A team called the Fermi Explorer Mission spent a year failing to find a trajectory that would carry a $15 million solar-powered probe to Alpha Centauri. An AI research system named Get Physics Done, built by PSI and running on models including Anthropic's Claude and OpenAI's GPT, produced one in a week, after three days of near-autonomous work and a billion tokens. The route is described in a paper that has not been peer-reviewed. Launch is planned for 2029. The flight could take 80,000 years.
The constraint that stopped the human team for a year was energy against mass. Alpha Centauri sits roughly 25 trillion miles from Earth. Voyager 1, among the fastest objects humanity has built, has been flying since 1977 and has covered less than 1% of that distance; at its speed the crossing would take more than 70,000 years. To do better you need power, power means larger panels and more propellant, and every kilogram added demands more of both. Philip Johnston, co-founder and president of Fermi Explorer Mission, and his team could not close that loop for a craft small enough to cost $15 million.
Johnston eventually described the problem on an audio program hosted by the physicist Alex Wissner-Gross, a co-founder of PSI. Wissner-Gross suggested putting it through Get Physics Done, PSI's open-source system: it takes a research question in physics, decomposes it into smaller tasks, and decides which calculations to run.
A week later it came back with a route that surprised Johnston. According to the paper, it combined known orbital maneuvers in a configuration the Fermi team had not tried. The craft first slows down, so that its orbit passes nearer the Sun than Mercury's. At each close approach it fires its engine. The panels receive four times more light there, and a short burst of thrust at that point yields more energy than the same burst anywhere else along the route. Because the engine only runs near the Sun, the panels can stay small and the craft stays light. The loop that defeated the human team for a year is not broken so much as sidestepped.
Matt Pines, co-founder and chief executive of PSI, said the system ran the investigation almost on its own for three days and consumed a billion tokens. An astrophysicist on the PSI team steered it toward the mission requirements, asked it for a cost estimate and for more readable charts, and then checked its output for errors. Pines said the system proposed an entirely different mission profile, creative and unconsidered by the Fermi team. He also said what the system lacks: human research judgment and taste. It cannot reliably tell which problems are interesting or which approaches deserve to be pursued, so it can stall in dead ends or fail to examine alternatives. PSI does not yet know how to teach a model to hold that instinct internally.
That admission is the most useful thing in the story, and it complicates the headline result. "Almost on its own" covers a human who supplied the requirements, redirected the work, asked for the costing, and audited the answer for mistakes. The claim that an AI found the trajectory and the claim that a physicist using an AI found the trajectory are both consistent with what PSI has described, and PSI has not said which one it is. Nor has anyone outside: the paper is unreviewed, and an orbital route is exactly the kind of result where a plausible-looking derivation and a correct one are hard to tell apart from the outside.
This is not the first attempt to reach Alpha Centauri. In 2016 the billionaire technology investor Yuri Milner announced Breakthrough Starshot, which would use powerful lasers to push tiny probes to one fifth the speed of light, enough to arrive in about 20 years. Milner committed $100 million to validate the concept. A decade later nothing has launched. Johnston says his team did not want to repeat Starshot and is aiming at an actual launch; the new mission is funded by private donors and is meant to cost $15 million in total.
Set those two side by side and the trade is stark. Starshot promised arrival inside a single career. Fermi's number is 80,000 years, and its organizers have deliberately declined to bound the project by a human lifespan; the goal, as they frame it, is to find a way to reach another star at all. That is roughly a four-thousand-fold retreat in ambition bought with a roughly 85% cut in budget. It is defensible, because Starshot's twenty years remain hypothetical and Fermi's eighty thousand would at least begin. It should still be named for what it is rather than presented as progress.
The team is also close to certain its probe will not arrive first. Propulsion improves. If an engine a thousand years from now is merely 20% faster than today's, a craft launched then still beats the Fermi probe by more than 10,000 years. Which means the probe is not really an expedition. It is a package: at least one kilogram of art and scientific material, messages, and a copy of the Golden Record, the gold-plated disc of Earth's sounds and images that NASA mounted on Voyager 1 and Voyager 2 in 1977. Fermi is sending the same gesture on a longer arc, and the team knows the science will be done by whoever follows.
The mission's name points at the reason. In 1950 Enrico Fermi framed the paradox: the galaxy holds hundreds of billions of stars, most of them much older than the Sun, and even a civilization that crawls between them could spread across all of it in a few million years, nothing against the galaxy's age. If intelligent life is out there, the signs should already be visible. Two explanations follow. Reaching other stars is too hard, or intelligent species simply never tried. Once the Fermi probe launches, the argument goes, humanity becomes a civilization that both can and wants to reach another star, so the emptiness must have some other cause.
That inference is thinner than the mission's framing suggests. A single kilogram on an 80,000-year arc does not demonstrate that interstellar settlement is achievable; it demonstrates that one civilization, once, threw something. Johnston's darker readings are the more serious part of the argument. Life like ours may be close to impossibly rare. Or intelligent life may be common and usually die before it spreads, in which case one candidate for the cause, he says, is the self-destructive character of superintelligence, and humanity may run into that great filter within the next 50 years.
Get Physics Done needed three days and a billion tokens to find a way for a gold-plated record to leave the solar system. It is the same class of system Johnston names when he lists the things that might stop anyone from ever arriving.