnot.

22:18

But I think the framing has to be around, hopefully we are, if we do our jobs well, we're moving towards a world that's highly abundant in everything that humans might actually need. I think that the number one way you change the NPS of AI is you make important things cheap quickly. Soon, okay? And we've all seen the famous chart that Mark has tweeted a thousand times, right? Which is the diffusion of products, prices on a per product category basis, right? So this is the famous one where it's 1970. Everything is essentially referenced to that date. And then certain things get more expensive. Certain things get cheaper.

22:42

The number one thing that gets cheaper is flat screen TVs. So flat screen TVs are asymptoting to essentially $0. Yep. The things that are getting expensive are healthcare, education, and housing. Okay? There's actually a little bit of math, and I did this math a few months ago, that you can do to show how you can make education and healthcare cheaper with AI very, very quickly. And by cheaper, I don't mean disinflation, which is a reduced rate of inflation. I mean actual deflation, cheaper than it was last year. Okay? So here's the math. Just consider it for a moment. Education is actually the easiest one.

23:21

Education, if you restore student-administrator ratios to what they were 10 years ago, and you make professors modestly more productive, modestly, then you can actually just have education and school getting cheaper every year. The explosion of administrators, not professors, not teachers, but administrators, is totally under-discussed, and it's insane. So you can make education cheaper. We could do it right away. We already have all the technology. We just have to make a different set of choices. For healthcare, 45% of healthcare cost is administration. Okay? It's all this overhead. And if you've done the healthcare thing, we've all done it, right?

24:03

The revenue cycle management, all the back office stuff, all the nurses phoning you to tell you what drugs to take the night before you get a procedure. All of that stuff is overhead, and all that adds to cost. If you can take a bunch of the cost out of that with models, and by the way, these are the numbers by category, the number one consumer of OpenAI models, for example, are healthcare companies and healthcare startups. You can make healthcare cheaper year over year. So I think our moonshot as an industry should be to make these two things way cheaper in the next five years, and that's how we're going to win the hearts and minds. Would you subsidize it?

24:42

What do you mean? As in, effectively, should the model companies give it away for free to these industries? Maybe. Yeah, maybe. I don't know how you subsidize it. Maybe at cost. Maybe at cost or whatever. Or something like that. But it's very interesting because actually, and Dixon said this a while ago, which really got me thinking, and he's like, how many of our problems in society are actually intelligence-bound versus being collective action problems? And that's why the third category I mentioned, healthcare, education, housing. Housing has nothing to do with intelligence or technology. It's entirely collective action. Totally.

25:24

We could just build skyscrapers in Marin tomorrow, and it would be abundant, cheap housing for everybody, but we have to decide to do that together. I wonder if giving stuff away or making it free, I wonder if people will realize it. I think generally the world has gotten cheaper and cheaper and cheaper. You can go to Walmart and buy a hairdryer for five bucks. That's ridiculous. I remember, not to do this again, but I tweeted about this. Of course I remember you. Let's go to the book. I'm an encyclopedia of signal. This guy's like, I swear to you, to Beth. I remember. No, it's like the billionaires drink the same Coke as you are. They're using the same goddamn iPhone.

26:09

They're using Claude and ChatGPT just like you are. Yep. And the underlying essence of equality or the access is pretty much incredibly similar. I mean. But they don't have the same healthcare you have, and we should fix that. That's true. If you get sick as a billionaire versus a normal person, what's the delta? Dude, you'd be surprised. I mean, look at New York State right now, right? Your beloved New York City. Assuming you have insurance, right? That's probably the... In New York State, they're about to make it illegal at the state level to give or receive health advice or financial advice via a model. Oh, my God. How fucked up is that, right? So what does that mean?

26:48

People who have lawyers and doctors already are going to be unaffected, and people who use the models for lawyers and doctors are once again set back enormously. How can we be okay with that? So a lot of these are own goals. It's crazy. If you look, the state of Massachusetts made it illegal to buy Apple stock because it was too speculative when Apple was going public. It was the early 80s. The home of Elizabeth Warren. It's like we must protect the consumers from these enormous financial gains. Like the number of ridiculous things that are done in the name of protection. And I think that is a fundamental underestimation of the average consumer. Yeah. Right?

27:17

I think people are pretty smart, pretty savvy. They talk. They'll figure things out. And if you don't prevent them from accessing the tools, they'll use those tools to make their lives better. I have a weird idea. Tell me. The fact that we don't allow normal people to have equity share or stakes in OpenAI and Claude. It's insane. Imagine if normal people were like, I own a piece of these things. Maybe that would feel much better. And they would have this ownership mentality. And right now, all this concentration is happening in Silicon Valley and a few people. Yeah. Give people access to create a sense of ownership. Yeah. Earlier.

28:01

And therefore, imagine if a billion people had stock in OpenAI in some way, shape, or form. Maybe that's a dumb idea. But would they be more bought in on AI? Would they have a positive view of AI? Or what if their kids did? What if you rolled it into the Trump accounts? And it's like, look, I have my job and I think I'll be okay. And now I know that my kids will be okay, too. Yeah. They have a stake in the future. That's right. Quite literally. You can market it that way. That, to me, has been a very weird development where I do think people perceive tech individuals as hoarding or concentrating resources and wealth. For example, we're all privileged in technology.

28:41

And it's wonderful. And a lot of us are exposed to this, whether it's via equity or whatnot or even just usage. But I think there's potentially a perception with all the power outcomes that happen that there is a concentration or a hoarding, if you want to use a negative terminology like that. Yeah. And that creates a weird dynamic. Right. Like, I'm going to get left behind. Yep. While the guys in San Francisco are going to be enormously wealthy. And that's probably one. I've heard people feeling this way. And especially, there's potentially a negative sentiment in technology. And the NPS score probably reflects that, right? And that is not.

29:17

I think that's something to be fixed. And ownership might fix that. But any, I mean, I love, I think the power law dynamic that Peter introduced is really fascinating to me because it was business outcomes. Yeah. And the internet drives power law outcomes in a variety of different other scenarios as well. Not just business. And I think people are starting to catch up and people are seeing this wealth discrepancy that exists. And technology is a crazy accelerator. We talked about this right at the beginning. Yeah. And this entity is accelerating returns. I think that's maybe something to look at.

29:46

I don't know if the right answer or whatever, but I did see that tweet about the concentration of the inaccessible private, the companies are staying private longer. Yeah. What does that really mean? That's crazy. Right. And the NPS score probably reflects that, right? And that is not. I think that's something to be fixed. And ownership might fix that. But, I love, I think there's this: the power law dynamic that Peter introduced is really fascinating to me because it was tied to business outcomes. Yeah. And the internet drives power law outcomes in a variety of different scenarios as well. Not just business.

30:08

And I think people are starting to catch up and people are seeing this wealth discrepancy that exists. And technology is a crazy accelerator. As we talked about this right at the beginning. Yeah. And this entity is accelerating returns. I think that's maybe something to look at. I don't know if the right answer or whatever, but I did see that tweet about the concentration of inaccessible private companies, with the companies staying private longer. Yeah. What does that really mean? That's crazy. Right. Maybe we should have a law that says you have to go public at some point before XYZ. We somewhat do have that with the way RSCs are structured and things like that.

30:51

Anything you want to tease or how do you want to? Oh, that's a great question. We are building a fun little consumer product and I'm excited to storytell on this. It's a little bit different. We're three people having fun building a fun little interface of consumer AI and what we think might be really interesting for average, normal people to experience and use. And it works out of the box. So I'm very excited for that. That's been really fun. It's what we've been up to. One of the fundamental things that I just don't want to talk the talk. I want to walk the walk. And I'm excited to be able to share what we're up to. It's small. It's fun. It's interesting.

31:35

We're going to, we're going to, we're going to see how well it lands. discrepancy that exists. And technology is a crazy accelerator. Like we talked about this right at the beginning. Yeah. And this entity is accelerating and returns. I think that's maybe something to look at. I don't know if the right answer or whatever, but I did see that tweet about the concentration of like the inaccessible private, the companies are staying private longer. Yeah. What does that really mean? That's crazy. Right. Maybe we should have a law that says you have to go public at some point before XYZ. We somewhat do have that with the way RSCs are structured and things like that.

32:17

Anything you want to tease or how do you want to? Oh, that's a great question. We are building a fun little consumer product and I'm excited to kind of storytell on this. It's like a little bit different. I mean, we're three people having fun building a fun little, you know, interfaces of consumer AI and what we think might be really interesting for average, normal people to experience and use. And it works out of the box. So I'm very excited for that. That's been really fun. It's what we've been up to. One of the fundamental things that I just don't want to talk the talk. I want to walk the walk. And I'm excited to be able to share what we're up to. I mean, it's small.

32:56

It's fun. It's interesting. We're, and, you know, we're, we're going to, we're going to, we're going to see how well it lands.

Reading tools

Type to find a passage

Appearance
Ask this transcript

Notebook

Export Markdown

Add a note