Somebody at PSI had to ask the model for clearer charts. That detail sits in the middle of the announcement and it is the most useful thing in it: an astrophysicist on staff steered the run toward the mission's requirements, requested a cost analysis and better charts, and checked the output for errors [18]. Around that supervision, the system worked mostly on its own for three days on about a billion tokens, according to PSI cofounder and CEO Matt Pines [17]. That is a claim with edges, which is more than most "AI for science" material offers. The problem got to PSI because Johnston mentioned it on a podcast hosted by Alex Wissner-Gross, the physicist who cofounded the lab, who offered to run it through their software [12]. Get Physics Done is open source, breaks a physics question into smaller tasks and decides which simulations to run, using models including Anthropic's Claude or OpenAI's GPT [13]. What it proposed was a recombination: slow the craft so its orbit swings in closer to the sun than Mercury, then fire the engine only on those close passes, when the panels take four times the light and a burst of thrust at high speed buys more energy than the same burst delivered further out [15]. Because the engine runs only near the sun, the panels stay small and the vehicle stays light [16], which is the wall the Fermi team had been hitting from the other side [11]. The individual maneuvers were well known; the combination was one the team had not considered, according to a paper that has not been peer reviewed [14]. Pines says the entirely different mission profile was the surprising part [19]. The year of failure is what made the three days tractable. It produced the $15 million cap [7] and the requirement that a small solar-powered vehicle carry at least a kilogram of cargo to another star [9][11]. Divide 365 days of human search by three days of runtime and you get roughly 122 to 1 [23], a ratio likely to end up on somebody's slide, though the model searched inside a fence the humans spent a year building. What actually happened, by this account, is that a well-posed problem was handed to a searcher, with an expert grading the returns, not that an AI did the research on its own. A benchmark score is marked by whoever picked the benchmark; this result has a payer and a date attached instead. The mission is budgeted at $15 million from individual private donors [7], launch intended by the end of 2029 [1], with a copy of the Golden Record that NASA flew on the two Voyager probes in 1977 aboard [9]. It can still be wrong in visible ways: review may not hold up the trajectory [14], the hardware may not close on the budget, or nothing launches at all, which is where Breakthrough Starshot ended up after Yuri Milner pledged $100 million in 2016 [5]. Johnston told MIT Technology Review that "we didn't want to do another Breakthrough Starshot" and that his team is "dead set on something actually launching" [6]. Fermi's budget is 15% of Milner's pledge [24]. The concession is time: Voyager 1 has been flying since 1977 and has covered less than 1% of the 25 trillion miles to Alpha Centauri [10], and Johnston says "we are not constraining ourselves to doing it in a human lifetime" [8], which is how you arrive at up to 80,000 years in transit [2], about 4,000 times Starshot's advertised 20-year trip [22]. Two questions separate the pitches worth taking seriously from the rest: can somebody outside the lab check the output, and is somebody committing money or schedule on the strength of it. Benchmarks are checkable and carry no commitment.
