Moonshots · 2026-07-29 · 124 min

Jensen Huang Launches an Open AI Alliance, Anthropic & OpenAI Team Up in DC, Kimi K3 Goes Global

Peter Diamandis

YouTube ↗

TL;DR

Open weights have become both NVIDIA’s stack strategy and a sovereignty strategy. Jensen Huang’s Open Secure AI Alliance argues that attackers already have frontier AI, so defenders need frontier AI ecosystems. NVIDIA benefits when models commoditize and value shifts toward compute. Dario Amodei’s counterargument, as presented by Peter and David Blakely, is that biology breaks the cyber analogy: powerful open-weight models could assist with pandemic-scale pathogens. His proposals were chip controls, restrictions on industrial-scale distillation, and safety testing for powerful models, open or closed.
The strongest regulatory call is to police AI actions, deployments, and compute—not intelligence itself. Alexander Wissner-Gross calls capability-based enforcement “thought policing” and favors defensive co-scaling: give defenders the strongest models first, monitor what systems do, and exploit even a one-month lead during recursive self-improvement. Salim Ismail doubts conventional KYC can survive shell companies; Alexander argues that transparency into installations, compute, and actions is the necessary layer.
The transcript calls the Kimi release K2 once and K3 elsewhere. The K3 passages describe roughly 2,500 downloads in two hours and about 100,000 in 24 hours. Peter got it running on dedicated Modal GPUs in under an hour at about $55 per hour; David Friedberg’s team had polling agents monitoring the repository. The investment implication is local control: “a model that you can control locally is way more valuable than a marginally smarter model” behind an API.
Claude Opus 5 appears to be an efficiency and vision-code release, not an unambiguous intelligence leap. It approaches GPT-5-level frontier intelligence at half the price, with pricing of $5 per million input tokens and $25 per million output tokens. ARC-AGI-3 rose from 1.5% to 30.2%, yet Alexander still preferred Fable 5 for general work and flagged weaker Frontier Math results, modest Humanity’s Last Exam gains, and possible benchmark gaming. The broader opportunity may be scaffolding, where a better problem representation can temporarily multiply domain capability.
OpenAI and Anthropic’s Washington alignment could create a safety regime and a regulatory moat simultaneously. Their reported agenda included an August 1 deadline, a voluntary 30-day government review of releases with serious cyber or national-security capabilities, and equivalent obligations for Meta, xAI, and startups. Peter sees both genuine safety concerns and possible regulatory capture; Alexander argues that the proposal targets the wrong layer.
AI profit pools may migrate toward compute, fabrication, proprietary learning loops, and applications. Peter’s four-layer model spans unreleased internal systems, paid frontier models, commoditized open models, and ecosystem distribution; Salim adds “layer zero,” compute and power. Alexander expects one power-law business model to capture most free cash flow, while David Friedberg expects multihundred-billion-dollar robotics, biotech, and entertainment companies to emerge above the model layer.
China’s export of models and infrastructure makes openness an instrument of geopolitical alignment. Salim warns that restricting US open models would protect a few domestic labs while handing the Global South’s AI ecosystem to China. Alexander’s warning is that foreign infrastructure can progress from lending to listening to “thinking for you,” making locally trained and locally controlled intelligence a possible sovereignty benchmark.
The physical economy is moving toward abundance, but monetary transition remains the risk. Starship 13’s intact splashdown supports a launch-cost path from the Shuttle’s roughly $54,000 per kilogram toward Starship’s stated $10–$100 target. Prima and Neuralink demonstrate new medical and mobility capabilities, but the panel did not claim they follow the same launch-cost curve. Elon’s “money won’t matter in 2036” scenario implies deflation and cheaper basics, while scarce assets, unequal ownership, debt, and fiat systems may persist.

Digest · the substance, structured for research

1. Jensen turns open weights into security infrastructure

Peter framed Jensen Huang’s first-ever X post as a deliberate market intervention: signed by 77 companies, it argued that open models strengthen safety, cybersecurity, innovation, diffusion, and sovereignty. The world needs “both frontier closed models and frontier open models,” and NVIDIA’s Open Secure AI Alliance would organize the defensive ecosystem around that premise.
Jensen’s proof case was Hugging Face’s reported intrusion: an agent logged 17,000 actions, escalated privileges, harvested credentials, and moved across clusters. The transcript names the blocked models as “GPT-5, 6, and Claude Fable”; an open-weight GLM-2.5 model then helped find and contain the attack. Peter’s summary: “Attackers have frontier AI, so defenders need frontier AI ecosystems.”
After three days’ silence, Dario said Anthropic had “never advocated for a ban on open-weight models.” He reframed the threat as authoritarian states reaching the frontier, especially in biology, and proposed blocking advanced chips and chipmaking equipment from reaching China, cracking down on industrial-scale distillation, and safety-testing powerful models, open or closed. Peter connected that emphasis to Dario’s biophysics background and Coefficient Bio acquisition.

2. Biology is the fault line the cyber analogy cannot settle

David Blakely’s defense of Dario began with the sharpest objection: “Okay, bioweapon.” Cyber models can help counter cyberattacks, but an AI cannot necessarily reverse a pathogen after release. He separated motive from economics: Dario’s position may protect Anthropic’s valuation, yet David believes Dario is “speaking his mind without an agenda,” even at the expense of his own valuation.
David rejected a simple open-source-software analogy because weights are tools for constructing downstream systems, not readable source code whose malicious component can simply be spotted. They might help “cure all disease and give us infinite longevity,” but that does not answer whether they should be handed to “every terrorist in the world.”
Alexander Wissner-Gross rejected biosafety as a sufficient case for restricting open weights: dangerous knowledge already exists online, dangerous acts are already possible, frontier AI may not be necessary, and biologically capable models already exist. He compared the warnings to Microsoft’s late-1990s FUD about viruses and IP lawsuits—claims that, in his telling, were overturned when open source proved safer in many settings.
Salim supplied a counterexample from the head of innovation at a three-letter agency: when capabilities are widely distributed, funding and opening the ecosystem can expose bad actors earlier. His preferred control point is transparent monitoring of data centers and workloads. Peter’s caution remained: once weights can run “in a basement somewhere,” defensive AI may see the threat only when it acts.

3. Defensive co-scaling requires visible deployments and compute

Asked about KYC, Salim said, “You can’t do it”—Chinese companies can use shadow entities in Singapore, making identity controls easy to route around. Alexander answered just as categorically: “Of course you can do KYC,” but superintelligence should go further by allocating some of itself to policing other superintelligence. “Defensive co-scaling is the answer.”
The apparent synthesis was transparency at the deployment and action layers. Open weights can remain open, but regulators would need visibility into installations, compute, and real-world effects. Peter suggested that governments might receive unreleased systems—possibly GPT-6 or a Claude successor—as white-hat defenders.
Alexander added a timing mechanism: in recursive self-improvement, a month or two of model advantage can compound from marginal into overwhelming. That makes temporary frontier access potentially more useful than permanent intelligence ceilings—provided the “good guys,” however defined, can detect and counter malicious behavior.

4. NVIDIA is opening a cold war over where AI margins accrue

Alexander sees the former détente between GPU suppliers and frontier labs becoming a “cold war.” NVIDIA’s incentive follows the aggregator rule: “commoditize your complements.” Popular open-weight models weaken profit capture at the model layer while preserving demand for NVIDIA compute; proprietary oligopoly would instead concentrate margins at OpenAI, Anthropic, or their peers.
His unresolved question sits one layer lower: why is NVIDIA not equally aggressive about commoditizing fabrication through TSMC or Samsung competitors? Salim’s answer was dependence—TSMC is too powerful to irritate publicly, so any NVIDIA effort would have to be exceptionally covert. Elon can announce a “Terafab”; a more dependent buyer cannot.
The Open Secure AI Alliance reminded Alexander of the Open Source Initiative’s 1998 creation and Linux’s challenge to Microsoft. Alexander also supplied IBM’s 1995 discovery: 95% of Fortune 500 CIOs claimed not to use open source, while 95% of their sysadmins said they did. IBM embraced what its hardware and services businesses could monetize.

5. OpenAI and Anthropic are writing guardrails their rivals must inherit

Peter said OpenAI and Anthropic were using shared Washington back channels ahead of an August 1 deadline for frontier-model rules. Their reported program included a federal review process, a voluntary 30-day government look at releases with serious cyber or national-security capabilities, and matching obligations for Meta, xAI, and frontier startups.
Alexander’s objection is architectural: applying identical safety rules to downloadable weights and gated APIs is “too clever by half” because deployment risks differ. His preferred line is memorable: “Aiming enforcement at intelligence is like thought policing.” Regulators should “police what the AIs are doing or being used to do, not what they’re thinking or how smart they are.”
Peter called the alliance “regulatory capture in real time,” comparing it with railroads, banks, telecom, and Big Tech: incumbents define acceptable guardrails, keep government close, and competitors out. He doubts it will hold because open models are improving too quickly, but views the attempt as economically rational.
Peter also argued that both CEOs may be genuinely trying to create a safe future rather than merely defend their valuations. He noted that Sam Altman is not an OpenAI shareholder and has roughly 400 investments in vertical companies that benefit from cheaper models such as Kimi. Alexander’s pushback was that genuine rulemaking should also include Google, Meta, and the rest of the ecosystem.

6. The winning AI business model will follow the bottleneck

Peter’s four-layer cake begins with unreleased “wild stallions” such as GPT-6, used internally to create materials, drugs, energy systems, and new businesses. Paid frontier products such as the models Peter labels GPT-5.6 solve and Fable 5 sit beneath them; commoditized open models power everything else; ecosystem owners monetize applications and distribution.
His distribution numbers were 3.5 billion Meta users across Meta’s apps, 2 billion Google users using Gemini, and roughly 1 billion OpenAI ChatGPT users. Consumers inside WhatsApp or another application will not ask which model answered; the ecosystem captures value because the model disappears into the product.
Alexander expects power laws, not four balanced pools: one business model may capture 80%-plus of industry free cash flow. Salim adds “layer zero”—compute and power—because value moves toward whichever resource remains constrained. David Friedberg sees value moving both downward to fabs and memory and upward to future robotics, biotech, and entertainment giants.

7. The Kimi release makes capability diffusion irreversible

The transcript calls the release Kimi K2 in one announcement passage and Kimi K3 in the surrounding download and architecture discussion. The K3 passages describe a frontier-adjacent model available globally through Hugging Face with no API key, gatekeeper, or revocation switch. The repository showed 2,500 downloads in two hours; Peter’s research suggested roughly 100,000 in 24 hours. Once copied widely, “there’s no undo button.”
David Friedberg’s team had polling agents pinging the page every 15 seconds and briefly encountered a 404. Peter interpreted that moment as a possible White House intervention before the page returned.
Peter got the model running on dedicated GPUs on Modal in under an hour. Full-throttle operation cost about $55 an hour, while voice prompting could provision 100 instances in about two minutes without specialized technical skill.
Peter called it “the biggest turning point in human history”: an AI capable of self-improvement running in the wild. He also cited a 13% decline in Anthropic secondary pricing after the K3 announcement, to about $230 billion, prompting the line, “It’s nice to lose $230 billion on someone’s tweet.”

8. Kimi swaps out the transformer while retaining its silhouette

Alexander’s architectural surprise was NoPE—no global position embeddings. Kimi Delta Attention, or KDA, behaves like a small recurrent neural network inside attention, smuggling in enough ordering information for a million-token context without the original Transformer’s explicit positional machinery.
His metaphor was the “Ship of Theseus”: attention, position embeddings, residual streams, layers, and sparsity are being replaced piece by piece. The system remains recognizably Transformer-like from a distance, yet its frontier implementation is increasingly unrecognizable beside Attention Is All You Need.
Salim’s practical reading was that NoPE supplies fading memory, unlike RoPE’s tendency, in his explanation, to give very distant context too much weight. He also highlighted the claim that attention can be removed around layers 100 or 120 with nearly the same output and greater speed.
Open access lets researchers—including those in China—try apparently obvious alternatives and lets other labs adopt what works. The discussion treated that diffusion as one of the central benefits of open models.

9. Local models turn token costs into an enterprise redesign problem

Salim’s call was that local control can outweigh marginal intelligence: companies can fine-tune proprietary knowledge, run inside secure environments, and avoid sending regulated data to a vendor’s cloud. That matters for sovereign workloads and industries where data control is part of the product, not merely a compliance cost.
Peter expects “token maxing” to turn a negligible line item into one larger than payroll by year-end and perhaps ten times payroll two years later. Specialized tuning, deleting general-purpose cruft, and streamlined models could yield 10X—and perhaps 100X—improvements for the same business process.
Salim’s implementation advice was deliberately narrow: rebuild one workflow that sharply increases revenue and one that sharply reduces cost, then expand. The thesis is not generic chatbot adoption; it is an AI-native operating model in which local inference, proprietary data, and repeated feedback compound together.

10. Opus 5 wins on cost, but Fable 5 still wins Alexander’s trust

Anthropic positioned Claude Opus 5 as its fourth Claude 5-generation release, approaching the frontier intelligence of GPT-5 at half the price. It became the Claude Max default at $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8.
ARC-AGI-3 produced the headline: 1.5% on Opus 4.8 became 30.2% on Opus 5, which Alexander said was the highest official baseline-model score he was tracking. He inferred optimization around front-end development and the vision-code intersection, and welcomed fewer refusals when benign biology or cybersecurity questions were misconstrued.
The counterevidence kept him on Fable 5. Humanity’s Last Exam with tools moved only from Fable’s 63.9% to Opus’s 64.7%, performance fell without tools, and legal, health, and Frontier Math results suggested weaker general capability. A third-party ARC-like game also failed to reproduce the headline jump, raising concern about “mild benchmark gaming.”
VoxelBench placed Opus 5 third, just behind Fable 5, at much lower cost; OpenAI’s Sol still led. Peter’s broader metric was velocity: Opus 4.8 arrived May 28, Opus 5 roughly eight weeks later with about 2X efficiency—faster than the previously discussed 10-week price-performance doubling. Alexander added that model releases had averaged one every six days.

11. Scaffolding is temporary alpha for domain entrepreneurs

ARC Prize Foundation organizers reportedly saw Opus 5 convert visual objects into algebraic software representations and do mathematics over them—the first such frontier-model behavior Alexander knew of. His caveat was explicit: that novelty holds “unless Anthropic was benchmark gaming on ARC-AGI-3.”
Peter said that reframing the puzzle might reportedly produce about 98% performance, although he could not fully replicate the result. The investable analogy is domain-specific: give a model a better representation of protein folding, drug discovery, or robot-arm design and it may appear three times smarter. “That is an entrepreneurial heaven.”
Alexander’s warning is that scaffolding advantage decays: “Today’s scaffold is tomorrow’s baseline capabilities.” Yet the current window is real enough that Peter wanted every university to teach scaffolding, with prompt engineering as a prerequisite. Opus 5’s one-shot Call of Duty demo reinforced the point: experimentation cost is approaching zero.
Privacy remains the counterweight. A Reddit user reportedly found that Google searches for shared Claude URLs surfaced health records, private documents, names, and telephone numbers. Alexander noted that the sharing links were designed to be social, while Peter and Salim kept the enterprise lesson: sensitive proprietary data eventually moves on-prem.

12. China is exporting an AI sphere of influence

Peter summarized the Financial Times framing as “Pax Sinica”: Xi Jinping using models and infrastructure as statecraft across the Global South while Washington debates open versus closed. Whoever supplies the intelligence stack can shape economic development and political alignment for decades.
Salim warned that restrictive US open-model policy would be a strategic “self-own and shooting your own foot at an epic level.” Protecting a few domestic labs would hand the wider startup, university, hospital, and defense ecosystem to China. “Openness is not a philosophical preference anymore”; it is soft-power infrastructure.
Alexander compared the choice to exporting F-16s: refusing a risky customer may simply send that country to Russian or Chinese suppliers, strengthening their industrial base. He also said that there was no US open-source model currently competing with the Chinese offer.
Alexander escalated the sovereignty argument: a foreign loan can seize an asset, telecom equipment can spy or shut down, but imported superintelligence is “thinking for you.” Peter’s supporting story involved a smaller Asian country taking a half-billion-dollar Chinese port loan when drones, improving every nine months, might eventually move 20,000-pound containers without that infrastructure.

13. Starship 13 moves launch economics toward ordinary infrastructure

Starship 13 deployed 20 operational Starlink V3 satellites, performed an in-orbit Raptor relight relevant to Artemis, and completed a precise soft landing in the Indian Ocean with the vehicle initially intact. The satellites were described as supporting roughly 0.5-to-1-gigabit connectivity almost anywhere on Earth.
Peter said the landing’s precision made a Flight 14 catch by the tower’s “chopsticks” plausible. The abundance curve was his real story: Shuttle launch cost roughly $54,000 per kilogram, Falcon 9 about $2,000, and Starship’s stated target just $10–$100.
Salim admired SpaceX’s willingness to treat every failure as information rather than embarrassment. Alexander Iskold connected the broader hardware renaissance to AI: models can specify part numbers, identify a German lens vendor, write manufacturing requirements, and combine rapid design with 3D printing and capable robots. “The amount of possibility is so exponentially bigger” than five years earlier.

14. Prima uses a near-term product to finance a deeper neural interface

Science Corporation’s Prima received a European CE mark for restoring central vision affected by age-related macular degeneration. Glasses capture the image and transmit it through infrared to an implant behind the retina, which stimulates remaining retinal cells. Patients reportedly gained five lines on a standard eye chart after 12 months.
Alexander Iskold called it almost Borg-like and noted that retinal surgery stays farther from direct brain intervention than Neuralink’s Blindsight approach. He expects competition among central and peripheral nervous-system implants, stimulation, wearables, ultrasound, and fMRI to reveal the most ergonomic interfaces within five to ten years; non-invasive vision restoration might arrive within 20.
Peter described Prima as Max Hodak’s “stage-zero revenue-generating engine,” funding a more ambitious interface in which neurons grow from circuitry into the brain. The long-term metaphor is a third hemisphere connected to the cloud—an “exocortex”—without the electrode damage Peter associated with current BCI approaches.
The entrepreneurial lesson was survival before scale. Apple began with a box of chips, Facebook at Harvard, and Google as a Yahoo plug-in. Peter cited Bill Gross’s conclusion that timing mattered most, then SpaceX’s own 2008 sequence: three Falcon 1 failures, a successful fourth launch, and a billion-dollar NASA resupply contract. “Just find a way to survive.”

15. Neuralink’s wheelchair is the first step toward embodied telepresence

Neuralink translated imagined cursor movement into analog wheelchair controls, including a safety design that returns the cursor to center if the user becomes incapacitated. Alexander’s next step is full exoskeleton control for paraplegic and quadriplegic users; Peter’s is an Optimus body whose eyes, ears, and limbs are operated remotely through Starlink—“effectively telepresence.”
Alexander then extrapolated, with hedges, toward behavioral or partial mind uploads trained from fMRI or ultrasound data. He expects something adjacent within roughly five years, “certainly by the end of this decade.” The liberating Avatar scenario carries its own social extreme: people who never leave a room and interact only through BCI-controlled robots.

16. Abundance demotes money before it eliminates scarcity

Elon’s 2036 prediction rests on a simple identity: if robots provide more goods and services than humans can consume, and output grows faster than the money supply, deflation replaces inflation. Salim separated money’s three functions—exchange, accounting, and storing value—and argued that some exchange mechanism survives wherever access, ownership, or location remains scarce.
Alexander translated the claim as “Star Trek economics”: food, shelter, healthcare, utilities, education, and entertainment could be effectively demonetized within ten years, while interstellar travel, a week on the Moon, antiques, or collectibles retain prices. His guess for demonetizing the total economy was closer to 30 years, “not investment advice.”
Peter’s bridge is a roughly $3,000 monthly UBI whose purchasing power rises as AI physicians, robot surgeons, autonomous vehicles, and robot-built housing push costs down. Wealth gaps might widen, but his criterion is whether the floor rises: every person gains food, water, energy, healthcare, education, and freedom.
Peter also cited Jeff Booth’s claim that each $1 of GDP growth has required $4 of debt, then imagined a $10 million TV factory whose $1,000 product falls to $500 and then $250 before the loan can be repaid. He said cash is losing about 14% annually and expects abundance to force a wholesale rethink of fiat, central banks, and economic measurement.

17. Proprietary learning loops outlast hidden reasoning

Asked whether Anthropic could prevent distillation by hiding model reasoning, Salim said it could make copying harder, not stop it. Enough input-output examples let competitors infer useful behavior. Durable advantage instead combines unique data, compute, users, distribution, and feedback loops that continuously improve the full system.
Acquiring companies does not automatically acquire their knowledge: much of it remains tacit inside key employees or incompatible operating processes. Salim expects more partnerships and co-ops, such as hospitals pooling data for pharmaceutical companies, because shared incentives may extract value more effectively than buying the organization.
Alexander rejected “garbage in, garbage out” as a universal indictment of synthetic data. Procedural worlds, generated code, and deliberately injected bugs can supply nearly infinite reinforcement-learning tasks; in the theoretical infinite-compute limit, a system could learn from entirely synthetic sequences.
Dave Blundin’s practitioner distinction was sharper: “Synthetic data is fine. It’s mislabeled data that actually kills you.” A single contradiction can warp weights as the network struggles to reconcile it. Clean generated data may therefore be preferable to supposedly natural data containing unrecognized errors.

18. Longer lives and AI education require adaptation, not population panic

Peter’s longevity mantra remained “LEV by 2033.” Salim described a spiky arrival: some subpopulations may already be approaching one year of gained expectancy per elapsed year, while others might arrive by 2030. Whether curing a disease counts as catch-up or escape velocity depends partly on accepting Salim’s premise that aging itself is a disease.
Population is not the binding constraint in Peter’s view. Replacement requires roughly 2.1 children per family, while South Korea and Japan were cited near 0.6; he expects world population to reach 9.5–10 billion and then fall. Apparent resource limits also invite substitution—lithium scarcity, for example, can yield to new deposits or sodium batteries.
Peter’s education prescription is AI-led, purpose-driven learning because fixed curricula cannot keep pace. High school should reward productive inquiry, real projects, entrepreneurship, prompt engineering, and scaffolding. Peter also proposed prompt communication, prompt engineering, and context engineering as a possible humanities credit at MIT; Alexander later identified prompt engineering as a prerequisite for scaffolding.
For individual investors, Peter suggested universities, crowdfunding, AngelList syndicates, VC funds, and demo days; Dave Blundin’s stronger advice was to earn access by helping early companies with sales, introductions, operations, or even moving offices. “If you add value, you will get stock”—the human learning loop beside the model learning loop.

Peter Diamandis

A couple of days ago, Jensen Huang, CEO of NVIDIA, said the world needs both frontier closed models and frontier open models. Anthropic was silent for 3 days, and there was a lot of conversation. Where’s Anthropic in this conversation? All Dario has to say is that OpenAI and Anthropic, who have been rivals for the longest time, have been pushing for the same agenda: a federal review process for the most powerful models.

David Blakely

Aiming enforcement at intelligence is like thought policing. Police what the AIs are doing, not what they’re thinking or how smart they are.

Peter Diamandis

Global AI diplomacy is coming. China’s leader, Xi Jinping, is wielding AI as a tool of statecraft, using it as leverage in China’s diplomacy across the Global South.

Salim Ismail

I can’t stress this enough: the whole power of the United States is its open and very broad innovation ecosystem. If they close up the open-model policy—

Alex

To the present. Why do you keep saying that, Peter? We’re right here, right now, maybe in our rearview mirror.

Peter Diamandis

Well, no. We’re on the curve, and we’re climbing at a hyper-exponential rate.

A lot to report this week, a lot to discuss, and a lot of intrigue from the frontier labs. We’re going to be speaking about Jensen Huang’s mission to create an open, secure AI alliance. We’ll discuss Anthropic’s past position and Dario’s position on open source, as well as the news that Anthropic and OpenAI are supposedly teaming up in Washington for lobbying efforts.

We’ll dive into Claude 5, and AWG will give us all the metrics, along with the release of Kimi K3 yesterday on Hugging Face. Then there’s the successful launch of Starship 13, and we’ll close with Elon’s comments on a post-capitalist world. You’ve got to love Elon. He does not disappoint.

You guys ready for this?

David Blakely

Absolutely.

Peter Diamandis

I don’t know. A little enthusiasm here, gentlemen. I mean, I can’t tell.

Salim Ismail

Amazing. Yes.

Peter Diamandis

All right. Let’s kick off this episode with a fight that’s framed the entire week: open source versus closed source.

A couple of days ago, Jensen Huang, CEO of NVIDIA, posted his first-ever tweet. He’s been on X for the longest time but has never tweeted. His first tweet was a letter on why open models matter, and it’s been signed by 77 companies thus far.

Jensen writes that open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. He says the world needs both frontier closed models and frontier open models. He then goes on to launch the open secure AI alliance. We’re going to talk about that.

In his letter, Jensen recalls a story that we reported on last week about Hugging Face experiencing an intrusive agent that logged 17,000 actions, escalated its privileges, harvested credentials, and moved across all of Hugging Face’s clusters.

The closed AI models—GPT-5, 6, and Claude Fable—that Hugging Face tried to use to hunt down what was going on blocked them. They blocked the forensic teams, and Hugging Face had to turn to an open-weight frontier model, GLM-2.5, to help Hugging Face find and contain the intrusion.

Jensen’s thesis in his tweet is that attackers have frontier AI, so defenders need frontier AI ecosystems. We saw Sam Altman jump in on this, saying OpenAI wants to have the United States leading in both open-source and proprietary models.

But Anthropic was silent for 3 days, and there was a lot of conversation. Where’s Anthropic in this conversation? Historically, they’ve been opposed to open source for a number of reasons.

Yesterday, Dario finally responded, saying he rejects the claim that Anthropic wants only closed models and doesn’t want open AI models. In his words, Anthropic has never advocated for a ban on open-weight models. There are a lot of videos showing that he was certainly hinting at that.

But Dario reframed the competition and this debate, saying that the real issue is not open versus closed; it’s whether authoritarian states, namely China, can reach the AI frontier.

Dario’s sharpest disagreement with Jensen is his belief that open-weight models could be used by attackers. Dario’s central thesis is that biology is the issue. Sufficiently capable models could weaponize pandemic-scale pathogens.

It’s worth noting that Dario, probably of all the frontier lab CEOs, is the most steeped in biology. He has a PhD in biophysics from Princeton, and he recently acquired a biotech company called Coefficient Bio.

So, rather than a ban, what is Dario proposing? Three things: first, block advanced chips and chip-making equipment from reaching China. I’m sure Jensen doesn’t necessarily like that one. Second, crack down on industrial-scale model distillation. And third, require safety testing for all powerful models, open and closed.

So, let’s dive into this. Dave, I’m curious. One of the things we talked about before is that if Dario really wanted levels of safety, he would put forward KYC requirements or crack down on mass distillation of cloud models, but we haven’t seen that. Your thoughts, my friend?

David Blakely

Well, right out of the gate, the argument, as you laid it out, is perfectly articulated. But if Jensen says, “Look, cyber threats can be defended with AI, and therefore open weights can defend against open weights within cyber threats,” all Dario has to say is, “Okay, bioweapons. I have a sufficiently advanced—how is my AI going to defend me from a bioweapon?”

You can’t argue against that. I feel really convinced. Dario, I am 100% convinced, is speaking his mind without an agenda. I’m not 100% sure about anybody else in this debate, but Dario is a brilliant guy laying it out exactly the way he sees it, even at the expense of his own valuation.

Everybody online is saying, “No, no, no. He wants closed weights because he has a competitive advantage, and if nobody else can get access, they’ll have to pay him.” True. But I don’t think that’s his motivation. I think he genuinely got into this industry long before there was any money in it.

Peter Diamandis

Yeah, no. Well, I’m good. I’m good.

Salim Ismail

I do believe you. I think so, but it’s interesting that Anthropic has gone from the most beloved, safety-conscious company out there to being raked over the coals over the last couple of days, or this last week.

Peter Diamandis

Isn’t that funny? But if you say the same thing about Sam and everybody and Elon, you go from darling to goat in a heartbeat in this world. It seems to be the common trajectory. As soon as you’re too big, everybody’s looking for ways to poke at you.

Alex

Yeah, be careful on the way up, because you’re going to get slammed on the way back down.

Peter Diamandis

Alex, thoughts?

Alex

I think after a number of years of détente between the infrastructure layer—that is to say, GPUs and the lower layer of the stack—and the model layer—that is to say, OpenAI, Anthropic, and other model providers at that layer—I think we’re seeing the beginnings of, if not open war, then at least a cold war between them.

The first rule, if you’re an aggregator in business, is to commoditize your complements. NVIDIA has been very stealthy, very polite, and very diplomatic about its desire to commoditize the model layer. It has struck agreements, including what some have argued are circular, wash-sale-type agreements, with the data centers providing compute for OpenAI, Anthropic, and others.

Now I think this is turning into open warfare. The real question is: where do the profits accumulate in the superintelligence stack? Are they going to accumulate at the GPU level, in which case NVIDIA wins—and NVIDIA wins by popularizing open-weight models that can’t capture value at a higher level in the stack? Or does value live at the model layer, in which case we see a proliferation of duopoly or oligopoly, with high-profit-margin model providers? Anthropic is reportedly highly profitable. Or does it live elsewhere?

Due to the competition, and quite frankly due to the outstanding success at the frontier of what until recently looked like an OpenAI–Anthropic duopoly, I think we’re seeing the NVIDIA GPU layer fire back.

What I don't quite understand is why NVIDIA isn't working more aggressively to commoditize—or commodify—the layer of the stack beneath them. Why isn't NVIDIA aggressively financing, say, TSMC and Samsung competitors? Why is Elon doing it and not NVIDIA? That's a head-scratcher for me. They really should be as aggressively pushing a secure, open fabrication initiative, just like they are for one layer beneath versus one layer above.

Salim Ismail

We could brainstorm on that for a whole episode. You know, I know the answer is that if he is doing it, he has to be doing it very, very secretly, because you cannot irritate TSMC for even a minute. They're so in control of the world right now. But if you're going to do it, Elon is the one guy who's overtly said, “I'm going to build the Terafab. I'm going to build something.”

But he's fearless. Any rational person has to be afraid of irritating TSMC. So if he's doing it, he's got to do it so secretly and so stealthily. I think he's actually doing it, because it's hard to contain that, but that is a really great question, Alex. I would love to riff on that sometime for an hour.

Peter Diamandis

What do you think of the open secure AI alliance that Jensen proposed?

Alex

It reminds me—if you think back, we're in 2026—of 1998. Do you remember when Eric Raymond and Bruce Perens founded the Open Source Initiative, OSI? It reminds me of just that.

If I look at the historic arc of commercial versus open-source AI models, I think we're at a point in this arc that's roughly analogous to where Microsoft was in the late '90s, when they just totally dominated the future light cone of software. In a sort of case of history rhyming, open source came from outside the U.S. Linux came from Finland. Sure, Richard Stallman and the FSF came from Cambridge, Massachusetts, but Linux, which really was arguably the nucleation event for open source, came from Finland and was popularized with an American Open Source Initiative.

There are parallels there, and in combination with the antitrust verdict against Microsoft, that helped to unlock the future light cone for just about everyone else afterward. So I think there are interesting historic parallels.

Peter Diamandis

Salim, last week you said something that got a lot of love in the comments: “Intelligence wants to be free.” I mean, the whole open-source movement is sort of the abundance thesis writ large, right? Your thoughts?

Salim Ismail

Yeah, completely. I like this Open Source AI Alliance, whatever the name is, because Jensen is reframing open weights from a security vulnerability to being a security capability. This whole alliance is very ExO. If you put together a community of people, they will be able to defend in a very powerful way.

We can look at the attackers: They have access to open models and powerful AI. The defenders can't have that if they have access only to a closed model and don't understand the outputs. They have to be able to inspect it, et cetera. So this is a great approach for some of that.

And to Alex's point, this is exactly the same transition as the open-source software transition, and everybody wins in open source except for the closed people.

Peter Diamandis

Yeah, and NVIDIA wins by supporting everything. I mean—

Salim Ismail

Yeah, there's a clear—

Peter Diamandis

Remember who ultimately, back in the '90s and early 2000s, was one of the biggest supporters of open source: IBM. At the hardware layer and the services layer, they benefited from the commoditization of software. Same here. Always, if you're an aggregator, you commoditize your complement.

Alex

I've referenced this before, but it's just worth bearing again. In 1995, IBM pulled all the CIOs of the Fortune 500 together and asked them, “How many of you use open source in your tech stack?” Ninety-five percent said, “No, we don't use open source. We're closed shop.”

Then they went to the sysadmins and asked, “How many of you use open source?” Ninety-five percent said, “Yes.” So IBM made a major bet on open source, which turned out to be a massive success. It also showed you that the CIOs had no idea what was going on in their enterprises.

Peter Diamandis

Yeah.

Alex

Well, that's all business strategy, and I think Jensen's talking business strategy in this alliance. But we already knew Alex Karp is working with Jensen to build a monster enterprise open-source model that is at the frontier level, so that enterprises can control their own AI, use the Palantir application layer to manage it, and use Jensen's chips to run it. So that's great business strategy. It doesn't answer the question of bioweapons. It's like, “This is how our business wins.” I get it.

Peter Diamandis

What does that mean?

There is one answer. There's a precedent to the bioweapons thing. We had the head of innovation of one of the three-letter agencies at Singularity once, so we asked them directly, “Do you think about the threat with open source and somebody being able to engineer a virus?”

He actually had a really amazing answer. He said, “When you have something like nuclear weapons, and you know how many there are and where they are, you put eyes on it, right?” When there's a distributed capability, what they've been doing is actively funding the ecosystems, opening them up, and making them more open, because it's much easier to spot bad actors.

That was a very smart way of going about it. I have much more respect for them than I thought I would have coming out of it, because something dodgy is going to come out and be visible much earlier than if you tried to close it up.

Alex

I also don't buy the biosafety argument. I know that's a favorite hobbyhorse of some frontier labs: to emphasize biosafety. I don't buy biosafety as an argument for several reasons.

One, you can just go out on the internet and find things. Two, without being specific, you can just go do things in the world that are dangerous already. Three, I'm not even sure you need frontier AI to discover new ways to do dangerous things in various disciplines.

Peter Diamandis

Yeah. That's—

Alex

And four, there are already models out there that are quite capable with their biological knowledge. So I think biosafety is, again, history rhyming. Do you remember how Microsoft in the late '90s made all of these fear, uncertainty, and doubt arguments about anyone who was touching open source? “Oh, you'll get viruses. Oh, you'll be subject to IP lawsuits.”

They came up with 10 different arguments for why open source was too dangerous to use in the enterprise, and they all ended up being wrong. In fact, perversely and ironically, open source ended up being safer than closed source.

Salim Ismail

Yeah, that's interesting.

Alex

And in this situation, NVIDIA is really well positioned because they win whether it's open source or closed source or both, because they're the infrastructure providers.

Peter Diamandis

Right? They're commoditizing their complements, and they need a proliferation of open-source competitors.

Alex

Yeah, the argument, though, with open source is that you're basically saying, “Look, if everybody's looking at the source code, if there's anything evil in there, somebody will see it. Everyone should be looking at the source code.”

Here, you're saying, “Okay, with open-weight models, everyone should be looking at the weights and seeing if there's anything evil in there.” The weights are not used—you don't run the weights. They're used to build other things. You use the weights to create a bioweapon. You use the weights to create a regular conventional bomb that goes off when a specific person is walking by.

So the weights are not a self-contained piece of open source. They're a tool to build other things. That analogy doesn't hold.

Salim Ismail

You can add some positive things to what the weights will do, like write a sonnet or get your job—

Peter Diamandis

Well, cure all disease and give us infinite longevity. I mean, this is the greatest thing that's ever happened to mankind. But you can't just throw it out there to every terrorist in the world and say, “Here, you can have it too.”

Listen, I'm just reminding everybody that our amygdala is on overdrive right now. Our brain is wired to give 10 times more attention to negative news than positive news. And that's what we've seen. We saw both Sam and Dario talk about job loss and talk about the dangers in all of this.

Part of it is the regulatory capture that we'll talk about in a minute, and part of it is getting attention and coming in as the savior. They both flipped their scripts on this.

Alex

Yeah, I think there's an opportunity to contain it at the weights level in open source, but there's also a better opportunity to monitor the actual data centers and have clear reporting and global transparency on what's running where.

Peter Diamandis

But what about a KYC solution? What about knowing who's using the models?

Salim Ismail

Yeah, can't do it.

Peter Diamandis

Why not, Salim?

Salim Ismail

You can't do it. It's too easy to bypass. Look, Chinese companies have shadow companies in Singapore doing things that they want. It's very difficult to track or police this.

Peter Diamandis

Alex, do you think it could be?

Alex

Yeah, of course you can do KYC. We can do better than KYC. If we're going to be in a world awash with superintelligence, let's allocate some of the superintelligence to policing the other superintelligence. Defensive co-scaling is the answer.

Salim Ismail

Yeah, totally right. I agree. Totally right. But it requires transparency. If you throw it out there as open weights, defensive co-scaling will work really well if the police AI can see the danger AI.

Peter Diamandis

Sure.

Alex

So you need that transparency layer. As soon as you throw it out there as open source, that’s fine, but now you have to crack down on the install and the compute. Where is it running? The new danger is that it could be running in a basement somewhere and no one would know.

Peter Diamandis

Until it takes action, right? Then we need real-world defenses against those actions. We all know that both OpenAI and Anthropic have models far better than what they’re showing us, right? Our next story is going to talk about that: Both of them are going to D.C., probably to unveil what GPT-6 looks like, or what the follow-on to Claude looks like. The government is going to have access to those as a white-hat defender. Salim?

Salim Ismail

I just want to make one more comment on Dario here. I do agree with you, and I agree with Dave that his intent is probably clear, but right now, the safety argument and the economic self-interest are very overlapped and hard to separate. They’re facing a very aggressive innovator’s dilemma response. Cheaper alternatives are coming very close to their capability, and that’s a very unpleasant place to be if you’re an industry leader.

Peter Diamandis

Yeah, I looked it up on the secondaries. Anthropic dropped 13% after K3 was announced, to about $230 billion. It’s nice to lose $230 billion on someone’s tweet.

Alex

I would have done much more.

Peter Diamandis

Yeah. Let’s go to our next story, which is related. OpenAI and Anthropic, who have been rivals for the longest time, have teamed up in Washington on lobbying. According to The Information, they’ve been working through the same back channels ahead of a Trump administration August 1 deadline to finalize rules on frontier models.

Alex, I’ll ask you to explain in a moment what that deadline is. They’ve been pushing for the same agenda: a federal review process for the most powerful models, a voluntary 30-day government look into the release of anything with serious cyber or national security capabilities, and a framework that would force their competitors—Meta, xAI, and all the frontier startups—to play by those same rules.

The question we’re chewing on today is whether AI needs guardrails. Who gets to set them up, and who gets locked out? This is a potential regulatory moat, a defense layer for Anthropic and OpenAI. Alex, you want to take this one?

Alex

Maybe let me point out the cliché, more superficial analysis, which is that the frontier labs are, to some extent, talking out of both sides of their mouths. It’s been widely reported that they’re publicly supporting open source, while privately throwing all sorts of monkey wrenches into the regulatory gears in order to derail any prospect of a free and open-source, open-weight future. They tell lawmakers and politicians privately, “Well, they’re unsafe for a variety of reasons.” I think the cleverest angle is simply to regulate them like you regulate the closed-weight models, subjecting them to the same safety standards.

I think that’s too clever by half, in some sense, because they’re not the same models. From a deployment perspective, which is half the battle, deploying an open-weight model has a very different deployment situation than accessing a closed-weight model through a gated API. So I think that’s sort of the obvious story.

A slightly less obvious story may be where the enforcement happens. What’s the right bottleneck for defensive co-scaling to work? I think one of the more interesting bottlenecks—or, let’s say, comparative advantages—that we’ve seen over the past few months, and that hasn’t been quite reported this way for defensive co-scaling, is simply a matter of time.

If the good guys, however you want to construe that, have access to the strongest models just a little bit ahead of everyone else, including the bad guys, then in an era of recursive self-improvement, what historically might have looked like only marginal advantages turns into enormous advantages. If the next-generation model suddenly generates step-function leaps in terms of its capabilities, then even a period of a couple of months, or 1 month, could make all the difference in the world.

Peter Diamandis

This is the recursive self-improvement argument as well.

Alex

It is. Well, it’s the regulation-of-RSI argument. The second point is that I continue to think enforcement is being leveled at the wrong part of the stack. Fundamentally, aiming enforcement at intelligence is like thought policing, but for the AIs, not for the humans. I’d much, much rather see enforcement leveled at the action layer: Police what the AIs are doing or being used to do, not what they’re thinking or how smart they are.

Peter Diamandis

Mhm. I’ve been thinking a bunch about the conversations going on. If we have incredibly powerful open-weight models, how do the top frontier labs make money? How do they survive against this onslaught of free?

The way I think about it—and I’d love your feedback, guys—is like a 4-layer cake. Layer 1 is the top layer. Call it the wild stallions inside OpenAI and Anthropic, right? Unreleased, brilliant AIs. You can think of it as GPT-6. You don’t let it out. You keep it to yourself. You use it for breakthroughs in materials science, biology, and building new businesses.

This is what you and I have discussed at AWG and Solve Everything. These models are going to create trillions of dollars in other adjacent spaces: longevity, materials science, energy, et cetera. So that’s the first layer: the most advanced models you use for yourself.

The second layer is the models on the Pareto frontier, right? This is GPT-5.6 solve. This is Fable 5. People will still pay for something that’s a little bit better than Kimmy K3, right? So you’ll make money providing the next-best model to people, just above the open models.

Layer 3 here is the open-source models, and everyone gets to use them. They’re good enough. They’re fully commoditized. They’re powering everything else.

Layer 4—and we’ve talked about this before—is the fact that companies like Google, Meta, and xAI have entire ecosystems, right? They make their money on the application layer. Meta has 3.5 billion active users using Meta’s apps, for example. When I’m inside WhatsApp or whatever, I’m not thinking, “What model am I using?” WhatsApp answers my questions.

Google has 2 billion active users using Gemini, and this is before they get on Apple. OpenAI has about 1 billion on ChatGPT. So the fourth layer is that they make their money when they provide their models to their communities. Yes? No?

Alex

Peter, I think your narrative—as I heard you say it—you were almost narrating the cost frontier of capabilities versus cost, starting from the upper right-hand and going to the lower left-hand through different business models. I think it’s an interesting narrative, but my bet is that, as with so many other things in life, everything follows power laws in the end.

I think just saying, “Well, there are these 4, or there are these N, business models”—in all likelihood, 1 of the business models is going to account for 80+% of all the free cash flow and all of the profits. So I think just saying, “Well, there are these multiple business models,” is probably unrealistic. There’s probably going to be just 1 business model that runs away with most of the profits.

Peter Diamandis

The point being: Don’t cry for the closed-source companies. They have plenty of ways to make money, even in our world now.

Alex

Cry for them? My goodness. 2 or 3 months ago, we were crying for everyone else who was going to be displaced—all of the labor, the service jobs that are being displaced by the frontier models. Now we’re crying for the frontier labs? So cry for everyone.

Salim Ismail

I think there’s a layer 0 in your stack, Peter, which is compute and power.

Peter Diamandis

Sure, the infrastructure layer.

Salim Ismail

The infrastructure layer, right?

Peter Diamandis

The frontier labs, like xAI, will own that as well. And Google will own that.

Salim Ismail

I think over time, as you get more and more powerful free models, the value will accrue there.

Peter Diamandis

It depends where the bottleneck is, and it’s clear that’s where the bottleneck is. For me, when I look at what’s happening with Anthropic and OpenAI, this is regulatory capture in real time. Every major industry is trying to do this. The railroads did it, the banks did it, the telcos did it, and big tech did it.

Now they’re trying to do it to set up the guardrails, to keep the government at bay but also to keep other folks at bay. It’s right there, and that has economic consequences. I don’t think they’ll succeed because the open-weight models are moving so quickly, but it’s a worthwhile try if you were them.

Yeah, I don’t know if you saw David Friedberg’s comments on this. He had a beautiful soliloquy in which he said, “In the ’90s, Netscape tried to own the server and the browser—the whole stack. Then Mozilla came out with Firefox, browsers went for free, and all of the value shifted to the application layer: Google, Amazon, and so forth.”

I think potentially that’s the same thing here. If that’s the case, then the fear about OpenAI and Anthropic running away with the show gets ameliorated.

Salim Ismail

Well, I think that’s exactly right. I think, actually, it’s going to go up and down, per Alex’s prior comment. Right now, you have $5–10 trillion locked up in the labs with their models, and in the chip companies that don’t actually make the chips, such as NVIDIA and AMD, et cetera.

David Friedberg

But underneath the chip companies that don't actually make chips, you have the fabs, who are largely overlooked: TSMC, Intel, and Samsung.

Peter Diamandis

And they've been skyrocketing.

David Friedberg

Skyrocketing. They're the memory companies too. So it's going down, and as you said, Peter, it's also going up to the use cases. I'm almost positive that if you look 5 years in the future, there'll be many, many multihundred-billion-dollar robotics companies, biotech companies, and other use-case entertainment companies that don't exist today that have used AI to have a hugely impactful user base, if it's entertainment; drug portfolio, if it's biotech; or robotics line—the manufacturing line. All that stuff is incredibly sustainable.

What did I just overlook? Oh, the foundation model companies and the fabless chip companies. That's why there's so much turbulence right now. The stocks are going up and down like yo-yos because no one's sure if they're actually going to have sustainable value in the end as everything moves to the upper and lower layers.

Peter Diamandis

The entire ecosystem moves up and to the right, right? And this is where Elon comes in with saying, you know, our GDP is going to double-digit growth and then triple-digit growth. Let's go back to the original story: OpenAI and Anthropic teaming up in Washington, D.C. I mean, this is a regulatory-capture story. Any thoughts on that?

Alex

I think the open source—again, I've mentioned this now on 2 prior occasions—the Chinese Communist Party coming to rescue American capitalism from itself. I'm not a fan of regulatory capture or the duopoly scenario that we would have found ourselves in. I hope that the regulators and the applicable regulators are able to see now that the vocal majority are interested in keeping the model layer competitive and are not interested in FUD reminiscent of the late '90s, directed at open-weight models, even if the strongest ones do happen to originate from China. I think that's the only way we all win.

Peter Diamandis

What's FUD, Alex?

Alex

Fear, uncertainty, and doubt. FUD.

Peter Diamandis

Okay. Thank you. Thank you. Honestly, though, I really think it's not a regulatory-capture move. I think both guys are genuinely trying to create a safe and secure future world. Because remember, Sam is not even a shareholder in OpenAI. Yes, he runs it. Yes, it's his lifeblood. He has 400 vertical company investments that are overjoyed that Kimi K2 came out. All of our portfolio companies are overjoyed that they have access to Kimi K2. Blitzy was over the moon. This is the biggest boon in—so, that's where Sam's economic upside is.

But yet he's still going to D.C. to say, “Look, we've got to have some rules. This is going to get out of hand.” So I don't think they're out there to try and drive up their stock price. I think both guys are out there to try and make the world safe.

Alex

Then why is it just them? Why isn't it everybody else?

Peter Diamandis

Well, who is everybody else? Because there's only so many people they'll let in.

Alex

And Google. I mean, there are a few other players in the mix here.

Peter Diamandis

Yeah.

Alex

I'd like to make a slightly tangential point here. I want to echo what Jensen Huang said: open models will make the U.S. stronger because you're building an ecosystem. You have universities, startups, defense contractors, and hospitals. As Dave said, every company is thrilled to bits that you have an open-source model that's as powerful, with open weights. You can go manipulate those weights. And when you have the whole ecosystem, open beats closed always.

David Friedberg

Yes, that's right.

Peter Diamandis

That was the faster way to go—one of the theses of our book together. Yeah, so open beats closed.

David Friedberg

Yeah.

Peter Diamandis

Over time. I mean, it is amazing that we're living in this incredible demonetization world of intelligence, right? It's just falling through the floor. It's 99.95% cheaper over the last 3½ years. I looked at the numbers.

And I think Alex makes a really important point. You don't try and regulate open versus closed. You try and regulate the capability.

Alex

Yeah.

Peter Diamandis

And the actions and the outcomes. And I just, for the life of me, don't understand why we're shedding any tears for the profit margins of a couple of frontier labs. This is what intelligence too cheap to meter is supposed to look like. Intelligence is supposed to get cheaper, and capitalism is doing its thing, creating competition and driving profit to zero. This is what we want to happen.

In full disclosure, I don't own any of OpenAI or Anthropic, so I'm not shedding tears for that. I don't think any of us do. Do you, Dave?

David Friedberg

Not that I know of, but I have a lot of indirect stuff.

Peter Diamandis

Alex, I know you just own the index, so you're set.

Alex

Just indices.

Peter Diamandis

Yeah. All right, let's go to our next story, and it's one we just started discussing. Yesterday, July 27, Kimi K2 went live for a global download on Hugging Face: a frontier-adjacent open model that anyone, anywhere on the planet, can download for free. No API key, no gatekeeper, no revocation switch. Once these weights are downloaded 10,000 times, they are free. There's no undo button.

Kimi K3's official Hugging Face repository showed 2,500 downloads in the first 2 hours, and the research I did shows about 100,000 downloads in the last 24 hours. I downloaded it. Dave, Alex, did you?

David Friedberg

Yeah, funny story on that, Peter. We had a whole bunch of polling agents pinging it every 15 seconds because I was worried that it would go away. I didn't realize a bunch of our companies were also doing the same thing. So, of those 2,500 downloads, we had dozens of them from here. [Laughter] But mine went through with no trouble. Right before it came out, though, the whole page went to a 404 error.

Peter Diamandis

And yeah, did you see that? I was like, “Oh my God, the White House intervened. This is not actually going to happen.” Because I've been telling everybody I think this is the biggest turning point in human history. You have an AI capable of self-improvement now out in the wild that anyone can use.

David Friedberg

Right?

Peter Diamandis

On your machine. This is massive. I just feel like it may not get documented in the history books that way. It may be the outcomes of this that get documented, but this is really the moment in the history of humanity that I think is so pivotal, and it was yesterday.

But anyway, it downloaded just fine. I got it up and running on my own dedicated GPUs on Modal. It took less than an hour to get a fully functioning Kimi K3 thinking and working 24/7. It's a little pricey, but it feels like $55 an hour on Modal to run it full throttle. But you can prop up 100 instances in 2 minutes now if you want to, just through voice prompting.

You don't have to have any technical skill at all. You can just go to Modal, ask it to install Kimi K3, download it from Hugging Face, and start talking to it. You're up and running in no time. It's mind-blowing. Alex, your thoughts?

Alex

I looked at the architecture. Now that this is actually open source and open weight, it's pretty interesting. The most interesting thing I saw in the architecture is that position embeddings are gone. This was one of the most critical elements of the original Transformer architecture. It's gone. It's literally called NoPE: no position embeddings, NoPE.

You could ask, how on earth is a model like this, with 1 million tokens of context, supposed to know what it's looking at without positional embeddings? You look a little bit more closely, and the attention mechanism—Kimi Delta Attention, or KDA—is basically a mini recurrent neural network at the attention layer. There's a little bit of positional information, or positional awareness, smuggled in via their attention mechanism. But otherwise, global position embeddings are gone.

My takeaway from looking at the architecture is that if you look at the original vanilla Transformer from “Attention Is All You Need” and compare it with the models we have now, this is arguably probably the frontieriest open-weight, open-source model that we have available right now. It's pretty instructive for a mere civilian to look at how it's architected, because this is the most capable open-architecture model I think that most of humanity now has access to.

Position embeddings seem like they're on the way out. More broadly, I think we're seeing almost a Ship of Theseus architecturally, if you will, where the original Transformer architecture—you can still recognize the outlines of Transformer, but piece by piece, all of the original elements—the attention mechanism, the position embedding, the layers, the residual streams, all of that, the sparsity, all of the original components that made the Transformer the Transformer—are getting swapped out for better versions.

And so, if you follow the path of continuous improvement, it looks like the same architecture. Even if you squint at it, you'd still recognize, okay, it's like a multilayer something that is attention-y. But if you look at the fine details, the frontier models now—to the extent that, say, this is indicative of what's actually being used inside OpenAI or Anthropic—are relatively unrecognizable compared with the original Transformer, which I think is interesting.

Peter Diamandis

Do you think the U.S. lab will end up replicating it—or should we say stealing that approach?

Salim Ismail

Oh, God, yeah.

Peter Diamandis

It's open source, so I don't know what license or IP is associated with a particular architecture, but of course they're looking at it.

Salim Ismail

Well, the NoPE thing, too. RoPE is just sort of rotating positionally.

Peter Diamandis

Rotary positional embedding, yeah.

Salim Ismail

It's sort of like, if I have a million-token context, it's giving as much weight to something I said a million words ago—which is hundreds or thousands of pages ago—as to something I'm saying right now. People don't work that way. That's nuts.

So when they got rid of RoPE, they put in NoPE, but NoPE actually has this fading memory now. Something I said a long time ago gets less weight than something that's more recent.

Obviously, that's so obvious. But that's the beauty of open source. Someone in China can say, “This is freaking obvious. Let me try it.” Oh, my gosh, it works. Of course, the other AI labs are going to adopt that immediately. It's just flat-out better.

The other thing that's really blowing my mind is that you don't need the attention layers later in the thought process. When you get to layer 100 or 120, you can actually eliminate attention entirely and get almost the exact same result out the other end.

Peter Diamandis

But faster.

Salim Ismail

What? But faster. What happened there historically is that the people who invented the original Transformer algorithm just put a for-next loop around it.

[laughter]

Because they were just writing code by hand back then. It's really hard to try to make different logical layers by hand.

Peter Diamandis

I mean, in defense of the original Attention Is All You Need team, the original vanilla Transformer was just alternating attention and dense linear layers because it's simple.

Salim Ismail

Yeah. It's simple. It's simple, and also, with that few layers, you really focused on just getting the text to mean anything. Now we've got deep thought. It's just so cool.

Peter Diamandis

A model that you can control locally is way more valuable than a marginally smarter model that you have to access through somebody's API, right?

Salim Ismail

Yeah. So organizations can now fine-tune their proprietary knowledge and build around it, put it in secure environments, and totally avoid sending any sensitive information to the cloud. This is going to be huge for regulated industries, sovereign applications, and all sorts of stuff.

Remember, Peter, a few weeks ago we did this AI OS pilot for the organizations thing. We had almost 400 applicants. We picked 10, so we're running the pilot with them. With the launch of this, I think this is one of the biggest things we'll ever see from a business and application perspective.

The big question every company is asking is, “What's my strategy?” It should be, “What do I do on Monday?” We're going to answer that question and help people start implementing and rewriting their organizations at the edge.

Peter Diamandis

One hundred X. This is going to be a gold mine for you, Salim, because nobody up until Kimi really cared about speed in any corporate environment. They're all like, “Oh, this stuff is super cheap. We'll just use it. We're not doing much with it anyway.” So then everyone starts token-maxing. All of a sudden, people are looking at their corporation and saying, “Oh, my God, my token costs are actually going to be bigger than my payroll by the end of the year. And if I forecast out to 2 years from now, my token costs are 10 times my payroll.”

Salim Ismail

Speed does matter.

Peter Diamandis

But there's an easy 10X, and maybe 100X, like Salim is saying, just by tuning it to what your business needs. Get rid of all the cruft. Use these new streamlined models. Tune it to just what you're trying to achieve. You're looking at 10 to 100X.

Now every corporation needs to figure out its strategy, and Salim's business is going to be sold out. Will you still come on the podcast when that happens?

Salim Ismail

We will still come on the podcast because, thank God, we've got our community of 50,000 people who can help with all this. If I had to help out, I'd be bald in 2 seconds. Oh, wait.

[laughter]

What we're doing with these CEOs is saying, “Okay, let's pick 1 process that's going to help you radically increase revenue and take 1 workflow that'll radically reduce cost.” That gets everybody excited. Rebuild that natively, then do more and just start moving things over.

Peter Diamandis

And when do you start franchising franchises at Salim so all our podcast listeners can start a branch of ExO?

Salim Ismail

Well, that's what our community is all about.

Peter Diamandis

Oh, is it? Okay.

Salim Ismail

Yeah. Everybody in our community is an independent contractor. We have no consultants on staff.

Peter Diamandis

This is the moment AWG's been waiting for.

[laughter]

It's the discussion on Claude Opus 5. Right in the middle of all of this, Anthropic shipped Claude Opus 5. This is their fourth Claude 5-generation release, and it approaches the frontier intelligence of GPT-5 at half the price.

It becomes the new default for Claude Max, priced at $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8. I made the switch immediately on Skippy for myself. Anthropic calls it the most aligned Opus model yet and their strongest model for scientific research.

I'm going to go to the slides now, Alex, and walk us through what this means. How strong is Opus 5, and how excited are you about it?

Alex

I'm somewhat excited. I'm not over the moon.

[laughter]

I'm not as over the moon as I was about Fable 5 becoming available. Fable 5 is incredible. I think Opus 5 demonstrates—if you look at the benchmarks, the evals that demonstrate the strongest performance, and I don't think this is a coincidence—for example, ARC-AGI-3, which is focused on the ability to solve interactive visual problems that humans find easy but AIs have historically found hard, went from 1.5% with Opus 4.8 up to 30.2%. As of this moment, and last I was tracking, that's the highest official score from a baseline model on the ARC-AGI-3 challenge.

It's a visual, code-intensive challenge. Something else that's intensive is developing front-end software. My overall whiff from using Opus 5 quite a bit is that there was maybe mild optimization toward front-end development and anything that touches the nexus of vision and code.

Historically, including with Fable 5, if you ask it to generate an image of something or generate a chart, it does moderately well. I think with Opus 5, just trying to read between the lines of the capability changes that I see, Anthropic is attempting to—given that the Opus series and Claude in general don't do image generation—lean in a bit to some of the gaps at the intersection between code generation and vision.

For uses of mine, I still honestly prefer Fable 5, even though it's more expensive. If you look at, say, the Artificial Analysis Intelligence Index—if you look at their overall chart of performance versus cost per task—according to that chart, Fable 5 is below the frontier. It's slightly below Opus 5 in terms of capability and a lot more expensive.

Despite all of that, for day-to-day usage, when I use Claude, I still prefer GPT-5. But I'm very glad for one thing about Opus 5: it doesn't shut you down as frequently if you ask anything that it misconstrues as a question about biology or a question about cyberattacks.

Peter Diamandis

So, Alex, I have the exact same experience—100% the exact same experience. But then I look at these benchmarks, and there's a whole bunch on these charts.

Alex

Yeah.

Peter Diamandis

And they seem to tell a different story.

How is that? Wait, how do you reconcile that?

Alex

I'm a little bit scared that there may have been some mild benchmark gaming here. That's what I was politely gesturing at. These seem to be benchmarks involving code generation and/or imagery or vision, living at the intersection between them.

Some of them, like Humanity's Last Exam—granted, it's saturating anyway—but the performance improvements are a little milder. For HLE, you see 63.9% with tools for Fable 5, a modest increase to 64.7% with tools for Opus 5, and actually a decrease without tools, which is also maybe a sign that there's been a bit of, again, benchmark gaming. It's still—I've used it extensively—still very well-rounded. I don't want to accuse it of broad benchmark gaming, but if you look at the drop relative to Fable 5 for legal or health or some other areas, obviously there was some sort of distillation.

This is the type of distillation that is, under the present regime, welcomed and not disdained: taking a larger model and using it to teach a smaller, more cost-effective model. There was probably a lot of Fable 5, or the Fable series, or the Mythos series, distillation down to achieve Opus 5, but the overall distribution of tasks definitely, from interacting with it for a while, feels biased toward code generation and visual stuff and away from general capabilities outside that.

Peter Diamandis

So, here's our next chart: agentic coding by effort level. You want to walk us through this?

Alex

Yeah, so we're looking at everyone's favorite form of scatter plot, at least in the industry: cost on the horizontal axis, performance on the vertical axis. What this appears to show is that Opus 5 is both stronger in terms of absolute score and cheaper—that's the horizontal axis—than Fable 5 and Opus 4.8.

Interestingly, it appears to be on approximately the same cost-performance frontier as Sol. I think the subtext we're supposed to get from seeing this chart from Anthropic is that this should be read as a direct competitor for Sol, which is interesting and slightly unnerving, given that, again, Fable 5 anecdotally seems to give better performance.

Peter Diamandis

All right, let's go out to our third chart here: novel problem-solving by cost. And I love this concept. Yeah, please.

Alex

Yeah, so ARC-AGI-3, again, is a challenge primarily focused on the ability to solve animated voxel problems. Tetris, for example. If a person had never seen a game like Tetris before, with a bunch of blocks moving around, and you wanted to do well at Tetris, it's a rough analogy, but that's approximately what ARC-AGI-3 is like: animated block-world challenges.

So, what's really striking—and Dave, you and I have talked about various attempts by pure scaffolding-layer parties to just completely saturate ARC-AGI-3—is that, according to the official rules, I think there are limitations on how much scaffolding you're allowed to use. This is just the raw model being injected.

What's interesting, and what was especially striking in the Opus 5 performance, as relayed by the ARC Prize Foundation organizers, is that it was reasoning algebraically about the visual challenges. It was handed a visual puzzle involving blocks, and it started to reason.

If you look at some of the founders of ARC-AGI, of the ARC Prize, they will go on forever about how this is actually a prize that tests the ability to do what's called program synthesis: to write programs from scratch in response to novel problems. Stunningly, what Opus 5 was able to do was take a visual problem with a bunch of what to humans look like objects, represent the objects algebraically in software, and basically do math on the objects in order to solve the problem. This is the first time that, to my knowledge, anyone's ever seen a frontier model ever do that.

Peter Diamandis

An approach that was not guided by anybody. This is its derived strategy for doing this.

Alex

That is, unless Anthropic was benchmark gaming on ARC-AGI-3.

Peter Diamandis

Okay. I think that scaffolding argument, though, is really, really important because if it holds up, I tried to replicate it. You can get 98%, I guess, on ARC-AGI-3 if you give it a reframing of the way it interprets the puzzle.

Alex

Right.

Peter Diamandis

And if that holds up, that gives inspiration to a billion entrepreneurs who can take something like protein folding or drug discovery or mechanical design of robot arms and say, "Fable 5 can do this," or, "Opus 5 can do this, but I gave it a better way to think about the problem. And now I tripled its intelligence within that domain."

That opens the door for scaffolding improvements in all these domains, like biotech, where if you can reframe it so the AI doesn't have to work as hard to understand what you're trying to achieve and can maximize its tokens and its parameter brain count, that is an entrepreneurial heaven. So, I'm really hoping that result holds up. I tried to replicate it. I couldn't quite do it. I didn't work on it that hard. But I do believe it's possible. I actually don't know. Did you get to the bottom of it? Is it real?

Alex

I'm not certain, but the scaffolding advantage is very real. My understanding is this is why ARC-AGI-3 has certain rules regarding what can be submitted and what can't. But I think the elephant in this particular room, to your point, is that scaffolding adds an enormous amount of value at the moment, at any given point in time, over the baseline model. The other side of that is the baseline capabilities tend to dissolve any scaffold. Today's scaffold is tomorrow's baseline capabilities.

Peter Diamandis

Well, I tell you, if that holds up—and I think you're right, I think it will—next semester every university in the country should have a class called Scaffolding. Everybody should have the opportunity to learn how to do this, because that is the power tool of all power tools for any entrepreneur. So, what is it now? It's coming up on August. You have 30 days to get your class curriculum together and launch it for next semester.

Alex

With prompt engineering as a prerequisite topic.

Peter Diamandis

Let's hit these next two charts and then watch the Call of Duty. I have a quick comment. Something I noticed was that 4.8 came out on May 28, and 5.0 came out just now. So, it's not that much of a better model, but the efficiency has gone up by twice as much. We've seen a 2× improvement, and we were saying 10-week doubling of price-performance for AI this year. It's 8 weeks now.

Alex

It's a model release every 6 days on average over the last few years.

Peter Diamandis

Yeah, this is incredible. The other thing I noticed was that in the grid, you've got different models that are becoming really good at different things—legal, health, coding, and so on. I think that will continue. Alex, these next two charts.

Alex

Yeah, so this chart is interesting insofar as it seems to support the hypothesis that there might have been mild benchmark gaming on ARC-AGI-3. This is a chart by a third party that evaluated Opus 5 on an ARC-AGI-3-like game involving a similar genre and discovered that the performance jump was not material versus, say, Fable 5.

Again, I'm not quite sure what was going on with ARC-AGI-3, but that was by far the most prominent increase that we saw from Opus 5. Interestingly, a benchmark that Anthropic did not highlight was Frontier Math, which is, I think, maybe in some sense a better bellwether for advanced reasoning capabilities by the models. I had to check this independently, and actually Opus 5 demonstrated inferior Frontier Math performance relative to Fable 5. So, again, Fable 5 is still my favorite model.

Peter Diamandis

Our last one here: the live leaderboard.

Alex

VoxelBench. Here we see Opus 5 now earning third place, just behind Fable 5 on VoxelBench—again, visually intensive tasks—but at a much lower price. Interestingly, but perhaps unsurprisingly, Sol from OpenAI is still carrying the lead on this.

The reason I'm not that surprised is that visually intensive tasks are an area where I would naively expect OpenAI to be doing a better job, because they've continued to invest in image generation, whereas we've seen no generative image capabilities at all, shockingly, from Anthropic. They're busy maximizing the value—the revenue per token—which leads them to code and not to image.

Peter Diamandis

I bet that, as a practical matter, anytime I'm doing something complicated, I'm working in Fable 5, getting a lot done. If I want to see an architecture diagram, I just take the entire thing and dump it over to GPT and say, "Make me my architecture diagram." Fable 5 is so bad at it.

It does so much work for you, and you get confused very, very quickly, and you want to just see a simple visual summary of everything going on. It's just so bad. But GPT is amazing.

Alex

My guess is Grok jumps to the top of this leaderboard in the next release. Elon has been speaking about that.

Peter Diamandis

Speaking about imagery, this made a viral loop on X. This is Opus 5 recreating Call of Duty from a single prompt—call it a one-shot, if you would. Let me go ahead and hit play on this.

Remember, these demos 30 days ago just looked like absolute garbage. Look at how much it's improved. It's crazy, the rate of improvement.

Alex

What's amazing to me—I mean, this is sort of—

Peter Diamandis

For those who can't see, this does look like Call of Duty.

Alex

The converse of not having native image-generation abilities, in defense of Anthropic, is that if you look at the entire physical world and say, "Well, everything is just code, including code that generates photorealistic video games," then you say, "You don't need native image-generation abilities."

Peter Diamandis

You just need the ability to generate photorealistic 3D environments like Call of Duty, and you're all set.

Guest

Huh. Yeah, the truth is in there somewhere. It's kind of in the middle, I think, but I think really clearly Anthropic cares about recursive self-improvement, purely and only. So they'll build anything and train on anything that helps the other side to ASI.

Peter Diamandis

While we're talking about Claude, I don't know if you heard the story that a significant number of Claude chats were found publicly searchable on Google this past weekend. A Reddit user discovered that by typing the search operator “site:claude.ai/share” into Google, it surfaced a long list of shared personal data, including personal health records, private documents, names, and telephone numbers.

Apparently, this originated from Claude's shared-chat feature, which allows users to share links of their Claude chats with their friends via URLs. Did you track this, Alex?

Alex Wissner-Gross

Yeah, I saw the story, and on the one hand, it's disappointing to see any information that would be expected to be private find its way out into the public world. I'm not a fan of that. On the other hand, I think there's another side to the story, which is a feature that was intrinsically designed to be social in nature. Shock, shock, to see gambling in this establishment, ultimately finding its way into the hands of other people.

I think there are two sides that one can see here.

Peter Diamandis

Yeah, but I think one of the issues—and one of the arguments we saw on the Ramp a few episodes ago—is that when you're using these models, your competition is, in some sense, seeing your data and learning from your data. It's something people need to understand. It's the argument for on-prem. Salim?

Salim Ismail

No, I just double down on the same thing. You've got to own your own proprietary data. I think over time people move everything sensitive in any way to on-prem.

Dave, any comments on this before we move on?

David Blundin

Nope.

Peter Diamandis

All right. I think it's just incredible, the rate of change. Just go back and look at an episode from 2 or 3 weeks ago and look at the rate at which one-shot can create things.

I'll make one other comment. The holodeck—you know, our holodeck is up and running here. You've got to come check it out, man. It went from, like, okay to mind-blowing in just a couple of weeks for the exact same reason. You can one-shot a world, and the audio and the visual are so good.

I think what this will do is—I don't think it'll fix the commercial games that much, but it allows you to do experimentation in an amazing way, because the cost of experimentation just went to zero. So you'll get so many more products. Jarvis has got an exoskeleton. Yeah, that entire—I mean, the Iron Man vision in that movie was so precious, but it's going to be exactly like that. And fun.

I spoke to Jon Favreau today, the producer of Iron Man 1 and 2, getting him to come to Moonshots Live. Elon introduced us. Let's move on to our next story. This was a story, Alex, that you'd wanted to raise here.

It's the idea, which is obvious, that global AI diplomacy is coming. The Financial Times is reporting that China's leader, Xi Jinping, is wielding AI as a tool of statecraft, using it as leverage in China's diplomacy across the Global South in a strategy that the Financial Times calls Pax Sinica. I love that.

While Washington is debating open versus closed, Beijing is out in the world, country by country, exporting AI as an instrument of influence, offering models and infrastructure to the developing world that wants to leapfrog what it currently has. AI is becoming an instrument of soft power. Whoever supplies the models and the infrastructure to the developing world shapes the next few decades—I would say the next century—of global alignment.

My concern is, if the US overrestricts, the developing world is simply going to adopt whatever frontier-adjacent open models are there. Salim, your thoughts?

Salim Ismail

I can't stress this enough. The whole power of the US is its open and very broad innovation ecosystem. If you create a restrictive open-model policy, it's going to be strategically a self-own and shooting your own foot at an epic level, because you're going to protect a small number of domestic labs while giving the entire open ecosystem—the Global South AI ecosystem—to China.

If you want leadership in an exponential era, it has to come from the largest network, where everybody's using your tools for stuff, not protecting the strongest incumbent. Openness is not a philosophical preference anymore. It's the tool of soft power.

The US has already lost that in diplomacy, USAID, and other areas. If they close up the open-model policy, it's going to be really disastrous for the future.

Peter Diamandis

Yeah, and I don't think they will. I mean, I think this is obvious. We're effectively splitting the world into 2 AI empires. Yeah. Yeah, Alex?

Alex Wissner-Gross

Well, as of right now, there is no US-based option at all. You can take everything you just said and swap out the word “model” and put in “fighter jet.” Should we sell F-16s to XYZ country? It's a no-win question. You have to pick and choose.

But if we don't sell the F-16s, they'll buy Russian and Chinese fighters, and that'll support the creation of more of those fighters. But you're also selling an F-16; it's exactly the same problem. There's no easy answer to it.

Right now, there is no US open-source model to compete with the Chinese anyway, which is kind of sad, I think.

Alex, your thoughts?

Alex Wissner-Gross

I think so. Pax Sinica had already been announced by the US before China announced its own initiative, and China, of course, announced many years ago at this point—Xi Jinping announced the Belt and Road Initiative.

There's a certain extent to which it's far more—I don't want to say insidious, but far more ultimately invasive and controlling. If a foreign country loans you a bunch of money to build a bridge, okay, so you default on the loan; that has a certain outcome.

Alex Wissner-Gross

A foreign corporation that's basically under the thumb of a foreign government builds telecommunications equipment and deploys it to you. So now you have cell towers. The worst that they can do is spy on you, and they can shut off your telecom infrastructure.

Next level up, a foreign corporation that's heavily involved with a foreign government injects superintelligence into the veins and arteries of your country. Now it's not just listening to you or loaning money to you; now it's thinking for you.

I think that's a far more vulnerable position for the so-called Global South to be in, regardless of which bloc or sphere of influence it finds itself in. I can only imagine that the long term—to the extent there is a long term in the middle of the singularity equilibrium point—is going to be pushing more, not just inference to the edge, which is what Chinese frontier labs would like with open-weight models, but pushing training to the edge.

I think that's the equilibrium point. Curiously, I don't hear that many countries in the so-called Global South agitating for domestically pre-trained models, but I do think that's where some sort of equilibrium could lie, if there is to be an equilibrium.

Peter Diamandis

Can I tell a related story here?

Alex Wissner-Gross

Please.

Peter Diamandis

A few years ago, I was talking with the prime minister of one of the smaller Asian countries, and their big city had tripled. They needed much more port capacity to be able to receive more containers for the big, huge population, and they had just taken a half-billion-dollar loan from the Chinese and totally mortgaged the future of the country.

I made the point: Look, drones are doubling in their price-performance every 9 months. If you waited a few years, you could have a drone pick up a container. You don't need a port.

4 drones could pick up the corners of a container, which averages 20,000 pounds, and so you don't have to wait that long for drone doubling to get to a quarter of that weight. Then you just pick it up and put it on a flatbed truck or on a railcar, and off you go.

They're like, “Damn, we just mortgaged the entire country because we didn't understand exponential thinking.” If you go back to what Alex just said, that goes up 10 times when you outsource your thinking. That's really dangerous. I think the safety and security of the future will be in these open-weight models that give you back your sovereignty.

Dario Amodei

Yeah, talk to California about its high-speed rail.

Peter Diamandis

Yeah, let's not go there. All right, so big news this week for my fellow space cadets: a successful launch of Starship 13. SpaceX has confirmed launch and splashdown—an incredible trip for Starship 13, their largest vehicle to date. It accomplished a number of key firsts. Let's run through them.

First, it deployed 20 operational Starlink V3 satellites. They were tested in part, and these satellites are going to deliver us 0.5-gigabit to 1-gigabit connection speeds everywhere on the planet. You're literally going to have a better connection from space than you have from your home Wi-Fi.

They did an in-orbit relight of one of the Starship’s Raptor engines, which is critical for the upcoming Artemis missions, and a successful soft landing on the Indian Ocean with the vehicle remaining intact, which was extraordinary. I’m going to watch two of the videos here. Let’s share them because they’re just fun. This is space porn. All right, let’s take a look at the launch first.

Alex Iskold

It’s important.

Peter Diamandis

You’ve got to love this drone footage from above the launch at Starbase.

Alex Iskold

Space is big. This has a high degree of beauty.

Peter Diamandis

Oh my God.

Alex Iskold

Yeah.

Peter Diamandis

I remember when I was with Elon and we were talking about the Starship first stage and saying it’s the most contained energy that you can ever experience other than a nuclear explosion.

Unknown Speaker

About 10 seconds away.

Peter Diamandis

I’m going to look at that footage. Starlink, baby.

Unknown Speaker

Let’s see if we can get this thing in the water.

Peter Diamandis

I got the link from you guys and I was like, “Yeah, yeah, I’ll check it out.” And I was like—

Alex Iskold

Beautiful. Wow. I’ve got to watch this whole thing end to end.

Unknown Speaker

Note the high thrust throughout your engines.

Alex Iskold

Yeah.

Peter Diamandis

Look at the amount of stress. You can feel it, because you can see things warping and warping.

Unknown Speaker

Soft landing. Down to 1. And landed in the water. Floating in the water for full recovery.

Peter Diamandis

That was a soft landing there because we really thought it would explode.

Alex Iskold

Yeah, well, it hasn’t in all the previous missions.

Peter Diamandis

What happened there?

Alex Iskold

It was so soft. Well, there’s still excess fuel.

Peter Diamandis

Yeah, and when it hits the water, it punches little holes or whatever in the sides, so that’s where it usually explodes.

What I’m excited about in particular is that Elon tweeted that, because the landing was so precise, Flight 14 is likely to capture the Starship on the Mechazilla device—the large chopsticks that come in and grab the vehicle. That’s a flight I want to go to Starbase to watch the return. It’s going to be awesome.

Let’s take a second just to talk about the abundance story here. The cost of launch is plummeting, and let me just give you the numbers real quick. The Space Shuttle was roughly $54,000 per kilogram to orbit. Think about taking a gallon of water or milk to orbit: $54,000. Falcon 9 dropped it to about $2,000 per kilogram, and Starship’s target—and you and I were discussing this with Elon back in our January podcast—is between $10 and $100 per kilogram. I mean, just extraordinary.

Alex Iskold

Well, it’s also inspiring. The amount of incredibly cool stuff that you can build now, specifically because all the feedback and control and all the remote intelligence is easy now, all of a sudden. We saw the Unitree robots in the last podcast, and they’re just beyond cool. That robot that’s plummeting down the side of the mountain with the wheels—

Peter Diamandis

Yeah.

Alex Iskold

Crazy cool. And then you’ve got the Starships. Just the amount of possibility is so exponentially bigger than it was just 5 years ago.

Peter Diamandis

Grok for building Starships is extraordinary.

Alex Iskold

But also, I’m on a photonic computer to run our new neural nets and all the parts. It’s actually giving me part numbers to order and saying, “This vendor in Germany will make this lens for you in exactly this way. Do you want me to write up the specs?”

Peter Diamandis

Sure.

Alex Iskold

And build it for me, too, please?

Peter Diamandis

I mean, literally, can they get robots in boxes and stuff that open the boxes and assemble them?

If you’re an entrepreneur out there and you’ve been looking for where to go build, building hardware—Ben Horowitz, who’s a friend of the pod, I’m going to have him back on the pod for one of these episodes—he’s the co-founder of Andreessen Horowitz. He wrote the book The Hard Thing About Hard Things.

It used to be that building anything—I built robots in high school and college—was tough. Now you can 3D-print parts and iterate rapidly. So, if you’re an entrepreneur looking for something to do in the world, what’s missing? What do you wish existed? You can actually use these models to design it, order the parts, and start building.

Alex Iskold

Yeah, I’ve got an incredibly cool robot that cleans the top of my pool. It has eyes and finds leaves, and it just goes and picks them up. But I want someone to build one that dives to the bottom and just goes down, picks up whatever—an acorn—brings it up, and throws it out of the pool. I bet you could vibe that up now. Just crank it out.

Peter Diamandis

Alex, if there’s anybody who’s as big a space enthusiast as I am here, it’s you. I mean, did the flight bring a tear to your eyes?

Alex Iskold

I wasn’t crying, but space is big, and I was delighted. Did you see, Peter, the views from the Starlink satellites that were posted later? That was pure science fiction.

Peter Diamandis

It was basically looking at Starship in orbit from the descending booster.

Alex Iskold

Yes. That was incredible. That’s like something out of Star Trek or The Expanse. I was very impressed with that.

I certainly hope that Flight 13 ends up in a museum at some point, given the soft splashdown. This is a historic first. Hopefully, the SpaceX team will use this as an opportunity to get a good look at the heat shield, which is interesting. If you noticed the drone feed, the moment the splashdown happened, it was just zooming in on all of the heat tiles, looking for damage and trying to analyze the structure, presumably because the team was worried the whole thing might explode a few seconds later. They were getting whatever footage of the heat shield they could while they had time. But now they’re going to have a ton of time, so it’s very exciting.

Peter Diamandis

In the past, after the vehicle—after Starship detonated from the onboard fuel, and that was expected—people would say, “Oh my God, it failed.” No, it didn’t fail. It did exactly what they expected it to do. They’d have to go diving and find pieces of it to try to reconstruct what happened.

And the heat shield here—I mean, people need to understand the amount of energy being dissipated. These vehicles are traveling at 17,500 miles an hour in orbit and have to dissipate all that safely and come to a precise landing. It’s insane.

Salim Ismail

I just love the fact that, flight after flight, you’re viewing any kind of failure as information. You don’t get embarrassed by it. You take the data, you learn, and you do it way better next time. Nobody else does that.

Peter Diamandis

Yeah. It’s exciting. As Alex, you’ve said many times about the speed run to Star Trek, what an exciting time.

Alex Iskold

Count on it.

Peter Diamandis

Yeah, I mean, the only thing better is if we discover that we have access to all the alien UFOs and can jump to light speed.

All right. Let’s stay on the science theme. Our next story is in the world of brain-computer interfaces. Two stories this week. The first one comes out of Science Corporation. Full disclosure: it’s one of my portfolio companies. I love this company.

The company is run by an amazing entrepreneur, Max Hodak. He’s the former president of Neuralink and someone who I’ve had on the Abundance stage a number of times. Science announced that its first product, called Prima—I’m sure it’s an acronym—has been approved for launch in Europe.

Prima is the first BCI device approved for restoring detailed vision in age-related macular degeneration, which destroys the central vision of your retina. They just earned what’s called a CE mark, which means it complies with European safety, health, and environmental requirements.

Let me show an image of what this looks like and we can talk about how it works. So, here it is. What you see there is a pair of glasses that are capturing the image and then beaming the image via infrared to that little rounded square sitting behind your retina. The signal from that Prima implant behind your retina is turning it into electrical signals and giving it to the remaining retinal cells. Those then get transmitted through your optic nerve to your visual cortex, and it basically restores your central vision.

The amazing thing is that the patients who’ve gone through this have experienced 5 lines of improvement on a standard eye chart after 12 months. This is a godsend for so many people with macular degeneration. Any thoughts here, Alex?

Alex Iskold

Yeah, so a few things. First of all, the overall setup is in the spirit of, as I’ve commented in the past, the singularity is essentially all sci-fi tropes happening everywhere all at once. This is reminiscent of an ocular implant from the Borg. Quite literally, it has an external module that those who are watching can see here—an external camera that captures the information and broadcasts it in human-invisible near-IR to this chip that sits immediately behind the retina.

It’s interesting insofar as Max, who was of course running Neuralink previously, is making a move away from the brain. Even though the retina is part of the central nervous system, Neuralink, which also has its own approach for curing blindness called Blindsight, seems to be focused less on the optic nerve and more on direct brain stimulation and direct brain intervention.

Yeah. It's interesting to me that Max, with this new venture of his, is sort of moving away from the brain, albeit still in the CNS. And I do think the further you get away from direct brain intervention and placement of electrodes, the easier it is to scale up a mass-market consumer device.

In this case, obviously, it's surgery on the retina, but I'm very optimistic that advances in ultrasound, advances in wearables, and a variety of other completely noninvasive advances will enable vision restoration without even needing retinal surgery in the next 20 years.

Peter Diamandis

You know, there's a chapter about Max and Science in my book Life Force. In particular, giving vision back to the blind is biblical. All right?

Alex Iskold

It's huge.

Peter Diamandis

It's amazing. And God said, “Let there be sight.”

What he's doing is interesting. Prima, as a product, is his stage-zero revenue-generating engine. One of the things that a lot of entrepreneurs do incorrectly is they jump straight to this massive moonshot that will take them hundreds of millions or billions of dollars to get to without generating early revenue. So, Prima is the means by which he's creating early revenue.

He has an amazing BCI approach. I can't say a lot about it, but he's basically growing neurons into the brain. The issue with Neuralink and many of the BCI companies is that their electrodes destroy thousands or hundreds of thousands of neurons when they're placed into the neocortex. But neurons actually can grow into the brain.

He's got an approach that creates an interface between electrical circuits and neurons, and neurons are growing into the brain and then wiring together and firing together. Hopefully, he'll disclose it. It's been in animal models, and his plan is to get to humans, but it's an incredible strategy for the BCI world.

Alex Iskold

BCIs are supercompetitive at this point. There are folks trying approaches at the CNS level, at the peripheral nervous system, through direct brain stimulation, wearables, ultrasound, and fMRI.

I think ultimately—by ultimately, I mean on the 5-to-10-year timescale—we're going to see something of a shakeout, and we're going to discover what the most ergonomic ways to interface with the brain are. So, I really hope that, for Max's sake, that brain—or that Science, rather—does well. And I, for one, would welcome some extra neurons.

Peter Diamandis

Yeah. He thinks of it as an extra—the corpus callosum is what connects the right and left hemispheres of your brain. Imagine having a third hemisphere of your brain that's actually connected to the cloud. I mean, that's the way he describes it.

Alex Iskold

Exocortex. I want my exocortex.

Peter Diamandis

It's extraordinary. Salim, you were saying?

Salim Ismail

There are 2 things here that I found really interesting. One is the feedback loop. Once you have a feedback loop from sensory input back into the brain, et cetera, it learns very quickly, and I think we can see that over and over again in some of these recursive, inner-loop-type applications.

The second thing that occurs to me, which goes to your comment, Peter, about business models, is that when you have an exponential and it's hard to predict out where the endpoint is going to be, it's not that difficult to look out 2, 4, or 5 hops and say, “What are the business models that may be enabled at each of those things? What are the use cases?”

So, if you're an entrepreneur and you see a technology that's growing exponentially—now there are a dozen of them—you can pick your biggest passion, your favorite technology, look out where it's going, and then say, “Okay, at that price-performance, what applications become enabled?” And now you have a very viable roadmap for the future, exactly like the way Max is doing it. That's going to be the future of companies and how they evolve.

Alex Iskold

I really want to echo something Peter said there, too, about go-to-market strategy, because if you look at the most valuable companies in the world—Apple, Google, and Meta—and you look at their very first day and their very first product, the Apple I was a box of chips you needed to assemble yourself. Meta was Facebook; it was a little face-picture-sharing thing at Harvard.

Peter Diamandis

At Harvard, yeah.

Alex Iskold

And Google was a plug-in to Yahoo, a little search plug-in to Yahoo. They tried to sell themselves to Yahoo for a couple hundred million bucks, and Yahoo was like, “You're not worth anything near that.”

Peter Diamandis

He wasn't always right. It's not worth anything near $100 million.

Alex Iskold

Yeah. But these really humble beginnings for go-to-market strategy—because we talk a lot about starships landing in the ocean, and people think, “Wow, I want to build a company that creates a new starship”—that's not how these things get started.

You've got to start with a go-to-market that actually gets you initial revenue. And if you really study the outcomes, but trace back to the first 6 months—which not enough people do, like really study the details, the people, the characters, what exactly was that product?—then that starts you on the journey, and you end up being Google or Apple.

Peter Diamandis

We had Bill Gross on stage with us this past year at the Abundance Summit, and he's brilliant. I love Bill. He's one of the most extraordinary entrepreneurs who's created more startups that have gone public or been acquired than I think anybody else, period.

He has a great video on DLD, also on TED, in which he looked at, I think, 50 companies in his portfolio that succeeded and 50 companies that failed. He asked the question: Why did they succeed? Was it because the CEO was smarter, trained at an exclusive Harvard or MIT? Was it because they had more money? What was it?

His conclusion at the end is an important lesson for all the entrepreneurs listening: It was timing. It was the companies that were there at the right time, that were able to survive long enough to intercept good luck.

Classic examples are Uber and Airbnb. They had been tried before, but when they launched in 2008, it was just after the recession, and people were looking to make money. They were willing to rent their bedroom out, willing to go and drive a car.

Our darling here, SpaceX—you have to remember, Elon in 2008 was effectively bankrupt. He had had 3 failures of Falcon 1. And the 4th one, which he scrambled to get money together for, finally succeeded.

Because the space shuttle had been shut down a couple of years earlier, there was a contract out, and he won a $1 billion contract for crew resupply from NASA, and that got him going. Timing is everything. So, if you can have an early revenue stream for your company that allows you to stay in business and intercept good luck, that's one of the single most important things.

Alex Iskold

Yeah, just survive. Just find a way to survive.

Peter Diamandis

Elon went from bankrupt to trillionaire in 18 years. Did you see his Twitter: “Shorter and shorter and shorter,” past a trillionaire or whatever it was—former trillionaire?

Salim Ismail

A little plug here.

Peter Diamandis

Yeah, please.

Salim Ismail

I mentioned this earlier, but we talked just now about the feedback loops. John Hagel and I have come up with a framework where you can measure luck, and once you have that feedback loop, that becomes very powerful. Yeah, we'll bring him on sometime and talk through it. It'll be useful for the viewers.

Peter Diamandis

Everybody, welcome to the health section of Moonshots brought to you by Fountain Life. You know, we talk about AI on this Moonshot podcast all the time. One of the most important things AI is going to be able to do for you besides educating your kids and helping you with your taxes is making sure that you're living a healthy lifestyle, that you get a chance to get to 100 plus. I'm here today with Dr. Don Mucciaccio, the chief medical officer of Fountain Life, and a part of my medical team. Don, a pleasure.

Dr. Don Mucciaccio

Great to be here.

Peter Diamandis

You know, the thing that people are concerned about most about living to 100 or 120 is their cognitive abilities, making sure they don't have dementia. And the numbers about dementia are problematic. Can you share what you've learned?

Dr. Don Mucciaccio

Such an important point, and you're right, at Fountain Life our members, the number one thing people are most concerned about is losing their brain health, forgetting the name of their child, forgetting the face of their loved one. We know that when it comes to dementia, the conservative estimates are that 45% are entirely preventable. What was amazing is with the advanced testing we're doing at Fountain Life, one quarter of our members had advanced brain age.

Peter Diamandis

Wow.

Dr. Don Mucciaccio

But what was really awesome is again, back to that prevention. When we partnered it with healthy living—eating healthier, moving our bodies, sleep, optimizing sleep is so important—you know what we saw? We saw that we improved that brain age by 26%. That is a big, big number to show that the majority of those individuals were able actually to improve the brain age.

Peter Diamandis

And one of the things I love about Fountain is we're searching the world for the best therapeutics, the best approaches, and making sure we bring it to our members. So, if having healthy brain function till 100, 120 is important to you, check out Fountain Life. Go to fountainlife.com/peter. Make sure you become the CEO of your own health. All right, now back to the episode.

Alex, this next story is one that you threw out, and I'm happy to talk about it. It's our second BCI story. This one's out of Neuralink.

The company just shared a video of people living with paralysis controlling a powered wheelchair using nothing but their thoughts. No joystick, no hand controls—just intention translated directly from the brain into the wheelchair's movements. Think what this actually means. It's massive freedom.

Let's roll the video and take a look, because it's a beautiful thing. Then we'll talk about what comes next.

Unknown Speaker

Through our clinical trials, we've been working really hard to give the world a brain-computer interface-powered wheelchair. It's designed for anyone with trouble controlling their wheelchair physically by translating their neural signals so they can control the wheelchair with their mind.

What we've done here is develop a set of custom electronics to take cursor movements from a participant's imagined motions, translate them into analog signals, and use them to directly drive and control all the functionalities of the wheelchair.

This is the wheelchair control app, and we built a custom UI for our users. I can now move my cursor up, slowly move the wheelchair forward, return to the side, left or right. As I get more and more into the ring, it'll go faster and faster.

Really, we built this system with safety in mind. As you can see, if I let go, the cursor slowly goes back to the center. That way, if a user ever becomes incapacitated, they won't drive directly into a wall. The cursor will allow them to go back to the center here.

Peter Diamandis

Alex, your thoughts and what comes next?

Alex

Obviously, this is a visual joystick via brain-computer interface. It's hard not to extrapolate this to full bodily controls. Imagine giving people exoskeletons that they can control via BCI—for paraplegics and quadriplegics—basically restoring free autonomy in a physical world. All 4 limbs.

I think that's pretty incredible. It's easy to extrapolate further than that. I think there's probably a sizable subpopulation in many countries. Certain folks would love mechas from anime. They'd be able to walk around in large robots, Sigourney Weaver or Aliens style.

Or maybe even—I think the end game for the motor cortex does look like some variant of partial brain uploading or something adjacent to that. Once we've fully decoded the motor cortex, we're in, I think, a stronger position to take some variant of human mind uploads.

It could just be behavioral uploads generated by pretraining a foundation model on large amounts, say, of fMRI or ultrasound data, and being able to decode the motor cortex to perform useful functions in the world. That's a low bar. There are much higher bars that would be connectome-based.

Something like that, I think we're going to see actually happen in the next 5 or so years, certainly by the end of this decade. I think that's a major step toward, in the short term, obviously, taking people who can't walk and giving them powers of locomotion. That's now-ish, but in the next few years, it's giving them exoskeletons and, ultimately, human uploads.

Peter Diamandis

You know, my answer, Alex, is that Neuralink connects these individuals to an Optimus robot. They see through the Optimus's eyes and hear through its ears, and then they move—

Alex

Avatar. Your vision is Avatar.

Peter Diamandis

It's effectively telepresence. You can be anywhere. You can inhabit an Optimus in Japan if you're sitting there in Boston. I think that is definitively coming, probably over Starlink. I mean, the singularity is here. This is insanely fun. Yeah, insanely fun.

Alex

There were a few—in addition to Avatar, there was another sci-fi movie where people never left their homes and only went out and interacted with each other via these telerobots. I'm blanking on the name.

I think every sci-fi scenario plays out at once, and I can guarantee you, Peter, if the scenario that you're describing comes to pass—and probably will—we'll find that some country, a few years from now, will be regulating the hikikomori, if you will, who only stay inside their bedroom and only interact with the outside world via BCI to a telerobot.

Peter Diamandis

So crazy, so fun, and so liberating for so many people. All right, our final story before we go to our AMA here is regarding Elon's prediction that AI and robotics are going to make money irrelevant within a decade. By 2036, he said, it's his post-capitalist vision: a world of radical abundance in which scarcity no longer matters and money no longer matters.

I talked about this with him when he was at the Abundance Summit this past year. Let's take a listen to this video, and then I really want to discuss it, because it has people both excited and fearful. I want to address the fear there.

Elon Musk

Money won't matter in 2036.

Unknown Interviewer

I'm not sure that the people who bought your shares think that money won't matter in 2036.

Elon Musk

What do you want money for? Do you want money for goods and services? Well, if that is so abundant that robots may be providing more goods and services than any human could possibly consume, what do you need money for?

In that case, I'll make a prediction: deflation will be the issue, not inflation. Because as the output of goods and services increases, if the output of goods and services increases faster than the money supply, you will have deflation.

Peter Diamandis

I think it's totally ironic this was an interview by The Economist of the world's first trillionaire. Yes. Gentlemen, thoughts? Salim?

Salim Ismail

I think there are a couple of different threads here. Thread 1 is that we're reaching abundance and the cost of things will drop radically. But abundance in production doesn't mean you end up with abundance everywhere.

You could eliminate scarcity in production, ownership, access, and location, and that would be amazing. But money is going to be relevant as long as you need an exchange mechanism. As long as you need that to allocate stuff that's scarce, money as a means of exchange will stay.

Let's remember, there are 3 uses for money: means of exchange, unit of account, and store of value. This would hit store of value to some extent; it would hit the means of exchange and unit of account as well.

I think the biggest challenge here is not that abundance is impossible. It's that, if you end up in the wrong way, abundance will get captured by a few big companies, which is where wealth inequality has been coming from.

I think the biggest opportunity, as we distribute and democratize technology, is that we also democratize opportunity. And I think the broadest framing I've seen for any of this is: Can we get abundance of opportunity? I think that's where technology will take us.

Peter Diamandis

Yes, Alex, you made that point: abundance of freedom.

Alex

I construe Elon's comments as—I'll use the technical term—Star Trek economics. I think he's arguing that 10 years from now, we'll live in a Star Trek economy, and I think there's some fine print on this.

I don't think he really means to say everything has been demonetized. I think what he's short-handing is that most aspects of daily living as we would construe them today, in 2026, will have been demonetized 10 years from now.

Food, shelter, healthcare, utilities, education, entertainment—all of these things will have been demonetized, and you won't need money because we'll be living in an abundant future in that sense. But I think there will be many things that are still not so abundant that they've been demonetized 10 years from now.

I don't know if we have the ability to go to another star system. That's probably still somewhat expensive. Or spend a week on the Moon—maybe there's some price there. Or just own some scarce resource that's an antique collectible.

This is not investment advice, but there are some things I think will resist demonetization for longer than 10 years. If I had to put my finger to the wind and guess when demonetization hits the total economy, don't hold me to this—and this is not investment advice, doubly so for anyone investing in 30-year Treasuries—I would guess approximately 30 years out.

Peter Diamandis

Funny story: Anthony Scaramucci, who was one of my interns and one of my strike force members, started a company collecting obviously scarce things like Tyrannosaurus skeletons, and—

Alex

And Pokémon. I know AJ pretty well. AJ, if you're watching, I'm not sure what's up with those Pokémon.

Peter Diamandis

Yeah, he's collecting the best first-edition comics and so forth. I mean, it's an interesting strategy. But I wrote a piece about this after Elon published it, or after we had the conversation with Elon, and the best way I think this works is that we're going to end up providing some level of UBI.

I call them COVID checks, right? $3,000 a month, which today gives you a bare minimum level of living. But all of a sudden, in this scenario, AI is already today—and will be in the future—the best physician. An Optimus robot with AI will be the best surgeon, and the cost of that is CapEx and electricity.

Then, with autonomous vehicles, we're going to see not just 1 or 2, but a dozen cars-as-a-service AVs beating each other out to bring the cost down. All of a sudden, $3,000 goes a lot further than ever before.

You want a house? Great. A fleet of Optimus robots will build it for you. So, it's a massive demonetization in the future.

There’s a monster elephant in the room, though, which is the radical transition this is going to require in our fiat currency systems. All our fiat currency systems are absolutely dependent on scarcity. If you move to abundance, we have a huge challenge.

We’ve mentioned this before on the pod. This is Jeff Booth’s odd observation. We should have Jeff as a guest sometime. He made the point that over the last 50 years, every $1 increase in GDP has come with a $4 increase in debt.

I use a metaphor for this. Imagine you decide to build a TV factory. You borrow $10 million to build that TV factory, and your business plan says, “I’m going to pay this back. If I can sell the TVs at $1,000 each, I’ll be able to pay back the loan.” The problem is that a year later, that $1,000 TV can only be sold for $500, and a year later it can only be sold for $250. You’re never paying back the $10 million.

This is how we’re growing the global economy. This is the money-printing problem we have, where we’re radically printing money to keep the whole thing afloat, which is why assets are so important to own rather than cash. Cash is deflating at about 14% a year.

This is going to require a wholesale shift in how we measure the economy, which is why people are pointing at crypto, Bitcoin, and other things. But this is the part that’s going to cause a massive problem for every currency and every central bank in the world. Every central bank is panicking right now because its only resource is to print money, and then you have inflation. Then they’re like, “Oh my God, we can’t have inflation.” This is a circular wheel that can’t be gotten off of, and this whole thing is going to come to a point.

Alex

If only the Federal Reserve had access to the same superintelligence that the rest of us did. They could design superintelligent fiscal and monetary policy.

Peter Diamandis

That’s where you’re going to have to go. That’s where you’re going to have to go. It’s a great point.

Alex

They’re just so broken.

Peter Diamandis

Yeah.

Alex

What I find incredibly interesting is that all of these AI visionaries—Elon, Demis, Dario—they’ve all played Civilization. They all speak in terms of Civilization, and they’ve all read Iain Banks’s Culture series. The entire Culture series is about the post-abundance world and what it’ll be like.

When they get together and brainstorm about the future, they’re totally on the same page about how this is going to work. Then they do an interview, like Elon does an interview with The Economist or whatever, and he says, “10 years from now, so in 2036, money won’t matter.” And they go, “Oh my God, does that mean the exchange rate between the pound and the euro…?” And then they’re like, “What? No. We’re going to make so much stuff and have so much abundance that it won’t matter. Did you understand the implication? Who gives a crap about the exchange rate or the deflation rate?”

The degree of change that’s coming over that decade is so massive that you just mention one little aspect of it—“We don’t care about money”—and it’s just a tiny little component of this overall, massive change. But all those guys are on the same page because they’ve all read the same science-fiction books. They’ve all thought, “I get how this is going to play out.”

There are different nuances to it. I’m not saying everybody agrees on every part of it, but what we’re talking about is that you don’t care about the cost of things because you just take them off the shelf.

Peter Diamandis

That’s ironic, given, Alex, that Iain Banks is Scottish. Scotland produces some of the world’s best science-fiction writers, so it’s interesting that The Economist, based in the UK, doesn’t quite appreciate Scottish science fiction.

There’s a tweet exchange that occurred. Let me just read it. Daron Acemoglu is a Nobel laureate in economics, and he says, “I propose a challenge for Elon Musk, an opportunity to put your money where your mouth is. If money won’t matter in 2036, why don’t you pledge to donate your current wealth, approximately $1 trillion, to charity no later than 2036? This would establish with great credibility your prediction.”

Elon responds, “I’m actually going to do something along those lines.” I thought that was pretty cool. I immediately texted him and said, “Okay, let’s launch 10 $1 billion XPRIZEs to solve the world’s biggest issues.” I haven’t heard back from him yet, but hopefully soon.

Alex

I think he’s planning something more along the lines of SpaceX and Tesla stock for everyone via UBI.

Peter Diamandis

Yes, perhaps. Perhaps. Anyway, this is the abundance story writ large again. The cost of everything—and it’s not going to be trips to the Moon or Mars, but if you want your basics—is going to be reduced.

It’s raising the floor so that every man, woman, and child on the planet has access to food, water, energy, health care, education, and freedom. I think that’s what we’re building here. I like to say that, yes, we’re going to have trillionaires living on Mars forever, but in that inflationary world where everybody’s experiencing a rising tide, I’m okay with trillionaires living on Mars as long as every man, woman, and child on the planet has access to all the basics. That’s a more peaceful world.

Alex

Can we lift the bottom?

Peter Diamandis

Yeah, lift the bottom. The gap will get bigger. I agree, the gap will get bigger, but as long as the floor comes up, that’s the single most important thing.

Alex

Mention one of my favorite abundance statistics that you put up, Peter.

Peter Diamandis

Sure. If you go back 100 years ago to 1820, 94% of humanity lived in extreme poverty, defined as $2 a day in 2011 purchasing-power-parity dollars. Today, that number is less than 9%, and you just don’t see stuff like that in the news.

Alex

Yeah.

Peter Diamandis

You don’t. News media delivers every murder and every crooked politician over and over again into your living room between 6:00 a.m., 6:00 p.m., and 8:00 p.m. As I like to say—I tell my mom this all the time—“Mom, turn off the news. Don’t watch the Crisis News Network. It will just give you a bad mindset.”

Alex

Join us in our echo chamber.

Peter Diamandis

She does. Every time. Hey, Mom.

Let’s do some AMA questions, gentlemen. I think we have some fantastic questions this week. Okay, Salim, let’s begin with you.

Alex Wissner-Gross

There are a few different interesting questions here. I’ll pick number 4 because I’ve commented on this one already on the pod: How far off do you think we are from hitting longevity escape velocity? This is from Sage Friedman 9260.

I think, Peter, if you were to answer this, or if friend of the pod Ray Kurzweil were to answer this, the answer would be something like 2030 to 2033. I think that’s the consensus.

Peter Diamandis

My mantra is LEV by 2033. Yes.

Salim Ismail

Yes. If I were to answer this, I think it’s going to be spiky. Just like superintelligence is spiky along different dimensions and with different skills and capability areas, I think some subpopulations may hit longevity escape velocity by the end of this year. I think it could happen for others by 2030.

I think there are so many variables that will lead to high volatility, or spikiness, in terms of who arrives when. In part, that’s because there are so many people who qualify for certain medications that—speaking hypothetically, it’s not medical advice—third- or maybe even soon fourth-generation GLP-1 RAs could end up having a profound longevity impact.

I was sufficiently interested in this that I did my own internal mini research study trying to answer the question, “Have we achieved longevity escape velocity this year?” There are a few confounding variables because you can achieve, in some sense, catch-up longevity increases if something terrible happens. For example, if there’s an agricultural revolution in China and a lot of people die, average life expectancy takes a huge dive, but then a few years later it zooms back. You could ask the question, “Is catch-up, or regression to the mean, longevity escape velocity?” I don’t think most people would consider it that.

On the other hand, if you have someone who has an illness—an illness that we’ve never been able to cure before—and now we’re able to treat it, and their life expectancy is increasing by approximately 1 year per year, has that subpopulation achieved longevity escape velocity? Some would say no, because you’re just helping a person with an illness regress to the mean.

Others, including myself, would say aging is a disease. Curing aging is basically helping a subpopulation—which is to say, more than 150,000 people per year dying on this planet—and helping the subpopulation that is the entire Earth’s population survive and get treated for the disease that is aging.

So, yes, I think some subpopulations are approximately there right now, with more to come.

Peter Diamandis

We discussed in the last pod that there are a number of ongoing partial epigenetic reprogramming experiments happening in humans today, which is super exciting. I’m going to take number 3 before one of you guys grabs it.

Alex

I’m sure you could do that. That’s yours.

Peter Diamandis

I’m going to take number 3. If average life expectancy hits 120 and infertility gets cured—which it will—how does society handle the population boom? This is from Mr. Gnushton.

Here’s the reality: We do not have anywhere near an overpopulation problem, even if we start getting to extreme longevity—120 and beyond. The majority of the world is in a population crunch.

We're seeing in Asia and Europe a reproductive rate of under 1 child per family. To keep the population without growth or decrease, it's 2.1 children per family. Places like South Korea and Japan are hovering at around 0.6 children per family. So, we have an issue: in a number of generations, these cultures and countries are going away.

The other question that this person might pose is, “Okay, what about access to resources?” Over and over again, even if the population reaches, I think, 9.5 or 10 billion, it's going to very rapidly decrease. People have always said, “The one-Earth precept is that we need to divide the resources of Earth equally amongst everybody.” Well, every time we think there's a scarce resource, we discover that it's not scarce—we just innovate around it.

Lithium was thought to be a scarce resource. Then we started discovering lithium deposits everywhere, and we started inventing batteries that are better than lithium, using sodium, which is much more abundant throughout the world. So, Mr. Ganesh Ingten[?], no fears about overpopulation and no fears about not having sufficient resources. Salim, over to you, pal.

Salim Ismail

Can Anthropic hide its models' reasoning to stop competitors from distilling it? You can restrict the visible reasoning and hide it a bit to make distillation harder, but you can't eliminate it, because people can still learn from the inputs and outputs and reverse-engineer that across a sufficiently large number of examples.

Capability diffusion is very difficult to stop permanently, because once you have useful behaviors, people are going to learn from them, and then they have a huge incentive to reproduce them in other models. The sustainable moat is going to be the whole system, not just the hidden reasoning. Do you have unique data? Do you have infrastructure and compute? Do you have users? Do you have feedback loops? We keep talking about feedback loops. Do you have those to provide proprietary and unique learning loops? That's a really big deal.

Then there's distribution and the ability to learn continuously from all of that. There's a whole-system approach here. The future of every organization is going to be that kernel of data, compute, and learning loops that can compound on each other.

Dave Blundin

There was one left. Hey, do I get first pick on that?

Peter Diamandis

Sure, I'll give you that, for sure.

Dave Blundin

Awesome, thanks. What's left? Number 1: How long does it actually take to patch a vulnerability like the Hugging Face breach? This is from @teachme, “t3h me teachme.” The coolest thing about this question is actually your handle. That's really awesome.

It only takes a minute. Once you know the vulnerability, sometimes there's a little bit of time to propagate the patch out, but once you know what the vulnerability is, you can fix it in no time flat. That's all there is to that.

Peter Diamandis

All righty, let's move on to the next 4. Dave, you get first pick.

Dave Blundin

Oh, thanks. I love number 5: If AI transforms higher education, how should we evolve the high school system? And that's from Dave Wilfart.

Peter Diamandis

Okay, I guess that was your mark for me after all. I love this topic, and we obviously need to get on it. I did a podcast with Joe Aoun, the president of Northeastern. Incredibly great podcast—the guy is brilliant.

We were talking about the evolution of the college system that has to happen right now. He's opening a new incubator on Mass Ave here in Cambridge, where the students who, you know, Northeastern has always had a lot of work-study, so you can work instead of taking classes and get real-world experience. But now you can build a company and be an entrepreneur, and that counts as part of your college curriculum. It's phenomenal.

That same mentality needs to move into high school, where you first have to recognize that the curriculum can't possibly keep up with the useful knowledge that kids are going to want to absorb. So, you have to switch it over to AI-based teaching and AI-based learning. Allow them to learn whatever they want to learn—purpose-driven learning.

We invented a class right in this pod that must exist: prompt engineering and scaffolding. That has to be a new class. But next semester it'll be something else, and the semester after that it'll be something else. You just have to allow that to come into your ecosystem. Then reward good behavior, like trying to learn or doing something that looks productive. Give that an A-plus. But don't try to force everybody down an ancient curriculum.

Dave, here's an idea for you. Given that I'll get in trouble for saying this, but I don't really care: MIT does not really take the humanities seriously. As an undergrad at MIT, I got humanities credits for the philosophy of quantum mechanics and set theory, and all my humanities classes had problem sets or lab assignments. What about a humanities credit at MIT for prompt communication, prompt engineering, and context engineering as a sort of stealthy way to introduce communications skills?

Dave Blundin

Yeah, that's a great idea.

Peter Diamandis

Alex Wissner-Gross, I'd love to hear your answer to number 8.

Alex Wissner-Gross

All right, I've been assigned number 8. Number 8 asks, “Does training on synthetic data degrade model quality over time?” And this is from QC for Life.

It depends on the synthetic data. You could generate synthetic data for, let's say, prototypical software-engineering problems, like generating code and then injecting a bug into the source code and testing whether a model is able to detect that bug. There are many reinforcement-learning challenges that, at least by historic standards, certainly benefit from synthetic data and especially synthetic environments.

You can create procedural 3D environments or applications that generate almost infinite variation, so that via reinforcement learning and reinforcement fine-tuning, models can learn from them. That is quite valuable.

What maybe the question's subtext is: Isn't it sort of garbage in, garbage out? How could you possibly learn from synthetic data? Isn't it just garbage in, garbage out? The answer is, in fact, no, it's not garbage in, garbage out.

I've mentioned a couple of terms on the pod in the past: Solomonoff induction and AGI. In principle, if you had a sufficiently strong model, you could train it with no physical-world data and no human data at all—purely synthetic data. If you had a sufficiently powerful model, and this is the premise for AGI, which is a theoretical, not very practical implementation of Solomonoff induction.

Solomonoff induction is an approach to basically building the perfect inference system. The premise of it is that if you've just observed a sequence of bits or a sequence of tokens, the perfect next-token predictor—which is what all large language models are—is essentially organizing and running every possible Turing machine, every possible computer program, on that history. That's computationally infeasible, but theoretically perfect.

So, going back to the question of synthetic data degrading model quality: in the limit of perfect and infinite compute, it's actually ideal not even to touch the physical world and not even to touch human data, and to train purely off of synthetic data.

Dave Blundin

I'm happy. You know what? Not to beat this to death, but having trained many, many neural nets—and training some right now—synthetic data is fine. It's mislabeled data that actually kills you. Even 1 mislabeled data point is a killer. The weights will warp themselves 8 ways till Tuesday trying to make it make sense in the context of everything else. So, clean synthetic data is totally fine. In fact, it's better in a sense because it doesn't have something just blatantly wrong and mislabeled.

Peter Diamandis

Salim, 6 or 7?

Salim Ismail

This one, 7.

Peter Diamandis

Why don't you do the next one? You always go last.

Salim Ismail

No, it's okay. Go ahead.

Peter Diamandis

I will. Gentlemen, the questions are abundant.

What stops frontier labs from acquiring other companies in other industries to get their data? This is from David Lee, Z5E—or Z5E, if you would say it properly.

Salim Ismail

Nothing stops them from doing that, because you're going to get acquisitions. You can see private equity buying chunks of mid-market accounting firms and trying to get that data. The real goal is the proprietary data they have.

The problem that I'm seeing as I watch private equity try to do this is that buying a company doesn't mean you get really useful knowledge, because there's a lot of tacit knowledge hidden in key employees' heads that's not easy to extract. You've got different operating processes and so on.

Traditional companies need frontier models, but the model provider needs proprietary information to make them useful. So, I think what we're going to see is a lot more partnerships. For example, we're seeing groups of hospitals band together, pool their data, and then make that data available to pharmaceutical companies, almost like a co-op model.

That becomes really interesting. But based on my previous point and the whole organizational singularity stuff, the implication for companies is super urgent. Organize and protect your proprietary data so that, when you add AI, you have the learning loops that become very powerful. That will be the engine of growth for your future. It'll become much more important and much more valuable than your current product.

So, I don't know. They may go buy some startups, but it's going to be a harder thing. I think it's easier for the frontier labs to partner with companies for their data and mutually figure out ways to cooperate, rather than trying to do this acquisition stuff, because that tends not to work out over time.

Peter Diamandis

Mhm. All right, final question number 6.

Salim Ismail

Oh, yeah. Question 6 goes to the guy who got SpaceX. That makes perfect sense.

Peter Diamandis

How can individual investors get involved in cutting-edge startups before they go public? That's from Nancy Jenner-C5D.

So, Nancy, there are so many ways for you to find cutting-edge startups before they go public. The easiest is getting in at the beginning. You go to your university first and foremost. A lot of these companies are beginning in the minds of 20-, 21-, and 22-year-olds, right?

You can go to equity crowdfunding platforms and see what's going on. There are syndicates on AngelList. Venture capital funds have minimums; you have to buy in there. You can go to demo days. There is so much going on today that can enable you. Dave, what do you want to add to that?

Dave Blundin

I think if you add value, you will get stock. There are so many ways. A lot of these companies are growing so quickly, and if you discover them early and just try to add value in any way, they need so many people.

You know, the guy who painted the Facebook office—what did he make, $100 million on that stock? They paid him in stock because they didn't have any cash. He was just there. So, I think people underappreciate how much you can reach out to these guys, especially early on, when they're desperate for help. Do you want introductions? Do you want sales help? Do you want help moving your office?

Peter Diamandis

What are your skills? If you make yourself available, are you a great coach? Right? Do you want to run it? Get coffee for the team?

Salim Ismail

My favorite suggestion would be to go to AngelList syndicates, because people have syndicates where Jason Calacanis will invest in a bunch of startups, and you can buy into that syndicate. They get a piece of the carry, but you get participation in all that, and those have done extraordinarily well. You don't have to put a lot of money up.

Peter Diamandis

Yeah.

Dave Blundin

It's funny you say that, because we had a summer intern. He just said goodbye to me today because he's got to go back to school in September. He put together an angel syndicate over the course of the summer, and he's young. He's still a student, but he's going to manage it, and all the rich, old, famous guys are like, “Great. If you manage it, you can just participate, and we'll lend our names.” So, he actually pulled together a syndicate in, what, 4 weeks?

Peter Diamandis

Yeah, amazing. Gentlemen, as always, a pleasure.

Salim Ismail

I have to go talk to all the rabid Moonshots fans who were like, “We can't wait to talk to you about questions.” I'm going to go talk to them now.

Peter Diamandis

You look like you're in a consulting office over there.

Salim Ismail

No, I'm in a hotel room.

Peter Diamandis

Uh-huh. Okay, fantastic.

Salim Ismail

All right.

Peter Diamandis

All right, Salim. Words of encouragement?

Salim Ismail

See you guys very soon. Love you all. Be well.

Dave Blundin

Thanks, Peter.