Claude Opus 4.5、白宫“Genesis Mission”与Amazon的500亿美元AI押注：Emad Mostaque、Salim Ismail、Dave Blundin与Alexander Wissner-Gross | EP #211

I've compared this moment to 1939. This is the Manhattan Project.

Similar to the Apollo project that put a man on the Moon in 1969, this is an all-in national effort to use the power of AI and the world's largest supercomputers to advance innovation and science.

This is just extraordinary. I think this could be the greatest accelerator for human knowledge in the U.S. yet, if it's properly funded and executed.

Genesis 1:1 says, “God created the heavens and the Earth.” Finally, we have the tools to actually be able to understand them properly. The application of large amounts of compute with the reagents allows us to unravel the mysteries of the Earth and the universe.

If we went back to the beginning of the year, did we predict that we'd be here, or is it moving faster than even the few of us could predict?

Now that's a moonshot, ladies and gentlemen.

Hey, guys. Welcome to our emergency pod for Thanksgiving week. A lot is going on here. Genesis Mission—we're going to be talking about that. We'll talk about Anthropic's new Claude 4.5. I'm here with AWG, Mr. Exo, and thank you, Emad, for joining us. I know this is Thanksgiving for you in England as well, isn't it?

It's Thanksgiving for everyone.

Yes, for sure. I wanted to start with a question: What does Thanksgiving look like in the year 2035? I'm curious—is it going to change at all? Salim, do you want to kick it off? Is 2035 far enough away to make a difference?

It's a hell of a difference. By that point, we should have the cost of Thanksgiving dinner dropping by 10X. It should be personalized to you for your nutrition, so that depending on your metabolism, the turkey or ham or whatever the heck it is is totally customized to your ability.

I'll have a little device inside me saying, “Whoa, whoa, whoa. Before you eat that turkey, I'm still metabolizing the cauliflower. Give me 3 minutes, please. Take a sip. Do not drink alcohol quite yet.” I think we should have gotten to the point over this hump of expensive energy where we have ultra-cheap energy and ultra-cheap food, and we're crossing right into the Alex Rubicon.

All right. My addition is that we'll have Tesla bots serving us everything. How about you, Alex?

I think if we're not celebrating it—or at least some subset of humanity is not celebrating Thanksgiving on Mars, some subset is celebrating Thanksgiving in the cloud in the form of uploaded humans, and maybe we have some uplifted non-human animals also celebrating with us at the table—then something's gone terribly wrong over the next 10 years.

Wait, so I get this: We're going to have uplifted turkeys arguing with their lawyers to keep a ceasefire against killing them all?

Yeah. If that doesn't happen, then something's gone wrong over these 10 years.

Emad, how about you? What are you going to see in 10 years' time for Thanksgiving?

Ten years is the pessimistic end of the AGI forecast, right? Assuming that humans don't end up like turkeys, where we get happier and happier and then, with AGI, it goes straight down, we'll have figured out how to make perfectly moist turkey by then.

Uh-huh.

As Alex has said, mathematics should be solved by then, along with science, et cetera. So you're in the post-abundance world, hopefully, with the robots and more, and there should be a lot to be thankful for if we can navigate what's coming.

1. The Genesis Mission

Yeah, if we can navigate what's coming. Okay, we're going to talk about that. But before we do, I want to jump into our first story, which is a doozy. Let's hear and learn about the Genesis Mission coming out of the White House—a very powerful concept. All right, let's dive into this with a video.

In every age, humanity invents new ways to see further. The telescope let us glimpse the stars. The microscope revealed the worlds within us.

For centuries, thinkers like Leibniz, Shannon, and Turing dreamed of making all knowledge computable. But today, knowledge grows faster than our ability to understand it. Trillions of data points, a universe of information still unconnected.

Now, a new instrument emerges—one capable not only of observing the universe, but of understanding it. Genesis Mission will transform how science is done in America, uniting our brightest minds, most powerful computers, and vast scientific data into one living system for discovery.

Built on artificial intelligence and quantum computing, it will radically redefine the scale, speed, and purpose of scientific progress in America. This is the work that will define our generation's legacy.

A new revolution begins, one guided not by competition alone, but by curiosity, imagination, and the belief that discovery is the truest form of progress.

Wow. Just wow. What an incredible story coming out of the White House. Again, the title here is “U.S. Government Launches Genesis Mission, Transforming Science Through AI Computing.” This is Trump's executive order to use massive federal scientific datasets to train powerful AI models. The Department of Energy will connect U.S. supercomputers and lab data into one unified platform intended to shrink the research timeline from years to days through AI-driven experimentation, focusing on biotech, fusion, and quantum. It's a big deal. AWG, do you want to kick us off?

I've compared this moment to 1939, and this is the Manhattan Project. In the Manhattan Project, as I've remarked previously, we turned the country into one big factory for nuclear weapons. In this case, the country is being turned into one big AI factory, and this is an incredibly ambitious moonshot.

We speak of moonshots. This is an incredibly ambitious moonshot—not just to turn the country into an AI compute factory, but also to supply some of the limiting reagents, as it were, like datasets. Federal datasets that are locked up in a variety of different enclaves are now, according to the EO, going to be unlocked and made available for pre-training. Probably software tools that right now are unavailable will be made available.

I think, to the extent that there may be a race dynamic with China, whose government is also collecting large amounts of data, the Manhattan Project positioning is probably pretty intentional. I think it's glorious to see the ambitious unlocking of scarce resources.

I'll also point out Dario Gil, who's been named as the mission director for Genesis Mission. I worked with him as an undergrad at MIT, and it's really great to see MIT in general, and that level of scientific influence positioning, in a 1939 moment—such an ambitious initiative.

I should just mention, by the way, that our other mate, Dave, is on a research mission in Italy this week. Let's leave it at that. We miss you, Dave. Wish you were here.

This is just extraordinary. I think this could be the greatest accelerator for human knowledge in the U.S. yet, if it's properly funded and executed. Emad, is this something that every country is going to have to follow through on and make a similar move?

I think you're seeing this. In the U.K., we had something similar with DSIT on a much smaller scale, along with new regulation and acceleration for nuclear reactors, et cetera.

Fundamentally, Genesis 1:1 says, “God created the heavens and the Earth,” and now, finally, we have the tools to actually be able to understand them properly. That's what this is really talking about. The application of large amounts of compute, with the reagents that AWG mentioned, allows us to unravel the mysteries of the Earth and the universe.

Obviously, that's a massive advantage. But I don't think it's any kind of coincidence that it's the Department of Energy that's running this.

Mm.

We've talked about how energy is so important, and the U.S. has been falling behind on energy compared to countries like China and more. You'll see more and more deregulation, more and more fusion, solar, et cetera, play into this.

The impact can again be immense if you can figure out any one of these things. I think we're in a good place to figure out almost all of them, again, if it's done properly.

Salim, your thoughts, buddy.

I think this is where you see the best of government, because they can leverage those global datasets in a powerful way. When you can do that, I think it brings out the best of what government is able to do, unlike the private sector. That's one really great point about this.

The second, I think, is that this is kind of catch-up, in a sense, because lots of countries use their federal datasets in different ways. China and France have been doing it for years, et cetera. So this is catch-up in one sense.

Taking the data, which is now a sovereign resource, and then applying it with all the AI capability the U.S. already has, I think really amplifies a huge outcome. The potential here is incredible.

It reminds me of the days when Silicon Valley started. They created secret labs at MIT, Harvard, and Stanford to figure out how radar could be blocked and came up with aluminum foil—tinfoil, chaff thrown out of planes. They had to create this global initiative, or this country-wide initiative, to protect and solve for World War II, and this is kind of like that initiative.

Mm.

I think this is that big. I love it.

I mean, this basically, in my mind, reframes basic science as a compute problem, and it's throwing everything we have at it.

Yes.

Yeah.

Absolutely.

I think that's the elephant in the room. I'll also point out that the Department of Energy has clarified that one of the goals of the Genesis Mission is to double American scientific productivity in the next decade. So when we speak of Thanksgiving 2035, I would say if we haven't 10x-ed or 100x-ed scientific productivity by Thanksgiving 2035, something has gone wrong. But I think having a 2x increase in productivity is an excellent baseline here.

I'm reminded of something. We had a very senior guy from the DoD at Singularity one year, Peter.

Yeah.

During the Q&A, he said, “This is all great. You guys all love these exponentials. VCs are all hovering at the knee of the curve, trying to catch a technology that goes vertical. You forget who funds the arbitrarily long, flat part of the curve”—which is government.

This allows government to really accelerate that part of the curve. So I think we'll see the exponentials moving forward in time in a pretty amazing way.

Yeah.

Yeah.

I think it's interesting how it's moved from the NSF and the classical grant-making, which has been disrupted over the last year, to a much more techno-optimistic approach. One of the key things that will determine the success of this is whether this is Manhattan Project-style—closed and private. Because obviously, even though they've announced it, it could be. Or is it open?

If it's open science, I think it'll be truly exponential. But if it's actually building up public-private partnerships with strong IP protections and doing a lot of stuff in private, I think it'll have a much lower impact on the other side.

Yeah. I'm excited to just watch this. This, for me—and you said it perfectly, Alex—is a moonshot. It's an extraordinary nationwide moonshot.

If not a shot at the moon, as I sometimes say.

And we'll talk about that. The only difference is it hasn't set an objective mission like, “Get to the moon and back before the end of the decade.” But this is America throwing its might and coordination at a massive opportunity.

Well, look at the areas: biotech, fusion, and quantum. Those are all moonshot domains that—

Mm-hmm.

—totally rewrite the rules of life. Amazing.

Yeah.

I would maybe just add that, as I've pointed out in the past, I think the next big thing after solving superintelligence—which arguably has either already been solved or is imminently solvable—is solving math, science, engineering, and medicine. This is what that looks like at grand scale. This is taking federal resources and applying them singularly to solving grand challenges.

Hmm. All right, spectacular. Let's go on to our second big story of this particular week. It's what's going on with the hyperscalers, but in particular, Anthropic. Nice to see Anthropic making some moves. Here is the story: Anthropic has released Claude Opus 4.5. It uses 76% fewer tokens to reach the same results as older models, outscored the entire engineering team, leads in 7 of 8 programming languages on industry coding benchmarks, and improves multi-agent support by 15%.

Alex, you want to kick us off? How significant is Opus 4.5?

Yeah. We're nearing, if not already at, the point of recursive self-improvement. Finally. The point of recursive self-improvement, many would say, is the point at which more compute and more infrastructure are being allocated by frontier labs to AI researchers than to human researchers.

I think the most important indicator isn't that the benchmarks and the evals are going up and to the right, although they are, and it's wonderful—I love benchmarks. It's that Anthropic has also announced, as you alluded, that incoming employees to Anthropic, in particular on the performance team, are now being outperformed on key tests and key homework assignments by the AI.

I think that's the canary that we're imminently, if not already—given that this model was arguably pretrained based on the data cutoff date several months ago—entering the moment of recursive self-improvement. But I think that's the bigger thing.

The smaller headline is that I, of course, have my evals whenever these code-generation models come out. One of my other non-cyberpunk FPS evals is asking it to see if it can one-shot a Mario-style side-scroller, and it did a beautiful job.

Amazing. Dario's been talking about being able to get to 100% or 90% of all the coding being done. So this is a big move in that direction. Emad, I'd love your thoughts here.

Yeah. I mean, from our tests, we got to the top of the SWE-bench Pro benchmark, which is Scale AI's benchmark, and it was really difficult, at 45%. Peter Diamandis

This is with Intelligent Internet, right?

Yeah. That's with Intelligent Internet framework using a combination of the other models.

This model, without reasoning, scored 52% without even using reasoning tokens, which I think was the most shocking thing for me.

Usually, the big breakthroughs we've had are that the models can think longer, they can check, et cetera. We didn't think it would be that way with just the straight output, and the quality of the code it outputs is actually really, really good. The average code base is 100 to 200,000 tokens, and this should be able to one-shot most code bases by next year.

The cost has dropped 67% from the previous version, so it's now $25 per million tokens as well. Coding and tokens will be ubiquitous, and it may not be that reasoning tokens are what are needed for those tokens. It was, again, completely shocking to me that it would score higher without reasoning than with reasoning.

Fascinating. Salim, any thoughts here?

It feels to me like we've moved geopolitics from nation-states to these hyperscalers. This is incredible stuff that's happening from each of these big four or five, and it's rewriting all the rules of everything. Then you have states riding on top of these, which is a much better way of doing it than the other way around.

Yeah. Alex, I want to take it back to our subscribers here. What does this mean for the average individual who's not using Opus 4.5 to code? What's the—

This is the inner— I mean, I think there are multiple levels of impact. The highest-level impact is that I've spoken in the past about what I call the innermost loop of civilizational progress. This is the innermost loop insofar as we're starting to see models that are so strong that they can conduct research and generate code for better versions of themselves. That's the innermost recursive self-improvement loop.

I've argued in the past that that's going to spin out and touch the rest of the economy. It's already in progress, but you'll see much more of it over the next 2–3 years as it fully solves robotics and physical-world automation, leading to, optimistically, radical economic growth. I would say that's the macro story.

The micro story is that, in the meantime, it's going to be trivial to generate programs, applications, and complex workflows on demand, implicitly and explicitly. I've mentioned in the past that with the length of a tweet, you'll be able to create a triple-A-level first-person shooter or video game. People are going to be creating so much more software, so trivially, that we'll be drowning in AI-generated software of very high quality. That's the narrow, micro impact in the short term.

Amazing.

The other thing that's interesting is that by itself it scores 75% in multi-agent when it's the same agent, Opus 4.5. When they combine it with Haiku, which is a very low-cost agent, or Sonnet, it gets up to 88%.

Mm.

So Opus is a really good orchestrator of agents, and this is the multi-agent support type of thing. Everyone was saying, “Well, agents can’t look after other agents.” This is the first AI that provably can, and that opens up the whole swarm nature of things that we’ve been discussing.

It reminds me of a competition that people do called the spaghetti competition, where they put uncooked spaghetti together and see who can get the highest vertical height. They found that the team that had a very efficient executive assistant as part of their team always scored the best.

This is the marshmallow challenge.

The marshmallow challenge, yeah.

Here’s the story: You get 20 sticks of spaghetti, a meter of tape, a meter of string, and a marshmallow, and you have to structure something so that the marshmallow is on top. Whoever gets it highest wins. Importantly, you’re right, Peter: The winners are the folks that have an EA on the team, but second place goes to kindergartners. Last place are MBAs, consistently.

They lie, they cheat, they break things, et cetera. It’s an amazing exercise.

You were going to make a second point on this one, Salim, I think.

If we went back to the beginning of the year, could we have predicted that we’d be here, or is it moving faster than even the few of us could predict? It feels like it’s moving faster than the few of us could predict, although Alex has probably—

Go ahead.

I’ll pat myself on the back narrowly for this and say, as Dave, who’s not here at the moment, would attest—and I think, Peter, you were in this group chat as well—at the beginning of the year, Dave challenged me to formalize a prediction for what end-of-year math solving would look like. I was banging the drum throughout the year: “Math is going to get solved. Math is going to get solved.”

I made a very specific prediction about FrontierMath Tier 4 and AI models passing that, and if anything, they’ve slightly overshot my very conservative baseline.

How dare they?

I know. I think we’re more or less where I expected we’d be by the end of this year in terms of the strength of AI models solving math, science, and engineering.

I’ll add one last thing on the Anthropic story here, which is that last pod we talked about their economic success as a business: They’re heading toward significant profitability in the next 2 years, and this is part of that equation. So congratulations to Dario and his team on Opus 4.5.

Let’s go to our next story. Again, on the leaderboards, I’ll turn to Alex for this: the ARC-AGI leaderboard update. This is not just performance; this is performance per dollar. Alex.

That’s right. The big story is that we’re driving the cost of intelligence to zero. The cost of superintelligence is being driven to zero as well. The ARC-AGI 1 and 2 benchmarks are really lovely benchmarks I’ve supported in the past.

The general theme is: Can AIs successfully visually reason and visually synthesize new programs to reason? What we’re seeing for the first time, between the Opus 4.5 results demonstrating breakthrough, state-of-the-art cost efficiency of visual program synthesis and an earlier result I don’t think we got a chance to touch on, in which a company named Poetic announced superhuman-level performance on the ARC-AGI 2 benchmark, is that visual program synthesis is starting to get solved. The world needs harder benchmarks.

The so-what for everyone right now is that so many problems in the world, especially in the physical world, rely on some sort of visual reasoning, some sort of intuitive ability to manipulate the physical world, spot patterns, and synthesize implicit programs, even if they’re never written down as source code. ARC-AGI and ARC-AGI 2 are excellent ways to capture problems that humans generally find easy but AIs have historically found challenging, and that’s all getting solved now and saturating.

You mean things like proprioception, for example?

Yes, and being able to, in general, recognize a pattern and solve it visually.

Emad, your takeaway from this one?

Yeah, I think even the authors of ARC-AGI are asking, “What on earth do we do now?” I think I’ve seen some tweets from them saying that. The next benchmarks are dollars. You have Vending-Bench and some of these other benchmarks where it’s like, how much money can they earn? You start to see trading benchmarks.

You’ve got the tipping point now where these models go from being like very smart people you tap on the shoulder who can do individual tasks to being able to do real economic work, and we’ll see many more benchmarks where the axis is literally dollars. That’s next year’s story, I think.

2. Agents Become Entrepreneurs

How far are we, guys, from the single entrepreneur with a set of agents building a billion-dollar business? What do you think, Emad?

I’d be surprised if it wasn’t within 2 years, probably next year.

There are some amazing entrepreneurs out there, and their only thing was, “How do we scale talent that listens to me?” Given they’ll be good at using these, again, it’s a year or 2 away at most.

Alex?

I think it’s now-ish, in the sense that right now, as I’ve remarked in the past, you see these poor baby AGIs that have some agency, peddling altcoins on X. I think the first zero-human or half-human billion-dollar startup is probably, for better or for worse, unfortunately most likely to be a baby AGI that pumps an altcoin and becomes worth a billion dollars.

I think we could do way better than that as a civilization than pumping altcoins, but I think that’s probably unfortunately where it’s going to happen first.

I would’ve gone for porn rather than pumping altcoins, because that’s such an obvious place where people have libidinal vicissitudes. But in terms of the broader picture, there’s a colleague of ours that we all know who launched 47 AI startups in a month a couple of months ago. People are now using this as a platform to really change the game, and whole incubators are just launching AI startups.

I’m going to make the point that that’s already in the works, and it just has to hit a market segment.

So here’s the question for you, Salim, Emad, and Alex: Is this just going to accelerate the rich-poor divide? In terms of the ability for single individuals who, let’s face it, are 21, 22, or 23 years old, just out of MIT or Stanford, to launch something great and create extraordinary wealth at a pace that doesn’t need other employees as part of their team.

Well, what’s going to happen is you’ll have that happen, but the ability to go from poor to rich has never been faster.

This is a really important point that I think you pointed out in Abundance, Peter: The richest people in the world used to inherit their wealth exclusively, and today the richest people exclusively have earned their wealth. That loop is going to just accelerate, and now you’re going to get 100 Vitaliks and Sam Altmans, et cetera—thousands of them just spinning off companies.

The bigger question, I think, is what happens to the broader economy when this happens?

Economy 3.0.

Alex, please.

I think all of this capital-substituting-for-labor discussion misses an important point, which is that these AI agents are arguably neither capital nor labor. They’re a new third category.

Everyone who’s hand-wringing—and I hear this a lot—says, “Oh, well, how are we supposed to survive? Not everyone wants to become an entrepreneur.” I would argue that a near future where everyone survives by, quote unquote, “becoming an entrepreneur” misses the point entirely. It’s not going to be the case that everyone becomes an entrepreneur. Everyone’s going to become an investor.

The entrepreneurs increasingly are going to be these AI agents that are identifying and solving valuable problems. The average human, the average unaided biological meat-body human, is going to be able to invest in fleets, entire economies, and indices of AI agents that are acting as the proximal entrepreneurs.

This is also the Accelerando premise.

Exactly.

I think Emad should say something about this because he’s been studying the social economy the most.

I am. I’m going to go to him next.

Oh, sorry.

Don’t mean to butt in on the hosting phrase here.

Oh, Salim, you do a beautiful job as well.

But this is the Accelerando playbook for people to read. Emad, over to you, pal. You've been thinking about this very deeply.

Yeah, Accelerando is a great book. Obviously, my book, The Lost Economy, is also great. But the problem is that it's going to be very difficult to outplan and outcompete something like Claude 5 when it comes to coming up with businesses, unless you have skin in the game and you care. This is the main thing, because they will try things dynamically, and they'll just move on efficiently. Whereas you can apply these agents to tasks, and the key thing is that, in economic terms, it's all variable cost.

Normally, when you had a company, you had to go and hire someone. That's a pain to do. You had to launch your own servers before the cloud. Now everything is variable cost, and it's also cash-flow positive because you typically pay the AI providers a month or two after you have an enterprise contract, and you charge people up front. So you can have brand-new economic models where you're taking information, organizing it, and adding value for people.

And I think that does close this rich-poor divide, because you won't know where companies are coming from. Compliance and everything can be done automatically now. We're actually seeing, around some of these big AI startups, entire systems that will do your tax compliance, your financial forecasting, automatically balance payments, and things like that. The stack is nearly ready. Again, it's about a year away before you can launch a business, probably in minutes—

Amazing.

—with everything there.

Can I mention something here?

Of course.

The plug for ExO here, which we stumbled across accidentally. Peter Reef quoted Jeremy Rifkin's book, The Zero Marginal Cost Society.

Yeah.

About three-quarters of the way through writing the ExO book, we stumbled upon this economic insight. When you're running a business, you worry about demand and supply, and hopefully the cost of demand and the cost of supply. Hopefully you're on the right side of that equation.

What the internet did was allow us to drop the cost of demand exponentially. Online marketing, referral marketing—every company is trying for a viral loop. If you get there, your cost of acquisition goes to zero, which is an amazing thing. We saw an initial wave of YouTube, Facebook, and so on explode out of the gate with that.

What exponential organizations and new models have done is drop the cost of supply exponentially. Think about Airbnb: the cost of adding a room to its inventory is near zero. If you're Hyatt, you have to build a hotel. With the launch of Amazon Web Services, you could take computing off the balance sheet and make it a truly variable cost.

To Emad's point, everything now becomes a variable cost, so you have almost no capital expenditure. Now you take out the denominator, the market cap explodes, and for the first time you have a breed of organization with a low cost of demand and a low cost of supply. That's like a magical holy grail for business. How we navigate that is going to be unbelievable over the next few years as this paradigm rolls out.

Love it. All right, I'm going to jump into our next story. A lot still to cover.

3. Shopping Agents Replace Search

So, ChatGPT introduces shopping research. It compares and searches, and provides recommended products that you're interested in. There's no question that this is coming out on Black Friday. They are moving this quickly.

This uses ChatGPT Mini, and their claim is that they're able to get accuracy of up to 64% for the best predictions of what you want to buy. For me, this is about replacing the search engine, affiliate blogs, YouTube reviewers, or Amazon's own recommendation engine. It's AI replacing the entire product-research economy.

In this one, at least from my perspective, the middleman is going to lose and the models are going to win. Alexander, I'm curious: what are your thoughts?

Critically, not just generalist models. This is a specialist vertical agent. To the extent that there was some expectation that we'd end up in a singleton in the near future, where there's one generalist agent that does everything, it appears that's not the case, at least to the extent that OpenAI is a leading indicator.

By my count, OpenAI has launched at least 2 major vertical specialist agents. They launched Deep Research originally, which is general research, and they've launched a coding agent, Codex. This is the third-ish vertical agent by my count that they've launched, other than the baseline model.

I think it's really interesting. Where are the generalist models? Yes, they and other frontier labs are launching generalist models, but we're starting to see a proliferation of specialist models. I think we're going to see many more. I wouldn't be surprised if we see more specialist post-trained models for finance, medicine, and management consulting, just picking off broad industry verticals one by one. In this case, this is going after consumer purchases.

But of course, I don't want to be calling on a particular model. I just want my AI to do this for me, right? And it will—

Well, think—think they're throwing a lot of resources at model routing and routers—

Mm-hmm.

—in general. So what you'll gain with this umbrella router, I think, will be a single pane of glass—a single UX surface that you talk to.

Hmm. Love that.

I have a question.

Yeah, Salim.

How far are we from a—Peter, you call it Jarvis, right?—a personalized layer that watches on your behalf, is totally secure from a privacy and sovereignty perspective, and navigates the external world for you?

If you've got shopping you want to do, or you need to buy something, or you need something that you may not even know you need, it's figuring out which agents to use and sorting it out. How far are we from that point?

Now.

Now.

Now. I mean, what you're describing, Salim, is, I would argue, a computer-use agent, a CUA. Microsoft and other major companies and frontier labs already have CUAs that are either about to be rolled out or have already been rolled out and are in beta stages to do just what you described for a desktop.

I think what Iron Man has is a CUA on top of a heads-up display, or HUD, but that can come as well, imminently.

You see, what I find interesting here is the notion that we're about to give our AIs access to everything we read, everything we say, and our intent, attention, and intention.

As soon as we get our heads-up glasses or augmented-reality glasses that are able to have not only forward-looking cameras, but also cameras that look back at our pupils to determine what we're staring at—if I'm staring consistently at that beautiful lamp over Alexander's shoulder, and my AI says, “Would you like a lamp like that?”—or if I make some side comment to somebody else, it may purchase it for me and ship it to my house.

This ability to understand what we truly want by listening to our conversations or looking at where we look empowers this AI to become our magical shopping agent in many ways. Emad, how do you think about this?

Andy Jassy, CEO of Amazon, recently said, I believe, that their agent, Rufus—which I doubt any of us have actually used—has 250 million users. That's their shopping agent. And they're estimating $10 billion in incremental sales from it next year, given conversion statistics up to 60% higher. Who would have thought?

I think the key thing is: where is it? Bing and Teams and other things had access to the user's eyeballs. The challenge for ChatGPT and OpenAI here is, how do you become that first intentionality in the shopping experience? And then what type of shopping is it?

If it's toilet paper, who cares? You just want your AI agent to do it—

Automatically.

—three-ply or whatever.

Yeah.

Automatically.

Yeah.

If it's super-discretionary, some people apparently enjoy shopping. Maybe not people like us, but definitely my wife and others. And then you have this middle bit: how many TVs do you really buy? You kind of know what TV you're going to get.

So I think the key thing is, who is the AI next to you? You go to Amazon for a shopping experience, you use Rufus. You Google search, you now have Google AI Mode up there. The key thing, going to Salim's point, is who comes up with the agent that's the most charming and engaging and licenses Paul Bettany's voice for a Jarvis? Because that can then disintermediate everything, and that's what the fight is on for now.

Fascinating. All right, let's move on. Let's get to Google here.

4. Compute Competition Accelerates

Google further encroaches on Nvidia's turf with its new AI-chip push. Google has launched the Ironwood TPU, its 7th-generation AI chip, with 4 times the performance of its previous version. Importantly, instead of selling hardware, Google is now offering its TPUs as a cloud service. For example, Meta is running on them without purchasing the TPUs.

In the photo here, we have Thomas Kurian, who's the CEO of Google Cloud, and he's been crushing it.

Google Cloud's been doing amazing, and this puts them directly in competition with Nvidia. Alex, how do you think about it?

There's been so much hand-wringing, Peter, over Nvidia's purported monopoly, or CUDA as a purported architectural monopoly. GPUs are now finally facing healthy competition.

We see TPUs that are being both purchased, according to this reporting, as well as licensed and rented. We see, obviously, AMD with their own stack. We see Trainium and other Amazon chips, and ASICs in general. I think what all of this is turning into is that accelerated compute is finally becoming a fungible commodity. It's not just a one-supplier commodity; it is a multisupplier, very healthy, very heterogeneous ecosystem of fungible accelerated compute, which is exactly the sort of competitive ecosystem we want to find ourselves in.

Hmm.

Emad, do you have a comment?

Yeah. We used thousands of TPUs a few years ago, and from the v5s, this now has 10 times the compute. The chip size, the single die, and the interconnectedness of the Google chips are beyond anything you've seen. So you've gone from 64 in a unit now to, I believe, 4,000—no, 9,000.

Hmm.

What Google's really good at is connecting lots of chips in one place and even across multiple data centers. We had runs of up to 50,000 of their low-energy chips, and what's important about that is context. Right now, actually, DRAM prices have gone up by about 5 times.

Yeah.

So if you want to get the DRAM for your gaming PC, it's gone up crazy. Google actually has the ability to use cheaper chips at massive scale to do large-context-window things, and that means that Gemini has 1 million to 2 million input tokens, from video to audio to other things, whereas it's still limited on other GPUs. That's going to become even more of a difference going forward.

Google originally built these chips to power Google Search, and now they've matured to a point where they can offer them to everyone, even hosted. The cloud service has been available for a few years, but now they're exploring actually saying, “Meta, you want it in your own data center? We can look at that.” That's going to be super interesting going forward, particularly as RAM versus FLOPs becomes the key differentiator in terms of performance, because context becomes almost everything. The models are already really fast, to be honest.

Oh. Well, kudos to Google. They continue to crush it week over week. Salim.

Two plugs for us here. About 3 months ago, we said 2 things. One is that Google would inevitably start to lease or sell the TPUs, and here we are. Second, I believe, Emad, it was you a few months ago on the pod who said, “Invest in DRAM companies because DRAM is going to become the short supply,” et cetera, et cetera. So it seems that we're typically 3 months ahead of the game.

Amazing, Salim. That's great.

All right. Let's move on to the next story here. Amazon is spending up to $50 billion on AI infrastructure for the US government. It's projecting that it'll add 1.3 gigawatts of new data center capacity, beginning construction in 2026. So what's the story here? AWG, do you want to take a shot?

Yeah. The government clouds, in many cases, including with AWS and otherwise, have their own availability zones, and they're notoriously undersupplied when it comes to accelerated compute with GPUs. I think it's surprising, depending on how you count, that the public sector is either half the economy or about a quarter of the US economy, yet it's historically been so compute-starved, or at least GPU-starved.

I think this is a welcome investment, at least from my perspective. We want a vibrant public sector and a public sector vibrantly supplied with accelerated compute, and I view this as a very positive step in that direction.

Nice. Some of the stats here in this article: AWS is serving 11,000 government agencies and expecting to spend $125 billion in capital expenses by the end of 2025. Massive support, right? AWS has really just dominated. Salim?

Is this essentially the federal government saying AWS is their cloud provider? That's a big deal if that's the case, because that's what it seems.

Well, the US government has multiple cloud providers. This is pretty well publicized and reported on.

Yeah.

But Amazon is an AWS key supplier to US government cloud resources.

All right, I'm going to move us along here. Also, another article on Amazon here: their data center tally tops 900. We forget the fact that Amazon, because of AWS, has been running a massive number of data centers around the world in over 50 countries. They're launching something now in Indiana. It's a 1,200-acre data center, and they're putting it up and getting it online faster than anybody else. Any particular thoughts on this one?

I was struck by the fact that the Indiana one uses 2.2 gigawatts of energy. That's an unbelievable amount of energy for a data center. That's a small country worth of power.

Yeah, I would maybe note that we're tiling the Earth with compute. That is what we're really talking about here, and this is just the opening act.

The Indiana data center, in particular, is— we were speaking about Anthropic a few minutes ago—that Indiana data center is the core computing facility for Anthropic, both for training and for inference. It's called Project Rainier, and it was farmland that was converted almost overnight—I mean, it took about a year, but almost overnight—into modern compute. This is 1939, when you see farmland in the Midwest being converted to compute resources.

Factories, yeah.

Yes.

Alex, you can't imagine the number of comments that I had from people saying, “What does Alex have... Why is Alex against the moon?” from our last podcast.

Is it obvious?

The moon has had it coming.

Is it obvious? The moon's had it coming for years.

We have an AMA section that we're going to hit in a few minutes, and that's one of the questions being asked: don't we need to save the moon?

It's lunacy, Salim.

Touché.

Oh, goodness.

All right.

All right, the third Amazon story here is that Amazon opens an $11 billion AI data center in rural Indiana. We've heard about this already. It's running 500,000 Trainium2 chips. So how do Trainium chips compare to the TPUs and NVIDIA's GPUs? What do you guys think about this?

We used a bunch of them previously. Trainium2 chips are equivalent to the Hoppers, and they're good for inference, but they're much more difficult to do large-scale training runs on.

If you look at the breakdown now, you have a core cluster for training, and Anthropic just announced another big NVIDIA deal: $10 billion with Microsoft and NVIDIA. But for serving up Claude, you always hit those capacity constraints, and Trainium is very solid for inference.

It's similar to how Amazon previously went all in with Graviton, which was their CPU equivalent, and now that runs massive workloads for Netflix and everyone around the world. So I think it's still one more generation until Amazon starts to catch up. Again, they're about a generation behind, but all those chips are going to be used, probably—

Yeah.

—for inference versus actual training.

I have a crazy question here. If your model is this closely bound to the chip, then if you did an inference model for any of these big hyperscalers on Trainium versus a TPU, would you get a very different result because the chip is different?

No. You typically use a framework like OpenXLA, which automatically translates it to different things once it's actually doing the inference, because the process of inference is quite straightforward: matrix multiplications.

The process of training can be really complicated in the way that things move back and forth, et cetera, and that's where you really need to have high resilience and high interconnect. Whereas if it's a single chip or a group of 8 to 16 chips, as these are, they're just doing forward passes. It's a lot easier to code and to have speed on.

But again, there are certain things, like Cerebras, for example, that will give you much faster inference, or a highly optimized Grace Blackwell, et cetera.

So that process is much simpler than training.

Yeah, maybe to expand on that, backprop is the key problem. If we could do away with backpropagation at training time and have some sort of magical—I remember Boltzmann machines were one sort of concept for how we could do away with global backpropagation. If we could do away with backprop entirely, then one could imagine a near future where training looks a lot more like inference, and training would be a lot more portable and a lot more parallelizable. But no one has yet in production figured out how to do away with backprop.

But aren't LLMs fundamentally anchored to backpropagation?

At training time.

At training time, yeah.

Not at inference time. Inference time is only—

Sure.

Forward propagation. So if we could figure out how to train—

But backpropagation is fundamental to training.

Backpropagation is fundamental to training of neural networks.

For now.

For now.

Yeah.

But there are lots of paradigms. There's a whole cottage industry of researchers trying to figure out ways to eliminate backpropagation entirely. If we could eliminate backpropagation, that would certainly eliminate a training-time compute bottleneck.

And, by the way, just a reminder: if you're listening and you've just heard a conversation that you think is being spoken in Greek, my suggestion is: take—

Join the club.

Take some notes and go to your favorite LLM and have a conversation with—

As Bruce Willis said in Die Hard, “Welcome to the party, pal.”

Yeah. I'm going to hit what you said earlier, Salim, and Alex. I think the most significant thing about this is going from farmland to 7 buildings in 1 year with 2.2 gigawatts. I mean, it's just the beginning, and we're knocking down regulations and capital is flowing in. This is continuing.

All right, I want to get to our AMA. I want to hit a couple of stories on the science side real quick. We've been talking about launch costs. We've been talking about launching data centers. We've been talking about going to the Moon. I want to give folks a little bit of an overview for a moment about how quickly the cost of launch has been changing.

5. Space Launch Costs Collapse

The space shuttle, which was originally supposed to cost about $50 million per launch and launch 50 times per year, ended up costing somewhere between $1 billion and $2 billion per launch and was launching anywhere from 1 to 4 times per year. Massively expensive—$50,000 per kilogram, super-high cost. Falcon 9 comes in and drops the cost at least 20-fold to $2,500 per kilogram by making the first stage fully reusable, right? It's got 9 Merlin engines, so you're recovering most of the engines on the Falcon 9.

And then here comes Starship, which is reducing it again by another 25-fold, to $100 per kilogram. So, you know, how many kilograms does each of us weigh? And what's your cost to get into orbit? It becomes affordable all of a sudden, right? So as Starship becomes fully reusable—

And then Elon comes and starts speaking about the work of Gerard K. O'Neill. Gerry O'Neill at Princeton University had actually designed and built, at least on the ground here, what are called mass drivers—electromagnetic rings that accelerate a bucket to lunar escape velocity. And just for the cost of electricity, which, by the way, on the Moon is relatively cheap because you've got all the solar flux, you can accelerate something and shoot it towards the Earth into an Earth acquisition orbit. And we get here the price coming down not 100-fold, but 1,000-fold, to 10 cents per kilogram.

So all of a sudden, we gain access to all the resources on Earth. I like to remind people that everything we hold of value on Earth—metals, energy, real estate, all these things—is in near-infinite quantities in space. So the 9-year-old space geek in me is super excited for what's coming. Alex, you want to add anything?

Yeah. I'll add that disassembling the solar system is going to require low cost to orbit, so this is great.

Alex, you're going to start protests outside our front door.

Yeah. Dyson sphere in the way. I like that this is generated by Nana Banana as well. You can see this little thing.

Yes, of course. All right.

I mean, basically, this turns lunar launches and rocket launches into software, right?

Mm.

There are 2 things here. One is, importantly, this is a log scale. So for folks watching, this is ridiculous orders of magnitude per level, and that's unbelievable in a very physical environment. This is not some social media gaming Silicon Valley play. This is getting out of the physical gravity of Earth's gravity well. This is nuts.

This is energy, baby. The one complaint I have about my conversations with Elon is he wants to get out of Earth's gravity well and then go directly back into Mars' gravity well. I'm far more interested in staying either in the Earth-Moon system or, better yet, building what some have called O'Neill colonies.

I mean, sure. If you want to start with the asteroid belt, we can start there. That's fine. Training wheels for solar system disassembly. That's great.

No. All right, listen. So disassemble asteroids and build large rotating cylinders called O'Neill colonies where you live on the inside—omega squared R.

6. Brain Interfaces Reach Humans

One more story before we get to our Q&A, which is a story from a friend, Matt Engel, the CEO of Paradromics. Paradromics has been one of a significant number of BCI companies, brain-computer interface companies. And what's interesting about them is they've just completed their testing in sheep.

Neuralink did its testing in macaque monkeys. Paradromics has done its testing in sheep, and they've been approved to go into humans, which they will do in about 2 months' time, in the early January-February timeframe. And I think what's most interesting is that they've been able to hit a speed about 10 times, or actually 20 times, faster than Neuralink.

So Neuralink's been at about 10 bits per second. The Paradromics implant is at 200 bits per second. So, Salim, you and I have always talked about whether Ray's prediction on high-bandwidth BCI by the early 2030s—2033—is going to happen. So we're seeing all these companies moving forward here.

A few years ago, I was a hard no on that. And now I'm like, “Oh, shit. He's right again.”

Alex, what are your thoughts here, buddy?

Yeah. I think we're seeing the BCI space become competitive, which is great. Yes, we should all get our “Ray was right” hats. Fine. But I think if you extrapolate this, one of my more fun thought experiments is: when do we actually get our nanobots in the brain for high-throughput Moravec-process-type BCI?

And it's interesting. You can look at the cost of producing a gigaflop of compute versus the typical size of a gigaflop of compute. When Apple introduced the iMac, the first iMac was about a gigaflop. The first iPhone and the first Apple Watch—there's something magical about rolling out a form factor with about a gigaflop.

You extrapolate that curve naively, assuming exponential progress for a gigaflop, saying, “Gigaflop—that's the threshold at which we have useful general-purpose computing,” including for the purpose of maybe even substituting for human brain cells in the context of high-throughput BCI and/or a very invasive uploading scenario. That curve gets you to 2045, which is again—#RayIsRightHat—that's when you get about a gigaflop, the size of a human brain cell.

So I do think we're very much on trajectory for Ray-is-right-style human mind uploading, invasive BCIs, and noninvasive BCIs. This is a quasi-invasive BCI. I think we're also going to get lots of wearable—

By the way—

We have Ray coming on soon.

Ray's going to be joining the podcast in early January to talk about his predictions for 2026. We'll also have Brett Adcock coming on the podcast to talk about—

Wait, wait, wait, wait. We can't ask Ray about 2026. We have to ask him about 2066. I mean—it's too soon; it's a waste of time.

Too soon.

Too soon.

We'll ask about all of it. And—

We're going to need a bigger podcast.

Yeah. Well, we'll get one by then.

By the way, we've crossed 400,000 subscribers, so thank you to all those who subscribed to push us over. Our next hill is 500,000, then we're going for that million. Why? Because Chet and Dax want us to get a million subscribers.

This is, whatever chemistry we have here in terms of processing the news and making sense of it for others, it's seeming to really resonate.

The number of calls, accolades, and feedback I'm getting is really—

Yeah.

I'm sure you're seeing the same thing.

Thank you to our listeners for the feedback. We do read all of your comments, and in fact, we process the comments and pull out the questions. We're about to jump into that segment with an AMA. But Emad, what are your thoughts on BCI?

I think this year has been a breakthrough year. Next year, you'll see even bigger advances. We've both seen what else is going on behind the scenes, and I think it'll probably be one of the biggest investment areas in the next 3 years, actually. What could be better than solving the issues and then augmenting humanity directly? I think, as Elon said, the only way you're going to be able to keep up with the AGIs is to plug in.

Yeah.

And so—

Yeah.

It's going to be of geostrategic importance as well as financial importance.

There's something fundamentally interesting about the brain because we still have very little idea how it works. But as long as we can interface with it effectively, that's very, very powerful. Our memories are already outsourced to our smartphones. We don't really use our memory neurons in the same way we used to, and therefore we'll start doing that with more and more brain-function capacity, releasing that load from the brain and using it for other things. So I'm really excited about what comes with this.

Maybe just to add quickly to Salim's point, this is admittedly a bit of a hot take, but arguably, we solved AGI. We solved superintelligence without actually having a good mechanistic understanding of natural intelligence. I think it's pretty likely we're going to solve brain-computer interfaces and maybe even whole-brain emulation without still having a detailed mechanistic understanding of the human brain, though you can get pretty far with phenomenology.

Hmm.

Do you, Alex, think we can use AI to solve the hard problem of consciousness, the whole qualia thing?

Yes.

Okay. We're going to have a conversation about that in terms—

Of how that goes about. But let's take that offline.

All right. Two quick points—

I'm not going to go there now.

Two quick points on this BCI. Number 1: amazing people are playing in this space, right? Max Hodak, who is the co-founder of Neuralink, now has a company called Science Go and check it out. They have a completely different approach to interfacing between the compute world and your neocortex. Brilliant. Basically, they're using neural stem cells to grow nerve endings into the brain that wire together and fire together.

Sam Altman invested in something called Merge Labs. It's still under wraps, but we'll be hearing a lot more about Merge in the next few months. So I have one final question. Ray Kurzweil's prediction on high-bandwidth BCI is really dependent on having nanotechnology. The question is: Where are we on that front? I'm still waiting to hear some good updates on the ability to assemble molecules atom by atom—not with wet nanotechnology, which is biology, but with assemblers like Eric Drexler spoke about. Alex, any thoughts there?

I spent so many years chasing nano-assemblers. I do think we're going to get to Drexlerian-style nanotechnology, although even Eric Drexler had a personal evolution. We've chatted—I’ve chatted a number of times with him about this—from pure diamondoid-style, quote-unquote, molecular assemblers to the nanosystems phase, where it's not about self-replicating nanorobots. It's more about desktop factories that produce things.

Here's what I think. I think by, at the very latest—and this is, in my mind, an ultra-conservative outer bound—2045, we get our Drexlerian nano-assemblers. I actually think we're far likelier to get them in some soft form. Maybe they'll look like DNA origami; maybe they'll look like AI solving the Feynman Grand Challenge, which includes both computational and nanorobotic challenges. I think we're likely to get some AI solution to early-style Drexlerian nanotech in the next 10 years. I don't think it's going to take that long, but at the same time—

There's our 2035 date.

Yeah, everything gets solved in the next 10 years.

I don't think you need that high bandwidth, to be honest. We did work at Stability on Mind's Eye, where we reconstructed images people saw from MRIs, which is incredibly low bandwidth. And if you look at the forward-backward diffusion processes, what you're likely to have is, before you get to the full bandwidth, partial bandwidth that can effectively reconstruct brain processes with very little information. Then you'll just run diffusion models to do that in a similar way to what Sunday Robotics and others have done going forward.

So I think—

All right. There was a project—

Hey, I have to throw this out. There was a project out of Japan called Dreamcatcher, and what they were doing was having you sleep in an MRI machine. They were storing the images coming off your optic nerve and then replaying your dreams back to you the next day, which was hugely unnerving.

Very quickly on this one, Peter: you get approximately 1,000,000 voxels per second just streaming off. You can do high-bandwidth decoding of thought with 1,000,000 voxels per second.

I just don't have my—

You don't need nanotech.

I just don't have my portable fMRI machine to carry around in my—

Yeah, but you will.

In my backpack.

By the way, there is a team I've been talking to that seems to have a credible path for molecular manufacturing. So—

All right.

If you're happy to connect them with you, I'd happily do that.

I can't wait. All right, let's get into some of our questions from our subscribers here. Let's jump in. The first one is from David Bowman6224. David says, “I'd like to hear AWG tackle Emad's thousand-day prediction.” So what does AWG think of Emad's AGI in a thousand-day prediction? Emad, do you want to state your prediction first, and then I'd love to hear Alex's commentary.

I was just saying that I think most human economic work is negative value within 1,000 days. Well, 900 days now left at most. It's not that it will replace all the jobs, but definitely it'll be there for just any job that can be done on the other side of a keyboard or mouse. So that's a weaker version of AGI than in some cases.

Alex.

Yeah. I think the central challenge, as always, is defining what we mean by AGI. I think if AGI means generality, I think we've had AGI since, at the very latest, summer 2020, when GPT-3 and the “Language Models are Few-Shot Learners” paper came out.

If AGI means some sort of economic parallel with humanity, I agree that either it is the case, Schumpeter-style, that we already have some sort of economic generality—for example, as parameterized by OpenAI's GDPval benchmark—or, if you believe that benchmark, economically general AI is either already here or imminent, like in the next few months. If you have some other preferred benchmark for human economic output, it's probably imminent, if not already here.

All right. Let's go to the next question. It's from Josh.

Insert my standard rant about AGI.

Okay, it's—

Incorporated by reference.

So acknowledged.

Thank you.

@JoshS5937 says, “What is the future of land ownership in a future without scarcity? Land is finite. Will it remain the final scarce resource?”

Josh, it's a good question. The way I answer it is 2 different ways. Number 1, we're going to be spending a lot of time in the virtual world, and there you'll be able to gain access to unique virtual real estate. The second is, you're thinking with a very Earth-centric point of view. There's the Moon, there's Mars, there are massive O'Neill colonies built out of asteroids, and we're going to start to see humanity migrate outside the Earth. Having said that, yes, Central Park West apartments are still going to be scarce. So—

Deeply disagree. I want to rant on this.

Okay, go for it, Emad.

I did some fact-checking here. It turns out there's about 16 billion acres of habitable land on Earth. That's about 2 acres per person, okay? That's a pretty decent number. And that's habitable. Let's note that passenger drones are going to make difficult-to-reach areas very habitable, so that goes up to about 20 billion to 22 billion acres of habitable land.

Technology will expand the amount of habitable and reachable land. But still, if we get to about 10 billion, we'll peak at about 10 billion people by 2050 before we start dropping off. That's still about 2 acres per person, which is a pretty decent number. All of this technology is allowing us to reach that land more easily and make that land more usable. And if you fly across India, the most populous country in the world, it's mostly empty.

Yeah.

Right? You see populations at the edge, on the coast. The middle—there’s nothing there. Same with Africa, same with the U.S. You fly across the U.S., and there’s nobody there in the middle. I’m Canadian; there’s nobody in Canada.

So there’s a lot of land that we can use, and technology makes it much more accessible: temperature, HVAC, heating, and cooling. The only constraint is energy and compute, as Alex would say.

Amazing. That’s a great point.

Yeah. I think—

Go ahead. You—

I’d like to give listeners a practical example. Waymo is now basically legal across San Francisco, right? And so that could completely change where people live, because you can just get into your Waymo and—

That’s right.

…it will just take you and your kids anywhere. And what’s coming—

Sorry, just to add to that—

Yeah, yeah.

The prediction for Tesla is that it’ll be about 30 cents a mile to get somewhere in a robotaxi. That’s near zero compared to where we are now. It’s a 10X drop from where we are now.

Alex, I need you to—

Continue.

Yeah, and—

Maybe, maybe just to add a bit of nuance to this. I think in the short to medium term, land is becoming post-scarce. As you say, we can build up, we can build down, and we can build on other planets.

One important use case that hasn’t been touched on is that we’re going to have so many humans, I think, uploaded in one form or another into the cloud. The cloud doesn’t have the same concept of land. So I think, in the short to medium term, land is post-scarce.

In the long term, I think the scarcity of land depends on whether AI economies have a better use for land than we do. If we do find ourselves taking apart the solar system, land could actually become really scarce in the end.

Yeah. By the way, let’s just talk for one second about uploading. When do you actually believe we’re going to start to see human uploads to the point where you, Alex, say, “Okay, upload me,” and this speaker comes over—this voice comes over—and says, “Hey, Alex, I’ve been uploaded. You can off yourself now. We don’t need your biological body anymore. I’m in the cloud”?

Well, thanks for the vote of confidence. I think we’ve already seen noninvasive uploading in the form of large language models. Large language models are arguably an upload of an ensemble of all of humanity.

Yeah, yeah.

In terms of individual uploading that’s noninvasive, I think we’re either there already in the form that Emad touched on earlier or alluded to, like constructing foundation models from fMRI scans. There are a number of groups that are training foundation models from fMRI scans. Arguably, those are low-fidelity, noninvasive facsimiles of human minds. I think we’re going to get to—

Wait, wait, wait, wait, wait, wait. Hold on a second. There are people training LLMs on fMRI scans?

Correct. A number of groups now, including Meta, by the way. Really well-financed, really talented groups.

Holy crap. Okay.

So the real idea about uploading—

The implications of that—

The implication is that LLMs are trained to reproduce the behavior of humans—fat biological meat fingers tapping keys on a keyboard, uploading text to the internet.

Yeah.

But with foundation models trained off of fMRI data—on the order of a million voxels per second—you can imagine pretraining a foundation model that basically encapsulates human thought. Certainly, for human thought-decoding purposes, you get that, but also—

Got it.

…fMRIs can—

And fMRIs can track a single neuron firing in real time.

No. fMRIs are both spatially low-resolution and temporally low-resolution. You get approximately 1- to 2-second temporal resolution and, at best, approximately 1-cubic-millimeter spatial resolution. Nonetheless, it turns out to be enough for thought decoding.

So, Alex, the concept around a true upload is: can I actually map your connectome? Can I map, for a human, roughly not only the 100 billion neurons but the 100 trillion synaptic connections? That’s typically done by slicing the brain into ever-thinner slices and using AI to map those interconnections. It’s a destructive process. Do you think we’re going to have a—

It’s really invasive right now.

Yes, very invasive.

Yeah. Here’s my brain. Slice it into 1,000 pieces.

Yeah, more like a billion pieces.

I think we’re going to have reference nonhuman organisms. Drosophila—major progress already.

Which is done.

Fruit flies.

Yeah, we’re done.

Fruit flies, done.

Yeah.

Mice—

Next.

…are about to be done. There have been 1- to 3-cubic-millimeter sections of mouse brain uploaded or scanned in some form or another, to the extent that the connectome is a proxy for uploading. Done. I think mice overall are going to be done shortly. And so—

Lobsters are next, right?

Lobsters are easier, interestingly. Mice are harder. I think we’re going to see the full human high-resolution connectome probably in the next 5 years.

Alex, aren’t you an advisor to Nectome?

I’m not a formal advisor to Nectome, but I am an advisor to a company, Eon Systems, that is working on solving human whole-brain emulation and uploading.

All right, I’m going to move us forward here onto our next AMA question. I want to close out with, “What are you most thankful for in 2025?” So start thinking about that in the background.

@SuccessCoach, Cody, writes, “How do we prevent a world where millions fall into poverty before AI-driven abundance arrives? What are the real solutions for people who may lose their jobs long before the long-term benefits of AI kick in?” @JNKind5 asks a similar question: “You often say AI will lift up people at the bottom. How exactly will that happen for those who can’t meet basic needs like food and healthcare today?”

We hit on this about 2 pods ago, where there are concerns—and we saw this in the data from the FII9 event—about poverty, losing jobs, and being able to support your cost of living. Salim, do you want to jump in on this one?

Wow. There are about 20 questions buried in each of these. I think we are in a difficult 10-year period as we transition all of the world’s systems from scarcity to abundance, right? Consider the fact that almost every business in the world is focused on scarcity. For the last 10,000 years, if we didn’t have scarcity, you kind of didn’t have a business.

We’re moving now to abundance models, and exponential organizations actually find business models around abundance, which is the starting point of that transition. But for society at large, we need to move to some model, whether it’s UBI or UBS—

UBS meaning?

Emad is pushing universal basic services, right? A similar type of concept where you just give basic capabilities and make them available to everybody, solving the bottom 2 layers of Maslow’s hierarchy.

The trick for UBI, by the way, for people who are naysayers, is that if you can find the balance where people can survive but not be happy, you still have a very thriving economy. Entrepreneurship explodes in that model, et cetera, et cetera. So we’re not far away.

The problem is governments, and getting governments to move from a union-labor, job-taxation model to that is such a big leap. We don’t have confidence in governments doing that. The problem with government all over the world is that they want to be needed. They ran a 2-year UBI program in Manitoba in the ’70s, and it was so successful that, at some point, the government realized, “We’re not even needed here,” and they canceled the program so that they could be needed. That’s the immune-system problem in government that has to be solved.

So I tend to be on the optimistic side. Technology uplifts people at the bottom. People are leveraging technology to make more and more money in the short term. We’ve got lots of data around that, and as we get technology democratized and demonetized for a broader population, everybody lifts up.

You and I talk—all of us talk—all the time, but forget the richest people. If you can lift the bottom, that’s the key, and the bottom is being lifted very, very appreciably. You just don’t see it that obviously, but it is being lifted.

People compare themselves against the Kardashians or whomever else. Emad, you’ve been doing incredible work with Intelligent Internet on this specific problem. Could you lay that out and give us your thoughts here?

Yeah. I think, as with many things in human life, this is a coordination problem, right? Again, we have enough resources—2 acres per person, food, healthcare, et cetera—to coordinate everyone.

But we've always lacked the capability to do so because our systems are dumb. So we have projects like Sage, which we launched at FII to do top-down policy, and really, the way that I've been thinking about it more is like AI social scientists. We talk a lot about AI scientists for biology, for chemistry, for quantum. AI social scientists to figure out economics, politics, and implementation are going to be so huge. That's basically our Sage project.

On the other side, I think you need to have universal AI given to everyone—a Jarvis that's looking out for people to help them navigate on an individual basis—because that's how they get access to food, healthcare, et cetera. The reason they don't now is because people are invisible, particularly the poorest of people. But the pace at which this is going to come over the next few years is going to be so intense that governments need to take a big step forward and say, “A, we need to use AI to coordinate this; B, we need to get AI to the people; and C, we need to look at historical counterparts.” I think probably you need to look at the 1933 New Deal that came out of the Great Depression and others, because you might see entire industries disappear within a matter of days or months.

Grok 4.1 Fast just scored like 95% on TaoBench, the customer service benchmark, and it's 50 cents per million words. It's better than any human. That would mean no customer service jobs within 2 years. Again, it takes a little while, but it's one way. So coordinate with AI, give everyone universal AI, and then layer services and coordination on top of that.

You're going to appear as the headline in some news article now, Emad, if you remember you said, “No more coding in a year.” Headlines across India for that.

I think the point that Emad makes right there is the really killer point: this is one way. We're not going back.

Yes.

We have to face the future that's coming, and let's get real about it. Let's get data-driven and evidentiary around it and just fricking make it happen. Left to itself, we've got these 2 futures: a Mad Max future or a Star Trek future.

Mm.

Right? You can see our politicians pulling us straight to Mad Max. We have the opportunity with technology to pull us in that direction. This is what this community is about. This is what we have to do.

Alex, closing thought on this one.

Yeah. Closing thought: I think the central policy challenge is growing the overall economy much faster than the value of conventional human labor is destroyed or obsoleted by AI. So I'm primarily focused on ensuring that we can achieve radical macroeconomic growth.

If we can do that, then making sure that UBI, UBS, or UBE—universal basic equity—or some other variant thereof, some door number 4, I think all become more a matter of policy decisions. But it's relatively easy to distribute abundance if we have abundance.

Yeah. All right, we're going to close out on a question aimed at you, Alex. This is from @xPhoenix96: “Hey, Alex, the Moon and Jupiter should be off-limits to mining. Don't they stabilize our environment?” What are you trying to do, Alex? Start a revolution?

Gosh, we're having an “All these worlds are yours except Europa. Attempt no landing there” moment, I think—if you've read 2010: Odyssey Two by Arthur C. Clarke. No, we don't need to stop mining the Moon and Jupiter to stabilize our environment. Jupiter does, at the moment, play an important role in protecting the inner solar system from Oort Cloud bodies and other objects from the outer solar system. The Moon does play, for the moment, an important role in the tides and other sort of atmospheric phenomena—

And romantic love.

But for the moment—

And romantic love.

“For the moment” is doing the heavy lifting in that sentence. So once we have the ability—which I think seems likely we will increasingly have—to disassemble the Moon and disassemble Jupiter, and assuming the solar system does go in that route, we will also have the ability to protect the inner solar system from a variety of asteroidal bodies and to recreate the tides artificially.

My favorite quote from you, Alex, is, “Saturn has had it coming for a long time.” That's got to be an all-time Alex quote.

It's true.

Oh, goodness. All right. Well, asteroids represent a significant amount of mass, and I think they can handle our needs for at least a decade or 2—

If you like that sort of thing.

I want to close out with a question here. What are you guys grateful for having happened in 2025? As a closing gratitude, I'll kick it off. I'm super excited that humanoid robots have made so much progress. The capital's being invested, the manufacturing plants are being invested, and my own version of Data or C-3PO is on its way. Alex, how about you?

So many things, but I'll pick one. I'm grateful that math is credibly and definitively being solved by AI. That is, in my mind, such a canary that this is going to work, that the singularity is in progress. We're going to solve all of the grand challenges in math, science, engineering, and medicine over the next few years, and math is just the tip of the iceberg. It's very exciting.

Amazing. Salim.

Again, a million things. I think 3 things pop to mind. One is I'm unbelievably grateful for this podcast. Peter, thank you for pulling it together.

Oh, I am too. Thank you—

Missing Dave a lot at this moment.

And thank you—let's just say it real quick—to Nick, to Nick Singh, to Dana—

The team.

And to Gianluca, who helped make this really excellent.

Yeah.

So thank you guys for that.

I think this radically optimistic, realistic view of the future is the most important kind of tonic for what's happening out in the world today. This kind of palpable relief from all the listeners going, “Well, thank God there's something I look forward to every week or few,” that's number 1.

Number 2, I think I'm starting to just wallow in gratitude on a near-permanent basis, thinking about the incredible future that's appearing in front of us, driven by that inner loop that Alex talked about. I'm still a fan of the Moon for the moment, so let's not put that too far aside.

Enjoy it while it lasts.

Enjoy. I think the third would be my ExO ecosystem finally gelling in a really powerful way. It's been like 10 years of building this ecosystem. If I ever say in the future, “I want to build an ecosystem,” please, somebody get a baseball bat and take me behind a woodshed. It's unbelievably difficult, but it's actually now coming together in a very, very powerful way.

Nice. Emad, what do you come out on your gratitudes?

It's a nice small one, Salim—a small question you're answering.

I go for niche projects.

Yeah. I think there's 2 big things. One is that I think we've had the technological breakthroughs and infrastructure breakthroughs to be able to build the AI social scientists, to improve our infrastructure and finally coordinate as a species. That is a huge thing that we'll start seeing rolling out and announced next year as well.

And number 2, I think minus the hard light, we have all the tools we need now for the holodeck.

Ah, awesome.

We just have to put that together.

I'm going to add one final gratitude to close us out here, which is the incredible progress being made on reaching longevity escape velocity. The focus by all the hyperscalers and model builders is: How do we understand how to add decades of health into our lives? How do we, as Dario says, double the human lifespan in the next 5 to 10 years? That gets me jazzed. You know why? Because I'm excited to see the Star Trek future coming our way.

As long as I get a phaser.

Ah, that was epic. I just don't like wearing a red shirt on some of those planets.

No, don't.

Yeah.

If I can be likened to Picard in any way, I'm good.

And Alex, of course, you're the science officer on all the missions here.

Obviously.

Everybody, I wish you an incredible Thanksgiving holiday. To all of our listeners and to my Moonshot mates—Dave, we missed you on this episode. Looking forward to seeing you recording again early next week.

A lot is going on, and we're going to be spending some time with Mustafa Suleyman as well, the CEO of Microsoft AI. We're doing a podcast with him. A lot of incredible things. Get ready—2026 is going to rock the planet. Hopefully not physically, but definitely emotionally and intellectually.

Let's all wallow in gratitude the next few days.

Yeah. Beautiful. And stuffing and turkey.

All right.

Take care, everybody.

Take care, folks.