The AI Wealth Gap: Why 40x Deflation Changes Everything w/ Dave Blundin, Salim Ismail, Dr. Alex Wissner-Gross | EP #208
Peter DiamandisDave BlundinSalim IsmailDr. Alex Wissner-Gross
AI intelligence costs are reportedly deflating 40× year over year, turning capability into the episode’s “nuclear core” for broader price declines. Moonshot AI’s Kimi models—described as within roughly 1% of the top of SWE-bench—reportedly cost $4.6 million to train, about 30–40× below earlier OpenAI and Anthropic models. If that curve persists while demand grows 1,000× annually, advanced intelligence stops being reserved for hyperscalers and starts pulling healthcare, food, energy, and other costs down with it.
The near-term risk is that abundance arrives after AI has already widened the wealth gap. A 60,000-person survey across 32 countries put cost of living first globally, followed by unemployment and social inequity; Dave relayed an Iranian intern’s example of her parents spending one-third of annual income on an iPhone and data plan because “you can’t live without information.” Peter expects long-run demonetization but sees the next two to seven years as the danger zone; Dave warned that adding AI could make “the gap ... really, really wide.”
Anthropic’s enterprise momentum is both a business-model signal and a bet that code generation unlocks recursive self-improvement. Dave said banks trust Claude with sensitive data and called AI management tools “the goldmine of all goldmines”; Peter cited projections of $70 billion revenue and $17 billion cash flow in 2028. Alex kept the technical call explicitly unresolved: code may rewrite models into an accelerating flywheel, or broad intelligence may still require “visual chain of thought” and physical grounding.
Power infrastructure is becoming the most legible picks-and-shovels trade in the AI buildout. One-gigawatt data centers are expected in 2026, while an $80 billion public-private nuclear initiative may build roughly ten 1.1-gigawatt AP1000 reactors—too late to answer the cited need for 92 gigawatts by 2030. Dave expects as much as $1.2 trillion annually to flow into data centers and power by 2030: “There’s nothing even close in the history of the world to that scale of money movement.”
AI science is moving from assistance toward bulk discovery, with timelines collapsing from decades to years. Sam Altman described GPT-5 as showing “tiny glimmers” of new science and said GPT-6 might deliver a GPT-3-to-4-like leap; Alex expects many or most grand challenges to start falling within three years. Edison’s Cosmos reportedly compresses four to six months of expert research into 12 hours, while disease-cure programs are shifting their stated horizon toward roughly 2030.
Regulation and labor resistance could determine which regions capture the productivity dividend. Peter cited European AI audits averaging €260,000 and 8–15 months, with European models reaching market 6–12 months later; Boston unions, meanwhile, want human safety drivers inside Waymos. Alex framed the missing technology as a system that achieves “acceleration and social cohesion at the same time,” because blocking deployment may preserve specific jobs temporarily while pushing investment elsewhere.
The same deflationary engine is opening physical and biological frontiers faster than governance can absorb them. World models can create traversable synthetic environments, coordinated drone swarms could act in the physical world, AI-guided weather control could steer storms, and embryo editing is moving from selection toward alteration. The episode repeatedly returned to legitimacy; in the weather-control discussion, Alex said deployment was “a social problem, less a technical problem.”
1. Anthropic’s enterprise wedge doubles as a bet on recursive AI
The episode’s Anthropic market-share chart led Alex to the decisive technical question: “whether code generation is the critical path to recursive self-improvement.” Anthropic has concentrated on code while OpenAI pursues modalities such as Sora; if code is the bottleneck, Anthropic’s enterprise gains may also indicate strategic proximity to a much larger capability breakthrough.
Alex’s honest non-answer preserved both cases. Code could rewrite “the core algorithms” and post-training architectures until “the flywheel spins faster and faster”; alternatively, superintelligence might require visual chain of thought or grounding in the physical world that cannot emerge from source code, text tokens, and limited imagery alone.
Peter took the aggressive side of that uncertainty, arguing Dario Amodei or Demis Hassabis could reach the next regime by directing enormous internal compute toward generating “the next test, the next test, the next test.” His marker was Humanity’s Last Exam: he said reports that it may soon be saturated matter because solving those questions and innovating in AI appear highly correlated.
Dave’s commercial read was simpler: banks and other enterprises use Claude because they trust Anthropic with sensitive data, while OpenAI leans consumer; “using AI as a management tool is the goldmine of all goldmines.” Peter cited Anthropic’s 2028 projections of $70 billion revenue, $17 billion cash flow, and a 77% margin, versus OpenAI’s $100 billion revenue target and losses through 2029.
2. Alignment becomes capability, while capital strategy determines reported profit
Alex described a “perverse duality between alignment and capabilities”: aligning a model closely with human intent is itself a valuable capability, while financing stronger alignment requires raising capital and training frontier systems. In his formulation, “every alignment project almost inevitably ends up as a capabilities project,” as happened first with OpenAI and then Anthropic.
Salim kept the institutional counterexample: Facebook began with strong privacy commitments before monetizing user information, so a safety-oriented origin does not prevent a company from becoming extractive. Dave nevertheless argued Anthropic has captured the moral high ground once associated with OpenAI—“if you want it to be guaranteed to be good for the world, come to Anthropic.”
Dave interpreted OpenAI’s prolonged losses as deliberate expectation-setting, not proof that its core service lacks margins. Sam Altman is effectively telling investors he intends to keep spending ahead of the curve; once a public company switches to profitability, asking shareholders to fund a trillion-dollar data center becomes far harder.
Peter compared that posture with Jeff Bezos’s early warning that Amazon would buy customers and revenue before “flipping the switch.” Dave expects Anthropic could similarly advertise a 77% gross margin, then launch an Anthropic-scale Stargate project that consumes it—sensible capital strategy, provided shareholders accept the bargain.
3. World models create a compute frontier beyond entertainment
World Labs’ Marble demo showed a photorealistic, traversable environment that Peter found both magical and troubling. His concern was not image quality but substitution: when users become “the god of your world,” an immersive simulation may compete directly with work, food, relationships, and ordinary reality rather than merely with existing games.
Alex called the consumer competition the beginning of the “Holodeck Wars,” but distinguished two architectures. Google Genie-3 generates every pixel on compute-intensive server GPUs; Marble generates 3D Gaussian splats—transparent blobs that accumulate into a 3D scene and can be rendered dynamically on the client with far less compute.
That creates an “efficient frontier” between versatility and computation, analogous to lightweight edge models versus server-intensive frontier models. Alex expects multiple tiers of generated worlds rather than one winning architecture, with economics determined partly by where rendering and generation occur.
Entertainment was “chicken feed” beside the addressable market Alex sees: synthetic environments for training vision-language-action models, robots, and scientific systems. Fei-Fei Li’s ability to demonstrate Marble on her phone in seconds reinforced Dave’s adjacent lesson—transformative products become easier to recruit, excite, and explain when their value is instantly visible.
4. Forgetting knowledge may make general intelligence radically smaller
Goodfire’s proposed unlearning method advances Alex’s vision of a “diamond-like perfect micro model” containing reasoning while externalizing almost all factual knowledge. The technique tests how small changes to weights affect overall loss: weights essential to generalization cause dramatic degradation, while weights holding memorized, possibly incorrect facts can potentially be pruned.
Peter initially framed the result as an enterprise privacy tool that could forget proprietary healthcare or financial records without losing the intelligence learned from them. Alex called that “half correct”: privacy may benefit, but the paper’s primary objective is separating general capability from memorized knowledge inside models containing hundreds of billions or low trillions of weights.
The holy grail would be a generally intelligent model with fewer than one billion parameters—or, aspirationally, around one million. Dave noted that 90% parameter reductions through distillation are already common: a 10× reduction could turn an assumed 100-gigawatt requirement into 10 gigawatts, making experimentation with open weights potentially worth trillions of dollars.
Google’s nested-learning paper attacks the complementary problem of continual adaptation. Alex described higher-order meta-learning—“learning to learn to learn to learn”—and a possible unification of models and optimizers as different layers of one information-compression process. Dave added that AI can connect its “brain” to notes and tools at immense bandwidth, making it functionally superhuman in learning workflows even before every internal faculty is superior.
5. Kimi puts frontier model development inside ordinary corporate budgets
Dave called Moonshot AI’s Kimi release “the biggest thing that happened in the last month.” He placed the models near the top of SWE-bench, roughly 1% behind Anthropic, while emphasizing their ability to run at high speed on Groq hardware—“G-R-O-Q, not Elon”—and their availability as open weights.
With Meta and OpenAI no longer providing their best models as open weights, Dave said the strongest editable systems are now coming from China. That matters to researchers and companies that want to “play with the guts” of a frontier model rather than merely buy API access.
The reported $4.6 million training cost was the key economic discontinuity: Dave estimated it was 30–40× cheaper than what OpenAI and Anthropic spent on their original models. Some savings come from drafting behind earlier innovations, but the result is that almost any company can now finance a serious training run.
Peter drew the geopolitical conclusion: if a trillion-parameter model costs about $5 million to train, efficient US capital markets cease to be a prerequisite for frontier work. Dave agreed the implications were “absolutely massive,” while adding a practical hedge: he had checked and was using Kimi, but users should still conduct their own spyware and security checks.
6. Forty-fold deflation is the engine pulling every other cost downward
Alex attributed to Sam Altman a 40× year-over-year decline in “cost of intelligence per unit of intelligence,” spanning training and inference. Salim contrasted that with demand rising roughly 1,000× annually, which explains why hyperscalers can keep increasing capital expenditure even as each unit becomes dramatically cheaper.
Alex called cheap intelligence the “nuclear core” of abundance: if sustained, it should drag down the prices of everything intelligence can materially improve. His forecast was that grand challenges in mathematics, science, engineering, and medicine would begin falling within two to three years under this compounding pressure.
Peter’s warning to skeptics was unusually categorical. People extrapolate from yesterday’s awkward model and diminishing benchmark increments, but two consecutive 40× improvements create a 1,600× change in economics; betting that this produces only a small capability gain is, in his words, “crazy.”
Salim’s counterpoint was historical rather than bearish: deflationary curves are normal in scalable technology, although 40× is faster than he expected. Carmakers dismissed electric vehicles because batteries were too expensive, then lithium-ion costs fell 90% over a decade; the lesson was to “go where the curve is pointing you.” Peter retained the condition: wider costs fall only where intelligence can solve the underlying problem.
7. Europe’s compliance burden is becoming a strategic tax on AI
Peter cited venture funding in Europe falling by as much as 30%, European AI models reaching market 6–12 months behind US counterparts, and mandatory audits averaging €260,000 plus 8–15 months. Those reviews reportedly delay 40% of projects while examining data, transparency, bias, documentation, and safety.
Salim argued Europe’s constraints are not merely cultural. His example was a postwar German constitutional restriction preventing one media organization from covering the entire country; regional fragmentation then left room for Google to aggregate the market. Undoing protections rooted in legitimate history is difficult even after their competitive side effects become clear.
Alex reduced the issue to sovereign choice: countries presently define “how much they want to participate in the superintelligence explosion.” Peter told European leaders they need energy plans, data-center sites, and deployable infrastructure within five years; Alex’s pushback was sharper—“more like five months than five years.”
8. The wealth gap may widen during the transition to abundance
Peter’s FII Priority Global Survey covered more than 60,000 respondents in 32 countries, representing roughly two-thirds of the world’s population. Cost of living ranked first globally, unemployment followed closely, and poverty and social inequity came third; Africa’s top concern was unemployment, while every other listed region led with living costs.
Dave grounded the charts in an Iranian intern’s example: her parents spend one-third of annual income on an iPhone and data plan. The intern’s explanation was that “you can’t live without information,” while an unusable currency made Bitcoin necessary and the phone essential for managing it. The spending then exits the country toward technology centers, concentrating wealth before AI adds another paid layer.
Salim argued that jobs are disappearing as a durable organizing principle while education still trains children into their early 20s for a labor market nobody can describe five years ahead. His preferred bridge is UBI; Peter leaned toward universal basic services and warned that the two-to-seven-year transition, before healthcare and education become abundant, is his greatest concern.
The disagreement was about execution, not the fear’s legitimacy. Salim wanted far more positive narratives because people are “10X more likely” to attend to fear; Alex wanted measurable targets—a cost-of-living benchmark, plus benchmarks for healthcare and crime—so 40× intelligence deflation can optimize toward outcomes rather than slogans. He separately suggested that social cohesion may need its own benchmark.
9. Coherent power is the trillion-dollar variable in frontier AI
With one-gigawatt facilities expected in 2026, Alex asked whether coherent training power eventually peaks. Distributed training innovations could cap clusters at a few gigawatts and then lower required density; further intelligence “phase changes” from ever-larger compression could instead drive facilities toward his extreme metaphor of a “desktop black hole computer.”
Peter cited an $80 billion US government, Brookfield, and Cameco partnership involving Westinghouse AP1000 reactors. Each Generation III+ plant provides about 1.1 gigawatts and may cost $7 billion, allowing perhaps ten reactors—but early-to-mid-2030s delivery does not answer the cited requirement for 92 gigawatts by 2030.
Alex positioned mature Generation III+ reactors inside a bridge from natural gas to nuclear fission and eventually fusion, with solar-plus-battery throughout. Unlike more experimental SMRs, of which Alex said there are perhaps only two or three, AP1000 is a relatively mature format; he cited at least six already built. The irony is that Westinghouse went bankrupt in 2017 building reactors before AI made dependable power strategically scarce.
Dave highlighted the financing template: private capital secures government-guaranteed loans, captures the upside if projects work, and limits downside if they fail. He projected $1.2 trillion annually flowing into data-center construction and power by 2030; with roughly 20 million GPUs manufactured in the cited year and every chip likely sold, the infrastructure requirement becomes unusually calculable.
10. Coordinated swarms may matter more than humanoid robots
China’s 16,000-drone display demonstrated what Dave thinks science fiction has underrepresented: perfectly coordinated swarms acting “down to the millimeter.” A swarm can assign two drones to a light object or 50 to a heavy one, making construction, yard work, gutter cleaning, and other physical tasks modular in a way one humanoid is not.
Salim said the Ukraine war is being prosecuted by roughly half a million drones “on either side,” while about 10,000 drones cross the Mexico-US border monthly: “drone technology beats wall technology.” Peter said Ukraine could become Europe’s drone-manufacturing capital after the war.
Tesla’s rumored Roadster demonstration was carefully bounded. Peter said gossip points to SpaceX cold-gas thrusters and perhaps 30 seconds of hopping or hovering, but stressed this is not an eVTOL; Archer, Joby, and EHang address actual multicopter transport, with Archer cited as holding a contract in Los Angeles for the 2028 Olympics.
Dave treated the Roadster spectacle as corporate strategy even if it never becomes mass transport. Tesla spends zero on conventional marketing and instead funds technically audacious projects that attract customers, employees, and attention: “What can I do to be inspirational and cool?” Peter added that the same formula makes elite engineers want to join.
11. Climate control and reusable launch systems turn engineering into governance
Elon Musk’s proposed solar-powered AI satellite constellation would make small adjustments to incoming sunlight. Peter’s sunshade formulation would reflect roughly one-quarter of 1%, acting as a planetary thermostat; reversibility is crucial, but conflicting national interests make any deployment a tragedy-of-the-commons problem.
Salim invoked Mount Pinatubo, whose atmospheric ash he said lowered global temperature by two degrees, and suggested existentially exposed regions might eventually launch interventions unilaterally. His rebuttal to anti-geoengineering objections: civilization is already geoengineering unintentionally through carbon emissions, while nation-state climate conferences have not solved the problem.
Alex’s broader proposal was a “global weather grid” that changes cloud cover through satellites, microwave heating, or other mechanisms, using AI weather models to mitigate or steer storms. Peter raised the liability problem—“Oops, we steered the hurricane in the wrong direction”—and Alex said deployment is “a social problem, less a technical problem.”
Blue Origin’s first New Glenn booster landing supplied the enabling-space counterpoint. The ESCAPADE mission launched toward Mars and the booster landed on Jacklyn; Peter cited SpaceX as launching more than 90% of US spacecraft and perhaps 70% of global launches. Alex welcomed a second reusable “railroad” to orbit because solar-system development cannot depend on one route.
12. Automation’s missing technology is social cohesion
Boston unions’ campaign against Waymo would require a human safety driver inside each vehicle. Dave called the policy “utterly insane” yet expected similar populist resistance everywhere; if major technology hubs cannot form workable transition policies, companies and their investment will simply deploy elsewhere.
Peter insisted the resistance begins with a rational survival calculation: “I need to feed my kids. I need to be able to afford my home.” Until society provides a credible safety net through UBI, universal basic services, or another mechanism, demanding that workers accept displacement as aggregate progress will not resolve their immediate loss.
Alex called regressionism one of the issues that keeps him awake: organizations may block technologies that save lives, generate wealth, and improve quality of life. The required “meta technology” must maintain cohesion while radically accelerating deployment—an optimal trajectory delivering both, not a choice between permanent stagnation and unmanaged disruption.
Salim framed institutional resistance as an “immune system problem.” He said his group has used a 10-week process around 100 times inside companies and a 16-week version in the public sector, then open-sourced the methodology; the harder next layer is adapting it to sector-wide immune systems in healthcare, journalism, and education.
13. AI science is shifting from isolated assistance to bulk discovery
Sam Altman’s framing set the progression: GPT-3 offered a first “spiritual Turing test” glimmer, while GPT-5 shows tiny instances of generating an idea or contributing usefully to a paper. GPT-6, he said, has a chance to produce a GPT-3-to-4-like leap specifically for science.
Alex’s stated maximum is three years for many, if not most, grand challenges in mathematics, science, engineering, and medicine to start falling. The anticipated output is “centuries of human capital” solved in bulk; Peter suggested tracking AI-generated patents and agent-to-agent transactions, including automated licensing, to see the exponent emerge.
The Chan Zuckerberg Initiative illustrates the timeline compression. Its 2016 ambition was curing most disease by the end of the 21st century; with Biohub, virtual cells, and a planned 10× compute increase by 2028, Alex read the new implied horizon as perhaps 2030, similar to timelines he attributed to Anthropic.
Virtual cells could expand into virtual organs and organisms, letting AI search intervention space rather than testing diseases artisanally one by one. Alex’s counterintuitive possibility was that “sometimes it’s easier to solve the more general problem than the more specific problem”—including curing all diseases before completing every separate cure program.
14. Cheap GLP-1s and AI scientists preview healthcare abundance
Peter cited a proposed GLP-1 price of $149 per month and results suggesting the drugs can cut repeat-stroke incidence by as much as half over three months. Alex called their broad benefits biologically mysterious: the “elephant in the room” is why a metabolic drug class appears to improve so many different forms of dysfunction.
Alex connected that affordability to universal basic services—roughly $150 monthly begins to resemble broadly available healthspan medicine. Peter supplied the critical warning: GLP-1 drugs are not a panacea; patients can lose muscle with fat, then regain fat without restoring muscle after stopping, making resistance exercise essential because “your muscle is your longevity organ.”
Edison’s Cosmos offered a different abundance mechanism: an agentic scientist using knowledge graphs and scaffolding to simulate context far beyond today’s millions-of-token windows. The ideal system would ingest trillions of tokens—the internet and every paper—then answer questions such as “What’s the solution to Alzheimer’s?”
The reported operating metrics were four to six months of expert research completed in 12 hours, 1,500 papers read, and 42,000 lines of code run per experiment. Peter argued the larger opportunity is feeding models raw experimental records, allowing neural systems to analyze interacting variables together instead of isolating one chemical reaction and barely clearing statistical thresholds.
15. Embryo editing marks a regime change from selection to alteration
Preventive and Manhattan Genomics were presented as companies developing CRISPR embryo-editing capability, potentially outside the United States and perhaps in the UAE. IVF already permits sequencing embryos and selecting which to implant; Peter called direct editing a “regime change” because it moves reproduction from choosing among existing genomes to altering one.
Salim’s premise was that “the human genome is essentially software” and a human therefore becomes a software-engineering problem; to him, editing embryos is inevitable, leaving “what do you want to design for?” as the central question. Peter argued parents already seek the best healthcare, education, and inherited traits, but acknowledged that extending this logic immediately raises eugenics fears.
Alex traced modern caution to the 1975 Asilomar guidelines and relayed a historical argument that fresh memories of Watergate influenced scientists’ desire for transparent limits. Fifty years later, he said he could not find one US federal statute categorically banning germline editing—only a patchwork of federal and state laws and regulations that strongly deter it.
The warning specimen remained He Jiankui’s 2018 CCR5 editing, intended to reduce HIV susceptibility, followed by global condemnation and imprisonment in China. Alex challenged the default dystopian reading of Gattaca, while Dave said the field urgently needs credible thought leaders; his final pushback was demographic: South Korea’s cited birth rate is already only 0.7 children per couple—“no one’s having any babies at all.”
Full transcript
The number one concern globally is cost of living, and tied very closely to that is unemployment: Will I get a job? The third concern is poverty and social inequities. We talk about a future of abundance, we talk about demonetization, but this is the reality of what people are feeling.
This is a story about preparing for the worst, maybe the worst thing any of us can imagine.
The wealth isn't going to go to the people who are doing the work or the people who get unemployed. It's going to making the rich richer and the poor poorer.
You have a third of your income going into your phone, your data plan, and all that money funnels out of the country and lands, like you said, in Silicon Valley and Boston. Then you add AI as a layer on top of that, and the gap is going to get really, really wide. That's the reality of a huge fraction of the world's population, though.
The question is, how do we help people believe in a hopeful and compelling future? Because if they don't believe it's a hopeful and compelling future—
Now, that's a moonshot, ladies and gentlemen.
Everybody, welcome to our episode on WTF Just Happened in Tech. Here with my moonshot mates, Dave Blundin. Hey, Dave.
Hey, morning.
Salim Ismail. Good morning, Salim. And Dr. Alex Wissner-Gross. Guys, it's been quite a week. You've been in Brazil, Salim? Just back?
Yeah, I was just back from 3 days in Brazil. It turns out we have a massive viewership and listenership there, demanding that we be able to translate this into Portuguese, so we should look at that.
Okay.
Mm.
We'll do that. I just got back from Milan and Madrid, right in the middle of Europe, talking about AI. We'll talk about it, but there's a lot of concern and angst about how far they're falling behind. We've got a subject to discover on this episode as well. Dave, how's your week been?
Phenomenal. I get to hang out with Alex face-to-face, which is a rare treat, and we brainstormed a ton of things going on. Can't wait to talk about them today.
Awesome.
I have a general complaint.
What's that?
Over the last few weeks since we recorded, there's been so much stuff happening. I need an Alex, Dave, and Peter AI next to me that, in real time, helps me interpret stories. We should maybe think about creating a GPT just for this.
Good news: it's in the queue, Salim. We're working on it. I budgeted it, and we're doing it.
All right. Alex, if this is the physical Alex or the AI Alex, I have no idea, but how are you doing, pal?
What's the difference?
Probably nothing. Probably not much.
Half the world's convinced I'm an AI already.
We'll figure that out soon enough. Let's open up with the hyperscalers. News about Anthropic, Google, and OpenAI is still going on. In fact, this week in particular, there was a lot of news about Anthropic, which hasn't hit our episodes in the recent past. Anthropic overtakes OpenAI in enterprise LLM API market share. All right, over to you, Alex. What's the significance here? We see this chart: OpenAI's dropping, Anthropic's rising. What does it mean?
I think the central question, Peter, is whether code generation is the critical path to recursive self-improvement. If code generation is the critical path, then one can expect amazing outcomes from Anthropic, which has quite publicly focused its strategy on code generation, perhaps to the exclusion of other modalities, like video generation, which we see from OpenAI with Sora.
On the other hand, if code generation turns out to be missing some special sauce needed for superintelligence, broad superintelligence, and recursive self-improvement, maybe this trend won't last. I think that's the core question here.
Dave?
I have a completely firm opinion on this. I don't want to lead the witness, but I can throw it out there first, Alex, or you can tell me. What is the answer? Is the LLM-scaling recursive self-improvement loop enough to crack the singularity and infinite intelligence? What's your guess?
I don't know. If I had supreme confidence on this one—
Mm.
—it would be far easier to make investments in this space. I can make steelman arguments on both sides.
The steelman argument in favor of code generation as the critical path to the singularity, to the extent it's a finite point, or a fixed point, and not an extended object, would be something like this: We leverage code generation to rewrite the core algorithms, the key models and architectures, and the post-training architectures underneath frontier models, and the flywheel just spins faster and faster.
The steelman argument in favor of code generation missing something looks something like this: Maybe we need visual chain-of-thought, or maybe there's some grounding in the physical world that's essential for general-purpose knowledge and general-purpose reasoning that you can't just get from looking at large source-code bases and the internet of text tokens, with maybe a little bit of imagery. So I'm not sure.
I'm firmly in the camp that Dario is on the right track, and it's not just Anthropic. Look at the chart. Google's also on the same trend line. So it's either Demis or Dario, and by scaling what they've already got, turning a huge amount of the compute internally, and having it generate the next test, the next test, the next test. I think the tipping point there is Humanity's Last Exam, and what we're hearing—and what you're hearing—is that it will be saturated very soon, which is mind-blowing given how hard those questions are. But to me, the solution to those questions and the innovation—
—in AI are incredibly correlated problems.
I took a different take on this one. I think what's interesting here is that AI is actually showing that it has a real business model, and that will be a really powerful feedback loop going forward.
The enterprises that I talk to—the banks and others—they're all using Anthropic because they trust it with sensitive data, while OpenAI is going consumer and not really going after that corporate market. So they're both going to thrive, but in a very different way.
Yep. The reliability that they're providing is amazing.
Mm-hmm.
All right, let's go to the next story here on Anthropic. Again, congratulations to Dario. Anthropic projects $70 billion in revenue and $17 billion in cash flow in 2028. We haven't heard a lot about Anthropic over the last couple of months, right? It's been OpenAI, xAI, and Google taking the headlines, but Dario is gaining ground here. How do you see them competing against the other hyperscalers, Dave?
I think he's positioned really, really well. The numbers aren't as big as OpenAI's, but the enterprise market is wide open. What OpenAI is doing is directly competing with Google, which is very aggressive and very cool, but also risky. I like the angle they're taking here because enterprises need AI. Using AI as a management tool is the gold mine of all gold mines, and I'd love to riff on that for hours on some other podcast.
Mm-hmm.
Everybody doing that is using Claude and Anthropic as their backbone, and so he's not facing a lot of competition right now in that market. It's not as fast-growing, it's not as sexy, but it's a really good strategy, and I think he'll hit these numbers.
When I was with my friends at Google, there was an interesting point: They view Anthropic as the other friendly AI company. They're obviously at each other's throats—OpenAI and Google, and OpenAI and xAI. Anthropic is sort of the friendly little brother to the other hyperscalers.
Isn't it ironic, though, that OpenAI's original mission was what Dario is now actually known for? “Machines of Loving Grace,” and the idea that if you want to work in AI but you want it to be guaranteed to be good for the world, come to Anthropic. He's really grabbed that high bar, that high ground, on that topic.
I think there's something fundamental, though. We've seen this happen over and over again: What becomes ultimately a frontier lab starts as an alignment lab. I think there's almost a perverse duality between alignment and capabilities. If you're the world's best lab at aligning AI with human interests, that immediately—whether it's for economic reasons, like you need to raise capital in order to train superaligners, or just for purely technical reasons—if you can align a model really well with human intent, that immediately itself is a strong capability.
I think every alignment project almost inevitably ends up as a capabilities project. So I think it's not just a coincidence that OpenAI started as an alignment-oriented effort to ensure that there wasn't just a global singleton in the form of DeepMind for superintelligence. Anthropic, similarly, in the long tradition of Silicon Valley, and Fairchild also started as alignment-focused, and then almost immediately pivoted to capabilities and superintelligence. I think that's just the law of economic nature here.
Mm-hmm. Totally right.
I have a counterpoint to that.
Go ahead, Salim.
Facebook started off being very aligned with protecting privacy—user privacy—and never leaking private information, and then they sold it for profit. It's not AI-related, but just in terms of business models.
Mm-hmm.
So at some point, this could become extractive, right?
One of the stories here that I want to hit on is the economics. Interestingly enough, when I did some digging, Anthropic is projecting 70 billion in revenue by 2028 at a 77% profit margin, right? That's pretty extraordinary if they can hold on to that. On the flip side, OpenAI is projecting 100 billion in revenue and is unprofitable until 2029, right? I mean, just their deployment of capital into data centers and model growth.
I'm super curious about the two business models—Anthropic versus OpenAI. What do you guys think about that?
Well, I think a lot of these companies are capable of having high margins on short notice, including OpenAI, and they're trying to tell the market—OpenAI in particular—that they intend to keep investing ahead of the curve. So if you'll give me the $1 trillion valuation—
Yeah.
And that's better. If you can pull that off, it's much, much better. I can tell you, having taken a company public, as soon as you switch to profitability, it's very hard to go back to your shareholders and say, "Oh, I want to burn $1 trillion building a data center."
Mm-hmm.
So Sam is declaring that up front, which is great strategy as long as the shareholders believe it.
Which is what Bezos did.
And what Dario's doing here might—yeah, exactly.
Do you remember—
And that's why Amazon is—yes, exactly.
Do you remember Bezos's famous letter at the beginning, when he started? He goes, "Listen, I am not going to be profitable. I'm going to be spending money. If you want a profitable company, go someplace else. Otherwise, I'm buying customers. I'm buying revenue." And he did, and then he flipped the knob. He flipped the switch.
Yep, that's exactly right. And what Dario will do here, in all likelihood, is declare this kind of margin—77% gross margin—but then, as that date approaches, he'll launch a new project under a new name and say, "Well, we're going to consume all that money building Stargate," or, you know, with whatever Anthropic Stargate or whatever it is, and that's a good strategic shift.
But before we leave this story, these numbers are much bigger than anything in the history of the world. Much bigger—the growth rates and the scales.
Yeah.
And I just want to make that point, because we get inured to these stories so quickly, and we go, "Oh, I already heard that."
We get numb to the trillions, right? Numb to a trillion here.
Oh, and also in terms of life plan, anyone who's doing something other than this, you've got to consider, "How do I get into this?" This is so much bigger than all other—
Mm-hmm.
—endeavors combined.
1. World Models Become Reality
The next article here, which is a fascinating one, is "Fei-Fei Li's World Labs Unveils World-Generating AI Models." We had this conversation with Fei-Fei backstage at FII. Let's take a look at the video, because the implications of this are absolutely huge.
Finally. I thought you bailed again.
Please, I wouldn't. All right. Let's reset to the next world.
All right. If you were listening to this podcast and not watching that beautiful video, what you saw is an extraordinary, immersive, photorealistic virtual world—a world model that Fei-Fei has been building. I'm both fascinated and concerned, right? I'm fascinated by these world models because it's magic. I'm concerned that it's going to be where we spend a huge amount of our time.
I just finished reading a book called The Unincorporated Man for the second time. In this book—which is a fascinating conversation by itself—the world has a crisis because everybody starts spending all of their time in these virtual worlds, to the exclusion of work and the exclusion of eating, and it decimates the population. I know our kids—my boys right now, maybe yours, your son, Salim, as well—a lot of time is spent in video games. But if they become so photorealistic and so immersive, and you are the god of your world, why would you want to spend time anyplace else?
I think it proves that we're living in a simulation.
Well, yeah. I tweeted that after I had the conversation with Fei-Fei, which was, "I have no question we're living in a simulation." But even if we are, what would you do differently? People are just so tired of this conversation about whether we're living in a simulation.
Before—
What's in—
Before we get—
What—yeah.
Yeah.
Go ahead, Dave.
Well, before we get too deep into the story, because I really want to hear from Alex on this—this is very, very different from what you think it is, but Alex will explain it to us.
On a personal note, when we were backstage with Fei-Fei in Saudi, anyone who's an aspiring leader, entrepreneur, visionary—anyone out there—we were backstage, she's one of the gods of AI, and she says, "Oh, Peter, Dave, let me—I'm so excited to show you." She whips out her phone, and she's showing us the product in 5 seconds, and she's like a kid in a candy store, excited about what it can do.
And that enthusiasm, but also the ability to show your thing in under 5 seconds, is so infectious. At her level, the fact that she's still doing that—learn from that. Everybody should be able to do that. No matter what you're excited about, you should be able to project it in 5 seconds or less on your phone, pull it out, have it ready to go. It's so cool.
All right, Alex, talk to us.
So the consumer story here is we're seeing the beginning of the Holodeck Wars. The technical story underneath the Holodeck Wars is that we're seeing different approaches for generating entire Holodeck-type simulations from scratch.
We see, on the one hand, the Google Genie-3 approach, where every pixel is being generated by a single model. On the other hand, I think what World Labs is demonstrating here with Marble—their Marble model—is sort of the opposite end of the spectrum. It's generating not individual pixels, but so-called 3D Gaussian splats, or 3DGSs, which are sort of transparent blobs that cumulatively build up into what looks like a photorealistic, 3D-traversable world.
But 3D Gaussian splats are so compute-efficient that you can dynamically recompute visualizations locally on your computer, on your client, whereas the pixel-wise generation approach of Genie-3 and other competing models requires compute-intensive, server-side GPUs. So I think we're starting to see the beginnings of an efficient frontier of trade-offs between compute and versatility.
There are going to be upsides and downsides to each of these. But in the same sense that we saw with frontier models, some models live on the edge and are relatively compute-light, while some are server-intensive. I think we're going to see a range of different levels of worlds that we're able to generate.
But all of this—I mean, the consumer use case—I think this is chicken feed compared to the larger addressable market, in my mind, which is using these models to generate synthetic training data for more capable vision-language-action models, for robots, and for scientific discovery. That's the much larger market, but this is fun in the short term.
2. AI Learns to Forget
Our next story here is "New Method Helps AI Forget Memorized Data Without Losing Reasoning Skills." This is a big deal. I'm going to go to you, Alex, first on this.
Yeah. I often speak about reaching, in the near-term future, a diamond-like, perfect micro-model of a frontier model that externalizes almost all knowledge. So it's a pure reasoning model, and all of the knowledge can live outside the weights of the model, in some external database or some external tool call.
I think this paper by Goodfire describes such a clever way to externalize all of that knowledge, to distill down—no pun intended—to the essence of a core model. Toward this vision, the basic idea is to look at the weights inside the model and distinguish which weights represent knowledge versus some sort of general reasoning capability.
You do that by looking at which weights, if trained over multiple examples and the model is run over multiple examples, impact the overall so-called loss—or the ability of the model to match the desired output—if they're changed a little bit.
The weights that are important to generalization will have a dramatic impact on the overall loss of the model if those weights are just tweaked a little bit.
Can I take a slightly different frame of this? So if you've trained your AI model on all of my company's healthcare data or all of my company's financial data, do I trust you then? And it has some incredible results coming out of that. Do I trust you to not give up my core healthcare data or financial data while still adding value, right? So apparently what they're saying is they found a way to make AI forget specific memorized content without retraining the model from scratch, without wrecking its intelligence. So you can forget the specifics about all the healthcare data, forget the specifics about all the financial data, and the model still delivers the same value. That's what I understood it to be. Alex, is that correct?
Yeah—half correct. It is correct that this is a pruning, or more generally a regularization, technique for helping a pre-trained model to forget knowledge that one might call memorized.
But really, the emphasis in the inflection is less about some sort of enterprise privacy or some sort of ML data firewall feature, and more about figuring out which parts of the model are actually needed for general capabilities. Models are huge. Models are, in many cases, hundreds of billions, if not low trillions, of weights. That's very compute-intensive. It would be highly desirable to figure out which of those weights are actually needed for general capabilities, and which of them are just for memorizing arcane and quite possibly wrong facts on the internet or from an enterprise environment. So this is more about generalization capability, less about filtering out enterprise privacy.
And is it also about making it a lightweight model to run on systems?
That's the holy grail here. The holy grail is: could we have maybe a sub-billion-parameter, maybe aspirationally a million-parameter model that's generally intelligent? That would be an incredible outcome.
Wow.
You know, if we solve memory in some of this, that's a huge breakthrough, and then maybe the next thing you could add in is some mechanism for adding curiosity into the model, because that feedback loop would be unbelievable.
Hmm.
There's a whole cottage industry focused on active inference and building curiosity from scratch as an instrumentally convergent motivation into these models.
Yeah, so cool.
Wait, “instrumentally”—can you repeat that? Instrumentally convergent?
Instrumentally convergent. Instrumental convergence is a very important term. Instrumental convergence is this idea that, in order to achieve a variety of different long-term objectives, you're almost forced to achieve common short-term objectives. For example, if I have a superintelligence and, on the one hand, it's instructed to build lots of paperclips, and, on the other hand, it's instructed to cure cancer, probably both of those long-term objectives require that, in the short term, it accumulate capital, maybe solve science, maybe build a bunch of factories.
Hmm.
So those are instrumentally convergent motivations.
Instrumentally convergent in the near term.
Yes. Convergence in the near term—divergence in the long term.
Yeah.
Well, high-level observation—
This is why this pod is so ridiculous. I learn something magical every single time. Every single goddamn slide, for God's sake.
Well, this topic is incredibly important and somewhat obscure, but there are many, many, many people working on outside-the-box chain-of-thought reasoning and vertical use cases, but very few dorking around with the weights. There's so much opportunity when you're messing around with the weights. These 90% reductions in parameter count through distillation are pretty common.
When you think about, “Okay, hey, we need 100 gigawatts of power. Oh, wait, I distilled it by 10×, now we only need 10 gigawatts,” the implications of that are trillions of dollars. So there really ought to be a lot more people playing with the open-source weights and trying things, because these things are really working, and this is a great case study.
This really gives you the brain equivalent of neuroplasticity.
And also, psychology researchers who for years have been trying to deal with surveys and outside-the-body tests: you can make so much more progress in understanding the nature of intelligence by playing with the weights of a big neural net and trying things like this. A lot more people in that area—cognitive psychology types—should be working on this too.
3. AI Learns to Learn
All right. Here's our next article, which is related but different, and I think more about neuroplasticity. Google introduces nested learning, a machine-learning paradigm for continual learning. As I read into this, I'm like, wow, this really is a big deal. Alex, I'll go to you first on this one.
Yeah. So this is a pretty dense, if I may say so, NeurIPS paper. NeurIPS is arguably the premier AI and machine-learning conference. Parenthetically, it's happening in early December. Definitely, I'll be there. Many folks I work with will be there.
The core thesis of Google's nested learning paper appears to be focused on higher-order meta-learning. Meta-learning, as a reminder, is learning to learn. It's of core interest to the AI and machine-learning research community. If you could learn to learn, that almost obviates a lot of machine-learning research. So this is focused on higher-order meta-learning, rather: learning to learn to learn to learn, and so on.
The core insight here is that models and model architectures, on the one hand, and optimizers that right now we use to train the models, may actually be 2 facets of a common object. Reading the paper, it almost seems aspirationally to be fishing for a grand unified theory of machine learning—an M-theory of machine learning, if you will—where all of these different processes are actually just facets at different levels, different orders of abstraction, of a single common paradigm, which is, as I and others have argued in the past, compression of information.
I like to say if we could send a message back in time and explain how we got to AGI, it turns out it's very easy. You just take a large amount of knowledge about the world, and you compress it. If you compress it beyond a certain point, you get some sort of phase transition, and general intelligence pops out. That's basically, I think, the story of AGI.
You know what? When I think about college, I think the entire college experience for me was learning how to learn. Everything I learned was irrelevant some number of years later. And so when I read this, Alex, what I'm seeing—and I'm curious—is that it's about enabling continuous human-like learning, right? Sort of a step toward—
Like lifelong learning.
Yeah, a step toward truly adaptive lifelong learning for AIs. Historically, we'd train up an AI system, and you would freeze it, and then you'd do inference on it. Here, this is something that's continuously learning in almost a human-like fashion. Dave—
My college experience wasn't about learning how to learn. It was learning how to avoid most of the bullshit that was taught there.
Okay.
Learning how to forget.
There you go. Dave, any thoughts on this particular article?
Not yet, but I think Alex's explanation was perfect anyway. One thing people overlook is that the machines doing this have access to incredible numbers of tools. When you're learning how to learn, you're thinking about never forgetting, but you write things down, you refer to them. You have your laptop, you have your phone, and the AI version of that has immense bandwidth between its brain and its notes.
A lot of the innovation is taking advantage of that incredible bandwidth. It's just clearly superintelligent before it's even intelligent, because of that ability.
Can I comment on this? That little explanation by Alex—I'm going to have to go back and listen to it, like, 4 times over just to parse all the stuff. We don't need to get into it here, but we may need a whole podcast episode just on that—
Oh, I'd love that.
—because it's pretty fundamental, right?
4. Open Source AI Gets Cheap
An Alibaba-backed company called Moonshot AI is launching a new ultra-low-cost AI model. As I read this, I'm like, wow, this is a big deal. Dave, let's—
They clearly watch the pod.
They watch the pod. Why? Because they called it Moonshot?
Yeah, the name.
Sure.
We're going to claim—yeah, I think Astro Teller claims Moonshot as the captain of Moonshots at Google. We're—
We'll try to enforce your trademark in China anyway. I don't think you have much.
Dave, what is your take on this one?
I hope people appreciate what a huge deal this is. This is, in my mind, the biggest thing that happened in the last month. These Kimi models are actually at the top of SWE-bench, right up there with Anthropic, just 1% or so below.
But they run on Groq hardware—G-R-O-Q, not Elon, Groq hardware—which we learned all about in Saudi Arabia, at blazing speeds and with incredible performance. They're the best open-source models. If you want to play with the weights, you've got to go with the Chinese stuff, because Meta stopped open-sourcing, and OpenAI stopped open-sourcing open weights. So now these are the absolute best models in the world, and they're all coming from China.
But the fact that you can train it for $4.6 million is within the budget of almost any company. It's not 10X; it's more like 30 or 40X cheaper than what it took for OpenAI and Anthropic to build their original models. Part of it is that you're drafting off their innovations, which is why those companies stopped open-sourcing. But you can grab these weights, grab this open-source model, and build on top of it.
Just be careful that there's no spyware or anything in there. I checked and I'm using it, and it seems to be okay, but do your own spot-checking. This is a huge deal in terms of the power for somebody who wants to play with the guts of one of these things. The amount you can get done on a limited budget just skyrocketed, and it's all about—
You know what I found most significant? There's been a conversation going on—we heard Eric Schmidt talk about this—that it's the US financial markets and their efficiency that are allowing these hyperscalers to raise billions of dollars to do what they do. But if all of a sudden the cost of training a trillion-parameter model is $5 million, you don't need efficient capital markets. That money's available from a lot of locations.
Isn't that wild? I mean, you're so right. Just think about that. I don't know if anyone appreciates what you just said. The implications are massive, just absolutely massive.
Alex, what's your take on this?
Yeah, a few takes. One, as Sam and others have pointed out, there's hyperdeflation going on right now on both the training and inference sides for AI. Sam's number is 40X year-over-year hyperdeflation. On the one hand, I'm not surprised—
Wait, wait, wait, wait, one second. Just—
40 times?
40X hyperdeflation year over year in the cost—
In the cost of intelligence per unit of intelligence.
That's insane. That's a big—
Mm-hmm. Yes.
That's—
So this is where abundance comes from.
This is many times faster than Moore's Law.
Yes. This is why I speak of the innermost loop as a catchphrase, because when you see—if we can see sustained 40X year-over-year hyperdeflation in the cost of intelligence, everything else is going to get dragged down. The price of everything else, rather, is going to get inevitably dragged down with that. So this is sort of the nuclear core, if you will, that's going to pull down the cost of everything else. That's 101.
And because the demand is growing 1,000X a year, that's just what justifies the capital expenditure.
Yeah. Pick your analogy.
When you walk around academia or corporate boardrooms, you find deniers everywhere, just absolutely everywhere. A huge amount of the denial is, “Well, I tried this yesterday. It was hard. It didn't work, so therefore we're nowhere near AI.”
When you talk about 40X hyperdeflation, deniers are saying, “Well, look, the evidence is that as we scale these things, they're getting decreasing improvements in intelligence. So I don't think there's anything. Forty X for 2 years in a row—2 back-to-back 40Xs—it's not going to do much.”
That's a really, really risky position to take, dude, because there is significant upslope in the data, and you don't know what 40X is going to do. But if you had to bet whether 40X is going to be mind-blowing or not much of an improvement, you're crazy to take the position that it's going to be a little bit of an improvement. Crazy.
Oh, absolutely.
Especially back-to-back 40Xs.
Yeah, no, I think we're going to start to see grand challenges in math, science, engineering, and medicine start to fall over the next 2 to 3 years, thanks to that sort of exponential—
So can I take the other side of this?
Yeah.
Please.
Because I will beat the crap out of you. Let's do it.
No, no. I think if you look at the demonetization curves over history with other things, like solar energy and the cost of compute, you'd almost expect this because you want to see that curve go in this direction. Forty X is way faster than I thought.
But I'll give you the macro counterexample. I remember when the Google car first came out, all the car manufacturers said, “Well, that's bullshit. The cost of that lithium-ion battery is too expensive. This is never going to work.” They all ignored it.
Then, over the decade, the cost of lithium-ion batteries dropped 90%. That's what Elon banked on for building Tesla: the cost of the batteries dropping. And he was right. All the carmakers and all the typical-minded folks were wrong.
There's a powerful lesson here: always watch for those deflationary curves and go where the curve is pointing you.
What's surprising is that this is the world's most powerful technology, by far. If I had gone back 5 years and described what an AI model could do today and how much you would charge for it per day or per month, I would have guessed millions of dollars or hundreds of thousands of dollars. I would never have expected it to be free. It's effectively free.
We have to eat our own dog food and say, “This is where the curve's going. Let's expect this at this point and see if we get it right.”
Yeah.
I'm constantly running around the office and saying, “Guys, within these virtual areas like computing and AI, it's really hard to visualize 40X. But imagine we had a factory that makes widgets or cars, and we 40X'd our production year over year. Everybody in the office would be going crazy. You'd see 40 times more stuff coming out the door. It would be obvious, dude. That's what's happening, and you have to really stretch your brain around the implications of this, because—
There's a great quote—
Remember my rant on the demonetization and the fact that the energy guys all got it wrong and whatever? This is that same issue, right? And to Alex's point about the inner loop, if you roll this out to the broader macro things that we do, you should expect a 40X drop in the cost of healthcare, food production, and everything over time as this bites, correct?
Yeah, I agree.
That's my expectation.
Assuming intelligence can solve the problem. There's a great quote from Gordon Moore, and I'm going to botch it. I don't have the exact numbers. He said, “If cars had improved at the same rate as Moore's Law, a Rolls-Royce—
Faster than the speed of light.
—would get a million miles per gallon, and you'd throw it away because it's so cheap at the end of your trip.”
Yeah, the statistic I heard was that if the top speed of a car had evolved at the same speed, we'd have a car today that went faster than the speed of light.
Yes. All right, I want to jump into this next article, especially coming back from a week in Italy and Spain. The story headline here is “Brussels to loosen GDPR rules to enable and feed the AI boom.”
This is a really important conversation. There was a lot of concern and angst when I was meeting with CEOs, consulting companies, and investors in Europe, in Italy and in Spain. The GDP of Italy is about $2.7 trillion, about the size of New York. The GDP of Spain is about $1.7 trillion, the size of Florida.
But there's serious concern about whether they can compete. One of the biggest issues that's keeping them from competing is the ability to access data. Salim, what's been your experience here?
You know, this is part of the problem with overregulation: you just slow this down.
The cost of compliance for stuff like GDPR has proven to be ridiculous, and this is a function of the historical issue with Europe. To be fair, it's not just mindset here, although there's a bunch of historical legacy that's really important. I'll give you a specific example. After World War II, the German constitution added a clause saying no media organization could cover the whole country, so that you could—
Hmm.
—never have another rise. That prevented a regional player from covering the country, and then Google came along and rolled up the whole thing, right? So they've got structural issues going back in history, and you have to figure out how to undo some of those. That's a really, really hard thing to do. They're doing the best they can in a difficult environment, but it's just—
Yeah.
—a massive problem.
For the longest time, you and I remember this at Singularity University, right? Europe in general prided itself on having the strictest privacy laws out there, and a lot of that simply meant, “Okay, we're going to exclude everybody from Europe in a U.S. product or service.”
I think what's interesting is that the data is showing venture funding in Europe dropped up to 30% as a result of this, and that the AI models in Europe have been 6 to 12 months slower to market than in the United States. And the compliance burden—so get this: there is an AI audit that's required before you put out a product, and those audits, on average, cost €260,000 and take 8 to 15 months, delaying—
That's nuts.
—40% of projects. So imagine that. You have to actually prove during your compliance audit that they reviewed the data sets, model transparency, bias, documentation, and safety standards. Some third party is auditing you and just putting sand in the gears. I understand why people want that to some degree, but you're trading that against your economy.
It's been really interesting. If I think about my grandparents—
I think it's really over-regulated.
—my grandparents assumed that there was no structure and the world was chaos. The military protects us, but other than that, it's just a zoo. And then when you look at the generations, if I talk to my kids, they assume that there's some rational thing out there that's thinking through these issues. But there isn't.
They assume we're adults.
We know there isn't.
They assume there are adults in the room. You know, isn't that a strange thing, when all of a sudden you say, “Oh my God, I'm the adult in the room”?
It's really, really freakish when you hit that.
Yeah. It's why I'm hoping for a benevolent superintelligence to actually be the adult in the room someday.
Yeah, we need Mo Gawdat's thinking: move as quickly to AI running the world as possible. That should speed up Alex's inner loop considerably.
Yeah.
Mm-hmm.
Alex, any thoughts here on the GDPR rules being changed?
Maybe just a broad comment that, under the current construct, it's up to individual sovereign countries to define the parameters of how much they want to participate in the superintelligence explosion, and maybe just leave it at that.
Yeah. The conversation I had with a lot of the leaders in the tech industry in Italy and Spain was, “Okay, if you guys are interested in playing this, you need to build out your energy sources and identify where you're going to set up your data centers.”
I think the timeframe for making those decisions and implementing them is the next 5 years. I mean, I think the next 5 years—
Yeah, it's now.
—are gonna set the objective for the next century.
Literally now.
More like 5 months than 5 years, but okay. Yeah, I think it's a little bit shorter than 5 years.
Okay, I was being generous, because you can't do anything in 5 months. Which is a big concern, and we can talk about that in the U.S. even when we get to the conversation on energy.
5. The Global Cost Crisis
This is my chance for a rant, and I want to share something that we found when Salim, Dave, and I were at FII. I'm on the board at FII—the Future Investment Initiative—and one of the things they do every year is something called the Priority Global Survey.
This is a survey that they do in 32 countries. They have over 60,000 respondents, and it represents two-thirds of the world's population. We've had some criticism on this podcast that we're not focused on the reality in different parts of the world, so I want to discuss the reality. What are people seeing and feeling outside of Silicon Valley, outside of Boston?
Here's some of the data, and I'd love to discuss it with you guys, because the data is important and concerning. They surveyed people and asked the question, “What are your top concerns?” We see this across the Global South and the Global North.
The number-one concern globally is the cost of living, by far. Can we afford to live in this world? Tied very closely to that is unemployment: Will I get a job? And if I get a job, will it pay me enough to live? The third concern is poverty and social inequities. This is what two-thirds of the world is feeling right now.
Here's the next chart, and we can look at it by region: Africa, Asia, Europe, MENA, North America, Oceania, and South America. We see that Africa's number-one concern is unemployment. For the rest of the world, it's the cost of living. It's just expensive to live.
We talk about a future of abundance, and we talk about demonetization, but this is the reality of what people are feeling. What are your thoughts here? Salim, you want to jump in first?
Yeah, this is an extrapolation of the basic nature of human reality. We've been living in fear since the beginning of time, right? In the cave-dwelling days, you were worried that a hyena would come and steal your baby at night. Now we're worried about jobs.
I think we need to flip over completely to a UBI-type structure to navigate the world going forward, because the concept of a job is going away. Think about the idea that all of our education systems are designed to take a young child, train them through their early 20s, and get them ready for a job market when we have no idea what a job looks like in 5 years.
This may be the thing that breaks the educational model and all the other models into a totally new reality, where most of these mechanisms for subsistence, like education and healthcare, are basically free. Take that 40x curve and apply it to some of these domains. That's, for me, the incredible opportunity.
I'd flip this into the massive opportunities there—but people aren't seeing what's going on, so they get stuck in the fear factor.
I got it, but it's not evenly distributed yet. This is the reality that people are feeling right now. They're feeling fear about, “Can I get a job?”
Understood.
And can I afford to live? It's a very real concern.
I think this is our job as leaders and podcasters and message conveyors: to show that. Take Amjad, a little developer out of Jordan—boom, builds a multibillion-dollar company. Take Vitalik. Take Elon, coming from nothing to building global, paradigm-changing things just from mindset.
And therefore, now, the inner loop—I'm gonna go back to that again—is just literally mindset and entrepreneurship. Peter, you talk about this all the time. I flip this around and see the opportunity in this.
I'll give you a little snapshot of what a big part of the world looks like. One of our summer interns—she's from Iran. Her parents are still in Iran. I grew up as a young child in Iran, and she said that her parents spend 1/3 of their annual income on their iPhone and data plan.
Mm.
She's like, “Look, you can't live without information,” and actually, the currency's no good, so everything's Bitcoin. How are you gonna manage your Bitcoin without an iPhone? So you've got a third of your income going into your phone and your data plan, and all that money funnels out of the country and lands, like you said, in Silicon Valley and Boston.
That wealth disparity, just from the phone—and then you add AI as a layer on top of that—the gap is gonna get really, really wide. So that's the reality of a huge fraction of the world's population, though.
Yeah.
I hear you. I don't know if you want to add anything, Alex, on this, but this is one of my biggest concerns, right? This data, for me, is worrisome. I'm clear that we are going to get to an abundant future. Maybe it's a decade out. We're gonna have continuous demonetization and all kinds of uplifting of healthcare and education by AI.
But I think in the next 2 to 7 years, that's what really concerns me, right? If young men aren't getting jobs, and if people are losing their jobs as a result of this before we sort of flip the economics into an abundance model, the question is: How do we help people believe in a hopeful and compelling future? Because if they don't believe it's a hopeful and compelling future, they're going to believe what they see from Hollywood, which is dystopian AIs and killer robots.
I think you've hit the crux of it, right? How do we get narratives out there that demonstrate that future, do it fast, and overcome the fact that people are 10X more likely to listen to fear stories than stories of a positive future?
Yeah.
And that has to be overcome. Therefore, you need 10 times more stories on the positive side
100%.
to overcome the natural balance.
People worldwide are really worried.
One of the positive, consistent pieces of feedback I get about this podcast is the fact that we're relentlessly optimistic about the future. Why? Because technology's a major driver of progress in the world, and maybe the only major driver of progress, and now that's moving exponentially.
Do you guys remember what we were talking about before we hit record on this podcast? The idea that it would be amazing to bring together a community of builders, coders, and entrepreneurs to work on uplifting humanity in the near term. I think we should do that. I think we should pull together this Moonshot community and see if they—
Do a summit.
—want to discuss how we make the world a better place, right? How do we build moonshots that really uplift mindset but help address unemployment and cost of living in the near term? I mean, Elon has built an incredible community toward going to Mars. Satoshi created an incredible community around Bitcoin and crypto. I'm talking about: Do we organize a meetup of the Moonshot listeners? Do we pull folks together—
Oh, yes.
—and talk about solving grand challenges together? Salim?
Yeah, I spend a lot of time with college undergrads and seniors, and they would flock to that mission like you wouldn't believe. You'd get incredible talent coming to that mission. When they're in their early 20s, mid-20s, before the scar tissue of life has accumulated too much, they are all in on that. And so you'd get really, really smart people working on it.
Bring together the builders, the visionaries, and the folks who want to really build. I like to say that the world's biggest problems are the world's biggest business opportunities. Want to be a billionaire? Help a billion people. I mean, that's the conversation, you know? And I know none of us have extra time to actually pull an event together, but if folks listening to this podcast—
I think it's mandatory because the only way you're going to change the world is to have people shift their mindset, listen to stuff like this, and then actually activate it and go do something. So imagine we did an event, brought everybody together, talked about things, and then people actually activated online, formed teams, and went off and did stuff, and then we tracked that over time. That would be pretty cool.
All right. So, Alex, are you in on that?
I think it's a benchmark problem. I think it's less about events and less about teams and more about just rigorously defining benchmarks for all of these problems. How about a benchmark for cost of living that then the world and this 40X year-over-year hyperdeflation of intelligence can optimize toward? The same with crime and delinquency, and with healthcare—the cost of healthcare.
We're going to be drowning in humanoid robots that are generalist in terms of their capabilities in the next few years—
Well, we'll talk about—
But benchmarks—
We'll talk about benchmarks—
We need the benchmarks.
—we'll talk about that. So, if you guys are in, my feeling is that none of us have time to put an event together, but if there's interest in the community, to everyone listening, this is what we talked about earlier. If you have an interest in joining us at some kind of Moonshot gathering, a Moonshot summit, whatever it is, if we can get enough of you—let's say 1,000—who say, “Yes, we want to do this,” and you want to spend time with the Moonshot mates, then we'll pull this together.
Here's what I'm proposing. We'll set up an email. Let's call it moonshots@diamandis.com. If you're interested in this idea of a Moonshot summit to bring everybody together, talk about the world's biggest problems, talk about the benchmarks, and talk about the moonshots required, send us an email. If we can get 1,000 people who say they want to be in on this, then we'll pull it together. We'll bring together Moonshot mates. We'll bring together the most exciting CEOs and Moonshot engineers and have an epic 2-day event. I think 2 days is the right length for this.
Can I riff on this for a second?
Yeah.
If you take Hans Rosling's work, which showed that over the last 100 years we dropped the cost of electricity, transportation, and telecommunication by thousands of times each, right? And then you say, “Okay, we want benchmarks that in the next 2 or 3 years drop those by—”
Mm.
—1,000 times each if you extrapolate the 40X, that then gives you the target to go after, to the benchmark comment that Alex made. And then you basically bring in the SAGE engine to say, “Okay, what policy changes do you need to make? If technology can reach this, how do you get this implemented?” You could bring that together and make that a showcase for the world in a very powerful way.
I think it would be incredible. I really would love to get everybody together and have that conversation, and really ignite a passion and interest among entrepreneurs to focus on this, because there are real challenges out there in the world.
All right. So here's the deal. If 1,000 of you who are listening want to join us, let's say sometime next fall, then send us an email, moonshots@diamandis.com, and if there's enough interest, we'll pull this together. All right. Let's get back. There's a lot more to cover.
6. Data Centers Race for Power
Our next segment here is data centers, energy, and space. Multiple data centers are reaching 1 gigawatt in 2026. We're tiling the world in data centers: Anthropic and Amazon, xAI, Microsoft, Meta, and OpenAI's Stargate. Alex, what's the story here?
Well, I think the trillion-dollar question, Peter, is: Will we see a peak in the amount of coherent power needed for coherent training runs of large frontier models? If we do, one could imagine, as incredible as it may sound looking at this curve where everything's going up and to the right in terms of total facility power, we might actually see a peak—maybe in a few gigawatts sometime over the next few years—and then decline if there are algorithmic innovations that enable us to do distributed training runs rather than needing one large, power-intensive, coherent supercluster to do it. Tiling the Earth could look like tiling the Earth with relatively lower-power-density compute.
Totally imaginable. On the other hand, I spoke earlier about AGI being essentially, as it turns out, the compression of information. If it turns out that there are further phase changes that we can achieve by compressing more and more and more with larger and larger facilities, then maybe eventually, in extremis, we end up in a sort of—I've spoken about this on the podcast previously—maybe more of a black-hole, desktop-black-hole-computer regime, where we're just building these incredibly power-dense facilities to train more and more and more. Again, I could go either way on this, but I think that's the trillion-dollar question: Will this peak or not?
Besides peak data centers, the question is: Are we going to see peak energy? That's a question for you, Alex. The US government, Brookfield, and Cameco have launched an $80 billion partnership to build nuclear reactors. As I researched this, what I found frustrating is that the timeframe for building out these nuclear reactors is still on the order of 5 to 10 years. Alex, what are you seeing here?
Yeah, I've gotten quite a surprising amount of feedback from the community and the audience reminding me that I shouldn't ignore existing Generation III+ nuclear reactors in favor of SMRs and fusion reactors.
So I want to make sure I just nail this point. There are right now at least 6 AP1000s. These are made by Westinghouse, which ironically went bankrupt in 2017 building a couple of these in Georgia and South Carolina. Now it's hot again because superintelligence is hungry for power, and now it's incredibly valuable. They cost maybe about $7 billion to build, so an $80 billion partnership could maybe build 10 reactors all across the US.
This is going to be a very big deal. Critically, unlike SMRs, where there are maybe only 2 or 3 and they're relatively emerging technologies, this is, by comparison, a relatively mature format for nuclear power. And I think when we talk about the bridge to power for superintelligence—from natural gas to nuclear fission to nuclear fusion, with solar sprinkled and solar plus battery sprinkled throughout—I think Generation III+ reactors like the AP1000 have a very important role to play. So, Salim, I said it—
So Alex, may I ask you—
—to appease the audience that I'm not ignoring Generation III+.
No, but I think that's really important. These AP1000s are 1.1-gigawatt power plants. When Eric Schmidt was testifying in front of Congress, he said we need 92 gigawatts by 2030, right? So this particular deal might put 10 of these 1.1-gigawatt data centers on the map, but they're not going to be coming online until the early to mid-2030s. So the question is, how do we build out an additional 90 gigawatts in the next 4 years? Where is that going to come from?
Yeah, I think the deal structure behind this is worth understanding too, because it generalizes to solar, fusion, and everything else. What's happening here is a company, Westinghouse, which got bought by Toshiba. This is part of America deindustrializing very stupidly for decades.
Mm-hmm.
Toshiba buys Westinghouse. Westinghouse tries to build nuclear facilities. The government is so bureaucratic and so onerous that it goes bankrupt. So then, in 2017, the private equity guys come in, led by Brookfield, and say, “Okay, we have very smart business school majors here. We'll try and revive this thing.” And the timing is 2017, right when the Transformer comes out, so the timing turns out to be perfect.
So now what's happening is the private equity firms and the econ majors from all these schools are going to the government and saying, “Give us 10, 20, $30 billion in loans—guaranteed loans—and we'll use that to build these facilities. And then if they're successful, we make a huge profit. And if they fail, we write off the loans, so there's not a lot of downside.”
Many econ majors and business people should be shifting into this area because the government is open for business now. But that's the structure. You go to the government, you get the loan, you build the next big thing. The next big thing, if it succeeds, you get the profit, you get the margin, and the government subsidizes it. So it's a golden era, because a lot of the people, when I'm lecturing on campus, all the AI people and all the computer science people know exactly what they want to do. But then all the econ majors and business majors are like, “How do I get in this? How do I get in this?” This is how you get in this.
The flow of capital is in the—it's going to be $1.2 trillion a year by 2030 coming just into data center construction and power for it. There's nothing even close in the history of the world to that scale of money movement. So just inject—
Well, I mean, the rail—
—yourself right into it.
—building out the railroads, right? Building out the telecom networks—those were significant, just not these dollar figures because of inflation.
I have a clarifying question to ask here. In our last pod, we talked about the fact that the US is building 5,000 data centers, 10X more than anybody else. Is that inconsistent with the amount of energy available, or is that—
So we have 5,000—
—building with the energy?
No. It's the data centers we have today of all types, not just AI data centers. We're at 5,000, cumulatively more than the rest of the world combined. That was the number.
To me, this is one of the few things that's easy to predict. Everything is changing so quickly, but the chip fabs are exactly what they are. We're building them at a certain rate. Every chip is going to get used. The chips have a certain power consumption. That's very calculable. You can assume that they're all going to be sold out. But we can't make any more of them than 20 million GPUs this year, and then it'll expand at some rate. So, working back from that, you can exactly predict the flow of capital required to build out this entire infrastructure. And it's usually undercapitalized.
We're entering the golden age of Generation III+ nuclear reactors. 20 years delayed.
Three plus.
Yeah.
Three plus. Fair enough.
And, for what it's worth, for those who are looking at this on YouTube, the visual format—the form factor—of these plants actually looks like some sort of hybrid between what you're seeing here with the conventional older-generation cooling towers and the newer SMRs. You could be forgiven for mistaking it for a normal building.
One of the issues here is the US public, remembering the original Gen I and Gen II reactors—Three Mile Island and Fukushima—and not wanting these plants in their backyard. But the III+ are fail-safe nuclear systems that, again, I'm happy to have in my backyard. We've got to change the narrative, and we've got to accelerate this. Even companies that are bringing previous nuclear power plants back online are taking 5 years or more to get them online. The timelines are just too long.
And Dave, the point you're making is that even with this, we're at one-tenth the rate that we really need. Therefore, this is a guaranteed boom.
This episode is brought to you by Blitzy, autonomous software development with infinite code context. Blitzy uses thousands of specialized AI agents that think for hours to understand enterprise-scale code bases with millions of lines of code. Engineers start every development sprint with the Blitzy platform, bringing in their development requirements. The Blitzy platform provides a plan, then generates and precompiles code for each task. Blitzy delivers 80% or more of the development work autonomously while providing a guide for the final 20% of human development work required to complete the sprint. Enterprises are achieving a 5X engineering velocity increase when incorporating Blitzy as their pre-IDE development tool, pairing it with their coding co-pilot of choice to bring an AI native SDLC into their org. Ready to 5X your engineering velocity? Visit blitzy.com to schedule a demo and start building with Blitzy today.
7. Physical AI Takes Flight
All right, let's jump into a conversation on robotics, and in particular, robot drones. China just broke the world record for the number of flying drones: 16,000 AI-powered drones flying together, controlled by a smart AI system. Let's take a look at the video. If you were watching, you saw a beautiful drone show in the sky. If you were listening, you heard some music. But here's one interesting concept: imagine being able to have 16,000 drones up in the sky. You can actually create a giant TV screen and watch a television program or a movie across an entire city. What's the significance of this for you, Dave?
Oh, big time. I think people visualize robots in humanoid form just because that's what's in the movies—constructing buildings, cleaning your yard, and whatever. But I think the swarm version is just as big a deal, if not bigger. It's been very hard for Hollywood studios to create swarm visual effects, so they don't use them much in science fiction, and therefore people don't really think about them.
But when you look at a flock of birds or a bunch of bees, they're actually not coordinated. They're a little coordinated, but they're not really coordinated. The AI version of it, as you saw in that video, is perfectly, down to the millimeter, coordinated.
Mm-hmm.
And it's a very effective way to do things like construction, yard work, cleaning your gutters, whatever, because you can put 50 drones in if you need to pick up something heavy. You can put 2 in if it's light, so it's very, very flexible. So I think it's going to be a big, big part of taking AI and making it affect the physical world much more than people are currently predicting.
If you look at our podcast, we must have had 50 videos on different humanoids dancing and fighting and whatever.
Mm-hmm.
But I really feel like the drone part of it, or the interactive thousands-of-drones part of it, is way underappreciated as a real-world thing to do right now.
Well, I still love the episode last week about a mosquito-killing drone that you're flying around and zapping mosquitoes in your backyard. Drones are becoming the frontline of warfare in Ukraine.
It already is. The war is being prosecuted by half a million drones on either side. It's not a man thing. It's just drones killing—
And those drones are being—
—each other, and drones killing men.
—to a large degree, being manufactured in Ukraine. And one of the things that's interesting is, when the war is finally over—God willing, and with the leadership of countries willing—Ukraine as a nation is going to be sort of Europe's drone manufacturing capital. There will be a dividend to it there.
They're already the world's best army.
Yeah.
Because they've had to be. And, just as a corollary, bring it back to the U.S.: there are 10,000 drones a month crossing the Mexico-U.S. border. I keep telling government people this: drone technology beats wall technology. What the hell are we doing?
Yeah. Oh, my God. There's probably a good slogan there someplace.
All right, here's a fun article: “Elon Musk: Tesla Might Unveil a Flying Car.” We've been talking about this as the next-generation Roadster. God knows what it's going to be called. This is not a Model Y mass-production vehicle.
I remember being backstage with Elon when he had funded the Global Learning XPRIZE. We were backstage talking about Tesla before we went out. Talk about the guy's level of intensity: he was worried about a Falcon 9 launch that night, with a mission carrying one of the largest payloads. He was concerned about center-of-mass issues and vibration issues, and then he was concerned about whether Tesla could, in fact, survive the next quarter. This was in 2017 or 2018.
I said, “Well, are you going to put out the next Roadster?” He goes, “Man, oh man, nothing matters other than the Model Y and the Model 3.” He said, “That's our mass-production car.” He was always looking at that. But here we are, and I think the interesting conversation here is whether this Roadster will include jet or rocket propulsion from SpaceX that will give it the ability to hover or hop. That's what people believe will materialize. Any thoughts here?
I'd like to point out a historical irony, if I may. There was a lot of hand-wringing about 8 years ago, circa 2017, about how we were promised flying cars and instead only got 140 characters.
But if you play the tape forward, as it were, the 140 characters were Twitter. That became X, which became xAI, and that's now the integrated technical and capital structure that's poised to give us a flying car. So it's a bit of historic irony that, in some sense, the 140 characters actually gave us a flying car after all.
So—
That's such a great connection.
Elon has gone on record saying, “By the end of the year, there will be an unforgettable demo.” I'm excited to see what that looks like. The current gossip is that SpaceX is going to provide cold-gas thrusters, propulsion systems that might allow the car to hop—
What?
—or hover.
Are you serious?
Yes, seriously. So, cold-gas thrusters: imagine you have 30 seconds of hovering time until you recharge them. Interestingly, you could recharge these thrusters by simply compressing air as you're driving along. Who knows?
Well, I'll tell you: the highways were originally designed to go 100 or 120 miles an hour, but then turned out to be way too dangerous.
Mm-hmm.
But if you added that cold thruster to the car and it had accident avoidance built into it, its ability to jump and hover over an accident scene would be incredibly valuable.
I'm curious how high it can hover. I'm guessing we're talking about a foot. I think once you get out of ground effect, you're not going to have much hovering capability.
You guys have heard me talk about this before, but I just spent the last 4 days commuting from cities into JFK or Guarulhos in São Paulo. This cannot arrive fast enough for my taste, and if Tesla is able to unlock this world the way it unlocked electric cars, it will be massive.
Okay, let's point out right now that these are not eVTOLs. These are not, quote-unquote, “flying cars.” These are, at best, short-hopping, hovering cars.
We still have Joby, Archer Aviation, EHang, and all kinds of other companies out of China that will be making multicopter electric transport vehicles. Archer has the contract here in L.A. for the 2028 Olympics, so that'll be fun. But I think this is more of a fun thing: the kid with all the toys wants a flying or hovering car.
Part of the brilliance of Elon, too, is that car companies spend something like 7% of revenue on marketing.
Yeah.
He spends zero, but he takes that money and does cool things that are far more valuable than marketing. So it actually is net profitable for the underlying companies to pursue these crazy, interesting science projects, some of which turn into real products and keep the market cap high. It also replaces the marketing budget brilliantly.
A lot of people should be copying this. What can I do to be inspirational and cool and use that to drive people to my product, to my company, to my team, to my mission?
It also motivates the team, right? People want to come work at the coolest places.
It helps recruiting; you get the best people. It all works. Elon has invented the formula and perfected it, and I think Steve Jobs kind of invented it, but Elon has taken it to the next level. Everyone should just be studying it. Whether you like Elon or not, study it. It clearly works. It's the right plan.
All right, I love this. I don't know if this was—I guess it was a tweet. Elon on how to prevent global warming:
“A large solar-powered AI satellite constellation that would be able to prevent global warming by making tiny adjustments to how much solar energy reaches the Earth.”
Salim, we've talked about this at the XPRIZE for ages. I call this a solar sunshade: being able to have something out at a Lagrange point that's able to reflect a quarter of 1% of the sunlight impinging on the Earth, and basically act as a thermometer to titrate solar flux on the planet.
The challenge with this is that it's a tragedy of the commons. There are going to be some countries, like Russia, that want global warming because it opens up the waterways north of their country, while for others it's decimating their agriculture, like Africa and parts of Europe. No one can take action. Salim, what are your thoughts here?
There's an old story called the Pinatubo effect. When Mount Pinatubo in the Philippines erupted in the early 1990s, the ash covered the whole atmosphere for a while, and it dropped global temperature by two degrees.
One of the thoughts I've been thinking over the years is that there are about 7 majorly threatened areas with sea levels rising: the Washington, D.C., river delta, Bangladesh—a little country—Florida, et cetera. I thought they would actually just start launching rockets without telling anybody to do something like this. They have an existential threat, they're going to do it, and the cost isn't that heavy. But this gives you a computable capability, and you can calibrate it much more effectively—
Yeah, it's reversible; that's the most important thing.
And be very comfortable.
Right?
That's really key.
It's geo-scale engineering. Alex, you probably thought of this at length.
And, by the way, just to point that out: people complain, “Oh, my God, you can't go around geoengineering,” and the response is that we have been geoengineering by default by throwing up all this carbon. We have to figure out technological ways of doing it.
The COP23 and all that stuff—nation-states will not solve this.
Mm-hmm. Good point.
This idea of a global weather grid is one of the biggest ideas I'd like to push forward, and I haven't seen either the political will or the technical will to push it forward. It doesn't even necessarily need to be a sunshade per se. It could be as simple as satellites using microwave heating or some other mechanism to increase local cloud cover in some areas and reduce it in others.
That could be enough to solve a broader problem than just global warming, which is weather control. Wouldn't it be wonderful if we could steer hurricanes in one direction versus another, or mitigate storms? I think this is, from a technical perspective, doubly so in the era of AI, when we can have planetary-scale weather models, including more recent, strong ones out of DeepMind, that can solve this problem.
It's more, I think, a political problem of simply deciding that we want to do it as a planet.
Yeah.
Well, it's—
Will we—
It's an insurance problem, too. Oops, we steered the hurricane in the wrong direction.
Right.
It'll be a—
It's a social—
Heyday for—
—social problem, less a technical problem.
Yes.
It'll be a heyday for conspiracy theorists.
Yeah.
And so if there's no mechanism for global consensus, we're screwed as a world, so we have to get over that hurdle.
Yeah. Well, we don't get to Kardashev Type 1 civilization status without a global weather control grid. It's as simple as that.
All right. Here's the next article. This is a fun one: “Blue Origin Lands New Glenn Rocket Booster for the First Time.” We see a video here of what we're used to seeing SpaceX's Falcon 9 do, but this is a Blue Origin vehicle. Blue Origin launched a mission to Mars called ESCAPADE. Congratulations to Jeff Bezos, and the booster touched down on the recovery ship called Jacklyn, which Jeff named after his mom. How's that? “Hey, Mom, I got a gift for you. I'm naming the recovery ship after you.”
So this is a big deal for me. This is doubling our chance of getting humanity out into space and not being overly dependent on SpaceX, which is, by the way, still launching over 90% of America's spacecraft and probably 70% of the world's launches right now. Any thoughts on this one, guys?
I just think it's great that we have a second capability aside from SpaceX. I think it's good for the world.
Yeah, and Jeff's been spending about $1 billion of his Amazon stock per year to fund this. It moved a lot slower. I used to bug him about why he wasn't going faster, but hey, he's here now, which is great. And of course, Blue Origin's going to be using its own booster now to launch its competition to Starlink, which is already being deployed. Alex, anything you want to add here?
Yeah, I think having multiple reusable railroads, if you will, to orbit is exactly the sort of space race we want to find ourselves in.
Bingo.
If we're going to colonize and develop the solar system, we're going to need multiple routes to orbit.
Yeah. It was nice to see Elon congratulate Jeff on this. Of course, Starship puts all of these other launch vehicles to shame. Elon very famously said once Starship is up and operating, he'll shut down the Falcon 9 line, and it will outcompete Blue Origin, Rocket Lab, and everything else. It'll be the big sucking sound.
Can I double down just on that for a second?
Sure.
I think it's so awesome that he tweeted the congrats because it just shows that they're all focused on the bigger picture. This is not about competition. This is about solving the problem, and I think that's just fantastic.
Yeah. Agreed.
All right. Talking about the opposite end of the spectrum, this just made me mad. But it's the conversation we had earlier coming out of FII. People are concerned. Labor unions in Boston are fighting Waymo. The Boston unions formed Labor United Against Waymo, and the approach here is that they're going to force Waymo to put a human safety driver in the right seat or in the left seat, God knows.
We've seen this before, right? When France made Uber illegal, lots of places were fighting to retain these unions in place. Dave and Alex, you live in Boston. How do you feel about this?
Well, it's a little disconcerting that our tech hubs—the best and biggest tech hubs in the country—are also the most dysfunctional governmentally. This is utterly insane, right? It's obvious to anyone involved in it. But the populist uprisings are going to be all over the place on all kinds of topics. We've seen the picketers outside the front of OpenAI, and so this is going to happen all over the place. But if the governments of those regions don't get on top of it and put some kind of rational system together, then people are just going to leave. Waymo will go elsewhere, and it is already going elsewhere.
That's just going to be really bad for Silicon Valley, Boston, and New York. They've got to figure it out. Alex, you can rant about that.
One of the things that keeps me up at night, as it were, is this sort of regressionist approach, where people, unions, and organizations that are worried about employment fight the advance of technology that will save lives, increase economic wealth, and just make quality of life radically better.
I think one of the things—this is almost a meta-technology that we need to develop—is a way to maintain social cohesion while at the same time radically accelerating technologies. We haven't cracked that yet. Maybe social cohesion tech needs its own benchmark, and if we solve that, there's almost an optimal trajectory where we get our acceleration and our social cohesion at the same time. But we haven't cracked the social cohesion part of that, and I'd love to solve that.
Sure. We can talk about that at the summit if it comes together. But here's the deal: people are worried for their jobs. That's it. It's survival. I need to feed my kids. I need to be able to afford my home, and this is going to take it away from me. How can you possibly do that?
Until we level up our capability to provide people that safety net—whether it's universal basic income, which I'm much more interested in universal basic services—anyway, Salim, you were going to say?
One of the things we noticed after the Exponential Organizations book came out was that you're going to see this massive Ludditeville-type stuff against new technology, because people would much rather be comfortable than happy.
We actually focused on this. We solved what I call the immune system problem. We created a 10-week engagement with big companies that solves this in big companies. We've done it 100 times. We even have a nonprofit that does this in the public sector where you need to change, where regulation and this type of construct are the immune system. It takes 16 weeks, but it works. We've done it a bunch of times. So anybody facing this, just give us a call. We found a way of hacking culture at scale in defined environments.
Where do they reach you, Salim?
Just ping me at salim@openexo.com, and we'll show you how to do it. We've open-sourced the methodology for doing it because—
Yeah.
A few years ago, when the book came out, with all of this technology, if we don't solve the cultural resistance to it, it doesn't matter what the breakthroughs are. We're going to be fighting this political problem, and the next level we're going after is how to solve the immune system problem in an institution like healthcare, journalism, or education. They each have their unique immune systems, and we're working on that now.
I mean, we're going to see this across every industry as white-collar AI, superintelligence, and humanoid robots come in. This is just a small peak at what's going to be coming. We've got to solve it now.
8. AI Starts Solving Science
All right, let's go into our final segment here, which for me is one of the more important and exciting ones: what's going on in the world of science. I'm going to start this conversation with a video clip from Sam Altman on his thoughts about GPT-6 and the science leap that's coming.
If GPT-3 was like the first moment where you saw a glimmer of something that felt like the spiritual Turing test getting passed, GPT-5 is the first moment where you see a glimmer of AI doing new science. It's very tiny things, but here and there someone's posting, “Oh, it figured this thing out,” or, “Oh, it came up with this new idea,” or, “Oh, it was a useful collaborator on this paper.”
There is a chance that GPT-6 will be a GPT-3-to-4-like leap that happened for Turing-test-like stuff, but for science, where 5 has these tiny glimmers and 6 can really do it.
All right. Alex, let's open up with you.
I've gone on record as saying I think we're going to see many, if not most, grand challenges in math, science, engineering, and medicine start to fall to AI over the next 3 years, maximum. I think this is very much on my anticipated trajectory. Science is going to get solved, and all of its disciplines are going to get solved, and AI is going to do it.
I, for one, am super excited about finding myself in a near-term, Star Trek-type future where it turns out that centuries of human capital—or the equivalent of centuries of human capital—just gets solved overnight, at bulk, at scale, by AI.
Yeah. I love it. I think the amount of patents being filed and the amount of Nobel Prize-winning science being done is going to skyrocket. You can actually see an interesting chart where, if you look at patent filings post-ChatGPT, there's exponential growth immediately thereafter. It's an aid to humans, but all of a sudden, if it's autonomously doing the science in sort of closed-loop cycles, it's amazing. Dave?
Those are 2 things really worth tracking: the AI-generated patents and also the agent-to-agent transactions, part of which are licensing the patents.
Mm-hmm.
But that whole agent-to-agent intellectual exchange world—
Loop.
...is starting to really take off, and you can track it by transaction count and see the shape of the exponent. That'll be something we'll track really closely. Salim?
This justifies why I didn't pay attention during my physics degree. The AIs will do it for me, which is just fantastic.
Oh my God.
Well, look, if history is consistent, GPT-6 and Gemini 3 will be about the same. They're just leapfrogging each other, and we think Gemini 3 from Google is within a week.
Yeah.
So we have to carve out probably a big chunk, maybe a full day, just studying its capabilities.
And we will. When Gemini 3 comes out, expect us to go live with an analysis of it as soon thereafter as possible. Seeing a lot, right? We just saw OpenAI's ChatGPT-5.1 come out. Mira Murati's company has just gone from a $9 billion or $10 billion valuation to $50 billion. There's a lot frothing right now.
All right, let's move on to the next one: Zuckerberg and Chan bet AI can cure all diseases. Zuckerberg believes AI could make cures much sooner while empowering scientists to take risks. The Chan Zuckerberg Initiative plans to boost compute 10× by 2028, shifting all science work under their Biohub brand. This is great. I love that. They've had an interest in medicine and biology for some time, but now they're doubling down and focusing. Alex, let's go to you first.
You'll remember when CZI launched in 2016, the goal was to cure all disease, or most disease, by the end of the 21st century, and now the messaging has radically changed. Now the messaging is that we're going to have generative-AI-based virtual cells, and presumably virtual organs and virtual organisms built on top of those, enabling AIs to search intervention space for cures to all disease.
I think the subtext is that you don't necessarily have to wait until the end of the 21st century to cure all disease. This could happen in the next 5 years—call it 2030. I think all of the timelines, not just CZI but other nonprofits that are working on AI for broad-spectrum, generalist cures for all diseases, have similar timelines. You see similar messaging out of Anthropic as well: 2030, cure all disease with AI.
Yeah, we saw that from Demis Hassabis: within a decade, cure all disease, right? So we're seeing a huge amount of talent, compute, and capital going toward that goal, which is good news for everybody.
What I loved about this is that, to Alex's inner-loop point, instead of working on specific cures, they're just focusing on generating more compute and making it available to everybody. I think that's great.
Mm-hmm.
Sometimes, in my experience, it's easier to solve the more general problem than the more specific problem. It may perversely end up being the case that it's easier to just cure all diseases with AI than to cure individual diseases one by one.
Yeah, I mean, that's the concept around age reversal. If you didn't have the disease when you were in your 20s and 30s but it develops in your 40s or 50s, how do you turn back your epigenetic clock so that your cells are younger and thereby not expressing the disease, since you didn't express it in an earlier state of your biology?
All right, next one in this area. This is a conversation about one of the first real therapies that people are calling longevity therapies. The U.S. government slashes the price of GLP-1 drugs, and we're finding that GLP-1 drugs are lowering the risk of repeat strokes.
One of the challenges has been that GLP-1 drugs have been expensive, and they are a go-to for most physicians when someone has a particular issue, especially obesity-related or diet-related issues. Here we see TrumpRx.gov looking at bringing this down to $149 per month, which would be pretty amazing. We also find that GLP-1 drugs in particular are able to cut the incidence of strokes by as much as half in a 3-month period. Who wants to jump in?
I'll comment on this one. It's so exciting. I guess the elephant in the room is the outstanding question in biology: Why are GLP-1-class drugs so seemingly miraculous? Why are they able to treat so many different forms of biological dysfunction, not just the metabolic issues that they were originally intended for?
Putting the question of biology and mechanism aside, I think when we talk about universal basic services and an abundance of healthcare, I think this is the beginning of that. Offering GLP-1-class drugs for $150-odd dollars per month to U.S. persons who need them starts to look like universally abundant healthspan drugs. I think this is a major step in the right direction.
I want to put out the warning again, just because I'm in this world: GLP-1 drugs are not a panacea. If you are obese and using these drugs, it's important to use them as a means to change the way you eat, change your diet, and change your habits, because if you stop the drug during this period of time, you're losing weight, but you're also losing muscle, and you need to be exercising throughout this process.
If you stop taking the GLP-1 drug, what happens is you gain the fat back, but you don't gain the muscle back, and that's a problem. Sarcopenia is a true issue as we're getting older. Your muscle is your longevity organ. It's super important to have that realization. Make sure you keep exercising and building muscle while you're using a GLP-1.
I just love all the side-effect benefits we're seeing without even realizing it. I think that's so great.
Edison launches Cosmos, the AI scientist. This seemed like a really big deal to me. Alex, do you want to walk us through it?
Sure. This is another scaffolding-based approach to agentic science. This came out of Edison Scientific, as mentioned, and I think this is almost a preview of the age we're about to find ourselves in. Maybe we're a few months into it at this point, an age of bulk discovery.
It'll look a little bit like, if folks remember AlphaFold 3, where essentially overnight a large chunk of structural biology was more or less solved. It's going to look a little bit like that, except much, much broader. With this particular agentic AI researcher, there were discoveries across a number of different subfields of biology, not just structural biology. As was published in this paper, we'll see discoveries relating to things potentially helpful for Alzheimer's and some other factors.
The core technical advance claimed here is effective increases in context length. That's the key. The frontier models right now usually have context lengths in the millions of tokens. But if you wanted to develop the world's strongest AI scientist, ideally, naively, you'd want a model that has a context length in maybe the trillions of tokens, so that you could, in principle, feed it the entire internet and every paper ever published and then just ask it, “What's the solution? What's the solution to Alzheimer's?”
The approach that Edison adopted here was a little bit more practical than some sort of algorithmic advance that advanced the context window to trillions of tokens from millions. It focused more on knowledge graphs and other scaffolding techniques to achieve effective context lengths that are much larger.
The end result is still essentially the same. You put as much information, as much scientific literature, into the context window as you practically can, and then you crank it and ask for discoveries. Discoveries and innovations pop out. I think one could imagine a near-term future where we can just scale our way, scaling-law style, to major discoveries across all of the important biological subfields.
Yeah, here's one of the metrics they threw out: “Completes 4–6 months of expert human research in 12 hours.” It can read 1,500 papers and run 42,000 lines of code per experiment. It's almost brute-force-like research. Would you say that?
We've talked about this before, because what happens when you take all the millions of research papers that have been written in the past, where people missed findings, and now run them through? We might find incredible things.
Yeah, not just the papers, but the raw test results that are in digital form.
Yeah.
Just the incredible amount of information this can assimilate. When I look at my biology friends—I was talking to them last night, actually—they're all like, “Well, you know, these things always take a lot longer than you think.” You're like, “How do you get so cynical at a young age?”
This is a completely new approach. It's brand-new, greenfield territory, and if I look at what they actually have been doing for the last 3 years, they try to tease apart a single chemical reaction or a single test, and then they run it through MATLAB or Mathematica to try to tease it apart. Then they draw these plots that say, “Well, you know, we have statistical p-values here that have significance just barely.” And it's like, what a waste of time, man.
All these are interacting, and if you take the neural-net approach and just bombard it with raw information, it's really good when there are multiple things going on concurrently. It tries to find the conclusion without having to tease apart every single element. It's a brand-new way to do things, and it could do anything. It could be mind-blowingly capable and quick. You don't know because it's a new thing in the world. So put all that cynicism behind you and think about the rate of improvement that might be possible and just embrace it.
9. Gene Editing Rewrites Parenthood
Yeah. All right, our final article here, one for a fun conversation: “Genetically Engineered Babies Are Banned, But Tech Titans Are Making One Anyway.” This is worth the conversation. There are a few companies now being funded that are building CRISPR capability for embryo editing. The Wall Street Journal reported on one out of San Francisco called Preventive, backed by Sam Altman. I mean, Sam is backing an incredible number of companies, right? A CRISPR company, a brain-computer interface company, and probably dozens of others.
There's another company called Manhattan Genomics that was just covered in Wired. So, in my mind, it's a regime change: we're going from selection to alteration. IVF clinics already allow you to screen your embryo, right? You can fertilize a number of embryos, and then you can do single-cell sequencing and find out which of those embryos are safest to implant, but you can't edit your embryo. That is verboten under FDA rules. The FDA is blocking any of that. They won't even support any research in that area, let alone allow it to be done commercially. So these companies are beginning to look outside the US. Where can they go and do it? There's some conversation that this is happening in, or will be happening in, the UAE. What do you guys think?
I remember very well a colleague getting up and saying, “Look, the human genome is essentially software, and we have 50 trillion cells in the human body. Essentially, the human being is now a software-engineering problem.” And when you can edit the embryo, you're basically starting it from scratch. It seems inevitable to me.
And do we get, again, the normal thing? The question is, what do you want to design for? That's the big question.
Mm-hmm.
Yeah. I mean, we give our babies the best we can, right? It's like you start genetic engineering when you pick your spouse, right? Do they, are they successful? Are they intelligent? Do they look good? And that's the first step that you take.
Can I just go back? This is an old, really important point. We used to talk about the shift from film photography to digital photography and then all the implications of that. Essentially, we've gone from breeding and genetic evolution to a digital model, which just accelerates the whole thing.
Right.
When the baby's born, just one more thing: you give it the best health care you can, the best education you can, the best clothing you can. You're giving your child the best you possibly can, so the question is, why not start with the best genetic stack? And the fear is the whole eugenics conversation. Alex, over to you.
A couple of comments. The first one is the elephant in the room: the movie Gattaca, arguably one of the best cinematic depictions of germline editing of human babies, or at least germline selection, I should say.
I have to watch that one again.
Mm-hmm.
Yeah.
It's an amazing movie. Many view it as a dystopian future. I think if you look at it the right way, it's arguably a more utopian future, in the sense that we get space colonization. There's a SpaceX-type movie that's named Gattaca in the movie, and we also get healthy babies.
But the second point is more historical: the Asilomar guidelines were arguably the inception of many of these bans, soft and hard, against germline editing. Those are from 1975. There's a historical argument that Asilomar was actually triggered by the Watergate scandal.
Really?
We're 50 years on from Asilomar. There was a concern at the time. Some historians argue that the Asilomar guidelines were originally proposed, or at least motivated in part, because Watergate was fresh in everyone's mind, and there was a concern among scientists that if there were recombinant-DNA experiments that were not very well advertised or forthright, according to some sort of public guidelines, something bad would happen to the scientific community in whatever form. This is one historical argument: Watergate helped to precipitate the 1975 Asilomar guidelines.
I remember I was in grad school. I was doing my joint MIT medical-school degree at the Whitehead Institute on recombinant DNA, right? The first restriction enzymes had come out that allowed you to edit DNA in a somewhat precise fashion—nothing compared to what we have now with CRISPR. The headlines of the magazines, the cover stories, were “designer babies,” and there was so much fear. That was, God knows, 40 years ago.
We're 50 years on from Asilomar—50 years this year.
I actually had to go back after I saw the story and look to check to see whether there is even, at least in the US, a single federal statute that bans germline editing, and I couldn't find one.
Mm-hmm.
Which is a little bit surprising. There's a patchwork of federal and state laws and regulations that certainly deter germline editing, but not a single one that actually bans it. So I wouldn't be surprised if, in the near-term future, we find a generational conversation about whether germline editing should in fact be allowed.
And we have to remember that in 2018, there was a Chinese scientist, He Jiankui, who did this kind of CRISPR editing. He was trying to target the CCR5 cell-surface receptor that would prevent a child from getting HIV, and the guy was just decimated in China, in the press, condemned by the world.
Arrested and jailed, isn't it?
Yeah, yeah.
They arrested him.
So that put a kibosh on this idea, seriously.
Well, these are really complicated topics, and they all need thought leaders. But the number of people who need to rise to the occasion—if just in this podcast alone, between fusion and AI breakthroughs, driverless cars in Boston, and now this—they all need thought leaders. They need to say, “Here's what we should do. Here's an idea.” Ethical, trustworthy people who know what they're talking about—the need for that is so backlogged now, and this is just one of those topics.
And I shout out—
And shout out here to Hank Greely, who's been working on bioethics for a very long time. You know, the long-term effects of this are really, really consequential.
One of my biggest concerns is that Hollywood is decimating our future, right?
Mm-hmm.
I mean, this is my pet rant. Every movie out there is dystopian genetics, dystopian killer robots, and AI systems. I mean, no wonder that people fear the future if, in fact, the only futures they see in TV series and movies are ones that they don't want for themselves and their kids.
Yeah.
All you see is this, and you immediately go to this negative vision of the future, which is pervasive in society. We need to retool that. We need to reset that. We need more Star Trek in our lives, or the following version of it.
Maybe, Peter, you're giving yourself a call to action to start a new Hollywood studio, this time powered by AI, that paints a much more optimistic future.
Yeah, do that.
Be the change agent.
There is—well, I can't say much about it yet, but there is a project in the works that I'm involved in.
The fundamental problem is human nature is so fear-based, per my earlier comment today, and so you're fighting against that. The way to solve for this is to give those embryos psychedelics and solve this right at the core.
Oh, God, no.
That's the way to solve for this.
No, we're going to reduce the size of their amygdalas so they reduce their fear.
Yeah, I'm serious. That's the way. Edit it out. You don't need an amygdala in today's world.
You know what's funny about this whole storyline is that the birth rate in Korea now is 0.7 per couple—0.7 children per couple. One topic is editing your baby; the other one is, well, no one's having any babies at all.
Crazy.
So doesn't that seem like a more urgent issue?
It's a harder problem, a harder birth problem.
All right. So we're going to wrap this episode. I'll remind folks, if you're interested in the idea of the Moonshot Mates pulling together a couple-a-day amazing event in the fall of next year, send us an email, moonshots@diamandis.com. Let us know you're interested. We're going to try and get to 1,000 people interested. If we can get there, then we'll pull the trigger, and we'll make that happen.
We have our next 10X Shift workshop happening on November 19th, so we'll be talking about immune systems and organizations and giving people directions on how to build the exos, so come join for that. It's 100 bucks. It's limited seats, so come enjoy.
Amazing. Dave and Alex, any closing thoughts today before I play our outro music today from Adam 822, which is amazing? I love the fact that our subscribers are sending us their musical—
It's awesome. They're so good as well.
They're so good.
They're blowing my mind.
Yeah. So—
Well, I have a closing thought, which is that we're closing out our best venture fund year by far. I mean, it's just crazy what's happened this year. The companies in our portfolio gained about $12 billion of value within the year. Most of those ideas came from Alex, so I wanted to give a shout-out to Alex for the vision.
One of the themes in this podcast has been that, if you look at what Elon’s been able to do—like, should we have a satellite at the Lagrange point shadowing the Earth by 0.01% or something like that?—if your track record of being right is perfect or very, very good, then you actually get to say, “Here’s the answer, guys,” and people will flock to it just because your track record is right. And I think Alex is right on that cusp now.
Mm.
And that’s why he’s a little cautious on the podcast sometimes, too, because he doesn’t know. He actually says he doesn’t know, unlike me; I just say something anyway. But Alex’s vision for what will and won’t work is becoming so honed and so beautiful. I really wanted to thank you for the gangbuster year we’ve had.
Yeah. Thanks, Alex. Alex, a closing thought from you.
Dave, it’s very kind.
Yeah. In addition to—thank you, Dave, for the very kind comments—I’ll just say I spend substantially all of my time thinking about how we solve the hardest problems on Earth with AI.
Amazing. All right, guys, this is labeled—
Unbelievable episode. I'm gonna have to listen to this at least three times over.
I know. I need to too. And we listen, we read your comments. First off, please subscribe, tell your friends. One of the things I was so heart-filled about when I was in Mexico City, Italy, and Spain, and when Salim was in Brazil, were all of our fans there telling folks that they share the episodes. Thank you for listening. We do love the comments. If you have questions, I really want to do some AMA episodes with our subscribers, so drop them in the comments. Our team will aggregate them. This is a piece of music labeled All Right, Folks, From the Moonshots to Math, the WTF Just Happened Crew, by Adam822. All right, let's enjoy.
I was up at 3:30 this morning prepping for this episode. There was just so much to cover, and we still have another episode we’ll record later this week.
We need to basically make this a full-time thing because the world is happening so fast. Just trying to process this episode is going to take hours and hours now.
Ah, all right, guys. Love my time with you, as with each other. Be well. Every week, my team and I study the top 10 technology meta trends that will transform industries over the decade ahead. I cover trends ranging from humanoid robotics, AGI, and quantum computing to transport, energy, longevity, and more. There's no fluff, only the most important stuff that matters, that impacts our lives, our companies, and our careers. If you want me to share these meta trends with you, I write a newsletter twice a week, sending it out as a short two-minute read via email. To subscribe for free, go to demandas.com/metatrends to gain access to the trends 10 years before anyone else.
This episode is brought to you by Accenture. When your advertising operations fall out of sync, everything else follows. Spotify and Accenture are working together to reinvent the rhythm of ad sales using automation, analytics, and smarter workflows to simplify campaign delivery and access better data across the business. The result? Less time spent on operations, more time connecting brands with the moments and fandoms that matter most. Learn more at accenture.com/spotify.