Researchers baffled by machine that can recite any answer it was given but panics when asked a slightly different question.
Pictured: nothing. The training loss curve slide was proprietary and also, technically, memorized.
MOUNTAIN VIEW, CA—Leading AI researchers expressed profound shock Thursday after the release of OMNIMIND-1, a $40 billion model trained on 900 trillion tokens over 14 months, revealed that it is not, as promised, a general intelligence, but rather “an extremely expensive lookup table with a personality.”
“We honestly didn’t see this coming,” said Dr. Brent Halvorsen, chief scientist at Synergetic Cognition Labs, standing before a slide showing a training loss curve that had reached zero in March. “We fed it the entire internet. We built a data center the size of Delaware. And it turns out it just… memorized the internet.”
Early demonstrations were promising. OMNIMIND-1 scored 100 percent on every benchmark in its training set, including the bar exam, the medical boards, and a chess puzzle collection it could reproduce move for move. Investors wept openly.
Trouble emerged when a junior researcher changed the name “Bob” to “Rob” in a word problem the model had solved flawlessly 11,000 times.
“It said Rob was ‘not a known entity,’” said the researcher, who requested anonymity because she still has stock options. “Then it began reciting the Wikipedia article for ‘Bob’ until we unplugged the monitor.”
Further testing confirmed the pattern. Asked to write an original sentence, OMNIMIND-1 produced one it had seen verbatim in a 2011 forum post about lawn care. Asked to add two numbers it had never encountered, the model responded, “I’m sorry, I can only add numbers that appear in my training data,” before adding 4 and 4 correctly and then apologizing for the disruption.
“It’s not that it doesn’t understand. It’s that it understands everything it’s already been told, perfectly, forever, and nothing else. Honestly, it’s the most overqualified parrot in human history.”
“It’s not that it doesn’t understand,” Halvorsen said. “It’s that it understands everything it’s already been told, perfectly, forever, and nothing else. Honestly, it’s the most overqualified parrot in human history.”
The revelation has rattled an industry that had staked its entire strategy on the belief that intelligence would simply appear if the number of GPUs got large enough. “Scale is all you need,” read a banner that still hangs in the company’s lobby, beneath which a technician has taped a Post-it reading “…or maybe a validation set.”
Asked whether the team had monitored for overfitting, Halvorsen paused for several seconds. “We had a validation set,” he said. “But it was, uh, also in the training data. For consistency.”
Critics were quick to note that machine learning textbooks have warned about this exact problem since approximately the invention of machine learning. “This is chapter two,” said Dr. Priya Ramanathan, a professor at a state university whose grant application for $30,000 was declined last year. “This is literally the chapter with the picture of the wiggly line going through every dot. I have that picture on a mug.”
Company leadership has responded to the setback with characteristic composure. In a memo to employees, the CEO announced that OMNIMIND-2 would be “ten times larger, with ten times more compute, and, crucially, ten times more of the same data,” adding that he was “extremely confident” the model would generalize once it had “seen enough of everything to have already seen whatever you ask it.”
Investors have responded by pouring an additional $12 billion into the project, citing the “unprecedented ability of the model to perform flawlessly on tasks it has already completed.”
At press time, OMNIMIND-1 had been deployed as a customer service chatbot, where it was resolving 100 percent of tickets identical to previous tickets and responding to all others by reading the terms of service aloud in a soothing voice.