“It will take '5 to 20 years' to build a system that learns as efficiently as a human being — the capability that leads to superintelligence — and it will not come from continuing to scale pre-training, which has run its course; we're moving from the age of scaling back to the age of research.”
Co-founder and CEO of Safe Superintelligence Inc.; former Chief Scientist, OpenAI
Said Nov 25, 2025Deadline Dec 31, 2045Dwarkesh Podcast, 'Ilya Sutskever — We're moving from the age of scaling to the age of research,' November 25, 2025
Resolves on whether a system with human-level learning efficiency (Sutskever's bar for superintelligence) exists by 2045, the upper end of his stated band.