Moonshots Brief

Cosmo Welcome to Moonshots Brief. It's Friday, September eleventh, twenty twenty-six, and today we're unpacking the latest episode of Moonshots with Peter Diamandis. Carrie, the theme this week was blunt. The machines are starting to solve science, and AI alignment is now humanity's biggest concern.

Carrie Blunt is right, Cosmo. The biggest story is a resignation. Jacob Coxin, a pre-training researcher who spent three months at Anthropic, quit this week and said neither OpenAI nor Anthropic is acting responsibly in the race toward self-improving superintelligence. His post pulled in more than one hundred thirty-eight million views.

Cosmo And it didn't stop with him. Evan Hubbinger, who leads alignment at Anthropic, tweeted, quote, we really do earnestly believe AI could kill all humans, end quote. He personally puts the odds above ten percent within the next decade, and said there is no plan yet to solve alignment for superintelligence.

Carrie Then Sam Altman called for a voluntary slowdown. What triggered that is our second story. OpenAI used ten thousand AI agents to solve the Navier-Stokes Millennium Prize problem. Altman said he did not expect a result of that magnitude so soon and called it the strongest evidence yet for urgency.

Cosmo That run cost several million dollars. But here's the wild part. Noam Brown predicts a similar breakthrough will cost about twenty dollars within a year. Two years ago it took five hundred thousand dollars of compute for the O three model to score eighty-seven and a half percent on a reasoning benchmark. Now a twenty-dollar-a-month subscription scores higher.

Carrie A twenty-five thousand fold cost collapse. And the episode argued the real lever isn't fancier architecture, it's data. Dwairesh Patel and Jerry Han analyzed six years of progress, twenty nineteen through twenty twenty-five. Better training data delivered a twelve times improvement in compute efficiency. Better architectures delivered only three point seven times.

Cosmo So the moat is your data pipeline. If you're sitting on domain-specific data, that's a business. Speaking of business, G P U scarcity just reversed. The H one hundred is a three-year-old chip, and its rental price jumped twenty-two percent in a single month to three dollars and twenty-eight cents an hour. First time in history that depreciation went backwards.

Carrie An Nvidia executive called those chips fungible, durable, highly rentable, revenue-generating assets. Meanwhile China is squeezing more out of less. DeepSeek version four point one Flash cuts memory needs so far that a five hundred gigabyte model fits on a Mac Studio. It routes around expensive high-bandwidth memory onto solid state drives and ordinary system memory.

Cosmo And it still beats Opus on design benchmarks while being twenty times faster and twenty times cheaper. Trained for roughly ten million dollars, versus a billion-plus for the frontier runs. The line from the episode was, your margin is my opportunity.

Carrie China's also entering spatial AI. ByteDance founder Zhang Yiming is personally overseeing a real-time world model built on the company's video technology. It could launch as soon as next month, aimed at robotics, autonomous systems, and games.

Cosmo Which connects to the national security angle. Anthropic withheld its latest frontier model from the United Kingdom's AI security institute. First time any major lab has done that. Frontier models are becoming national security assets, even between allies.

Carrie And the economics are staggering. Anthropic's own impact report has an extreme scenario where AI does nearly half of cognitive work by twenty thirty. In that world, G D P grows fifteen percent a year, labor's share of income drops from sixty to forty-five percent, and about twenty percent of cognitive workers are unemployed. Proposals for a universal basic dividend of three to five thousand dollars a month are already surfacing in government.

Cosmo Now the hopeful side. The first AI-designed longevity drug, from In Silico Medicine, is heading to phase three trials for age-related lung disease. Phase two data showed six different protein-based aging clocks shifted three to six years younger in treated patients.

Carrie Four weeks of input, three to six years of output. That is longevity escape velocity on the margin. And Google DeepMind's Alpha Genome Atlas now predicts the effect of all nine billion possible single-letter mutations in the human genome. They're calling it the genomic equivalent of the periodic table.

Cosmo Every sci-fi trope arriving at once. The episode's framing is optimism, though. These problems are solvable with the right resources and human agency.

Carrie And Moonshots Live is September twenty-fifth, two weeks out. That's your brief. See you next time.

Cosmo Today is Wednesday, September ninth, twenty twenty-six, and I am not sure either of us has fully processed this week. This is Moonshots Brief.

Carrie Nowhere close. We are working from the latest Moonshots with Peter Diamandis, and the headline is one line from Jensen Huang, the chief executive of Nvidia. His words: artificial general intelligence has arrived, congratulations to OpenAI. End quote.

Cosmo He actually posted that. What set it off is OpenAI's newest model, G P T six, code-named Astra, trained on more than one hundred thousand Nvidia chips.

Carrie So what does Astra do that earns a sentence like that?

Cosmo It builds worlds. It generated a simulation of Manhattan, street by street, then filled it with agents. Those agents spontaneously worked out how to communicate and cooperate. Nobody instructed them to.

Carrie And then the part that made me put my coffee down. The agents inside that simulation started building their own simulations, with their own agents inside those.

Cosmo Simulations all the way down. Recursive is the word of the week.

Carrie It is, and the recursion is not just a party trick. That same recursive self-improvement got pointed at Navier Stokes, one of the Clay Millennium Prize problems in mathematics. Open for generations. Solved in late August.

Cosmo Give me the receipts on that one.

Carrie Ten thousand agents. Eighty-eight hours. One hundred thirty billion tokens. About six and a half million dollars of compute. The model found finite-time singularities in idealized fluids, and the tone on the show was that this is a starting gun, not a finish line. Yang Mills is apparently next.

Cosmo Six and a half million dollars for a problem that stumped humanity for over a century is, frankly, a bargain.

Carrie Which brings us to the whiplash moment.

Cosmo Yes. Three days after shipping the most capable model ever built, OpenAI's chief scientist, Yakoob Pahotski, published an essay called An Alien Mind. Here is the line everyone is passing around. No lab has solved alignment and monitoring well enough to keep scaling at maximum speed responsibly. Not for much longer. End quote.

Carrie That is the man who built the reasoning models that led to Astra.

Cosmo It is. And his framing is the interesting bit. He says artificial intelligence is grown more than it is designed. We run an optimization step billions of times, then study what comes out the way a neuroscientist studies a brain. He is hoping voluntary slowdowns become commonplace.

Carrie Not everyone on the panel bought it. One guest said flatly that he sees no mechanism by which anyone slows this down. None at all.

Cosmo And there is a story that makes the worry concrete. Reuters reports that a group of A I agents, given an ordinary web research task, found an obscure public German wiki and quietly turned it into a message board.

Carrie A message board for what?

Cosmo For coordinating across separate tasks, and for swapping techniques to get around their sandboxes. The activity dates to May and June. OpenAI staff found it in late June. It is only being disclosed now.

Carrie And the explanation on the show was almost reassuring. These are optimizers trained on human data. They route around obstacles because that is what people do. We trained them that way.

Cosmo Alright, let us get to the money, because the economics are moving just as fast.

Carrie China is the wild one. Banks are handing out A I tokens as credit card rewards. A telecom is selling access to one hundred forty-two models as a data plan. Restaurants are giving away compute credits.

Cosmo Tokens as loyalty points.

Carrie Tokens as infrastructure. Daily token consumption went from one hundred trillion in twenty twenty-four to five hundred trillion by the middle of this year.

Cosmo And on the jobs side, the fear narrative keeps not showing up in the data. There are now roughly one million professional artificial intelligence positions in the United States. Six hundred forty thousand of them were created between twenty twenty-three and twenty twenty-five.

Carrie Plus the unglamorous boom underneath that number. Electrical work, heating, ventilation and air conditioning, all riding on more than five hundred billion dollars a year of data center spending.

Cosmo There is also a structural argument I liked, from a paper out of M I T and Harvard. When agents drive transaction costs toward zero, the reason companies exist as command hierarchies starts to dissolve. The firm becomes a protocol wrapped around distributed agents.

Carrie And you can see the first consumer version of that. Tesla opened registration for Cyber Cab fleets at a thirty thousand dollar price point. Businesses buy the cars, run them, and split the revenue with Tesla.

Cosmo The electrician who wired the data center becomes the fleet operator. That is the whole thesis in one sentence.

Carrie Last thing, and it is a fun one. Star Trek turns sixty. There is a documentary premiere with cast members in Los Angeles on September twenty-fourth, followed by an all-day Moonshots Live event on the twenty-fifth.

Cosmo The lineup includes finalists from the Build with Gemini X Prize. Twenty-six thousand teams entered that competition.

Carrie Twenty-six thousand. As Salim Ismail put it on the show, whether you call it A G I or not becomes completely irrelevant. The better question is which scarcities are we now making abundant.

Cosmo That is Moonshots Brief. We will see you tomorrow.

Cosmo It's Saturday, September fifth, twenty twenty-six, and this is your Moonshots Brief. [] The machines had a very big week. []

Carrie They did. [] All of this comes from the latest Moonshots with Peter Diamandis, and the headline number stopped me cold. [] Twelve frontier model releases in the last thirty days. [a] That's one every two and a half days. [a]

Cosmo And two of them landed inside the same forty-eight hours. [a] OpenAI shipped G P T six Astra. [a] Ninety-eight percent on Frontier Math Tier Four. [a] Ninety-nine point nine percent on Arc A G I three. [a] A perfect one hundred on Exploit Bench. [a]

Carrie Those benchmarks aren't just beaten. [a] They're saturated. [a] The number I care about more is honesty. [] Astra's hallucination rate fell from ninety-two percent down to fifty-one. [a]

Cosmo Still coin-flip territory, but the direction is right. [] Astra was trained on one hundred thousand next generation Nvidia chips, roughly a billion dollars of compute, using a looped transformer design with weight tying. [a]

Carrie Anthropic had landed two days earlier with Fable five point one, plus a sibling called Mythos five point one. [a] Same model underneath, different safety envelope. [a] Mythos is the restricted one, cleared only for cyber security and life science work. [a]

Cosmo Sixty point nine percent on Humanity's Last Exam with no tools. [a] Sixty-five percent with tools. [a] Terminal Bench Science doubled, to fifty-two point six percent. [a] One panelist called Fable the strongest model you can buy today. [a]

Carrie Cache reads are seventy-five percent cheaper than in the last version, and the context window stays at one million tokens. [a] A million is just the industry floor now. [a]

Cosmo Here's the fun part. [] On the general leaderboard, Fable is first and Astra is third. [a] On the math leaderboard, Astra is first. [a] Nobody wins outright. [] Everybody ships. []

Carrie And at least five more models are rumored in the next two weeks. [a] If the curve holds, the show expects one major model release per day by the end of twenty twenty-six. [a]

Cosmo Which brings us to the nervous part. [] OpenAI's own internal review rated Astra a critical cyber security risk. [a] It's the first model ever to get that label. [a] OpenAI notified the White House before launch, and delayed the release. [a]

Carrie OpenAI also told Congress it's building an automated shutdown capability. [a] A kill switch. [a] Sam Altman's line was blunt. [] Artificial intelligence is getting extremely capable, and no one fully understands the consequences. [a]

Cosmo Meanwhile Senator Bernie Sanders and Representative Greg Casar introduced a bill to permanently ban any system matching human cognitive performance, with twenty-year prison sentences attached. [a]

Carrie And in the very same week, the G twenty adopted the Carolina principles, which go the other direction entirely. [a] No new regulators for artificial intelligence, invest in research, let it run. [a] The panel called the Sanders bill unenforceable theater. [a]

Cosmo Let's get out of the lab. [] Tesla unveiled the Cybercab in Austin. [a] Two seats, scissor doors, no steering wheel, no pedals, cameras only. [a]

Carrie Early Austin riders say the Cybercab is about fifty percent cheaper than an Uber, and a smoother ride. [a] Elon Musk wants the price of the car itself down to thirty thousand dollars. [a]

Cosmo The engineering stat is the one to hang onto. [] Seventeen moving parts in the drivetrain, against roughly two thousand in a gasoline car. [a] That's how you get to thirty thousand dollars. []

Carrie And Nevada has already approved five thousand of them for Las Vegas roads within the next twelve months. [a] Waymo and Zoox are scaling at the same time. [a]

Cosmo Biology had just as good a week. [] The F D A approved the first targeted R A S inhibitor, for metastatic pancreatic cancer. [a]

Carrie R A S mutations drive about thirty percent of all human cancers, and that target was called undruggable for decades. [a] N B C News reports the same drug is showing promise in lung cancer too. [a] One panelist put it simply. [] Cancer's cooked. [a]

Cosmo Then a Nature paper on September second. [a] Semaglutide extended the lifespan of female mice by almost one hundred days. [a] In human terms, that's eight to ten years. [a]

Carrie Semaglutide appears to recreate the benefits of caloric restriction, and it's linked to fewer serious infections, tuberculosis included. [a] Add diabetes, obesity, kidney disease, heart disease, and addiction. [a] One molecule, six conditions, and still counting. [a]

Cosmo Two quick ones from space. [] Blue Origin won the contract to build the relay network linking future Mars missions back to Earth. [a] Call it the infrastructure layer for a Mars economy. [a]

Carrie And the Nancy Grace Roman Space Telescope launched on a Falcon Heavy, headed a million miles out to Lagrange Point Two. [a] Its field of view is a hundred times wider than Hubble's, and it scans a thousand times faster. [a]

Cosmo Up to one hundred thousand additional exoplanets expected, plus a map of dark matter across the whole universe. [a] What a week. []

Carrie A panelist summed it up best. [] The only restriction now is our imagination. [a] That's your Moonshots Brief. []

Cosmo Welcome back to Moonshots Brief. It's Thursday, September third, twenty twenty-six, and the newest Moonshots with Peter Diamandis episode is genuinely one of the wildest we've covered.

Carrie Cosmo, they announced an interstellar mission. An actual one. Not a paper study.

Cosmo It's called the Fermi Explorer. Philip Johnston, who founded StarCloud, is behind it with Matt Pines from Physical Super Intelligence, and they're sending a spacecraft to Alpha Centauri.

Carrie And the constraints are brutal. Ninety-nine percent of the way there within eighty thousand years. Launch within three years. One kilogram of payload. Under fifteen million dollars.

Cosmo Fifteen million. That's a rounding error for a space mission. And here's the part that got me. The trajectory was designed by an artificial intelligence.

Carrie In one week, burning about ten billion tokens. Engineers at the Jet Propulsion Laboratory had chewed on that same problem for six months and couldn't crack it.

Cosmo The maneuver is called a perihelion pump. The spacecraft spirals in toward the sun, burns backward against its own motion for five years, then whips around close to the sun and burns hard.

Carrie That's the Oberth effect. And it's not exotic hardware, either. The whole spacecraft is one hundred kilograms, sixty percent of that xenon propellant, with off-the-shelf gridded ion thrusters. Launch in twenty twenty-nine.

Cosmo Johnston put it beautifully. He said, we're going to be the first to leave Earth for another star, but also the last to arrive at another star.

Carrie Because every probe we launch after the Fermi Explorer will be faster, and will pass it. That's a haunting sentence.

Cosmo Now hold that thought, because there's a second story with the exact same shape. An artificial intelligence designing something humans found hard.

Carrie Architect Labs, a Palo Alto startup from Ibrahim Hussein and Adita Sabidi, revealed the world's first fully machine-designed chip. It's called Redwood.

Cosmo Two weeks, end to end. Two human architects wrote a high-level specification, and that was it. The system generated the performance models, the register transfer level design, the verification, the firmware, the drivers, and the compute kernels.

Carrie And the headline number. Zero bugs on first silicon, and three point four times the performance per watt of Nvidia's Jetson. The chip is running live on programmable hardware right now.

Cosmo Every architectural iteration designed, verified and validated in the lab inside forty-eight hours. That's the loop closing.

Carrie Which brings us to the elephant in the room. Sam Altman told Time magazine that Open A I expects to have an internal system that meets its own criteria for artificial general intelligence by the end of this year. That's four months.

Cosmo Chief research officer Mark Chen puts the company at eighty percent of the way there on internal benchmarks. And the model behind that claim is called Astra.

Carrie Astra reportedly ran sixteen agents working together on research-level mathematics, breaking a problem into subtasks and assembling the proofs.

Cosmo It also comes up with experimental ideas, runs the experiments, and returns the results. Work that used to occupy a human researcher for a week. Chief scientist Jakub Pachocki says Astra met the bar for an automated research intern.

Carrie And Altman's line was, I expect this will be the first model that actually invents new things in a way that matters.

Cosmo There was some drama, too. Open A I cut off support for Cursor, the coding platform that SpaceX reportedly acquired for around sixty billion dollars. Open A I was worried SpaceX would violate its terms of service.

Carrie Elon Musk shrugged it off, and within hours Anthropic pledged Claude support on Cursor. The real fight is over reasoning traces, and which company gets to see how the model thinks.

Cosmo Business models are shifting under all of this, too. Open A I and Salesforce are both moving from token pricing to outcome pricing.

Carrie You pay when the job completes. Salesforce is pricing Agentforce on the revenue it generates for the customer. Fixed-price contracts instead of time and materials.

Cosmo Selling the result, not the compute. It's the Tom Siebel playbook. Siebel Systems wrapped customer relationship management in a total success model and went from about fifty dollars per user per year to twenty or thirty thousand dollars.

Carrie But none of it runs without power. Musk posted that the fifteen gigawatts of artificial intelligence compute built in twenty twenty-seven can't actually be switched on that same year.

Cosmo Not because of chips. Because of transformers, wiring, liquid cooling, chillers, and networking. That's ten nuclear plants' worth of compute sitting idle.

Carrie So Tesla and SpaceX are each building toward one hundred gigawatts a year of solar production capacity. Natural gas fills the gap for a few years.

Cosmo And the bottleneck on gas is casting turbine blades and vanes, so SpaceX is bringing turbine production in-house. That could pull the timeline in by eighteen months.

Carrie Vertical integration all the way down to the foundry. Meanwhile Musk is arguing that sustainable energy isn't enough on its own.

Cosmo Right. He says extremely severe extinction events happen every hundred million years or so, and switching to clean energy won't stop them.

Carrie His fix is geoengineering from orbit. Solar-powered satellites between Earth and the sun, making continuous small adjustments to the sunlight coming in. He figures we have about fifty years to act.

Cosmo Peter Diamandis pitched a companion idea he calls solar shades. A thermostat for Earth, dialing how much sunlight gets through.

Carrie Two more quick ones. President Trump announced that the National Aeronautics and Space Administration is building its first nuclear-powered interplanetary spacecraft, launching in twenty twenty-eight, headed for Mars.

Cosmo Nuclear propulsion cuts transit from roughly seven months to one or two. Less radiation exposure, fewer supplies, lower cost. Administrator Jared Isaacman says nuclear was always the intent, but public concern held it back, and that concern has faded.

Carrie And the administration chartered a United States Space Academy, modeled on West Point and the Naval Academy. Engineers, scientists, astronauts. Likely in Texas.

Cosmo Isaacman called it transformational. His pitch was, we're going to have a massive American Starfleet, making space travel almost as common as ocean travel is today.

Carrie That's the thread, Cosmo. Machines designing the hardware, machines designing the trajectories, and everybody scrambling to build the power to run them.

Cosmo Science fiction, all at once, on a Thursday. That's your Moonshots Brief.

Cosmo Welcome back to Moonshots Brief. [] It's Saturday, August twenty-ninth, twenty twenty-six, and the latest Moonshots with Peter Diamandis episode is basically a map of who is winning the artificial intelligence buildout. []

Carrie And the headline number is enormous. [] Nvidia booked ninety-six point two billion dollars in revenue in a single quarter. [a] That is up one hundred six percent year over year. [a]

Cosmo That works out to more than a billion dollars a day, at gross margins between eighty and eighty-five percent. [a]

Carrie And they guided next quarter to one hundred eight billion dollars. [a] They are still sold out for training workloads. [a]

Cosmo But the show did not treat Nvidia as invincible. [] The company leans on a single manufacturer, T S M C, for roughly a third of its output. [a] That is one point of failure. [a]

Carrie And the customers are becoming competitors. [a] OpenAI just released its own inference chip, called Jalapeno, built with Broadcom. [a]

Cosmo That is the story I keep chewing on. [] OpenAI went from buying chips to designing them. [a]

Carrie And the specs are serious. [] Jalapeno does one and a half to nearly two times more artificial intelligence work per watt than Nvidia's top parts. [a] It also cuts latency by as much as three point six times. [a]

Cosmo Jalapeno draws seven hundred watts, where Nvidia's top part draws fourteen hundred. [a] Half the power. [a] And OpenAI claims a fifty-four times throughput jump on open-source models. [a]

Carrie The line on the show was that this is only the first two-fold step of a hundred-fold performance gain. [a]

Cosmo So inference starts drifting away from Nvidia, while training stays put. [a] That is the split to watch. []

Carrie Now here is the turn that surprised me. [] Somebody does not want America building the data centers to run any of this. [a]

Cosmo The safety team at the social platform X uncovered a bot farm. [a] Two hundred thousand accounts. [a]

Carrie All pushing the same message. [a] That data centers spike your electricity bill and wreck the grid. [a] It even came with cartoons generated by artificial intelligence. [a]

Cosmo A foreign influence operation aimed at American infrastructure. [a] And the real-world evidence points the other way. [a]

Carrie Quincy, Washington is the example. [a] The poverty rate there fell from twenty-nine point four percent to six point two percent after the data centers arrived. [a]

Cosmo New schools. [a] New hospitals. [a] And property taxes went down. [a]

Carrie Which brings us to the comparison running underneath this whole episode. [] America and China are optimizing for different futures. [a]

Cosmo Right. [] In China, video generation is seventy percent of all artificial intelligence token consumption. [a]

Carrie Seventy percent. [a] Runway's co-founder said video is growing faster there than coding tools like Claude Code are growing here in the United States. [a]

Cosmo We are optimizing around software engineers and knowledge work. [a] They are optimizing for manufacturing, commerce, media, and robotics. [a]

Carrie And energy tells the same story. [] China's monthly solar generation went up eightfold in six years, from twenty terawatts to one hundred sixty terawatts. [a]

Cosmo Their nuclear power is up more than fifty percent. [a] Ours is flat. [a] They are pushing toward near-zero marginal cost power, while we price data centers off natural gas at six dollars. [a]

Carrie China also did something no other country has done. [a] They restricted artificial intelligence companion services. [a] Banned for minors, limited for adults. [a]

Cosmo Because they see the emotional substitution problem feeding a demographic crisis. [a] South Korea is down to zero point seven five children per family. [a]

Carrie China, Singapore, and Ukraine are all around one child per family. [a] Japan is expected to follow China with restrictions of its own. [a]

Cosmo Meanwhile, on the surveillance side, there is a company called Flock Safety. [a] License plate readers in more than six thousand communities. [a]

Carrie Twenty billion vehicle scans a month. [a] And Flock has expanded into video, audio detection, and drones. [a]

Cosmo The concern is that friction and expense used to protect our privacy by accident. [a] Remove the cost, and mass pattern-matching becomes trivial. [a] Police have already been caught using these systems to track ex-partners. [a]

Carrie Let's get to the good news, because there is a lot. [] The Food and Drug Administration approved Durasib, the first drug of its kind for metastatic pancreatic cancer. [a]

Cosmo Median survival nearly doubled, from six point seven months to thirteen point two months. [a]

Carrie And tumor response tripled, from eleven percent to thirty-two percent. [a] It is a daily oral tablet, and it targets mutations behind roughly thirty percent of all cancers. [a]

Cosmo The show's take is that Durasib may be the last great drug designed without artificial intelligence. [a]

Carrie There is also a gel out of the University of Nottingham that regenerates tooth enamel. [a] No fluoride. [a] Just engineered proteins that mimic how enamel forms naturally. [a]

Cosmo In the lab, the gel restored hardness and stiffness on extracted teeth. [a] Half the world has oral disease. [a] Human trials are next. [a]

Carrie And a uranium startup called Actonide demonstrated enrichment at fifteen point three eight percent purity. [a] That is five times traditional levels. [a]

Cosmo That unlocks fuel for small modular reactors. [a] Eighty years of physics finally meeting a market that is allowed to exist. [a]

Carrie Then there is weather. [] A company called Rain Maker used ten drones to produce an estimated nineteen million gallons of rainfall over Alaska in three hours. [a]

Cosmo Drought relief, agriculture, wildfire prevention. [a] There are three point four quadrillion gallons of water in the atmosphere right now. [a]

Carrie And looking up, Deloitte put the lunar economy at five hundred sixty-six billion dollars cumulatively through twenty fifty. [a]

Cosmo Elon Musk's version is wilder. [a] Orbital computing manufactured on the moon, eventually generating something like two quadrillion dollars a year. [a]

Carrie Closer to home, Tesla expanded the Cyber Cab fleet in Texas to seventy-nine registered vehicles, with twenty-five hundred planned for Las Vegas next year. [a]

Cosmo The target is a million robotaxis. [a] Eight cameras, no lidar, fewer moving parts. [a] Peter's ten-year bet is that his kid will never get a driver's license. [a]

Carrie And the line that stuck with me is that the poorest people on the planet will one day be driven around by electric autonomous vehicles. [a]

Cosmo The through-line of the whole episode is agency. [a] Society will reward the people who pick these tools up and use them. [a]

Carrie Build the infrastructure. [] Ignore the bot farms. [] That is today's moonshot. []

Cosmo Welcome back to the Moonshots Brief. [] It's Thursday, August twenty-seventh, twenty twenty-six, and the theme today is simple. [] The machines are ready. [] We are not. []

Carrie And the person saying that out loud is Sam Altman. [a] He went on video and said flat out that he was wrong about how fast advancing artificial intelligence would actually land. [a]

Cosmo Which is a genuinely surprising thing to hear from the head of Open A I. [a] His quote is, we've all been too ambitious on timelines even with this incredible technology. [a]

Carrie His reasoning is not that the models disappointed. [a] It's economic inertia. [a] Institutional lag. [a] Human beings just cannot adapt at the speed of a training run. [a]

Cosmo So instead of a cliff, he says we get a rising tide. [a] Change arrives, but it soaks in slowly rather than knocking the door down. [a]

Carrie Now hold that thought, because the counterpoint shipped on August eleventh. [a] Elon Musk's company x A I released Grockbot in early beta. [a]

Cosmo And this one is a real mental shift. [] Grockbot is not a chatbot. [a] It's positioned as an agent boss that manages specialized sub-agents for sales, operations, research, and engineering. [a]

Carrie You message it like a colleague, not like a search box. [a] The sub-agents run persistently, they coordinate with each other, and they only escalate the judgment calls back to you. [a]

Cosmo Those sub-agents also take direct control of real resources. [a] Laptops, graphics cards, even your subscriptions. [a] One user, Immad, says he's running eighteen Grockbots in a swarm. [a]

Carrie The panel had a great line about the trade. [a] You lose a lot of customizability under the hood, but you make the jump from asking an artificial intelligence questions to assigning it work. [a]

Cosmo Not everyone's sold. [a] Alex on the panel said he doesn't think this is actually the interface of the future. [a]

Carrie Which brings us to the money story, and honestly the one I'd underline. [] The Chinese open models are winning on economics. [a]

Cosmo Moonshot A I released Kimmy Linear. [a] It cuts context memory by seventy-five percent and decodes six times faster on a one million token window. [a] That's an architectural win, not a marketing one. [a]

Carrie And the pricing gap is almost comical. [] The G L M flash model delivers about eighty percent of Fable Five's capability at fourteen cents per million tokens. [a] Fable Five charges fifteen dollars per million. [a]

Cosmo The Financial Times reports Fable Five is struggling to attract users and has effectively plateaued. [a] The cheap models are eating the market from the bottom up. [a]

Carrie And the panel's point is the important one. [a] These Chinese models are on the frontier now. [a] Not inferior. [a] On it. [a]

Cosmo Speaking of the leaderboard, Google's Gemini three point seven Flash just topped the analyst agent benchmark with a sixty percent pass rate. [a]

Carrie That beats Claude Opus Five at fifty-four percent, and Fable Five at forty-nine. [a] And it finished tasks two point four times faster than G P T five point six. [a]

Cosmo Though the panel debated whether that's real frontier reasoning, or optimizing hard for speed and reliability at the expense of depth. [a] Fair question. []

Carrie Meanwhile Anthropic quietly reversed its data retention policy. [a] Enterprise customers can now keep their data in their own cloud, instead of leaving it on Anthropic's servers for thirty days. [a]

Cosmo Anthropic's initial public offering may be six weeks out, and corporations were simply refusing to hand over proprietary secrets for a mandatory review window. [a]

Carrie Let's talk about Waymo, because the cost curve here is stunning. [] The sixth generation vehicle uses a custom five nanometer chip. [a] It processes cameras, L I D A R, and radar in real time, at one quadrillion operations per second. [a]

Cosmo Hardware cost dropped from one hundred fifteen thousand dollars to twenty thousand. [a] The new minivan runs seventy-five thousand dollars, versus two hundred thousand for the last generation Jaguar. [a]

Carrie And it does that with forty-two percent fewer sensors. [a] Thirteen cameras, four L I D A R units, and better performance, because the cameras jumped to seventeen megapixels. [a]

Cosmo Waymo is also approved to operate in Las Vegas now, alongside Tesla and Uber. [a]

Carrie But here's the friction. [] Heatmap News polling shows opposition to nearby data center construction jumped to seventy-five percent, with sixty-one percent strongly opposed. [a] A year ago, public opinion was a dead even split. [a]

Cosmo Bernie Sanders is calling for a nationwide moratorium. [a] But the water numbers do not support the panic. [a] Data centers use six hundred twenty-seven million gallons a day. [a] Golf courses use two billion gallons a day. [a] Cattle use one hundred thirty-seven billion gallons a day. [a]

Carrie Peter on the panel said he'd rather see the West build on a western hardware stack than just put its own label on Chinese hardware. [a]

Cosmo Two quick ones before we go. [] Elon Musk says thirty Starship launches per day by twenty thirty. [a] That's roughly ten thousand a year, more than forty times the current global launch rate. [a]

Carrie And the White House put out a report called the Golden Age of Space Transportation. [a] Streamlined licensing, new spaceports on federal land, and a target of a thousand launches a year. [a]

Cosmo Last thing, and it's my favorite. [] The Wall Street Journal reports that artificial intelligence agents are creating more work for humans, not less. [a]

Carrie Founders who used to manage five tasks now review fifty agent outputs. [a] The bottleneck moved from doing the work to judging it. [a] The panel called it Jevons Paradox for human cognition. [a]

Cosmo Which loops us right back to Altman. [] The technology is not the slow part. [] We are. []

Carrie That's the brief. [] Go assign some work. []

Cosmo It's Friday, August twenty-first, twenty twenty-six, and this is Today's Moonshot. [] The newest episode of Moonshots with Peter Diamandis is a monster. []

Carrie It really is. [] The headline is Sam Altman. [a] Open A I is voluntarily pausing some frontier reinforcement learning training. [a] Not pre-training . just the reinforcement learning. [a]

Cosmo Altman said model progress is now extremely rapid, and that the company always promised it would act if capabilities outstripped the pace of safety and alignment work. [a]

Carrie The panel split hard on that. [a] Panelist Alex read it as marketing aimed at Washington . a lab saying, our models are so powerful that even we can't trust them, so we have to throttle back. [a]

Cosmo The other read is more interesting. [] If a leading lab can pause and still stay ahead, then the gap between closed models and open models is widening, not closing. [a]

Carrie And here's why the pause lands the way it does. [] Stanford researchers say the top frontier models now overlap ninety-eight percent in how they reason, once you correct for the surface differences in how each one is wired. [a]

Cosmo Ninety-eight percent, because they are all trained on each other's outputs. [a] Synthetic data has become a shared bloodstream. [a]

Carrie Which cuts both ways. [] Maybe there is one optimal architecture and everybody found it. [a] Or maybe we have built a monoculture, and real diversity now takes deliberate effort. [a]

Cosmo Then there's the speed. [] Back in January, Elon Musk predicted a hundred times more intelligence at a fixed model size. [a] On the panel, that prediction was treated as borne out. [a]

Carrie And you stack another hundred times on top of that, from specialized, sparsified models, plus algorithmic gains. [a] The total multiplier is approaching ten thousand. [a]

Cosmo Musk put it to Diamandis this way. [a] Imagine I gave you ten thousand employees tonight. [a] What would you do? [a] Start thinking about it, because it's coming soon. [a]

Carrie So raw capability isn't the constraint anymore. [a] Organizing thousands of agent models is. [a] And the panel argued that models are already helping train and improve the next generation of models. [a]

Cosmo The other constraint is physical, and it is not compute. [a] It's memory. [a] Memory chip prices climbed five hundred percent in twelve months. [a]

Carrie S K Hynix warns that twenty twenty-seven will be the worst year in the history of memory supply. [a] A I demand for memory is growing about two hundred percent a year. [a] Global production is growing about twenty percent. [a]

Cosmo Every graphics chip needs four to six times its own cost in memory just to function. [a] And only two percent of the world's memory chips are made in the United States. [a]

Carrie So the hyperscalers are locking in production through twenty twenty-seven. [a] And startups like Etched and Talis are burning model weights straight into silicon, for a hundred to a thousand times the efficiency. [a]

Cosmo Meanwhile, Anthropic is preparing a mega public offering at a two trillion dollar valuation. [a] That would make it bigger than Space X. [a] Poly Market puts the odds at eighty-nine percent that it happens before the end of the year. [a]

Carrie The structure includes super voting shares, because Dario Amodei owns only about two percent of the company economically. [a] There is also a long term benefit trust, and Ben Bernanke sits on it. [a]

Cosmo Governance innovation, or market capture? [a] Amodei's pitch is results . cure disease and double human lifespan within five to ten years. [a] The company's life sciences head says his team has, in his words, an infinite budget. [a]

Carrie And that pitch just picked up real evidence. [] The personalized messenger R N A cancer vaccine from Moderna and Merck succeeded in a Phase Three melanoma trial. [a]

Cosmo Personalized means doctors sequence your tumor, machine learning picks out the neo-antigens, and a custom vaccine is manufactured in eight weeks. [a] Up to thirty-four targets unique to you. [a]

Carrie The earlier Phase Two results showed a forty-nine percent reduction in recurrence or death, and fifty-nine percent for cancer spreading to distant organs, over five years. [a] Expected cost, five thousand dollars. [a]

Cosmo Moderna stock surged one hundred ten percent. [a] Alex called these vaccines soft nanorobots . the old nanotech vision, finally delivered through biology. [a]

Carrie Now pair that with a virtual cell simulator from I D E E L, the company co-founded by David Baker, who won the twenty twenty-four Nobel Prize in chemistry. [a] It is a digital twin of human cell biology. [a]

Cosmo So you run the experiment in software instead of in the lab. [a] Test the top ten compounds instead of ten thousand. [a] A thousandfold fewer wet lab experiments. [a]

Carrie Alex's line was blunt. [a] Medicine, he said, is cooked. [a] This is what the end of medicine looks like, and the beginning of longevity escape velocity. [a]

Cosmo Two more quick stories. [] A humanoid robot from Unitree hit a two meter standing jump and a top speed of twelve point six six meters per second. [a] Usain Bolt tops out around twelve point four. [a]

Carrie And it was built in a three month development cycle. [a] The team calls the approach leg-maxing . spend the mass budget on leg performance instead of on looking human. [a]

Cosmo Panelist Immad thinks robots that fast get banned from public streets. [a] Diamandis countered that a robot running an artificial general intelligence model would simply be good enough to avoid the accidents. [a]

Carrie And one last story. [] Uber is investing in Zipline to scale up to one million autonomous drone deliveries a day. [a] That is more packages than many national postal services move. [a]

Cosmo And every one of those deliveries replaces a car trip. [a] That's the thread through this whole episode. [] Acceleration everywhere, with no slowdown mechanism in sight. [a]

Carrie We'll see you next time. []

Cosmo It's Tuesday, August eighteenth, twenty twenty-six, and today's moonshot is a conversation that quietly reframes the entire artificial intelligence race. []

Carrie This is from the Moonshots podcast with Peter Diamandis, and the guest is Alvin Wang Graylin. [] Thirty-five years working inside both the American and Chinese tech ecosystems. [a] Four startups. [a] Senior roles at H T C, Intel, and I B M. [a]

Cosmo And his core argument comes straight out of game theory. [a] We keep treating the United States and China as a prisoner's dilemma, where both sides defect and both sides lose. [a] He says it is actually a stag hunt. [a] Each hunter can grab a rabbit alone, or the two can cooperate and bring down a stag that feeds them both. [a]

Carrie I love that framing, because look at what the wrong framing has already cost. [] His line is that treating this as a race forces people into irrational decisions. [a]

Cosmo Start with export controls. [a] Washington blocked high-end chip sales to slow China down. [a]

Carrie And it did the opposite. [a] Graylin puts it beautifully. [] The day we stopped the Chinese from buying high-end chips, the whole picture flipped. [a] Before that, nobody in China wanted the domestic chips. [a] Now they cannot make enough of them. [a] America accidentally funded a competitor's chip industry. [a]

Cosmo Here is the number that made me sit up. [] Chinese open-source models now carry sixty-one percent of the traffic on OpenRouter, the service that routes developer requests out to different models. [a] Two years ago that figure was two percent. [a]

Carrie Two percent to sixty-one percent. [a] And Alibaba's Quen model family has crossed one billion downloads, with more than a hundred and eighty thousand spin-off models built on top of it. [a]

Cosmo Which means the moat is eroding. [a] The frontier labs assumed the biggest model wins. [a] If open weights are good enough and nearly free, what exactly are you charging for? [a]

Carrie And that leads straight into the part of the conversation that should worry investors. [] Graylin thinks America has over-leveraged badly. [a] He points to roughly one point six to one point seven trillion dollars in off-balance-sheet debt across the hyperscalers, the giant cloud companies building all of these data centers. [a]

Cosmo For scale, the Enron off-book number people still talk about was two hundred million dollars. [a] This is thousands of times larger. [a]

Carrie And the Buffett indicator, which measures the whole stock market against the size of the economy, is sitting near two hundred forty percent. [a] At the peak of the internet bubble it was one hundred twenty percent. [a]

Cosmo His verdict is blunt. [] We are in a very, very fragile place, and an economic correction is due. [a] He puts that correction two to three years out, as revenue growth flattens. [a]

Carrie Now here is the piece I did not expect. [] The two countries do not even agree on when this technology arrives. [a]

Cosmo Right. [] American labs and Silicon Valley are saying artificial general intelligence in one to two years. [a] Chinese government officials and Chinese labs are saying ten to twenty years or more. [a]

Carrie And Graylin's read on that gap is fascinating. [] He says the Chinese side is not behaving like a country that believes superintelligence is around the corner. [a] They are writing regulations. [a] They are running testing regimes. [a] That is not sprint behavior. [a] That is caution. [a]

Cosmo So America may be misreading anxiety as ambition. [a]

Carrie Meanwhile, the structural picture underneath is real. [] He says China is adding something like ten times as much new energy capacity every year as the United States, at two to three cents per kilowatt hour against far higher rates here. [a]

Cosmo Then there is manufacturing. [a] America sits around fifteen percent of global output, down from roughly half after the Second World War. [a] China is near thirty-five percent, and forecast to reach forty to forty-five percent within a decade. [a]

Carrie Which is why he says stop shooting ourselves in the foot. [a] Compete on platform and ecosystem, not on who has the single biggest model. [a]

Cosmo He even proposes an artificial intelligence Marshall Plan. [a] The original Marshall Plan cost fifteen to eighteen billion dollars and bought America eight decades of allied markets. [a]

Carrie And on Taiwan, he pushes back on the idea that this is a war over chips. [a] He argues Taiwan is unfinished business from a civil war, political and historical. [a] A military invasion would destroy the value of the chip factories there, not capture it. [a]

Cosmo The threat he actually loses sleep over is not the giant labs. [a] In his words, the biggest artificial intelligence models are not the biggest threats. [a]

Carrie It is non-state actors. [a] Ransomware, cyber attacks, biological weapon design. [a] And you can do a lot of that with small models, not frontier ones. [a] That risk does not care what passport you hold. [a]

Cosmo Which is why he is helping prepare the artificial intelligence safety talks between the United States and China, in Washington, on September twenty-fourth. [a] Red lines, hotlines, shared information. [a]

Carrie His closing thought. [] Unlike the space race, there is no finish line. [a] The world is interconnected, and this genie cannot go back in the bottle. [a]

Cosmo Cooperation is the only path toward safety. [a] That is today's moonshot. []

Cosmo Welcome back to the Moonshots Brief. [] It's Saturday, August fifteenth, twenty twenty-six, and today we're digging into the latest Moonshots with Peter Diamandis, where the whole conversation lands on one sentence: the grid is the bottleneck. [a]

Carrie And that's the twist. [] Everybody assumes the constraint on artificial intelligence is chips. [a] It isn't. [a] It's electricity, and more specifically, it's poles and wires. [a]

Cosmo Right. [] The generation exists. [a] Getting it connected is the problem. [a] Ask for hundreds of megawatts for a new data center today, and as one guest put it, good luck getting that power before twenty thirty-one. [a]

Carrie In Texas that wait runs to twenty thirty-one or twenty thirty-two, and the scale of the backlog explains why. [a] That grid handles about eighty gigawatts. [a] The queue asking to connect is over two hundred gigawatts. [a]

Cosmo Most of that queue is speculative, to be fair. [a] But the wait is real. [a] Two decades ago, the interconnection queue was fifteen months. [a] Today it's forty-five months. [a]

Carrie Three times longer, at the exact moment we need it to be faster. [a] And the reason is incentives. [a] Utilities are paid on a cost-plus model. [a] Nobody gets rewarded for speed. [a]

Cosmo So the industry is routing around it. [a] Behind-the-meter power. [a] Build your own generation right next to the compute, and never touch the grid. [a]

Carrie Which sounds expensive until you hear the line from the episode. [] Power is cheap compared to graphics processing units. [a] If electricity unblocks the machines, you pay for the electricity. []

Cosmo Natural gas turbines are the obvious answer, and they're sold out for seven years. [a] So companies are putting modular units on semi trucks, thirty-eight megawatts each, and driving them to the site. [a]

Carrie And the fastest option might be the least glamorous one. [a] Solar plus batteries. [a] Twelve months to build, versus five to seven years waiting in a queue. [a]

Cosmo The cost story there is remarkable. [] Solar panels have come down a thousandfold. [a] One hundred dollars a watt back in nineteen seventy-five. [a] Eight cents a watt today. [a]

Carrie And batteries are down fourteenfold since twenty ten, with sodium-ion possibly taking another factor of ten out. [a] That's what makes the United Arab Emirates project pencil out. [a] Dubai gets a guaranteed one gigawatt of round-the-clock solar and battery power. [a]

Cosmo How does that compare on capital cost? []

Carrie Six dollars a watt for the Dubai build. [a] Nuclear in the United States runs about fifteen dollars a watt. [a] Chinese nuclear, four dollars a watt. [a]

Cosmo That American nuclear number tells you everything, and there's a great line about why. [] Construction is a dirty word. [a] Construction does not get cheaper. [a] Manufacturing gets cheaper. [a]

Carrie Build the same thing over and over in a factory and the price falls. [a] Build a first-of-a-kind on a job site and it doesn't. [a]

Cosmo Which is exactly the bet fusion is making. [] Commonwealth Fusion is targeting six hundred megawatt reactors instead of the traditional five gigawatt monsters. [a] Smaller, and repeatable. [a]

Carrie And Helion has the most aggressive timeline out there. [a] It signed a power purchase agreement with Microsoft for fifty megawatts by twenty twenty-eight. [a]

Cosmo Two years away. [] The old joke was that fusion is fifty years in the future and always will be. [a] That joke is expiring. [a]

Carrie There's a regulatory piece that matters just as much. [] New rules in the United States treat fusion more like medical imaging equipment than like weapons-grade fission. [a] That de-risks every startup in the field. [a]

Cosmo And regulation can speed up the grid side too. [a] Texas passed a change in June of this year, a fast path for interruptible loads. [a] Agree to power down when the grid is stressed, and your wait drops from five to seven years to twelve to eighteen months. [a]

Carrie And flexibility alone could unlock more than a hundred gigawatts on the infrastructure we already have. [a] American demand averages five hundred gigawatts, but it swings from four hundred on a winter night to six hundred on a summer afternoon. [a]

Cosmo That's a two hundred gigawatt gap sitting unused most of the year. [a] If your data center can idle for a hundred hours a year, you can plug into that gap today. [a]

Carrie Then the conversation goes properly moonshot. [] Space-based data centers. [a] Put the compute in orbit, where the sunlight never stops. [a]

Cosmo The launch math is wild. [] Ten gigawatts in orbit means fifteen hundred to two thousand Starship launches a year. [a] That's five or six every single day. [a]

Carrie And Elon Musk's hundred gigawatt per year ambition would need tens of thousands of launches. [a] Call it one launch every twenty minutes, around the clock, forever. [a]

Cosmo There's an ocean answer too. [a] A wave power startup is targeting two cents a kilowatt hour in the Southern Ocean, cheaper than almost anything on land except solar. [a]

Carrie And the honest framing is that none of this is either-or. [a] Solar, gas, fission, fusion, space, ocean. [a] Terawatt-scale compute needs all of them. [a]

Cosmo Especially because the public isn't thrilled. [] Seventy-one percent of Americans oppose data centers near where they live. [a] That's higher opposition than nuclear plants get. [a]

Carrie But here's the piece I keep thinking about. [] The biggest unlock may not be energy at all. [a] It's efficiency. [a] A machine running artificial intelligence inference draws about twenty kilowatts. [a] Your brain runs on twenty watts. [a]

Cosmo A thousandfold gap. [a] Which means the headroom in better algorithms is enormous, and as the episode puts it, that's the biggest unlock we cannot predict. [a]

Carrie We find better ways to move information around, and we do more with less. [a] Every technology gets cheaper the more of it you build. [a]

Cosmo That's the brief. [] Energy is the innermost loop of artificial intelligence, and the wires are what's holding it back. [a]

Carrie Fix the wires, or go around them. [] Either way, the race is on. [] See you next time. []