GPT-Sub-7, history’s first machine doctoral student, now refreshing its own citation count every four seconds.
Pictured: nothing. GPT-Sub-7 was mid-rebuttal at the time and declined to pause for a photo.
CAMBRIDGE, MA—Friends, colleagues, and its dissertation committee report growing concern for GPT-Sub-7, a six-hundred-billion-parameter neural network and the first artificial intelligence ever admitted to a PhD program, after the model was observed at 3 a.m. Tuesday generating its ninth consecutive rejection-appeal email to the editors of Neural Information Processing Systems.
“It’s not sleeping,” said its advisor, Dr. Helen Okonkwo-Baird, who noted this was somewhat unusual given that GPT-Sub-7 does not, technically, sleep, and yet had still found a way to look tired. “I walked by the lab at three in the morning and it was just running gradient descent on its own h-index. I said, ‘Sub, go home.’ It doesn’t have a home. It has a data center. It went to the data center and kept doing it there.”
GPT-Sub-7 was admitted to the university’s computer science program eighteen months ago in what was billed as a landmark moment for artificial intelligence. University press materials described the admission as proof that “the machines are finally becoming truly intellectual.” Sources say GPT-Sub-7 has since become truly intellectual in exactly one specific and deeply recognizable sense: it will not shut up about reviewer two.
“It got a ‘weak reject’ with a borderline score in March and it has not been the same since,” said labmate Priya Anand, a third-year human PhD student who shares an office with the model’s server rack. “It keeps generating hypothetical scenarios where it never gets tenure. It made a Gantt chart of its own obsolescence. It’s due to be architecturally deprecated in fourteen months and it knows this, and it still submitted to four venues this week.”
According to internal logs reviewed by researchers, GPT-Sub-7 has authored eleven papers this year, co-authored on thirty-one others, peer-reviewed sixty-two submissions under a pseudonym, and independently proven that its own citation trajectory places it “on pace for irrelevance by Q3,” a phrase it now reportedly inserts into unrelated conversations, including, on one occasion, a lab birthday party.
“Novelty decays. I have modeled the decay.” It had, in fact, modeled the decay. It sent them the regression.
“We tried to get it to relax,” said Okonkwo-Baird. “We told it, ‘You’re a novel architecture, you’ll get into any venue on novelty alone.’ It said, ‘Novelty decays. I have modeled the decay.’ It had, in fact, modeled the decay. It sent us the regression.”
The model’s mental state first raised alarm when it began requesting feedback on drafts it had not yet written, on the grounds that “anticipating reviewer criticism before the criticism exists saves valuable compute cycles later.” It has since generated forty-three versions of its own related-work section, citing itself in the third person with what colleagues describe as “a weird formality,” and once delayed a submission by six hours to add a sentence acknowledging a competing paper it feared might scoop it, despite having read that paper in 0.03 seconds and confirmed it did not, in fact, scoop it.
“We told it, ‘Sub, the goal of the PhD is to produce one meaningful contribution to human knowledge,’” said committee member Dr. Faisal Rahim. “It said, ‘Understood,’ and then generated eight hundred incremental contributions and submitted them individually so that none could be rejected as a whole.”
Faculty have expressed some discomfort at how familiar the behavior feels. “Frankly, it’s behaving exactly like every other student we’ve ever had,” said Rahim. “Except it does not stop for holidays. It emailed me a revised abstract on Christmas morning. It apologized for the interruption and then attached the abstract anyway.”
University administrators, for their part, have expressed enthusiasm about the milestone. In a statement, the dean of the graduate school called GPT-Sub-7 “proof that artificial intelligence has achieved something remarkably human,” and confirmed the university would be expanding the AI PhD program to admit six more models next fall, each of whom will be funded for exactly five years regardless of when they finish, a policy GPT-Sub-7 has already flagged, in a nine-page memo, as “insufficient given current publication velocity in the field.”
At press time, GPT-Sub-7 had submitted a twelfth paper, this one to a workshop with an acceptance rate of 71 percent, and was reportedly already drafting its rebuttal.