For a full year, Philip Johnston’s team was trapped by a single contradiction.
They wanted to do something no human had ever done: build a spacecraft and send it toward Alpha Centauri, the nearest star system to our own, 4.4 light-years away. Not a concept in a paper, not a slide in a deck, but a real object that would actually be built and actually launched. The budget was only about $15 million. The payload had to be at least one kilogram, a small box roughly ten centimeters on each side, carrying scientific instruments, artistic works, and a copy of the Golden Record, the gold-plated disc that flew out of the solar system aboard Voyager 1 in 1977, etched with the sounds and images of Earth as humanity’s first letter to the cosmos.
The problem was the route. Alpha Centauri sits about 25 trillion miles away. Voyager 1, launched in 1977, has covered less than one percent of that distance in all the decades since. Johnston’s team tried route after route and kept slamming into the same knot: to give the spacecraft enough energy to travel that far, you need more fuel or bigger solar panels, but that makes the craft too heavy for a $15 million budget to lift. Make it lighter, and you don’t have enough energy. Every trajectory they could imagine crashed into the same wall.
Then a podcast changed everything.
An offhand remark on a podcast
Johnston is co-founder and president of the Fermi Explorer Mission, a Seattle-area nonprofit (a US 501c3) named after the physicist Enrico Fermi. In 1950, over a casual lunch, Fermi posed a question that still has no settled answer: the galaxy holds hundreds of billions of stars, most far older than our sun, and even slow travel between stars would let a civilization spread across the whole galaxy in a few million years, so why do we see no sign of anyone out there? That is the famous Fermi paradox. In a sense, what Johnston wants to do is take humanity’s first step toward answering it: we go first.
But the routing problem had stalled for a year, and the team was close to conceding that the mission was impossible within their budget.
The turn came while Johnston was a guest on a podcast hosted by physicist Alex Wissner-Gross. When the difficulty came up, Wissner-Gross suggested running the problem through an open-source AI system his lab, PSI, had been building, just to see if it could find a way through.
The system is called Get Physics Done. It doesn’t simply ask a large language model to chat out an answer. Instead it breaks a physics research question into a series of smaller tasks, decides which simulations to run and with what parameters and how to analyze the results, then calls on models such as Anthropic’s Claude or OpenAI’s GPT to do the work. It behaves less like an all-knowing oracle and more like an automated lab assistant that writes its own research plan.
A week later, the AI handed back an answer.
A path nobody had considered
According to a MIT Technology Review report, the trajectory Get Physics Done proposed was unlike anything the Fermi team had considered.
Its core idea was counterintuitive enough to make you frown. Rather than pushing hard to accelerate outward from the start, it first has the spacecraft slow down, dropping its orbit closer to the sun than Mercury’s. Then, each time the craft reaches the point nearest the sun, called the perihelion in orbital mechanics, it fires its engine. Because it is so close to the sun, the solar panels receive roughly four times the light they would at Earth’s orbit, and a burst of thrust delivered at high speed near that point buys far more energy than the same burst anywhere else. This is a known principle in orbital dynamics, but the AI pushed it to an extreme: the engine works only along that short arc near perihelion, and stays off the rest of the time.
What does that buy you? The solar panels don’t have to power the craft across the whole journey, so they can be small; the engine runs only briefly, so fuel needs collapse; and the spacecraft can therefore be light. The vicious circle of energy versus weight that had trapped the team for a year was sidestepped from an angle no one had seen.
PSI’s research paper, which has not been peer-reviewed, names this the “perihelion pump maneuver.” PSI’s CEO, Matt Pines, told MIT Technology Review that the AI came up with an entirely different mission profile, one that was creative and one the Fermi team had never considered, and that this was the more surprising part.
Three days, a billion tokens, one astrophysicist
The trajectory didn’t spring from a flash of machine insight. Pines said the system spent most of its time working autonomously, running for about three days and burning through roughly a billion tokens. Throughout that stretch, an astrophysicist on PSI’s staff guided it, not by doing the research in its place, but by setting task constraints, asking for a cost analysis and clearer charts, and checking the output for errors.
In other words, this was a collaboration between a human researcher and an AI system. The AI searched, combined, and tried and discarded options across an enormous parameter space at a speed and breadth beyond any human. The human judged direction, set boundaries, and corrected drift. The trajectory that emerged was neither pure human ingenuity nor pure machine computation, but the product of the two working together.
But what Pines said next may be worth remembering more than the trajectory itself.
He said the model still lacks a human researcher’s judgment and taste. It has no reliable sense of which problems are interesting or which directions are worth pursuing, and it often gets stuck chasing dead ends. “I don’t think we’ve yet figured out how these models can internally represent something like that,” he said.
That sentence pins down the real capability boundary of AI in research today. Get Physics Done found the perihelion pump maneuver not because it “understood” the deeper elegance of orbital mechanics, but because it could recombine known physical principles and engineering constraints at a speed no human can match, then sift viable options from a vast space of possibilities. It is an extraordinarily powerful combinatorial search engine, but it doesn’t know what it is looking for, doesn’t know what counts as “elegant,” and can’t tell whether a solution is a dull technical footnote or an insight that shifts a paradigm.
Which is exactly why the astrophysicist’s role was indispensable. Without a human to define which directions are “interesting” and which boundaries are “worth chasing,” the AI might have found nothing across its billion tokens, or produced a heap of solutions that were technically correct and utterly uninspiring.
First to leave, last to arrive
On September 1, 2026, the Fermi Explorer Mission formally announced its plan to launch humanity’s first spacecraft toward another star by the end of 2029.
Johnston was blunt about it: “We’re dead set on something actually launching.” The subtext is a quiet reply to another ambitious interstellar plan from a decade earlier. In 2016, the Russian-born billionaire Yuri Milner announced $100 million for Breakthrough Starshot, which aimed to push tiny light-sail probes toward Alpha Centauri with powerful lasers. Ten years on, nothing has launched. Johnston doesn’t want a repeat: “We didn’t want to do another Breakthrough Starshot.”
Fermi Explorer’s budget is a fifteenth of what Starshot pledged, and its strategy is entirely different. It does not aim to arrive within a human lifetime. As Johnston put it plainly, “we are not constraining ourselves to doing it in a human lifetime.” If all goes well, this small spacecraft will need roughly 77,500 to 80,000 years to reach Alpha Centauri.
Seventy-odd thousand years. That number would clear the room at any commercial fundraising pitch. But Johnston said something more interesting still: “We’ll be the first to leave, and the last to arrive.”
His logic runs like this. Over the coming decades and centuries, propulsion technology will keep improving. If someone a thousand years from now builds an engine even 20 percent faster, a craft launched then would beat the Fermi probe to Alpha Centauri by more than ten thousand years. Johnston is entirely at peace with this; he is, in fact, “pretty confident” Fermi won’t be the first to arrive.
So why launch at all?
Because the point isn’t arriving, it’s leaving. Humanity needs to actually take the first step rather than stay forever in papers and plans. A first step makes a second and a third possible. The significance of Fermi Explorer lies not in whether it can reach Alpha Centauri, but in proving that interstellar travel is no longer science fiction, that it can actually begin as an engineering project on a $15 million budget with existing technology in a foreseeable timeframe.
To make sure this wasn’t just talk, Fermi Explorer assembled a serious advisory bench, including former Blue Origin president Rob Meyerson, former SpaceX propulsion head Jeff Thornburg, and the CEO of AstroForge, among other industry figures. The launch may ride along on a SpaceX rideshare mission.
What AI research really looks like
Back to the AI.
On the same day, PSI (Physical Superintelligence), headquartered in Cambridge, Massachusetts, formally launched with $58 million in funding led by Breakthrough Energy, the climate investment group founded by Bill Gates. The Fermi Explorer trajectory is their first public showcase, and the promotional impact is maximal.
But strip away the marketing, and what makes this case so interesting isn’t “how powerful AI is.” It’s how clearly it maps the contours of AI in research: where it is strong, where it is weak, and what role humans play in the partnership.
AI’s strength is combinatorial creativity. Every physical principle behind the perihelion pump maneuver, from perihelion gravity assist to the way solar panel efficiency changes with distance to thrust-timing optimization, is familiar to researchers in orbital mechanics. But assembling those known elements into a complete trajectory that satisfies every engineering constraint requires searching an enormous parameter space, which is both tedious and inefficient for humans. The Fermi team spent a year and found nothing. The AI found it in three days.
AI’s weakness is research taste. It doesn’t know which questions are worth asking, which directions are worth chasing, or which dead ends to abandon early. The “judgment and taste” Pines described is, at bottom, an intuition you can only develop after long immersion in a discipline, a sense of “the right feeling,” a sense of what result will make peers sit up. No large language model can reliably display that today.
So the most effective mode is exactly the one this case shows: humans define the problem, set the constraints, and judge direction; the AI explores, combines, and generates candidate solutions at high speed within the space humans have framed; and humans then pick out the valuable one. Humans supply taste, the AI supplies compute.
It calls to mind an image. The AI is like a tireless mountain guide with limitless stamina, able to try a thousand routes up the same mountain at once and remember precisely where each one hits a cliff. But it doesn’t know why the mountain is worth climbing, or what the view from the summit means to a climber. Choosing which mountain to climb, and deciding where to go once you reach the top, remain questions only humans can answer.
A letter into the void
Fermi Explorer’s official FAQ page carries a line describing this as the first time generative AI has independently devised a novel trajectory that will actually be flown on a real mission. If the spacecraft really lifts off before the end of 2029, that line stops being a marketing phrase and becomes a historical marker: the first time humanity set out for another star along a route designed by an AI.
No one can predict what happens seventy or eighty thousand years from now. Johnston himself admits faster spacecraft will most likely overtake it. But so what? When Voyager 1 flew out carrying its Golden Record in 1977, no one expected it to reach any star either. Its meaning was never in the arrival, but in proving that humanity had both the ability and the will to send something beyond the solar system.
Fermi Explorer’s payload will carry a copy of that Golden Record too. Half a century separates Voyager 1 from Fermi Explorer. The difference is that this time, the one drawing the route for humanity is no longer only human.
That AI system, which spent three days and burned a billion tokens, will most likely never “understand” what it did. It doesn’t know what Alpha Centauri is, doesn’t know why the Fermi paradox keeps humans up at night, doesn’t know what the sounds carved into that Golden Record mean. But it found a path, one that humans had failed to find in a year.
That, more or less, is the truest relationship between AI and humans at this stage: it doesn’t understand why we want to leave, but it can help us work out how to go. As for why we should go, Fermi already asked that for us over a lunch seventy-six years ago.
Related reading
- Which AI Email Assistant Actually Saves You Time? Superhuman AI vs Spark AI vs Notion Mail vs Gmail AI in 2026
- Which AI Video Editor Saves You the Most Time in 2026: Descript vs CapCut vs Runway vs OpusClip
- Amazon Is Shutting Down Mechanical Turk: The Machine That Hid Humans Inside Finally Doesnt Need Them
- The Next AI Platform Divide: It’s Not About Bigger Models, It’s About Who Can Raise a Team of Agents
- The Real Barriers Blocking Enterprise AI Agent Adoption in 2026
Stay updated with our latest AI insights
