This jagged, spiky shape from Silicon Valley writer and analyst Tomas Pueyo nicely illustrates how AI intelligence has developed and the situation we find ourselves in today: models that excel at specific tasks while failing unexpectedly at other simple challenges. Gemini and Claude ace PhD-level reasoning tests, yet both still make mistakes a child would not.
The moment of the week was undoubtedly ex-OpenAI co-founder and leading AI researcher Ilya Sutskever’s first major public pronouncements since last year, when he declared the end of model pre-training. On the Dwarkesh podcast he doubled down on that view. Pre-training plus ever more compute took us astonishingly far, but he argues the easy wins from “just scaling” are largely behind us. We are now discovering the limits of this recipe through that jagged capability frontier. Sutskever’s claim is that closing those gaps is no longer an optimisation problem, it is a science problem. To move forward, labs will need new ideas about representation, memory and learning, and crucially better generalisation.
The path to AGI requires closing the gaps between the spikes on this chart. Around a year ago, mathematics looked like one of the most difficult areas, yet this week DeepSeek released an open-weight frontier Math-V2 model that can compete at Olympiad level. (Just to put this in perspective, the model scored 118/120 on the Putnam test which far exceeds any of the thousands who try this prestigious competition each year). This is a reminder that some gaps can suddenly disappear when a new idea lands, while others may remain frustratingly wide.