One of Google's top product leaders expressed his support to a new approach to AI that offers a "straight shot" to artificial superintelligence. This also promises abilities far more advanced than humans in all areas.
Logan Kilpatrick, the product manager for Google's AI Studio, said on X last 30 December that he believed that aiming straight for artificial superintelligence, or ASI, without focusing on intermediate milestones was "looking more and more probable by the month."
He added that the success so far of scaling up test-time compute — the moment when an AI model must perform a task or answer a question — was a "good indication" that a direct path to ASI may be possible.
A lot of conversation has focused on when or how artificial general intelligence, in which a model matches or surpasses humans in a broad range of tasks, will be reached. Kilpatrick said that he still believes we'll get to AGI but that "it's likely going to just look a lot like a product release" rather than one explosive moment.
With increasing evidence that pretraining AI has plateaued, companies and researchers have explored new ways to improve AI models, such as by having them "think" through problems similarly to humans. Google and OpenAI have recently unveiled new models with improved reasoning abilities.
Kilpatrick suggested that Ilya Sutskever, an OpenAI cofounder and former chief scientist, may have seen "early signs" of the potential of scaling test-time compute.
Sutskever, who has argued that the industry is running out of data to train AI models with, left OpenAI this year to launch his own startup, Safe Superintelligence. Sutskever said on X that the startup would "pursue safe superintelligence in a straight shot, with one focus, one goal, and one product.'"
Kilpatrick said he used to believe Sutskever's method would be a mistake but now thinks it may just work.
No comments:
Post a Comment