When engineers at NASA’s Jet Propulsion Laboratory could not crack a trajectory that might ship a probe to a different star utilizing present expertise, the group behind a brand new mission referred to as Fermi Explorer handed the issue to an AI. It solved it in every week.
That is the story Philip Johnston and Matt Pines advised on the Peter Diamandis podcast. The mission — humanity’s first deliberate try to achieve Alpha Centauri — will launch by 2029 with a price range of $15 million, roughly the value of a modest Collection A spherical. The spacecraft weighs simply 100 kilograms, carries a 1-kilogram payload, and can take about 80,000 years to get inside 99% of the space to the closest star system. That timeframe is deliberate: by accepting an 80,000-year journey, the group avoids the billion-dollar propulsion analysis applications that killed earlier efforts like Breakthrough Starshot.
However the actual revelation is not the mission itself. It is the pace at which the design work occurred. “If we might gone to a bunch of astrophysics PhDs and given them a billion {dollars} in 5 years, I am certain they’d have give you this trajectory,” Johnston stated. “It is extra that this was completed in every week.”
AI Has Turn into the Fee-Limiter Remover
Johnston’s group spent six months making an attempt to design a workable trajectory utilizing photo voltaic electrical propulsion, together with consultations with two JPL engineers who could not remedy it. After Alexander Wissner-Gross related him to Bodily Tremendous Intelligence (PSI), an AI-first physics lab co-founded by Matt Pines, the system produced a viable trajectory in seven days utilizing roughly 10 billion tokens of computation.
The important thing innovation was a maneuver PSI named the “perihelion pump.” As a substitute of utilizing planetary flybys to decrease the spacecraft’s orbit towards the solar, the craft fires retrograde burns at aphelion for about 5 years to gradual itself and pull inward, then fires thrusters at perihelion the place the Oberth impact — which delivers extra vitality for a given thrust when transferring quicker — maximizes effectivity. The AI designed and ran its personal Monte Carlo simulations, layered in multivariate optimization throughout price and launch home windows, and shocked even PSI’s personal workers with the consequence.
Pines framed the broader thesis bluntly: “These types of extremely technical scientific challenges are bottlenecked by people which have been pre-trained for 22 years, post-trained in grad college… These are the speed limiters of what our scientific and technical ambition is.”
The numbers inform a placing story about what AI-first physics truly means in follow:
| Mission | Estimated Value | Design Time | Payload |
|---|---|---|---|
| Voyager 1 (1977) | ~$865 million | Years | 825 kg |
| Breakthrough Starshot (2016) | Billions (proposed) | By no means accomplished | Grams-scale |
| Fermi Explorer (2029) | Underneath $15 million | 1 week (trajectory) | 1 kg |
PSI introduced its $58 million seed spherical the identical day, led by Breakthrough Power Ventures, utilizing the Fermi mission as proof of idea for its “get physics completed” mannequin. The episode’s hosts framed the mission as a concrete demonstration that the bottleneck on grand scientific ambition is not physics or funding — it is human organizational pace.
The Battle for Reasoning Traces
The episode’s second main thread is the escalating battle between OpenAI and Elon Musk, triggered when OpenAI minimize off Cursor — the AI coding platform Musk’s SpaceX acquired for $60 billion — from entry to its GPT fashions. OpenAI’s said rationale: “We can’t be assured that SpaceX will use our expertise throughout the phrases of service based mostly on our expertise with Elon Musk’s firms violating contracts.” Musk responded by calling Sam Altman and Greg Brockman “completely untrustworthy,” and inside hours Anthropic’s Dario Amodei pledged Claude help for Cursor.
Wissner-Gross provided a principle that goes past the general public cleaning soap opera: “I feel that is all in regards to the reasoning traces. It is at all times about who will get the reasoning traces.” His argument: SpaceX’s acquisition of Cursor was primarily motivated by having access to the reasoning hint information from customers interacting with frontier fashions — the post-training useful resource that determines mannequin functionality. He famous precedents: Anthropic beforehand minimize off Windsurf after Google DeepMind acquired it, and Chinese language frontier labs have been alleged to make use of proxies to siphon reasoning traces from Anthropic.
Dave Blundin contextualized the transfer as strategic relatively than purely combative. OpenAI’s launch of GPT-Soul — which he described as “extremely nice” — instantly prior gave Altman the boldness to chop off Cursor. “The mixture of Codex on GPT-Soul operating on Amazon Bedrock is now phenomenally good for enterprises,” Blundin argued, noting that Altman is executing a Invoice Gates-style playbook: use a genuinely superior vertically built-in product relatively than counting on partnerships.
Blundin additionally recognized Anthropic’s structural weak spot: Dario Amodei is “trapped in his ethics,” and Anthropic’s phrases require transmitting all buyer IP to Anthropic for 30 days for security evaluation — a situation firms more and more hate, even when utilizing Anthropic fashions by means of AWS. That, Blundin argued, makes OpenAI’s vertically built-in stack extra engaging to enterprises regardless of the controversy.
The hosts predicted the alliance between Musk and Amodei is pragmatic however probably unstable. Anthropic desperately wants entry to Musk’s Colossus compute cluster in Tennessee, whereas Musk wants Anthropic to maintain Cursor’s put in base joyful after OpenAI’s withdrawal. Blundin speculated the connection might finish with Musk buying Anthropic outright at a $1–2 trillion valuation — a state of affairs that relies on the end result of Amodei’s proposed super-voting management construction, which stays in limbo.
The Chip That Designs Itself
The third growth is probably essentially the most consequential for the semiconductor trade. Architect Labs, a Palo Alto startup based by Ibrahim Hussein and Adita Sabidi, introduced Redwood — described because the world’s first absolutely AI-designed chip. Two people wrote a high-level specification; the AI autonomously generated the efficiency mannequin, RTL design, UVM verification, firmware, drivers, and customized compute kernels with zero human intervention. The chip took two weeks to design, had zero bugs on first silicon, and delivers 3.4x the efficiency per watt of Nvidia’s Jetson.
Wissner-Gross, who disclosed he’s an adviser to Architect Labs and that it’s a Hyperlink Ventures portfolio firm, framed the importance straight: “I feel we all know how this recreation ends. It ends with recursive self-improvement on the chip layer.” The argument is that as Moore’s legislation ends, the abstraction boundaries between software program, working techniques, chip design, and underlying physics should collapse. Architect Labs’ strategy — burning AI fashions straight into silicon in “quick style” model — represents the post-Moore’s-law aggressive technique.
Blundin assessed the risk to Nvidia as extreme: “Is that this an unimaginable risk to Nvidia? Yeah, completely.” He famous that Nvidia’s Jensen Huang is producing roughly $1 billion per day in revenue, giving him monumental acquisition firepower — “if he can stretch the lifespan of Nvidia by every week that is $7 billion.” The query is whether or not Nvidia acquires Architect Labs and buries it, or whether or not AMD or one other competitor acquires it to speed up their catch-up. Nvidia had introduced its personal inner chip-design basis mannequin (ChipNemo) two to a few years in the past however by no means made it usually obtainable, leaving a market hole for democratized AI chip design.
The compounding effectivity features are the deeper story. Blundin estimated “in all probability seven layers of issue of 10 in effectivity which can be all compounding get unleashed,” probably a million-fold efficiency achieve as layers of abstraction are eradicated.
Final result-Based mostly Pricing and the New AI Enterprise Mannequin
OpenAI adopted Salesforce’s lead in adopting outcome-based pricing, permitting massive clients to pay solely when the AI truly completes the job — not for tokens, compute time, or API calls. Salesforce had pioneered this with Agentforce, pricing based mostly on buyer income generated relatively than tokens consumed.
Blundin traced the lineage again to Siebel Techniques’ “Whole Success” mannequin, which wrapped CRM software program in a assured end result and raised worth factors by an element of a thousand whereas making clients happier. The structural drawback this solves: at $2 per million tokens, AI utilization is so low cost that it generates trivial income for OpenAI and does not justify the corporate’s consideration in serving to enterprises implement AI successfully. Final result-based pricing aligns incentives — OpenAI solely will get paid when a financial institution truly achieves “thrice extra customer support at half the value,” and all of the sudden the corporate cares deeply about implementation success.
Wissner-Gross predicted the market will converge on a construction analogous to digital promoting:
| Pricing Mannequin | Digital Advert Analogy | AI Equal |
|---|---|---|
| CPM | Pay per thousand impressions | Pay per GPU-hour / flop of compute |
| CPC | Ppc | Pay per token |
| CPA | Pay per conversion | Pay per accomplished end result |
This creates a worth discovery mechanism: clients throughout industries successfully bid {dollars} per activity end result, revealing which purposes generate the best income per token and directing OpenAI’s consideration accordingly. Diamandis raised the chance that sturdy optimizers are “unimaginable reward hackers” — they may discover intelligent methods to fulfill said standards with out delivering real worth. Blundin countered that almost all real-world enterprise processes are so trivially easy by AI requirements that the low-hanging fruit is successfully limitless: “Stroll into any customer support heart of any firm on the planet and say, ‘Are you utilizing AI but?’ 99.999% likelihood the reply is not any.”
The Bottleneck Has Shifted from Considering to Constructing
Elon Musk posted that consensus estimates counsel 15 gigawatts of AI compute produced in 2027 can’t be turned on in 2027 on account of infrastructure constraints — equal to 10 nuclear crops sitting idle. The bottleneck isn’t electrical energy per se however transformers, wiring, liquid cooling, chillers, and networking infrastructure. His response: SpaceX and Tesla are every constructing 100 gigawatts per yr of photo voltaic manufacturing capability, with pure gasoline as a bridge gas. The limiting think about pure gasoline turbine manufacturing is casting blades and vanes, so Musk introduced he’ll carry that casting in-house at SpaceX, accelerating turbine availability by as much as 18 months.
Peter Diamandis’s take was characteristically direct: “By no means ever ever wager in opposition to Elon. He does not like being quantity two. He does not like dependencies both.” The sample is vertical integration throughout the whole stack — from vitality innermost loop to orbital compute.
Blundin framed this because the clearest funding sign within the episode: “If I take a look at our portfolio firms that get into the information heart stack… each a type of firms is creating billionaires out of its founders.” The returns differential between information heart infrastructure firms and vertical-use-case AI firms is, in his phrases, “night time and day.” He urged listeners to search for alternative contained in the infrastructure buildout — land acquisition, transformer sourcing, architectural design to squeeze worth from present chips — relatively than asking the best way to take part from the surface.
Wissner-Gross predicted an ironic end result: “Elon is on a trajectory to grow to be the king of liquid pure gasoline.” SpaceX’s two Star bases on the Gulf Coast require pure gasoline for launches, and Musk is constructing his personal pipeline. As soon as he has fossil gas provide functionality, the logic extends to powering terrestrial information facilities.
Geoengineering Returns to the Mainstream
Musk went “absolutely existential” within the episode, arguing that sustainable vitality alone is inadequate for humanity’s survival. His reasoning: “Extraordinarily extreme extinction occasions occur each 100 million years or so, and simply switching to sustainable vitality won’t be sufficient to cease them.” His proposed resolution includes sentient, solar-powered AI satellites positioned between Earth and the solar that make steady small changes to incoming photo voltaic radiation. He estimated humanity has about 50 years to behave.
Wissner-Gross was enthusiastic: “I am an enormous fan of geoengineering… We have been doing it for lots of of years. We simply have not been doing it very effectively. We’re about to start out doing it very effectively.” He argued that with international AI climate fashions and 1000’s or tens of millions of low-Earth-orbit satellites with actuation functionality, “that is a recipe for international climate engineering.”
The governance problem is the unresolved rigidity. Wissner-Gross instructed climate might grow to be a tradable commodity — “Municipality A trades with municipality B for rain” — probably enabling Sahara regreening by means of trades with North America. He recognized the previous half-century as a interval when the West grew to become “allergic to drastic utilized bodily engineering,” relationship roughly from the late Nineteen Sixties Silent Spring period, costing 50 years of potential progress in fission reactors, geoengineering, and continued lunar exploration. Musk’s curiosity in geoengineering, he argued, is “a return to kind.”
The deeper through-line of the episode is that AI functionality is not the binding constraint on human ambition — bodily implementation is. Whether or not it is designing an interstellar trajectory in every week, designing a chip in two weeks, or planning a Mars mission, the AI can produce the plan nearly immediately. The bottleneck is constructing the photo voltaic panels, casting the turbine blades, laying the transformers, and launching the spacecraft. Considering is now low cost and plentiful; doing stays costly and gradual.
The hosts’ recurring prediction is that this hole will shut — that AI will design the bodily infrastructure it wants, from chips to vitality techniques to spacecraft, and that the ensuing compounding impact would be the defining financial and technological dynamic of the following decade. The info level that captures this second: throughout unrelated complicated AI tasks, groups are selecting a ratio of about 10 billion tokens of pondering per 100,000 tokens of ultimate output. The reasoning is occurring at a scale that dwarfs the deliverable. The query is who controls the reasoning traces — and who will get to construct the bodily world these traces describe.
