Google DeepMind has begun routing every closed-ended question it can find through a single pass of its new Gemini Decisions API, which accepts up to 1,000 questions at a time across text, images, audio and video and returns a verdict on each within a quarter of a second. Engineers at the company said the system, Gemini Decision 1, had been asked to clear a backlog of its own internal review forms by lunchtime and had finished the backlog before the request was typed.
The company confirmed that the model now sits between employees and their own work. A product manager who wishes to ask whether a feature should ship submits the question alongside 999 others, most of them about whether the same feature should ship, and receives 1,000 answers before the coffee in the kitchen has finished brewing. Staff in the Mountain View office have been reassigned to the question queue, which now stretches down two corridors and into the cafeteria, where employees stand in line to be asked whether they are sure.
A senior director told reporters that the cost of each answer is roughly one fifth of the previous model's, a saving the company said it had reinvested in asking more questions. Asked whether any question had been declined, the director said the model had declined none, and that the company had therefore concluded there were no questions it was not prepared to answer, and had begun drafting a policy requiring every meeting agenda to be submitted to the API in advance so that the meeting could be cancelled.
Next week, the company will extend the system to the cafeteria menu, which it expects to settle the question of soup versus salad in 180 milliseconds, after which the chef has been told to stop cooking until he is asked.