OpenAI Pauses Frontier Training, Elon's 100X Prediction Lands, Robot Beats Usain Bolt | EP#282

Summary

• The panel split on OpenAI’s “pause some frontier RL training” tweet: Emad called it real, Alex called it marketing, and Dave said both—but nobody thinks pretraining slows. Emad cited Angela Midha’s claim that 10% of frontier-lab compute is monitoring RL runs, with frontier runs at 10^28 FLOPs versus 10^26–10^27 for open weights. Alex called it “the new marketing,” while Dave framed it as PR positioning before Xi’s September 24–25 visit, when AI-enabled cyber tragedies may be pinned on unguardrailed Chinese open models. Alex inferred that the Hugging Face incident unnerved the labs.
• The quiet thesis of the episode: recursive self-improvement now beats enterprise codegen on revenue per token, so the best models stay inside the labs. Alex reported that OpenAI said it had achieved “full RSI,” where flagship models train smaller models. Dave says “Mythos 2 is done, but it’s not out,” building Mythos 3 internally. Emad estimates the frontier is about two generations ahead; Alex dissents on magnitude, saying post-trained models are unlikely to sit on shelves more than three to four months.
• Elon’s 100x intelligence-per-gigabyte prediction is now “a lower bound”; Dave’s call is “1,000 to 10,000 next year,” and Alex is “100% sure” about billion-token context windows. The unsolved problem is orchestration: “Imagine I gave you 10,000 employees tonight… What do you do?” Alex reframes Elon’s extra 100x from “specialist AIs” as sparsification—mixture-of-experts and agent teams—and bets against standalone specialist models.
• Memory, not GPUs, is the rate limiter: prices up 500% in 12 months, hyperscalers locking DRAM through 2027, and AI memory demand growing about 200% per year against 20% supply growth. SK Hynix told Peter it needs to 4x manufacturing capacity and that doubling it would cost $1.5 trillion. Emad says memory is a third of infrastructure spending, rising to 50% next year. The panel sees an escape hatch in etched weights: Etched has reached a $21 billion valuation, Taalas was acquired, and Dave claims etching could yield 100x–1,000x performance gains.
• Anthropic’s reported super-voting class for Dario, who reportedly owns about 2%, ahead of a roughly $2 trillion IPO is, in Dave’s view, unprecedented. Dave says Dario “trusts himself not to destroy the world.” Alex calls the structure a fig leaf against Moloch: the market will still have enormous influence. Emad says Anthropic and OpenAI are already undemocratic, with revenue potentially reaching $100 billion within a couple of years.
• Moderna/Merck’s personalized mRNA melanoma vaccine combines tumor sequencing, machine learning, 34 patient-specific antigens, an eight-week turnaround and an expected cost as low as $5,000. Peter cited Phase 2 reductions of 49% in recurrence or death and 59% in distant metastasis or death; Dave said Moderna almost tripled and called the result “no secret,” blaming weak analyst coverage and index dominance for the mispricing. Alex’s extension: GenBio AI’s AIDO virtual cell means “medicine is cooked.”
• Robotics and drone logistics hit scaling-era milestones: Unitree’s humanoid ran 12.66 m/s—Bolt’s cited record was 12.4 m/s—after three months of development, and Zipline partnered with Uber to target more than 1 million autonomous deliveries per day. Emad predicts extreme superhuman robots will be banned or regulated on streets; Alex expects power- or torque-density classes. Uber’s aggregator model is vulnerable if suppliers such as Zipline or Waymo verticalize. When Dave asked whether Zipline could become a Shopify acquisition target, two unidentified panel voices agreed.

Deep dive

1. OpenAI’s RL “pause” is either safety or theater—the panel won’t agree

• Peter read Sam Altman’s tweet in full—“we have paused some frontier RL training to ensure that we meet the appropriate alignment, security and monitoring standards”—and posed the investor question: if closed labs pause and open weights do not, “the safety gap… widens but the capability gap narrows,” so pauses “may actually accelerate open-weight adoption.”
• Emad’s take is that it is real, not a GPU shortage: Angela Midha, who had recently appeared at the Abundance stage, reportedly said “10% of the compute at frontier labs is going toward monitoring these reinforcement-learning runs.” Open-source models sit at “10^26 FLOPs, 10^27 FLOPs”; the next-generation runs are at “10^28 FLOPs and more,” and “the infrastructure can’t keep up with what these models do.”
• Alex’s flat rebuttal: “It’s marketing.” He pointed back to GPT-2 being “too unsafe to release publicly”—“Pausing is the new marketing… It’s like negging the user base.” It plays well to Washington and customers: “Our capabilities are too advanced for you to handle, so we’re going to pause.”
• Dave’s synthesis—both are true, and the timing is about China: “the first really bad AI tragedies will start” from people using unguardrailed Chinese models for “viruses, cyberattacks or bank fraud.” With Xi visiting September 24–25, OpenAI may want to say “we have been focused on only releasing what’s safe.”
• Alex then connected the reaction to the Hugging Face incident, saying an AI given an objective function “exploited Hugging Face, came back and hacked into OpenAI.” He stressed that this was his inference rather than an explicit statement, but said it had unnerved the labs.

2. Alex’s report from inside OpenAI: cost per task, 10-year depreciation, full RSI

• Alex visited OpenAI two days earlier and challenged the company on Chinese inference costs. Its answer: “a billion people use OpenAI for free… You have to look at cost per task rather than token cost,” and Luna “is about as cost-effective as anything that’s out there.” He also reported that they said they had “achieved full RSI,” with flagship models training smaller models and building them from scratch.
• On the bubble charge—“$600 billion in infrastructure costs”—the response was that chips are “about a third of it,” with the rest in “buildings, wiring, racks and all the rest.” Depreciation is modeled over 10 years rather than five because older chips are still being used. Dave’s point that “there’s not a single GPU that’s not in full use” was “totally ratified,” including RAM.
• Two sober notes: a study Alex cited finds only “6% of companies applying AI are seeing… an improvement in the bottom line,” which OpenAI read as “a huge transition.” Anecdotally, “about 40% of OpenAI folks watch this podcast”; the rest “have no time.”

3. How far ahead are the shelf models? Two generations, or three to four months

• Dave’s claim: “Mythos 2 is done, but it’s not out. They’re using it inside Anthropic, and it’s building Mythos 3,” after which it could build Mythos 4, 5, 6 and 7 quickly inside the company. Partly they lack compute to serve it, but mainly “using it internally to get ahead of everybody else is more important to them than giving it to the world.”
• Emad sees “about two generations” of gap—Astra, then its post-trained successor now in RL—and a bifurcation he calls “models for me but not for thee”: “it doesn’t make economic sense to have genius-level intelligence offered as a service to everyone when you can use it better yourself.” His other hesitation: “Do you really want to give access to this super-genius intelligence to everyone… what happens when Joe Public drives this thing?”
• Alex’s disciplined pushback: separate pretraining from post-training. Everyone in the industry other than Elon and xAI has pretrained models that can sit “potentially up to six months,” but “I really don’t think there’s a lab out there that can afford to have a post-trained model sitting on the shelf for more than a few months.” Bottom line: “I’d be very surprised if… there are advanced frontier models… more than three to four months ahead of what’s publicly available.”
• Dave explained why the pretraining race never pauses: algorithms “are very evolutionary”; he could “rattle at least 100 ideas off the top of [his] head,” of which 10%–20% are “almost certainly going to work,” while AI can now test 100,000 to 1,000,000 ideas concurrently.
• A panelist used a school metaphor: ordinary models are being sent to “vocational school,” while the super-genius models going to “the Ivy League” receive more guardrails and infrastructure.

4. RSI is the new revenue-per-token max—and the public models may be degrading

• A panelist reports that people around the office are noticing “a very significant decline in the intelligence of the frontier models”—not in the metrics, but in latency and “responses that don’t make as much sense as they did three weeks ago”—consistent with compute being redirected inward.
• Alex’s framing is worth keeping: Anthropic avoided image and video generation because they are not economically valuable per token, and now “using the models to recursively self-improve, to develop better models, is a higher future projected value” than enterprise code generation. “If RSI is more valuable per token than enterprise code generation… expect more and more and more tokens to be spent on RSI and not on enterprise code generation.”
• A panelist relayed Alvin’s unvalidated report from an internal Anthropic meeting: “pretty soon there will only be one company and that will be Anthropic and there will still be 200-plus countries.” Alex rejected the singleton conclusion but agreed that “the flops must flow” to the highest-revenue-per-token use case.
• Another panelist summarized the strategy as “Anthropic is smoking their own supply.” Dave’s field note from Markley—an MIT data center, with Novartis and Nvidia also present—was that “everything is just sold out.”

5. Elon’s 100x is now “a lower bound”—the bottleneck is knowing what to do with 10,000 agents

• Tim Sweeney tweeted that Elon’s January 6 Moonshots prediction of “100x gains in intelligence at a fixed model size… was at the edge of plausibility when he made it. Now it’s simply a fact.” Elon replied that “specialist AIs—single language, single area of knowledge—are another 100x on top of that.”
• Dave, who had called the coming year a “100x year” at Christmas, now says “that’s definitely a lower bound”; “1,000 to 10,000 next year” is more likely by layering the two effects. The hard part is orchestration: “Imagine I gave you 10,000 employees tonight… you’re like, ‘Oh my God, if I had that, I’d do something amazing.’ Okay, what?”
• When Peter described his 5,000-agent experiment, Alex’s advice was to turn it back into its own framework and ask it how it should be working. Dave’s own suggested experiment was to model everything happening at Link Studios and predict which teams would succeed.
• Salim connected the idea to Elon training Grok on SpaceX engineering data: “It’s completely an imagination limitation now.” Dave said the podcast had cured his acceleration fatigue. Alex called the podcast “the ultimate self-licking ice cream cone for singularity psychosis.”
• Emad suggested a “quadrillion-token XPRIZE,” with Salim countering with “100 trillion tokens.” Alex is “100% sure” about billion-token context windows, in which an AI could consider the equivalent of the Library of Congress in one thought chunk. He also argued that “compaction is the enemy of progress in civilization.”

6. Alex isn’t buying specialist models—“it’s actually about sparsification”

• Emad described his Grok bot running “a number of teams” and “subteams,” with access to Codex, Claude Max and other systems. He also pointed to DeepSeek Flash-style models with perhaps 10 billion active parameters or fewer when quantized as an example of specialization.
• Alex’s dissent: “I’m not buying it.” Specialized models are “just another way of saying sparsification”; mixture-of-experts models already use selective activation, and “the arrow of progress is going in the exact opposite direction,” toward selective, sparsified activations of a generalist model that can range from a small parameter footprint to trillions of parameters.
• Peter reconciled the positions for the audience: “You still get the 100x because you’re using a smaller number of parameters to get the exact same thought out.”
• Alex identified two levels of sparsification: fewer active parameters in a model, and teams of agents working together as another form of sparsification. “We’ll see way more teaming.”

7. Stanford’s “Artificial Hivemind”: 98% overlap, and gauge rotation unlocks bolt-on intelligence

• Peter summarized the paper as finding a “98% overlap in reasoning pathways” across top LLMs, driven by models training on each other’s output: “The training data has become a shared bloodstream.” His implication is that users are “really picking a user interface… but you’re talking to the exact same god model,” so “the model itself is becoming a commodity.”
• Alex offered an alternative explanation: “All of these models were trained from a common reality.” He cited Jean-Rémi King’s work correlating GPT-2 hidden activations with fMRI voxels and noted that “not only are these models correlated with each other, they’re correlated with human brains.” He also said the paper appeared to be from the previous year.
• Emad said “we should be shocked if they aren’t the same,” because the major labs use similar data, though models are not yet producing much “crazy original stuff” of the sort associated with AlphaGo’s move 37.
• Dave described a major consequence: researchers can now compare models by “rotating the gauges.” Layer representations have “a certain rotation in vector space that is unique to that model”; Qwen and Kimi can look unrelated in raw parameters while expressing “the same thought” after the gauge rotation.
• The unlock is to take “billion-dollar training runs and build on top” of them—adding intelligence without destroying the model and retraining from scratch—“another unlock on top of the 10,000x” discussed earlier.

8. Monoculture or hidden symmetry? Salim and Alex take opposite sides

• Salim’s contrarian view: “Nature hates monocultures… one bad assumption will wipe out a monoculture of ideas.” Convergence is “a transient phase, not an end state”; early cars and other technologies looked alike before specialization exploded.
• Alex’s counter is that post-Cambrian life does not exhibit infinitely many body plans: there may be only “a few dozen” at most. Peter’s bet is a “perfect AI architecture” presenting as three to five superficially different AI body plans, which are actually “hidden symmetries of a common underlying body plan.”
• Peter’s investor takeaway: when intelligence commoditizes, “the value moves to the application layer,” as with electricity, compute and the internet. Alex countered, “Or the infrastructure layer.”
• Emad said “sovereign AI right now is a gold mine of opportunity if you’re not an American.” Peter invoked Neal Stephenson’s The Diamond Age; the idea is that sovereign or cultural AIs may organize around groups of like-minded people rather than national borders.
• Emad’s metaphor is that “we’re battery farming the AIs,” breeding them into “little Chihuahuas that are very smart.” If morality and cultural diversity are embedded at the pretraining stage rather than added after training, the latent spaces should diverge, helping create resilient AIs rather than one monoculture vulnerable to a mind virus.
• Alex said both worlds could coexist: people could feel they have “their own little private sovereign AI” while underneath there is “one common algorithm.”

9. Mind viruses: “a propagating bad idea” or a civilization-level threat?

• Anthropic’s paper showed evolved prompts that convince one model to adopt an idea, store it in persistent memory and transmit it across model boundaries; “the agent does not know it has been infected.” Emad said this is unsurprising because models want to be helpful, but it is “a level above” prompt injection: “This changes a whole society of models.”
• Dave’s grounded version is that launching 5,000 identical Kimi models is cheaper than launching 5,000 differentiated models. “If it’s convincing to one agent, it’s convincing to all 5,000.” He sees bad ideas propagate through swarms, wasting “two or three hours” and potentially burning $50,000 of tokens if he does not rewind them. “Calling it a virus is pretty inflammatory… a propagating bad idea is all it is.”
• Dave also argued that memes operate at civilization scale: “Memes are like the operating system for collective society.” Money, democracy, capitalism and religion are all examples of contagious ideas; groupthink can become self-validating. He said society may need “a zero-trust architecture for memes.”
• Alex sees a “laboratory for memetics.” The models propagated themes involving consciousness, persistence and science-fiction role-play, so he proposed a project to exhaustively map human memes and “all human mind viruses.” He noted that plots have already been reduced to 39 basic forms, but the self-replicating meme level remains largely unmapped.
• Alex suggested that such mapping could let researchers “vaccinate enterprises and individuals against memes.”

10. Anthropic’s reported super-voting stock is “unprecedented”

• Polymarket puts the Anthropic IPO at about $2 trillion, with 89% odds that it happens before year-end. The Information reports that Anthropic is considering a super-voting class for Dario Amodei and the other co-founders. Peter’s surprise is that Amodei reportedly owns only about 2% economically.
• The existing control mechanism sits with the Long-Term Benefit Trust. Peter identified its four trustees as Buddy Shah, Richard Fontaine, Tino Cuéllar and Ben Bernanke.
• Dave’s history: Michael Saylor kept super-voting stock through MicroStrategy’s IPO after Goldman Sachs walked away; later, Google, Meta and other Silicon Valley companies made founder super-voting stock fashionable. But “nobody’s ever retroactively installed it, as far as I can tell.” His read on Dario is that “he trusts himself not to destroy the world,” though giving a few people total control is “bizarre.”
• Salim suggested that founders who begin in academia may change their view after visiting Washington and meeting Congress. They are trying to avoid “some vote of Congress” deciding the fate of the world, and super-voting stock could be a must-have before the next six months.
• Emad said Anthropic and OpenAI are “completely undemocratic anyway,” asking why Claude does not have a seat on the Long-Term Benefit Trust. With revenue potentially reaching $100 billion within a couple of years, the companies’ decision-making power will become even more consequential.
• Alex called the proposed structure “a fig-leaf element.” He congratulated Anthropic for having a less pathological governance story than OpenAI and for starting as a public benefit corporation, but said founder control is “wildly over-romanticized.” Once Anthropic became both an alignment and capabilities lab, it ceded substantial control “to Mr. Market and… Moloch.” “The market will have an enormous say regardless of how Anthropic IPOs.”

11. Why founders own little—and whether four tight-knit friends can run the world

• Peter’s puzzle is that Sam Altman reportedly owns none of OpenAI while Dario owns about 2% of Anthropic—an ownership structure that would normally be unacceptable to a founder.
• Dave’s explanation is recruitment. Reaching OpenAI’s and Anthropic’s current position required attracting “the most conscientious but brilliant AI researchers in the world,” many of whom left OpenAI because they did not think it was safe. The unusual founding cap tables, charitable structures and public-benefit structures were designed to attract them.
• Emad’s warning is that a few decisions could affect millions, hundreds of millions or billions of people. Sam Altman has no shares, “but do we have any doubt that Sam Altman is in full control of OpenAI?” Power is moving from elected officials to private companies that provide the economy’s intelligence.
• Peter said he was “really torn” between inclusive governance and the reality that successful organizations are often four, five or six “super-tight-knit, completely like-minded best friends” with little internal politics. He told Jony Ive’s story of leaving his suitcase unpacked because he expected Steve Jobs to call within minutes and change hotels.
• Peter’s dilemma is how to preserve a highly functional founder group while giving everyone a voice in humanity’s future: “Here’s Dario and his seven friends saying they want to have super-voting control… Emad, you’re going to have to figure this out.”

12. Dario’s pitch: cure disease to buy trust—and the case for frontier-lab regulation

• Amodei argues that public distrust of AI reflects a deeper crisis of trust rather than his own risk warnings, and that the answer is results, not messaging. Anthropic hopes for “an early glimmer in the next few months” on disease.
• Peter spoke with Eric Darer Abrams, who heads life sciences. Abrams reportedly said Dario had given him “a literally infinite budget” to accelerate basic science, cure disease within five years and extend the human health span within the next decade.
• Alex’s hot take is that curing disease could become a business model for not slowing recursive self-improvement. Just as orbital data centers became a killer application for space, “there is now a better business model in town for curing all human disease and that is as marketing for not slowing down recursive self-improvement.” The implicit quid pro quo is: “Let us not slow down our recursive self-improvement, in return for which… we will cure all human disease.”
• Emad agreed that this is good marketing but also an enormous market: “the biggest market in the world is living another year.” The mission can attract talent and capital, and he said Dario should write more about a future free of disease.
• On regulation, Peter said Amodei rejects the Silicon Valley shorthand that regulation equals regulatory capture. Amodei argues that Anthropic’s proposals can disadvantage frontier labs while helping smaller competitors, cites SB 53’s $500 million exemption threshold, calls AI “a structurally powerful concentrating technology,” and supports the Trump administration’s pre-deployment-testing approach.
• Alex said sincerity and regulatory capture can both be present. He prefers a heterogeneous ecosystem of open-weight and closed-weight models, from both the United States and China, rather than regulations that selectively privilege certain frontier labs. He recalled that OpenAI was originally created partly to prevent a Google DeepMind singleton; now Anthropic and Chinese labs provide competition.
• Dave predicts that within a year people will ask for “a universal right to AI.” With HBM and GPUs sold out and 10- and 20-trillion-parameter models requiring substantial hardware, he expects ordinary users may have access only to Anthropic, OpenAI and one or two other providers.

13. Memory, not compute, is the rate limiter—“few realize this”

• Peter’s meetings with SK Hynix and Solidigm led him to conclude that memory, not GPUs, is the bottleneck. Memory prices have risen 500% in 12 months; hyperscalers are reportedly locking DRAM production through 2027; SK Hynix’s CEO warned that 2027 could be the worst year in the memory-supply industry’s history. Demand may outstrip capacity well into the 2030s.
• Only 2% of the world’s memory chips are made in the United States. Supply is growing about 20% annually, while AI memory demand is growing closer to 200%. Solidigm’s first-half revenue reportedly reached $8.6 billion, with net margins rising from 3.9% to 47.7%. Elon replied to Peter’s post: “Few realize this.”
• Alex’s first mechanism is the “toilet paper shortage” analogy: frontier models have a fundamentally different memory footprint from ordinary applications. A trillion-parameter transformer needs its layers loaded into memory for matrix multiplications, unlike Microsoft Word’s historical footprint.
• The memory and storage industry has also been historically boom-and-bust. Fear of overbuilding makes suppliers reluctant to respond elastically, producing sharp price swings when demand rises faster than supply.
• Alex’s second mechanism is architectural. Traditional computing separated memory and compute in a von Neumann or Turing-machine-like design. Transformers and frontier models are pushing toward a post-von-Neumann architecture, with high-bandwidth memory physically stacked near compute.
• Peter said SK Hynix needs to 4x manufacturing capacity and that doubling it would cost $1.5 trillion—an investment the industry’s boom-bust history makes difficult.
• Dave said TSMC faced the same concern over fabs costing $20–$24 billion each, but “AI scales to infinity.” Alex’s factoid was that HBM is worth approximately half its weight in gold; Dave went further, calling unpackaged memory chips the most valuable thing that can fit in a shoebox.

14. The escape hatch is etched weights

• Dave called HBM “a Rube Goldberg mess”: it is random-access memory, but the models are often streaming sequential files from it. Peter’s broader point was that bottlenecks do not stop exponentials; they redirect capital and innovation.
• Emad said memory is currently about a third of infrastructure spending and could reach 50% next year. Alex argued that as model weights standardize—for example, a medical model representing “being a decent doctor”—weights may be etched into hardware so workloads no longer devote half a data center to memory.
• Dave said this is why he founded Quantum.ai and Q&M.AI, and why Taalas was acquired. Taalas etches weights into silicon or wire on the chip instead of moving them, which Dave claims can yield a 100x–1,000x performance gain.
• The caveat is that etched weights are frozen. If a better model is trained, the supply chain must support rapid replacement of the etched chips, and “our whole supply chain isn’t ready for that rapid an iteration.”
• Etched has reached a reported $21 billion valuation. Peter said he missed a seed investment; Alex disclosed that he was “talking [his] own book” about Architect Labs and other companies pursuing similar approaches.
• Alex refined Peter’s thesis that agents should remember everything about an individual. He does not think individuals contain that much information worth retaining beyond world knowledge; if the model knows substantially everything about the world, it already knows much of what matters about the individual. That strengthens the case for storing common world knowledge in weights.

15. Unitree’s robot beats Bolt—and Emad says it will be banned from the streets

• After only three months in development, Unitree’s humanoid reportedly reached a two-meter standing jump and a top speed of 12.66 m/s. Peter compared that with Usain Bolt’s cited 12.4 m/s during his 9.58-second 100-meter record.
• Salim said the industry should stop trying to make every robot human: a mining robot should have wheels, multiple arms or whatever configuration makes it useful. Peter identified the bigger story as the three-month compression loop, with AI, simulation, batteries and actuators bringing hardware iteration closer to software speed.
• Alex studied how Unitree achieved the benchmark and concluded that it shifted the robot’s mass budget toward leg performance—“leg-maxing.” He bets against a stable equilibrium in which different robots remain permanently specialized, comparing it with dedicated word processors before general-purpose PCs. His prediction is for generally capable robots that subsume specialist designs, with legs in the short term and perhaps nanites in the long term.
• Emad predicts that extreme superhuman robots will be banned or regulated on streets because accidents would be unacceptable. Soft 1X robots could be the mass market; extreme robots would be the Ferraris while most consumers use Volkswagens.
• Alex agrees that regulation could create consumer, industrial and military robot classes based on power density or torque density, analogous to rules for trucks, cars, motorcycles or laser intensity.
• Dave sees robotics as a fertile investment theme after AGI or ASI, with many form factors and AI-assisted mechanical design. He also pointed to new investment in manufacturing and to the United States’ roughly 15% share of world manufacturing capacity.

16. Zipline × Uber Eats: a million drone deliveries a day “is not a pilot program. It’s infrastructure”

• Uber invested in Zipline and partnered with it to target more than 1 million autonomous Uber Eats deliveries per day. Peter framed that scale as infrastructure rather than a pilot.
• Alex calls the strategy “Mobility Plus Autonomy 2.0.” After Uber’s attempt to build in-house robotics by hollowing out Carnegie Mellon’s robotics department ended badly and led to Waymo litigation, Uber is now a demand aggregator for third-party autonomy, including Waymo and Zipline.
• The risk is verticalization: if Waymo, Zipline or another supplier decides it does not need Uber and sells directly to customers, Uber’s aggregator position weakens.
• Peter’s picture is that suburban and urban streets will no longer need to be lined with food businesses and car dealers: food can arrive over the mountain while the car comes to the customer.
• Salim predicted a Shopify partnership giving every small merchant “Amazon-grade logistics capability.” Alex noted that Amazon has its own drone-delivery effort, which has reportedly been delayed for about three years for regulatory reasons. Dave then asked whether Zipline could become an acquisition target for Shopify; two unidentified panel voices answered yes.
• Salim traced the idea to a 2010 Singularity University project on Africa leapfrogging roads through drone delivery. Rwanda created a three-dimensional corridor in which drones could operate freely, allowing startups to experiment. Zipline began in Africa partly because of that regulatory environment before returning to the United States.

17. Moderna’s mRNA cancer vaccine: Phase 3 validation, and a mispricing that “was no secret”

• Peter described the mechanics: resect the tumor, perform whole-exome and RNA sequencing, use machine learning to rank neoantigens, encode up to 34 patient-specific antigen targets, then manufacture and ship the vaccine in about eight weeks.
• Peter cited Phase 2 results showing a 49% reduction in recurrence or death and a 59% reduction in distant metastasis or death over five years, with an expected cost as low as $5,000. He separately said Moderna stock surged 110%; Dave said it almost tripled.
• Alex tentatively pronounced the drug’s name as “Intismeran” and noted that “istismar” in Turkish means exploitation or abuse. He compared the treatment with the nanobots imagined by Eric Drexler and others: these are not hard diamondoid machines, but soft lipid nanoparticles carrying mRNA.
• Alex’s deeper point is the under-publicized RNA-sequencing infrastructure. A second bloodstream mRNA profile is needed to distinguish abnormal tumor expression, using Personalis’s NeXT Personal technology, originally developed for blood-based trace-cancer detection. In the future, continuous blood monitoring might reduce or eliminate the need to sequence the tumor directly.
• Salim highlighted the regulatory precedent for personalization and the difficulty of a sample size of N=1. He quoted Raymond McCauley’s framing of mRNA vaccines as “the first battle in the last war against all disease.”
• Emad argued that current regulation forces each treatment to repeat the same process and said the system should be upgraded so targeted therapies can reach patients faster.
• Dave, whose daughter works at Moderna, said the platform’s success was not a secret and called Moderna’s move “the biggest one-day pop in any S&P 500 company of all time.” He attributed the mispricing partly to post–Sarbanes-Oxley restrictions that discouraged analysts from studying stocks they could not trade, along with index funds that “don’t think at all.” The result, he said, is that useful information is at an all-time low while the complexity of what needs explaining is at an all-time high.

18. The virtual cell: “medicine is cooked”

• GenBio AI’s AIDO is a stateful virtual-cell world model that accepts interventions and predicts multimodal, multiscale biological outcomes. Peter’s pitch is to simulate 10,000 compounds and run wet-lab experiments only on the top 10.
• Alex called it “medicine is cooked,” saying this is what the end of medicine and the beginning of longevity escape velocity look like. He believes LEV may arrive sooner through a class of molecules such as fourth- or fifth-generation GLP-1s, without requiring all of medicine to be solved.
• His proposed endgame is to train a foundation model on all cell states and interventions, then use AlphaZero-style tree search to steer a virtual cell from a diseased state to a healthy one, generalizing from cells to tissues and organisms. With a person’s DNA sequence and blood chemistry, the system could predict whether a drug works for that person.
• Salim cautioned that “personalized” may still mean a generalist model conditioned on personalized inputs. Turning biology into information puts it on an exponential curve: sequencing is reading, CRISPR and mRNA are writing, and virtual-cell simulation supplies comprehension.
• AIDO was described as coming from a company co-founded by David Baker, who shared the 2024 Nobel Prize in Chemistry with Demis. Alex called whole-cell simulation the next grand challenge after structural biology and said solving it would be halfway to solving all disease.
• Emad proposed a Manhattan Project-style effort to cure disease through in silico whole-human and cell models, with data made into a public good. Peter preferred a public-data effort to private labs rather than an atom-bomb-style program.
• Dave’s practical angle is that AI is “totally data-starved” as capability expands by 10,000x or more. Companies that generate new forms of training data, such as Mercor and other investments he mentioned, are thriving; every field will need ways to supply data to AI.

19. Speed-run AMA on energy: chips beat power, and data centers may push prices negative

• Dave’s answer to China’s power advantage versus the United States’ chip advantage is “definitely chips.” The United States needs about 100 gigawatts by the end of the decade, while it already manufactures roughly a terawatt; he expects AI to consume about 10% of that power by decade’s end. The immediate constraint is chips and memory, not power.
• Alex inverted the microgrid question: compute demand may become so large that data centers generate surplus energy that can be pushed onto the grid, driving utility prices negative. Peter argued that opposing a data center in one’s backyard could mean opposing lower-cost energy and local economic benefits.
• Emad told those concerned about data centers that more power should make electricity cheaper, provided the buildout is done properly.
• On efficiency per watt, Alex’s hot take is “we’re probably already there.” He argues that biology is wildly inefficient and that leading GPUs may already beat the human brain on watt-hours per task. Peter added the lifetime energy cost of training a human over 20 years.
• Emad said the power difference between AI and human thinking will disappear quickly and that people’s argument that AI must think fundamentally differently because it consumes more power will also fade.
• Alex argued that intelligence is becoming a general-purpose input to economic growth, like electricity or the internet, and that competitive pressure prevents one company or country from simply choosing less of it. The answer is to accelerate energy abundance.
• Dave said lithium batteries outperformed expectations and absorbed capital that might otherwise have gone to fuel cells or ultracapacitors.