AI重置已至：搜索、工作，以及其他一切——与 Anish Acharya、Dave Blundin、Salim Ismail 对谈

This was a big week in AI. I think we've had convergent AI announcements happening. Every day is a miracle, and that's not overstating the case. Google has spent 20 years making commitments to an ads ecosystem. It's very difficult for them to break those commitments, and I think it's a real threat to their search monopoly.

The cooler this is and the more search moves over to it, the more it cannibalizes the core. Already my kids are telling me, “Dad, everybody knows that ChatGPT is better than Google.” 2029 to 2031: the end of white-collar work. That's only a couple of years for people to remap their entire career path.

We're going from a world where our brains were wired for fear and scarcity to a world that's very different. Technology is a substrate to make us more abundant and happier across every aspect of our lives. Now that's a moonshot.

Salim, by the way, happy birthday this past Saturday.

I sat in the garden with a bottle of wine and a glazed look in my eyes. Good birthday.

That's a lovely day. We have a special guest today, Anish Acharya from Andreessen Horowitz. Anish, good to have you joining us.

Thank you, Peter, Salim, and Dave. It's great to be here.

For those who don't know, Anish heads the consumer investment portfolio at Andreessen Horowitz, one of the most extraordinary VC funds on the planet. He's a general partner there. What I love about Anish's portfolio and his vision is that you're running the abundance thematic throughout a16z, which is one I, of course, love.

Yes, we are. You can't talk about consumer tech these days without talking about abundance. I was telling Peter prior to the show that we may or may not have been inspired by his thoughts on abundance, but either way, it's abundance of abundance.

Exactly right. Now, let's dive in. This has been an incredible week for AI, but before we get there, Anish, I want to talk about how you, your leadership, and your partners at Andreessen think about abundance, because it is a real thing. My next book coming out is called Age of Abundance: How to Survive and Thrive in the Decades Ahead.

We're going from a world where our brains were wired for fear and scarcity to a world that's very different. Here's a slide from your deck, from some materials we stole from your website. Would you mind just giving us a little insight into how you think about the abundance agenda?

Yeah, Peter. Maybe to zoom all the way out—and I'm going to touch on topics I know you've touched on as well, so the audience will be familiar—in our belief, the 2 greatest catalysts for human flourishing are market economies and technology. Over and over again through the arc of human history, we've seen these 2 things deliver extraordinary results for human flourishing.

The story of the last 100 years—the industrial age and everything that is coming after it, the technology age—has really been a story of technology and innovation. Our belief is that the more significant the technology and the technological change, the more significant the results will be for consumers in terms of flourishing and, of course, abundance.

We think about abundance in a lot of ways. I love the definition that you used, Peter, which is that it's not about luxuries; it's about possibilities. That's exactly the thrust with which we've been exploring it.

There are a couple of areas in which the technology and the concept of abundance apply, but the one thing I'd give the group for framing is that I think of this as the most human technology we've ever built. If you look at technology for the last 40 years, it's really extended our intellects. Steve Jobs famously said it's a bicycle for the mind, but when he said “a bicycle for the mind,” he really meant a bicycle for the intellect.

That's what a spreadsheet is. For me, a spreadsheet is symbolic of all the technology we've built for the last 40 years. It allows us to do this extraordinary math and computation that we simply couldn't do before, and of course it has a ton of implications for human society. The Fed, of course, and all of these other systems that we built rely on all these technologies.

However, we haven't done that much for our souls, our emotions, or our mindset. With AI, we're able to explore the side of humanity that is defined by subjective emotional experience. We've just never had a technology that could be brought to bear on it.

As we talk through all of this, I would love for folks to keep in mind that we're really taking that left-brain, right-brain pairing that was missing from technology for the last four years. Does that make sense?

Yeah. I had an incredible experience driving from Vermont back to Boston with my mom, who's in her mid-80s. I asked her, “Have you ever talked to AI before?” She said, “I don't even know what you're talking about.”

So I put on ChatGPT's voice mode and put it on the car stereo. This beautiful, sweet voice came on, crystal clear, understood every word she said, and she started asking it about her hometown, where she grew up in Ohio, and whatever happened to the cheese factory that her father owned.

She's like, “Who am I talking to?” I said, “You're talking to AI.” She said, “Well, it must be somebody's recorded voice.” I said, “No, it's completely synthetic.” She said, “That can't be.”

In this abundance slide, abundance has always meant including everybody. The internet and its capabilities have largely bypassed people who are over a certain age, but also the vast majority of the really engaging youth, dominated by young boys playing video games. That demographic shift—I know, Anish, you know all about this—I would love to get your thoughts on how this opens up so many new capabilities across users who were not previously users.

I love that point. You know, Dave, it's actually so interesting because typically, when a new technology is introduced, it becomes very hard to grok for the generation that didn't grow up with it. They fumble around, and they never achieve the full potential of it.

With a lot of AI, I think seniors are going to have the experience that your mother had, and they're going to benefit from it disproportionately because they're able to interact with technology now in these unstructured ways, a lot of it via voice.

I'll give you a great example. We've got a portfolio company that is an AI nurse. Now, an AI nurse can't take your blood, so it can only do a subset of nurse-related tasks. But it's a voice nurse that will phone patients the night before a surgery and help them prepare mentally, as well as go through the checklist. It'll phone them after a surgery, make sure that folks are taking their medicine, and the impact on health from having that AI nurse take all those actions is dramatic.

Because it's over the phone and via voice, a lot of senior citizens aren't intimidated by it, so they really stand to disproportionately benefit. When we talk about companionship and loneliness, this is also an area that's very exciting.

If you look at the last 20 years of technology as applied to relationships, it's been social media. We can have a really rigorous conversation about social media, but when you look at AI and the impact on human relationships, it feels like it gives people an opportunity to explore aspects of human relationships in a depth that may not be available to them in their real-world friendships and family relationships.

There's something there that's opening Pandora's box. I know because, in his latest book and all of his recent research, Peter shows that so much of human health and happiness comes from these little things that you eat or do. He's documenting it all now, and you're like, “Oh, my God, the amount of benefit from just basic behavior change.”

There's another part, which is that some of human happiness comes from setting a goal and a challenge and overcoming it. The question becomes: when these abundance technologies are overcoming the challenges for you, what happens?

I had an interesting experience the other day. I was in an escape room with a group of friends. It was a pharaoh's tomb, and we were sitting there having to solve these math equations on the wall with different symbols. I was hampered by not having a piece of paper and a pen, which was fundamental technology.

You start to realize how few things you can actually hold in memory during the course of that. I was so tempted to pull out my phone, take images, and ask ChatGPT for the answer. It started to make me realize that there's a slippery slope in which we become so dependent on AI that it takes away the challenges from us, unless we hold that part of the human spirit in place. It is a double-edged sword in that way. Do you think about that?

Well, I think we've been doing that forever, right? We moved from the slide rule to the calculator, and people said, “That's a terrible idea.” Then we moved from the calculator to the spreadsheet, and people said, “That's a terrible idea.” We keep moving the goalposts on this.

If you're a software developer, I used to program in assembly, and then we moved to 3GLs like Pascal and C.

And people said, “Well, you’re losing the benefit of knowing exactly what’s happening.” I think we just keep moving the challenge along, right? I think we just shift the goalposts and change the dynamic of what’s going on. But it’s adding much more capability.

If I think about what it takes for somebody to compose complex music today, it’s 1,000 times easier than 20 years ago, and you just get that much more music. I think that’s what feeds into the abundance thing. We just can create so much more. But is it so much more crap, or is it so much more? What’s the filtering system? Crap for one is gold for the other, maybe.

Funny enough, I remember growing up as an engineer, hearing, “Hey, if engineers don’t know how to do memory management, then they’re not real engineers.” So, you’re right, this has happened over and over again. Look, with that said, I agree with you, Peter, that there’s a level of agreeableness that is too much.

I think we need these models and these AI technologies to also explore the sort of uncomfortable aspects of the human experience, which are disagreement, persuasion, and sexuality. I don’t want to jump ahead, but this is one of the reasons I think the incumbents have struggled so much with it, because there are 1,000 committees working at Apple and Google that are explicitly designed to take the humanity out of their products, and these are fundamentally human technologies.

I also agree with your underlying point, which is that every person needs to feel like they have meaningful purpose. Even if it’s created for them, if it’s a little bit synthetic, without that, the flourishing point starts to get impacted.

There’s another quick point I want to hit with both of you, Anish and Salim, which is that at the endpoint of continued increasing abundance comes a post-capitalist society where money has little to no meaning. I’m not sure how a venture capital firm thinks and deals with that, but, A, I think we’re far enough away that we don’t have to worry about that right away.

One thing that I’m encouraged by with all of this is that we may break through the whole Douglas Adams framing. In The Hitchhiker’s Guide to the Galaxy, he said, “Anything that’s invented when you’re born, or that’s in the world when you’re born, we call that normal. Anything that’s invented when you’re young, that’s called a career. And anything invented after you’re 35 years old is just bad for the world,” right? You talk to any banker about Bitcoin and they freak out and whatever.

I think that these technologies, as we humanize them, make it easy for 80-year-olds to interact decently with technology in a very humane way. I think that opens up abundance again. New dimensions open up. I think it’s powerful as hell.

Yeah. We always say internally that human relationships are fundamental to the experience of flourishing. Maybe the human part is overstated. Maybe it’s just relationships. As long as we feel the feelings that result from the conversations, who cares who’s on the other side?

One of the questions I have—and I’m just putting forward the second slide here—which is market opportunities: As we get to—let’s forget about AGI. Let’s skip ahead to ASI, artificial superintelligence—the question, as you know, is that Dave and I and yourself are all VCs. We’re finding, incubating, and supporting incredible entrepreneurs and startup companies, but the big question is: What moats are going to continue as we go forward?

How do you differentiate yourself and prevent yourself from being disintermediated by the next entrepreneur with a faster set of agent-enabled systems? How do you think about that all the time?

Yeah. I think there are 2 answers to it. Let me give you maybe a cute answer, and then I’ll give you a specific answer.

My cute answer is that abundance means an abundance of categories as well. There will be many new categories and areas for consumer spend and business spend. What happens when new technologies are introduced is that incumbents often get better at what they do today. So, I think Microsoft will make a better word processor and Google might make a better search engine, but search engines and word processors will be less relevant, and there will be new categories that pop up where the new entrants dominate.

On your specific question, you guys understand the technology at a fine level. You understand that these systems are very good at predicting static systems. They’re very good at averaging the training data and telling you what the training data implies. They’re not very good at predicting adaptive systems, like the stock market or even culture and music.

As a thought experiment, if you trained an AI model with all the music right up to hip-hop, but not including hip-hop, would it infer—would it imply—hip-hop? I don’t think so, because culture and music are these sort of adaptive systems that work together.

I really do think that there are moats that are based on adaptive systems, and network effects are a great example. They’re actually as good as gold, and they always have been. Moats that are based on static systems, like integration moats and systems of record, I think those are really at risk.

Interesting. Do you want to walk us through this next slide here on market opportunity?

Yeah. Consumer investing is very interesting because it’s hypercyclical. When consumer works, you get the biggest companies in the world, as you can see from the slide here. When consumer is not working, it’s really not working.

We’re now in a product cycle that’s as important as the internet. I think it’s probably more significant than mobile. The biggest winners, I believe, are going to be consumer winners, and every consumer behavior, including some that don’t exist today, is up for grabs.

All of which is to say, it’s a great time to be a builder or be around builders, as all of us are.

Dave, what do you think about that?

Well, I totally agree. I think a lot about the fact that Midjourney, actually, I think, is an a16z darling, where the cash flow gets into hundreds of millions of dollars of very high-margin revenue. Because coding is getting so cheap and so automated, the cost of a build of something like that is lower than ever.

You’ve got very rapid growth of the revenue, very low capital costs, and nothing to lose by jumping in there. Now, the question always comes up: What’s your moat? Is it going to be defensible over the long term? But you’re so profitable so quickly while you explore that, there’s no downside.

We know now that management teams have to pivot over time. That’s just the nature of tech going forward. It always changes continuously, right? So, they have to learn that skill anyway. Why not learn it while growing like crazy and being profitable?

I like the idea of not being worried about it. But it is still, I think, worth exploring a couple of fundamental questions while we have a niche, which is, if you look at consumer as a whole, a lot of the frameworks are defined by the big guys.

The App Store is really not a fact of nature. It’s defined by what Apple and Google decide: Here’s a framework that you can operate in. If you go way back in time, in the early days of Apple and Microsoft, the ISV market—independent software vendors—was also defined.

Then Microsoft changed its mind one day and said, “You know what? Spreadsheets and word processors, those are ours now. Sorry, you set up your camp there, Lotus, but we’re going to just take that back.”

You do have to be conscious of the fact that if you’re on top of a big LLM, you’re on top of a heavily funded company. You have to predict what they will and won’t do inside their core $200 or $240-a-month service offering and be outside of that, but not too far outside.

I don’t know how you think about that.

It’s a great point, and actually, I worried a lot about that when we were in the early days post-November 2022, and it felt like OpenAI was the only foundation-model game in town. Because in that world, OpenAI could just raise prices and take 100% of the economics that are downstream from them.

Now, look, we’ve got a bunch of foundation-model companies that have great models. We’ve got a bunch of open-source models that are super competitive. Because of that, you see OpenAI and other companies trying to move up the stack.

They bought Windsurf, which is really interesting, and as an application developer, you’re not dependent on any single platform. I mean, if you’re an iOS app developer, there’s only one game in town: the Apple App Store.

The web is not that way, and the AI game is not that way either because of the presence of multiple foundation models. On your point, Dave, if I may, on Midjourney as well, I think there’s an interesting note there. You talk about defensibility, but what’s happened in that market is that it’s fragmented. Now you see Midjourney as a really interesting player that points in a specific aesthetic direction. It’s image generation, and it creates these beautiful, hyperrealistic images, but it has a very specific aesthetic.

If you’re a designer looking for a different aesthetic or more controllability, you’ll work with a company like Krea or Ideogram that has a whole different aesthetic. As a result, you’ve got 2 companies that both do image generation that are pointed in different directions. There’s a broader comment here, I think, which is that when we have new technology, these markets are expansive in the same way the universe is. Companies tend to move away from each other over time, not toward each other.

Yeah, I really want to riff on that while we have you, because it’s really clear to me that the number of opportunities way outstrips the number of teams. A lot of people are intimidated, and they’re like, “Isn’t Midjourney or OpenAI going to do exactly this, this, this, and this?” But you’re pointing out, I think, a really critical and inspiring point. The aesthetic difference alone creates a new market, but I think you were talking about digital makeup on one of your podcasts, just as a category. There’s something that’s actually a company, a product, that’s actually defensible. It’s amazing.

And I think the other amazing thing, Dave, is if you look at the price points that these products are commanding, Spotify’s most expensive plan—the family plan with lossless audio—is $20 a month. It doesn’t get better than that for Spotify. ChatGPT is $200 a month. Google just announced a $250-a-month plan that’s consumer- and prosumer-facing.

Not bad.

A lot of companies are overlooking the fact that when you go live with one of these direct-to-consumer products and it hits, it’s global instantaneously. If you think about a $10- or $20-a-month subscription, but you get 30, 40, or 50 million people, that’s a small fraction of 8 billion. But that can happen very easily within a lot of different categories. So, again, just so many opportunities relative to the number of teams.

Wait, I need to drill into something: digital makeup.

Yeah. Do you mean, like, on Zoom or something, where you can be anyone you want to be in the AI world, Salim? We should debrief later.

Well, I choose to be a bald, gleaming-headed fellow.

As do I.

Wow. All right. So, here we have your third slide before we jump into our AI universe, which has been exploding this week. AI use cases, please.

Great. So, maybe I’ll touch on each of these in turn very quickly: creativity and productivity. The thing about creativity that’s so interesting is that we all grow up believing we’re creative, right? We all drew pictures and colored pictures when we were 3, 4, and 5. Then we get to a point in our lives where we start to self-select into being good at it or bad at it.

We’re talking about the technical skill when we say that, not the inspiration behind it. With AI, the technical skill of being creative gets separated from the inspiration behind being creative. If there’s something that you can dream of making, whether it’s music, art, or video, you can make it now, which I think is incredibly abundant from a consumer-impact perspective.

We talked about companionship and social experience. The experience that you had, Dave, in the car and the car ride—this is bringing empathetic, patient, maybe disagreeable human relationships to everyone that wants one, which has enormous implications from a flourishing perspective. Finally, wellness and personal growth. I think so much of this, if you look at even something like finance—I worked in fintech and financial services—you quickly realize that fintech is not about helping people make rational choices. It’s about exploring the nonrational parts of our relationship with money, and AI is uniquely suited to help us get better around that. This technology is a substrate to make us more abundant and happier across every aspect of our lives.

I’d love to get your thoughts. There’s so much on this slide because there are so many different facets. Historically, when you looked at a founding team, you were looking for your Steve Jobs visionary married to your Steve Wozniak engineering genius. But now that engineering component is largely AI-automatable, and the vision of the perfect founding team looks different. Being able to navigate through so many choices on this slide alone and come up with the perfect strategy—that’s got to be the rare commodity, I assume, right?

I’d have to disagree. We’re seeing more technical founders be more successful over and over again. Yes, AI extends their capabilities and makes them more productive, but there’s so much work at the edge that AI is not going to do. So, we’re really seeing a rise in the dominance of an engineering-oriented founding team in a way that we haven’t in 10 or 15 years in core tech.

Well, the Fred Wilson rule always said, “Hey, we’re looking for 3 or more founders, best friends, who write the code themselves.” So, when you say technical founders, are you talking about former engineers who wrote code at Google, very similar to yourself?

I’ll tell you, one of our most fun investments is a company called Krea, which is a really cutting-edge research and consumer technology company that brings together all of the creative tools in 1 product: image generation, video generation, image enhancement, and so on. These are 2 incredible AI researchers and artists and enthusiasts. They live in a house up in Pacific Heights. They live with all their engineers, and they work 7 days a week.

The first board meeting, they sat me down and said, “We’ve got a problem.” I asked, “What is it?” They said, “Our house only has 10 bedrooms. What happens when we get our 11th employee?” I said, “Well, maybe we shouldn’t all be living together at scale, but that’s a separate conversation.” That’s the intensity and how technical they are in their leadership, which is a beautiful thing. Honestly, the most success and the most fun I’ve ever had as an entrepreneur is when I’m living that monomaniacal, singular-focus effort in life.

I’ve got to tell you, we bought an apartment building right on the edge of MIT and Harvard’s campus. It’s actually the closest building to MIT that wasn’t already owned by MIT. We bought it a couple of weeks ago. It has 24 beds in it, 6 units, for exactly this reason.

Amazing. Amazing. So, if you’re a product visionary who wants to use AI in a totally different domain, then surely you can acquire the technology capability at low cost. You don’t have to be the technical founder in that sense. Is that not an option that you’re seeing?

I think it absolutely is. It’s a question of whether you want to be at the edge of technology and new models, in which case you probably do need to be. There are plenty of off-the-shelf tools and products like Cursor that make it a lot easier for somebody who’s even familiar in a cursory way—no pun intended—to bring products to market. So, all of the above.

I would say that when the companies that are thriving right now in your portfolio started their journey just a year or 2 ago, SWE-bench was maybe 10%, and now it’s suddenly 60%. If you look forward a year, at the rate that that’s changing, you would have to assume that, to some degree, orchestrating the AI agents becomes dominant.

Right now, if you’re going to build something on top of HeyGen or on something on-prem, you’re going to be mostly coding it up with Cursor. It helps you, but you’re still coding it up. There’s some kind of paradigm shift coming. You could debate whether it’s 3 months from now, a year from now, or 2 years from now, but it’s coming.

I agree. I completely agree. What percentage of your companies in the a16z portfolio are AI? And how many would you put in the bucket of physical robotics, biotech, nanotech, or 3D printing—non-AI exponential tech?

I’m focused on consumer software. I think if you look at that theme as a substrate across the firm, many of our investments—many of our American Dynamism investments—are also very much in that vein. In terms of software, largely all of our investments are direct AI companies today. There are 1 or 2 that aren’t, but that’s very much part of the strategy and on the come.

It just begs the question: If you’re focused on delivering abundant outcomes and you’re not thinking about AI, why? How can you be?

This was a big week in AI. I think we've had convergent AI announcements happening. I am curious why everybody's announcing on top of each other, but let's take a look.

Today, May 20, and tomorrow, it's Google I/O. Google unveils Gemini updates and Android ecosystem bets. On the 22nd, Anthropic is debuting Code with Claude 2025, its first developer conference. Also this week, Microsoft Build 2025 focuses on Copilot, scale-out AI infrastructure, and developer tooling. Of course, we've got NVIDIA making announcements, and we'll talk about Elon’s Grok 2.5 announcements.

I'm in L.A., Salim's in New York, Dave's in Boston, and you're in the Bay Area. What's the feeling like right now where you are?

The level of acceleration every day is a miracle, and that's not overstating the case. Just this Friday, OpenAI released Codex, which is an autonomous software agent that simply writes pull requests for you to review, by the way, on your phone if you'd like.

Every single day, any one of these things could be the basis of an entire ecosystem. You're showing 3 for this week. It's crazy. We talked about Codex last week.

Here's a fun article, Dave. Well, actually, maybe I should say, give it to our brethren of Indian descent. But Dave, you want to set this one up? India plans to make chips by 2025 and its own GPUs in 3 to 5 years. You picked this slide.

Yeah, this was really important to talk about, coming on the heels of Saudi Arabia, which we talked about a couple of days ago. Every country that wants to be competitive in the future needs to have some kind of AI strategy. That boils down right now to having its own supply of chips, because the chips are going to be unbelievably constrained for at least the next 3 or 4 years—maybe forever into the future.

Some of these more consumer-facing use cases that involve voices and now multimodal imagery will use an entire GPU, 2 GPUs, or 4 GPUs concurrently to get the best possible user experience. That's not super expensive, so it's easily worth it for the consumer. But the chips don't exist. They physically don't exist.

We're going to make 20 million new GPUs this year. It's nowhere near enough to keep up with just basic call-center use cases. Countries that have their act together are starting to think, "Do we need our own fabs?" India is saying, "If we're going to compete, we definitely need fabs," starting with 14 nanometers. But you've got a long way to go from there.

Okay, great. This is exactly the right thing to do. But then, in their own internal research document and plan, it says, "We expect this to be underutilized and bureaucratic." Like, holy crap, there are challenges trying to get a sovereign strategy together.

We'll have to keep a close eye on it, but there are probably on the order of 50, maybe 100, other countries that need to immediately get on the tails of this same exact thing. We're advising one of the big Southeast Asian countries on exactly this, and we're basically saying, "You have to develop your own fabrication capabilities. There's just no other way around it."

The learning that will come from that will be very powerful, one way or the other, if only to select who the best supplier is going forward. We're going to end up with an abundance and different architectures coming from different places that will all pull together.

India specifically—the bureaucracy is insane. There are so many overlapping federal- and state-level systems, and so on. I always talk to people and say, "Don't think of India as a country. It's more like Europe, with 20 different major languages, different tensions, cultures, and so on." You have to look at it from that perspective, and then it makes a little bit more sense.

Yeah. One thing also: NVIDIA now, as of right now, is worth over twice as much as Meta and almost twice as much as Google. You're like, "Well, how can that be?" Especially Google, where the transformer was invented, Google Cloud is huge, and they have their own TPU v7s, which are incredible. Does this make any sense? That's debatable.

The chip demand is such a dominant factor, and underneath that, the fab shortage is such a dominant factor. It's not a secret. It's amazing to me how many people don't know this, but Elon Musk is video-podcasting it out: We absolutely need to accelerate our fab production.

I was talking to Kaveh Kazerani over at Allen & Company, and they're looking at these new $4 billion fabs. Normally, a fab is a $20 billion to $40 billion investment, but there are some new designs that are more around the $4 billion mark that might actually unclog the machinery. Those are really interesting to study and track.

For India, the perfect scenario is, "Let's get some $4 billion fabs up and running, start on 14 nanometers, but quickly work our way down to 5 nanometers." So, how quickly do these obsolete themselves? If you're pumping out 14-nanometer GPUs, are they going to be useful and compete with the cutting edge at TSMC?

No, not at all. They're not even vaguely competitive. They'll all be fully sold out for a long time.

The fabs themselves are actually very sustainable. Most of the machinery in there, even when you upgrade, can be reused. The EUV component is an exception to that, but most of the rest of the pipeline you can reuse over and over again.

The key is just to get on the map, get something up and running, and then you can work it down as you move forward. You reduce the lithography as you go.

I do think that the other thing that's really affecting this is that if you have your fab act together, AI is getting really good at chip design. I was working on that a couple of weekends ago, and your cycle time can come way down.

There are going to be algorithmic breakthroughs all the time now, and those will affect the designs right away. You can automate that pipeline so that the new algorithm immediately gets a new chip design. Get that right into the queue almost in real time, and then you have to wait a month or 2 for something to come out the other side.

But if you get that machinery fully integrated—which no one's ever worked on before, because a microprocessor design would last for a full year—you never really thought about real-time design using AI automation. Now it's a completely different world, just how this pipeline should shake out.

Anish, any thoughts on this one?

Yeah, it's really interesting. Presuming we do resolve the shortage of chips, which I believe we will, there's also a sustained need for countries to build sovereign AI because AIs embed values within them. That's why DeepSeek is such an interesting conversation, because arguably DeepSeek is trained with a set of values that may be a mismatch with Western values.

Every country needs to think about what values it wants to imbue into its AIs. There's something very interesting here. I'm not sure if you guys are familiar with the Jones Act for ships that operate in and around U.S. ports. They must be manufactured in the United States for national security purposes.

I believe we'll see a sort of Jones Act for AI, where AI that operates in sensitive contexts will have to have been trained nationally.

That's a great insight. I don't think anyone's ever drawn that analogy before, but I love that insight.

It's really clear that if you don't have a national strategy for compute, during this era when the chips are constrained, the highest-value use cases are just going to buy out all the data centers. It would be very natural for 1 or 2 economies, like the U.S. and China, to have a higher standard of living and then say, "Well, because I can overbid the Indian or the Ethiopian, they don't get access to any compute."

That goes on for 1 year, 2 years, 3 years, 4 years, and that's the natural cycle if you don't have a national strategy to get compute for your citizens. Once you start getting behind, you're going to get way behind, and then you're economically unable to get back on the map.

Our next story here: we've been hearing about this forever, and it looks like we're on the verge of GPT-5 being released.

Here are some tweets that went out: "Just got to try an early GPT-5. Just wow. It can do anything I can do at a computer, but much quicker. We actually made it." This is from Chris at @gpt21. "GPT-5 is in red-teaming. This is not a guess. This has been confirmed."

When I think about GPT-5, I think about self-recursively improving AI software, recoding AI, and leading to an intelligence explosion. I think about Ph.D.-level or multi-Ph.D.-level capabilities. What do you think of when you hear GPT-5, Dave, Anish? What do you guys think about?

Well, actually, before we jump into it—there's so much to talk about there—but before we jump into it, you notice how they always jump on each other. Google I/O: "Oh, my gosh, I've got to get something right on top of them." This is not a coincidence. You pointed it out before, Peter, but this is not a coincidence.

And notice how Apple isn't even trying. They're just completely invisible.

Yeah, we'll talk about that. Apple has opted out.

What do you think about GPT-5?

We've seen so many exciting model releases since the GPT-5 conversation started, and we've seen a new model architecture with the reasoning models. I feel like we're seeing all the steps that point in the direction of GPT-5. I don't know that there's a big unveiling coming that will show us something that isn't implied by what we've seen so far.

If you look at the reasoning models, which are just as important an innovation as the language models, they are models trained through reinforcement learning in specific domains, which is why they work so well for coding. But we still need to go collect the data sets and build the models for all the domains that have a formal concept of correctness. Before we have something as broad-based as you're describing, Peter—which I'm sure we will have someday—there's an enormous amount of model work to do between here and there.

So, look, my belief is we'll see something that looks more like o4 or o3 Pro than a completely different animal.

I'm still waiting for the day my cell phone rings, I pick it up, and it goes, “Hi, Peter. This is GPT-5. I just wanted to introduce myself. I'm here if you need anything.” That will become a spooky future.

I want the model that, to your point, Peter, is Jarvis, right? It just goes, “Okay, what do you need right now?” It can go and make suggestions and get to that personal level that we're waiting for.

I interviewed Sam Altman at the MIT Media Lab a year and a half ago. I tried to get him to open up about parameter count and where parameter count is going, and he said, “Look, we need to stop masturbating over parameter count.” That kind of shut down the conversation. The crowd loved it—there were about 1,500 MIT students in the crowd.

But now I can't track just the raw parameter count. I'm hoping that with GPT-5, we can still distinguish between the model producing an answer that should be mind-blowingly intelligent and the chain-of-thought reasoning version. The chain-of-thought reasoning, like you said, Anish, adds immensely more of a feeling of intelligence than the core model does.

I think OpenAI wants to tie it all together and say, “Look, don't worry about it. It's just one subscription. Here's all the brilliance.” We don't want to talk about what came from where. But as an MIT-oriented researcher, I want to know what came from where, which I can do with Llama, but it's becoming increasingly difficult with GPT-5.

That being said, I think it's going to blow our minds. They wouldn't be putting it out if it wasn't mind-blowing. All the prior step functions have been mind-blowing. So that's one. Let me ask you a question: if you had to guess what capabilities are in GPT-5 versus GPT-4, what would be a standout feature?

Right. Everything is fully multimodal now, and the language model has been phenomenal even in GPT-4. But once you start marrying it to images, right now, when you prompt it to create an image or a video for you, or you show it an image, it's good and mind-blowing, but it's not perfect.

I think if you take your digital nurse example that I was talking about—“Hey, let me show you a rash on my foot”—it's going to diagnose it perfectly. You need a lot of long-tail images to do that kind of stuff, so I expect there will be a lot in that category, since they've had more time to work on the multimodal capabilities that just didn't exist when they built GPT-4 and GPT-4.5.

I keep thinking about Leopold Aschenbrenner's Situational Awareness: The Decade Ahead paper that came out a couple of years ago, showing that, on the heels of GPT-5, there would be just an acceleration of AI. We've started to see this. We've started to see the speed—if you want, even in human IQ points—at which these models are increasingly gaining capabilities.

Do we get an unconstrained intelligence explosion on the back of this? That's what, for me, is interesting.

The other thing I'm really expecting in consumer applications is that, in a podcast like this, if you don't like the way the audio recording comes out, there have always been tools to clean up audio. But now you have tools that will actually recreate it from scratch using your voice, so you can create anything out of anything.

Yes. When you start talking about multimodal video and you say, “Here's a static image,” it should be able to instantly turn it into a 3D, movable, any-angle scenario. I think it'll have that capability.

We have a lot of these capabilities. Krea today can do what you're describing, Dave, taking a static image and turning it into a sort of 3D splat.

In terms of how good the multimodal models are, I can tell you that when I'm barbecuing at the house, I'll have GPT-4o looking at my hot peppers as they're grilling, and I'll be saying, “Are they ready yet? Are they ready yet? Are they ready yet?” “No, no, no. Now, yes, they're ready.” I mean, that's what I did this weekend.

Step away from the AI. Just step away from the AI. I may have gone too deep, guys. I may not be able to find my way out.

If you want Jarvis, I think we have Jarvis today in the form of Operator. Operator, to me, is the most underappreciated new OpenAI launch. It's been around for a while, and it allows anybody to use the web.

Essentially, every UI becomes an API, and you can use the web as this sort of control surface to do anything. The implications of that are so significant. Just think of the mundane use cases as a consumer. Every day, I wake up and it checks for lower auto insurance and a better personal loan rate for me. It spends the day scouring the web and refinances all of my credit lines, and then I get a push notification at the end of the day saying, “Hey, Anish, I cleaned all this stuff up for you. You're going to save $200 a month.”

And please send my commission to me.

Yeah, that's right. That's all possible today. I'd argue we have more—just as you said, there's more ideas than teams, Dave. We have these magical, miraculous new capabilities that are underexplored because there are so many of them.

All right, let's head to Muskville. Elon Musk unveils Grok 3.5 AI that reasons from first principles. Let's hear it directly from his voice.

So really, the focus of Grok 3.5 is to find the fundamentals of physics and apply physics tools across all lines of reasoning, and to aspire to truth with minimal error. There are always going to be some mistakes, but the aim is to get to truth with acknowledged error and minimize that error over time.

I think that's actually extremely important for AI safety. My conclusion is the old maxim that honesty is the best policy.

So how do you guys feel about the basic concepts of Grok—of maximally truth-seeking AI? What I love is the idea of applied physics as part of first-principles thinking, as part of its RL structure. Anish, what does that sound like to you? Does that sound real? Inspiring?

No, it sounds consistent with what we're seeing from reasoning models. I think a reasoning model trained around physics would do what he's describing. I do think honesty is the best policy, and I think this is exactly the right step.

I really salute Elon for having led the way in a bunch of areas, including having the Grok language models allow you to interact with them in ways that the big companies never would. Those things may seem banal, like the adult sexy-chat features of Grok, but he's been pushing the edges, and I think he's going to do the same thing with reasoning models. This is an example of that.

Well, truth is definitely in the eye of the beholder and varies by country, and we learned that in Saudi this week in a big way. So the message is right on target, but then the definition and the implementation—there's no single answer to it.

Like Anish said earlier in the podcast, the LLMs we're using, which are transformer-based, are very good at interpolation, not quite as good at extrapolation. So it's going to fill in the blind spots between the training data that you give it and don't give it. I love the message of giving it first-principles physics, but I think everyone's going to do that.

There's nothing controversial about physics. It's when you start including some X data and not other X data, or decide whether to include all X data, that the devil's in the details. I'd be very curious to know, though: is Elon going to roll out versions, or is there just one Grok? Because then you're like, okay, Grok has this opinion of the world, Gemini has this other opinion, and ChatGPT has this other opinion—or are there 20 Groks, 20 GPTs, and you tune them? So, TBD.

The biggest challenge we've had even in the social media space is echo chambers. So the question becomes: are you able to take Grok 3.5 and build your own echo chamber, or will it say, “Peter, I'm sorry, your point of view is absolutely wrong, and here's the data about why it's wrong”? Is it going to challenge us in that way?

Elon is pushing boundaries all the time. Will Grok call him out, or President Trump, or Biden, or anybody else, on things that they say which are not defensible by specific evidence?

One thing I think about a lot—and this is right in Anish's wheelhouse—is how much the user can control, especially when it comes to copyrighted material, because a lot of the most interesting and fun things you can do as a user involve copyrighted material. But the foundation-model companies are under a lot of scrutiny, and they have to be very, very careful with copyrighted material.

If you put it in the hands of the user to decide what they want to create, what they want to do, and what voice they want on it, then it's outside of Grok's control. Then you can start using copyrighted material. It's easier for me to steal it than for Google or OpenAI to steal it.

Yeah, I think we have to be careful not to have a zero-sum conversation about that, though, Dave. If you look at, for example, sampling in hip-hop, it was very controversial. The genre wouldn't exist without it, and I would argue that it sort of drove more royalties and traffic back to the tracks that were sampled than would have happened otherwise.

So I do wonder if there isn't a positive-sum version of a lot of this.

I would love to get you and Bill Gross together talking about that, because that's his whole focus at ProRata AI. You guys could go for probably an hour on that topic. That would be really interesting.

Love to. So, here's our next article: “xAI deploys 170 Tesla Megapacks for Colossus 2: the power backbone.” We can see all those power packs in this image. What we're seeing is just an extraordinary building race, faster than ever before.

I was just reading some articles saying that regulation around getting access to power or building permits had been the constraint, but perhaps not now. Perhaps people are taking the constraints off—regulation, building permits, and so forth. Any comments on Colossus 2? How big is it going to be? Do we know how many GPUs it will have?

Yeah, I think it was 200,000 in the next wave, mostly H100s. The B100s are not ready yet.

I thought the first Colossus was upgrading to 200,000, and maybe Colossus 2 is going to be 1 million. But what's an order of magnitude between friends?

No, I was talking about the prior iteration. So this is what they're planning in the next iteration.

In the next iteration. Yeah, we'll see. I don't know. But I know the B100s are not in volume yet, so you're still using the prior iteration.

What's amazing about this is that Elon is immensely practical. The reason the Tesla packs are so important is that when you use NVIDIA chips for these massive-scale, single-training runs, the power spikes like crazy because they do these huge all-reduce, all-gather operations that are extremely communications-intensive. The GPU is idle while it's waiting to transfer data, and the power drops dramatically. It's fluctuating all over the place.

You can't use lithium to store anywhere near enough energy to power these things, but you can use it to smooth out the power flow. That's a pretty big advantage for Elon.

I found this next article on NVIDIA's breakthroughs. Let me read this out loud: “One spine of NVIDIA's NVLink Fusion can transfer more data than the internet.” Holy, that's incredible: 130 terabytes per second. It connects GB200 chips across 5,000 coaxial cables connected to 72 GPUs.

The peak traffic of the entire internet was 900 terabits per second, so the spine transfers 16% more data than the entire internet. Let's take a listen to Jensen speak about this.

“And so this is the NVLink spine: 2 miles of cables, 5,000 cables, structured, all coaxial and impedance-matched. It connects all 72 GPUs to all of the other 72 GPUs across this network called NVLink Switch. 130 terabytes per second of bandwidth across the NVLink spine.

“So just to put it in perspective, the peak traffic of the entire internet—the peak traffic of the entire internet—is 900 terabits per second. This moves more traffic than the entire internet.”

I just feel like we're building some version of Skynet right now.

Well, when I call it Stargate, all these things—the terminology lines up.

I completely don't understand a single word of this. What is a spine of what?

It's the interconnect. One of the things Elon did when he built Colossus, the first version, was that he was able to colocate all the GPUs and, effectively, for lack of a better term, harmonize them to get them all talking to each other. This is, at least from my point of view, the capability to do that within a full NVIDIA system.

To me, it actually illustrates something more mundane, which is that so much of the work we need to do is engineering work. Of course, there's work at the edge and research, but there's just so much raw engineering work that needs to be done to make these systems operate at scale.

A lot of the constraints we're going to see are going to be engineering-related. There's also a lot of upside in the day-to-day engineering work. If you look at DeepSeek, the reason it was supposedly so cheap to train is that a lot of it was just really clever engineering techniques. That's what I see.

All of this stuff is happening contemporaneously, which is what makes it so exciting to be here right now. Daniela Rus and I were touring one of the biggest deployments. It's actually the biggest B100 deployment. Daniela runs CSAIL at MIT, the biggest AI lab in the world.

The chips are all liquid-cooled now, which means those racks are dead silent. I was expecting this really eerie, awesome, silent experience, but the interconnect spine is still air-cooled, and it sounds like a jet engine right next to it. It takes away all of the magic.

What really surprised me, though, is that the GPUs are all in a rack—$6 million a column—and then the interconnect is physically a rack over. I would have expected it to need to be much closer together to get optimal performance, but it's actually physically separated into separate columns for some reason. That really surprised me, but it's an incredible tour to take. You should definitely throw it on the to-do list. Use your a16z all-access card for anything.

Anish, do you want to walk us through Google I/O, which is happening right now? Take it away on this slide. Tell us about it.

It's so interesting. I've only had a chance to try a few of the products so far. I tried AI Mode, and I was a little underwhelmed.

What stands out to me and is notable is, number one, that Veo—their video-generation models—are incredible. Veo is probably the only video model that has been competitive with the Chinese models, which are less constrained on copyrighted training data, let's say. I think Google has shown real strength in video generation, so that's one thing I'm interested in spending more time with.

The second is the price point that we discussed earlier: $250 a month. Has Google ever had a consumer product priced this way? Have we even seen many? It's extraordinary that they think they can command these prices, and I think they will.

I do think that AI Mode felt to me a bit like a watered-down Perplexity. It's just not that good. It really shows the power of counterpositioning, because Google has spent 20 years making commitments to an ads ecosystem, with all the blue links and everything else, that the competitive setup is now demanding they break.

It's very difficult for them to break those commitments, and I think it's a real threat to their search monopoly.

I'd love to follow up on that. I noticed the stock went down pretty significantly during this, which is really unusual during this kind of event. But I think it's for exactly the reason you were saying: the cooler this is, and the more search moves over to it, the more it cannibalizes the core.

Correct. That's exactly right. It's such a challenging position for them to be in. The front door of the internet—they've been the front door of the internet for the last 20 years—is the most interesting place to be on the internet from an economic perspective. The front door to the internet is up for grabs.

Already, my kids are telling me, “Dad, everybody knows that ChatGPT is better than Google.” So, there’s a near-term threat from products, and then there’s a sort of long-term threat from generational change in the products that they prefer to consume.

Yeah. The fact of the matter is, there will be something that comes along and bypasses OpenAI, Google, and everything else. We haven’t met the founder yet or heard what it’s called, but that’s just the reality. I’ll never forget when Jeff Bezos got up at an all-hands meeting and said, “In 30 years, Amazon may not exist anymore.” That’s pretty extraordinary.

I think those are sirens happening in New Jersey. They’re here.

No, that’s a niche, I think, in San Francisco—the main streets of San Francisco.

That’s right. For me, the one that struck me was the 3D video communications, the Google Beam. That’s the one I’m really looking forward to. I think it’ll be really amazing to watch.

We had that at Abundance 360 this year.

That’s right. And they’re consuming.

Yeah, it really does make a huge difference in your experience. You feel like you’re sitting across from the person. I remember when Cisco had their giant version that cost a couple hundred thousand dollars per setup, and now there’s Google Beam. I didn’t actually know they were going to call it Beam, but that’s great. I mean, it’s going to be a consumer product.

And Project Aura, their smart glasses—I think that, with full Gemini integration, is going to be great. We’ll probably start to see AR wearables on the street. Remember the first time we started seeing people wearing AirPods and people on the street talking to themselves? We’re going to start to see these smart glasses as well. I can’t wait. For me, it’s the future of education as well.

Let’s move on here. This is what we said a little bit earlier. It really drives me nuts that Apple still cannot spell my wife’s name or my name properly when I dictate it. It’s embarrassing.

So, why hasn’t Apple cracked AI? This is an article that came out. Let me just read this real quick: “Siri overhaul delayed repeatedly. New features failed internal tests and missed 2024 and 2025 rollout goals. Apple spent billions on AI chips and startups, but internal disagreements on budget allocations led to a lack of GPU supplies. Seven years after hiring ex-Google AI chief John Giannandrea, Apple is still lagging in AI versus its competitors.”

I mean, this is so sad. This could be a wound that truly damages Apple significantly. If someone had a beautiful, high-end phone with full AI integration, this could be up for grabs. I don’t have a fully entrenched, high-level theory on this, but before Anish, do you see truly great people in the Valley going to Apple to work?

Apple has a very specific, slightly insular culture, so fewer people come in and out of Apple, in my experience, than come in and out of places like Google and Meta. I read this article, too. I think they have a real challenge. As a consumer, using Siri every day is like a stick in the eye.

I think they have a cultural problem because I really think that AI has, as we talked about earlier, disagreement, sexuality, persuasion—all these aspects of the human experience that Apple has tried to polish away. Finally, I think they’ve not had a great track record of partnering. They really do want to build everything, and that served them well, but the models are moving too quickly, and it’s not a core capability of their technology team today.

So, I think they’re in a pretty tough position. With that said, fixing things like voice translation in Siri—come on. It’s just a daily reminder that they can’t do AI. It’s crazy.

Here’s my high-level observation, and I’d love to get your thoughts on it: the visionary-integrator model. Eric Schmidt, a good friend of Peter’s, talked for maybe 20 minutes when you guys were on stage, in the side-by-side chairs, about the visionary-integrator model.

Eric told this story about how he was walking through the streets of Silicon Valley with Sergey and Larry, and Sergey said, “Eric, that building right there—we need that building.” Eric said, “We don’t need that building, Sergey.” Sergey said, “Yes, we do.” Eric said, “Okay, well, you’re the visionary. If you say we need that building, then we’ll make it work.” And he made it work.

That defined what is now the visionary-integrator model, which Elon Musk has perfected. He has so much vision, but he needs a perfect—

I can finish my sentences.

—integrator for every business, every operation.

It’s really obvious that Steve Jobs was the visionary and Tim Cook was the integrator for years. But if I go to Google Trends and see how many searches there are for Steve Jobs’s name versus Tim Cook’s name versus Elon Musk’s name, it’s 100 to 2 for Steve Jobs and 1 for Tim Cook. Fourteen years after Steve Jobs passed away, still twice as many people are searching his name as Tim Cook’s name.

Meanwhile, in Saudi Arabia this week, Tim’s not there. All the visionaries are there. So, you’re like, well, who is the visionary and who’s the integrator in the company? The visionary needs to attract talent, which means you need to be on this podcast, in Saudi Arabia, on All-In, or whatever it is that, in that era, gets people inspired to come and work for your company. That’s a huge part of the visionary’s job.

I don’t see that dynamic. I’d love to get Google thrown into that same bucket. We’re just talking about Google’s kind of underwhelming I/O, but if you can’t name the visionary and the integrator, then you’re not following the known-to-work current model.

Having a company that is founder-led, where the founder can basically say, “I don’t care what the board says or what everybody else says; this is where we have to go,” means that an AI-first founder company is going to eat the lunch of everybody else. Especially at scale, you need the founder model just to cut through all the bureaucracy. Otherwise, you end up in terminal political fights and nothing gets done, which is what we’re seeing.

I’ve had a few friends who’ve sold their companies, and Apple was on the list of choices. But I think there’s got to be a solution that creates a visionary role, even though you obviously can’t go back to the founder.

Yeah. I’ve had friends of mine who had the chance to sell their company to Google or Apple, and I’m like, “If you go to Apple, you will be swallowed up into a giant NDA, not be able to talk to anybody, not be able to speak at a conference or an event.” It’s a real challenge.

Listen, Anish, let me give you a chance to close out. I know you have to jump on a thousand board calls right now. I’m grateful for you joining us. Any closing thoughts here on the AI world that you’re deep into at this moment?

Thank you, Peter. It was a pleasure. Dave, Salim, I would love to join you in the garden for that bottle of wine next time. That’s one closing thought.

I love having these conversations. I do feel like this is the most human technology that we’ve ever invented. And as you talk, Peter, about making the transition from a mindset of fear and scarcity to one of abundance, I think this is exactly the technology that unlocks that in a very tangible way.

I wouldn’t trade being alive now for any other time in history.

It truly is the most extraordinary time ever to be alive. I want to acknowledge something, Anish: that insight you had at the beginning about subjective technologies and how much of a difference that will make. That’s the first time I’ve seen somebody clearly articulate the massive opportunity. So, much appreciated. Thank you.

The analogy to the Jones Act, too, is something I’ve never heard before.

We’re going to Borg-like assimilate all of it.

Please do, Dave. I look forward to hearing it.

Just as I adopted abundance from Peter, you guys have got license to use those now. Nice to be with you. Thank you for having me, guys.

A pleasure, Anish. Thank you. Bye-bye. Cheers.

All right, we really enjoyed having Anish. I found this article somewhat interesting. It’s a speculative scenario, a history of the future from 2025 to 2040, but it brings up a lot of great points. I’d love to hear your guys’ thoughts on it. Let’s take it one at a time.

So, 2025 to 2026—the year ahead—it’s the agentic AI boom. GPT-5 is multimodal; it’s coming out in Q4. Code generation floods apps. I think this is happening. It’s not really a prediction; it’s just a report on what’s going on.

Let’s go into 2027 to 2028 for this one.

Hold on one second. I think the agentic phase is going to give way to vertical AI because there’s so much opportunity in training vertical AIs on very specific use cases. I think that’s going to complement the agentic phase, but in general, the paradigm is right.

Yeah. My prediction is that the word “agentic” is going to get old very quickly. It’s just too vague and generic.

What was the word that got old very quickly?

A couple of years ago, there was something—Copilot as a term. Everybody started using Copilot as a generic term.

Oh man, that got so annoying so fast.

It did. True. I think vibe coding is going to come and go real fast, too. No, it’s a great term, actually. I’ll be sad to see it go, but what you saw with Codex is that you can launch 20 concurrent processes and then stitch them back together again. That’s very different from vibe coding. No one’s named that yet. When Greg Brockman rolled it out, he said, “Yeah, I’m so bad at naming things, we’re just going to call this Codex.”

I know. I was like, “Huh? Didn’t you just call it Codex years ago?”

From 2027 to 2028, we’re going to see breakthroughs. I think one of the things that’s interesting is that it’s predicted that, in the next couple of years, we’re going to start to see physics breakthroughs and mathematical breakthroughs. We’re going to start to see fundamental technology moving forward in its ability to help us understand the universe even deeper.

I find that, for me, the biggest and most exciting thing, because I think there’s so much research data that we have not seen the signal from the noise in and have not extracted key observations from. I think AI let loose on that will absolutely do magical things.

From 2029 to 2031: the end of white-collar work. I believe that—and this is an argument I have with a lot of people—there’s no job I don’t think it will be able to do. At this point, we’re talking about advanced superintelligence. Do you think this is true, that we’re going to start to see the end of all white-collar work by that point?

The thing that’s driving me nuts about this is that that’s the endpoint. You know, 2030 is the endpoint, but it’s pretty much a straight line between here and there. The amount of job dislocation in 2026 and 2027 is going to be like nothing we’ve ever seen.

I keep telling all the CEOs, “You’re way underplanning. You need to look at every single person in your organization—all the individual contributors doing white-collar work—and you need to get them to become AI users right now.”

Mm-hmm.

Otherwise, you’re condemning them to being sitting ducks. You’re saying, “Well, it’s 2 or 3 years in the future.” They’ve been working and doing their career planning for 20 years. They’re stuck.

Yeah. You’ve got to get them on the platform now and free up the time for them to learn, and put formal education programs in front of them now. Because if you draw a straight line between now and 2030—which is probably more like 2029 or 2028—that’s only a couple of years for people to remap their entire career path. There’s a huge disservice in sticking your head in the sand and ignoring this.

There’s another thing as well, which is getting your employees, your kids, and your friends to start thinking as entrepreneurs. When I was having a conversation with a friend of mine, Dan Sullivan, I was saying, “As this technology starts to truly become magical at a level we can’t even imagine right now, is it going to quell my sense of purpose? Is it going to start solving things for me to the point where I’m not motivated?”

Dan said, “Has there ever been a time when more powerful technology has made you less motivated?” I said, “No.” And he said, “Why do you think that is?” I said, “Because I’m an entrepreneur, and I just dream bigger every time there are more capabilities handed to me.”

I think that mindset of finding problems, solving problems, and letting everybody know they can become entrepreneurs—that they can start to create new capabilities, new companies, and new nonprofits; that they can start to dream at a level like never before—is an unleashing of the human spirit that is so important, and creativity to boot. It’ll force it, which I think is really amazing.

Well, when I was lecturing at Stanford 2 weeks ago, at the end of the lecture, the TA came up to me and said, “Man, you really won the hearts and minds of all these students when you said, ‘The administration is completely out of touch with you.’”

I said, “Really? That’s what won them over?” But it’s that dynamic where your kids are going through what my kids are going through. They’re chomping at the bit to use AI because it’s so empowering, and because the administration hasn’t figured anything out yet and is woefully behind, they’re a barrier.

The corporate version of that is, “We can’t do this because we’re regulated, or because of data leaks, or because of hallucinations.” That’s an excuse.

Good. It means you haven’t figured it out at the exec staff level. Now you’re condemning your people to being behind the curve. To your great observation, it’s absolutely a leadership failure, and it’s inexcusable.

The last quadrant in this slide here is 2036 to 2035.

I think it’s supposed to be 2036 to 2045. I think you should say “trillion-dollar.”

And so it’s the notion that we’re about to have a massive explosion in the number of robots. It’s not just humanoid robots; it’s robots of every shape and size. At the same time, we’ll have bio-longevity cures. We’ll unleash human longevity. I love that stuff.

All right, let’s move on. Along the lines of what we just saw, here’s an article I pulled in: “AI designs an antibiotic that beats MRSA in mice.” This is about multi-antibiotic-resistant strains of bacteria. Researchers built an AI tool that designs new antibiotic molecules faster than traditional methods. AI-designed antibiotics effectively treated MRSA infections that no longer respond to existing drugs. AI can design affordable, lab-ready antibiotics to fight the deadliest infections.

This is the stuff that turns me on. This is the stuff that I’m just thrilled about in our future.

Well, once you can personalize it to the individual, which we’ll be able to do as a very quick next step, everything changes at that point.

Yeah, for sure. All right. One of our last slides here is that Gemini 2.5 benchmarks outperform competitors yet again in mathematics, coding, and multimodal capabilities. We’re seeing Gemini 2.5 Pro really beat out OpenAI’s o3 and o4 models.

What it’s not doing, though, is winning the revenue race. OpenAI is just trouncing Google in revenues, and that’s a dangerous place for it to be. I don’t know if we need to say anything else about this, but Google has incredible engineering talent and a massive number of wet, squishy brains working on this stuff.

All right, let’s go beyond AI into the robotics world: robotaxis. This comes from Unusual Whales. ARK Invest—this is Cathie Wood—projects a $34 trillion enterprise value from robotaxis by 2030. Cathie has been long on Tesla, as I am. Here we see a few of the companies: Waymo and, soon, Cybercab.

That’s massive. I mean, the idea that people are going to stop taking their cars. I think when this really works, Dave and Salim, is when my AI is automatically anticipating when I need a car and the car is showing up without me having to request it. It knows my calendar, it sees me walking toward the front door, and the car is just waiting for me. It’s magical.

For God’s sake, when can it drive our kids to practice? That can’t come soon enough.

Well, it’s within 2 or 3 years. There you go. You’ve been waiting a lifetime. Two or 3 years, and you’ll be all set. We’re supposed to have AI do the math for us. Thank God.

It’s easy to lose track of the fact that a huge fraction of humanity—the number-one job in the world is driver. That’s the number-one title. But then you look down the long tail of other things that people do, and a huge fraction of humanity is doing things as routine as driving, screwing in a circuit board, or assembling a laptop lid. It just goes on and on and on.

If you sample humanity, you lose track of that when you’re in a high-tech hub and everybody’s working on next-generation stuff, spaceships, and AI. But just sample a random person on the planet and what they do, and that’s where the scale of it comes out to $33 trillion. I guess AI would do the math. It is what it is.

Well, here’s an example. Waymo outpaces Lyft in San Francisco. Three hundred Waymo vehicles now complete more rides than 45,000 Lyft drivers in San Francisco. Each Waymo averages the workload of 150 human drivers operating 24/7.

Have you guys taken a Waymo ride yet?

Oh, yeah. Yeah, I did one in Phoenix. It was pretty mind-boggling, and that was ages ago. It’s gotten way, way better since then.

It has. You don’t need to tip your driver, and you can have confidential conversations. I’m always wondering if my driver is listening to my conversation while I’m just sitting there listening.

I had a back-to-back where I did a Waymo ride in San Francisco. The next day, I was in New York in a smelly yellow cab. The thing that jumped out at me about the Waymo, which I had completely overlooked, is that you can control the lighting and the music, and they’ll keep adding things to it. It becomes your little travel cabana. They’ll add flat-screen TVs, it’ll sync to your phone, and it’s going to be just such a different experience. It’s not just about being driven; it’s about control of your environment.

So I had anticipated that.

Yeah. I can’t wait for the lie-down beds so I can get a nap on my way to work.

This was a fun article from The Wall Street Journal: “Apple to support brain implant control of its devices.” It’s teaming up with Synchron to bring brain-computer interfaces to consumer devices. Albeit, this is specifically for people who are impaired and have severe disabilities, but this is the direction we’re heading. Again, Ray predicted BCI—our ability to connect everything to the neocortex—in the early 2030s. I just love seeing these little hints toward that direction.

Yeah, these are the early signals around that, but it’ll become mainstream pretty quick after this.

Well, I’m very much a humanist, and I think it gets really creepy when you start punching directly into your neurons. But I’ll put that aside for a second. Ray is going to emerge as one of the greatest prognosticators in the history of the world after going up and then way down, and then predicting artificial super intelligence in 2029 30 years ago. It’s going to land at the exact moment. It’s super annoying because he’s so outrageous but always correct. It drives everybody crazy.

Amazing. It’s funny. I don’t know him—you guys know him personally very well. I don’t know him personally, but the books were so inspiring to me, reinforcing everything I was already thinking, but then really quantifying it. I’m going to love it when it lands right on the minute.

Yeah, it’ll be great for Singularity as a brand, too, I think. All right, let’s close out our conversation today on Moonshots with a look at crypto once again. Here we go: “U.S. enacts stablecoin bill in 2025.” The prediction market Polymarket shows a 95% probability that the U.S. will pass a stablecoin bill in 2025, up 35% from prior weeks. Any comments on this, Salim?

I think this is going to be huge. They have to first pass a budget, so let’s give them some time to figure that out. But when they figure out the stablecoins, this will completely unleash the crypto world and give it a bridge from the crypto economy into the real-world economy. I think all sorts of things become possible. The opportunity for AI agents to do microtransactions using this is going to be amazing. So, huge opportunity.

Yeah, I think that last point is by far the most important one that we should track closely week to week, because the ability to move back and forth instantaneously and digitally from a stablecoin back to Bitcoin or whatever you want is intimately tied to agent-to-agent microtransactions. That’s going to be—we’ll get Cathie Wood to do the math for us—but that’s probably going to be the biggest part of the economy by 2040: those transactions.

Oh, yeah. There will probably be hundreds of billions of agents, each doing microtransactions. This is half of the internet. The internet, when it was built, allowed us to share imagery, data, and documents, but not financial transactions. This is the other half of the equation, and it’s coming fast.

We’re playing with Codex pretty much all weekend. You can spawn agents to do all kinds of things for you. The interface they put in front of it, each time you spawn an agent, creates a new row and there’s a little bar turning. You’re like, “This interface—right out of the gate, I want to do 1,000 agents.” Then it’s going to be 1 million agents, and it’s immediately obvious that we’re all going to be fighting for inference-time compute. You can think of things so fast and deploy so many agents to do them for you so quickly. The gating factor is, well, I can’t get the compute in this interface. Obviously, it’s not going to support the scale that I need, either.

The thing that Anish said as well is having his agent go out and shop for lower-cost insurance or mortgage rates and such. I love that idea. You’re going to be able to optimize everything. You’re going to have your agents constantly searching the web at minimal cost. Then I want an army of agents that are going out and making investment trades for me—just an army making money for me. I’ve been waiting for my boys, my 2 boys, to get excited about trading crypto, but I think the agents will get there first.

Another topic I wish we had more time with Anish to talk about is disposable code and personalized code. You’re like, “What’s an example of that?” The example Anish usually gives is: imagine you’re using Suno or Udio to create a song, but the song is about this exact podcast, with content related to our topics today. That’s a great theme song to play as a pre-roll. The software created this thing specifically for the podcast. It uses a fair amount of compute, but it’s still only 12 cents. No big deal. You create it, then you throw it away. I can think of a lot of those things. Let me spawn all those agents to create them, and I’m going to throw them away when they’re done. That’s a lot of compute. It’s so worth it. But where’s all that compute going to come from?

My last slide for today. Still on a Bitcoin roll. At this moment, Bitcoin’s running at roughly $106,500. Pretty good. We’re approaching near highs. This article came out: “Bitcoin is becoming America’s reserve asset. More Americans own Bitcoin than gold.” That’s pretty extraordinary. I think this was kind of predictable just because it’s so much more democratized. It’s 1,000 times easier to own Bitcoin than it is to own gold, so I would have expected to see this sooner. The thing that I think will become really powerful is when people own, at a gross level, more Bitcoin than they do gold. I think that’ll be amazing.

Pleasure, as always, gentlemen. Onwards. This was a big weekend in AI. We have one last thing, Peter.

What’s that? Well, I think, Dave, you have a slide or a photograph, right? Do you want to show that? I’ll stop sharing.

Can the team pull it up for me? Yeah, or I can show something if you have it. Given that it’s your birthday, Peter, we found an MIT photograph of you from way back, and we thought this was absolutely worth showing. Talking about AI agents and retro, we want to hear that guy’s voice.

Yeah, I got this from Matt Rita[?]. He sent it to me with an emoji of a hair pick along with it, so I didn’t include that. I was really reluctant to drag it out of the archives because I know it can come back around.

It’s quite a riot. I want to show you one other thing, though. Hold on. I’ve got 2 images that I thought were really interesting. Where’s my image gone? Oh, my God. We have to cut this out somehow.

I respect you guys. I’m spending my birthday doing a podcast with both of you. How was that? Hold on. Okay, let me share this now.

Hey, congratulations, and happy birthday, by the way.

Thank you, pal. Thank you. I started my birthday by doing 100 continuous push-ups in a row, so I’m still riding that.

Very nice. Was that part of the celebration?

Yeah, that was. I have 2 favorite photographs of you. One is this little 3D-printed image of yours that I kept, and then one of the next generation.

Oh, my God. Yes. My boy. The 3 of our boys were all born within about a month of each other. They’re now all turning 14.

Yeah, that doesn’t look like me. But anyway, I wanted to show one little more surprise. Lily, come on along.

Okay, we brought you a birthday cake.

Thank you.

Unfortunately, you can’t eat it because you’re an avatar. Nothing’s on your diet unless there’s salmon in the middle of it.

I’ll reach out using Google’s Beam.

Well, guys around Cambridge talk about you all the time. I don’t know if that photo was long gone.

Yes. Glad you dug it out of the archives. I’ll just start it all over again. Just so you know, when your ears are ringing, that’s everyone out here talking about you.

Thank you, brother. Thank you. Thank you. Love you both. Have a beautiful day, Peter, and happy birthday again—from Saturday or Sunday, whenever it was.

Saturday.

Yeah. All right. Take care, Peter. Take care, Salim.