Another super interesting conversation with Ilya Sutskever at the Dwarkesh Podcast

Another super interesting conversation with Ilya Sutskever at the Dwarkesh Podcast[1]which made me fall short on my exercises in the gym last night as brain detached from body to process the input. Still on it.

One interesting aspect is Ilya talking about how essentially from his perspective it’s again time for research as “just” scaling the current set of technologies likely will generate a large amount of revenue with a considerable amount of cost, but research is required to get to the next level of AI.

Another thing which keeps me busy is his view on pre-training, AGI and humans: “The second thing that got a lot of traction is pre-training, specifically the recipe of pre-training. I think the way people do RL now is maybe undoing the conceptual imprint of pre-training. But pre-training had this property. You do more pre-training and the model gets better at everything, more or less uniformly. General AI. Pre-training gives AGI. But the thing that happened with AGI and pre-training is that in some sense they overshot the target. If you think about the term ‘AGI,’ especially in the context of pre-training, you will realize that a human being is not an AGI. Yes, there is definitely a foundation of skills, but a human being lacks a huge amount of knowledge. Instead, we rely on continual learning.”

That is something which is largely missing in today’s mainstream pre-train-then-deploy recipe—continual learning is an active research field, but deployed frontier models don’t yet keep learning the way humans do.

There are many, many more things in this interview. Providing you a short summary as a glimpse:

The interview covers a broad discussion about the current state and future of AI, focusing on Sutskever’s argument that the age of scaling large models is likely over and research-driven methodologies are regaining priority. Ilya Sutskever explains key issues in pre-training large models, why current AI models generalize less efficiently than humans, and the role of reinforcement learning (RL) and value functions in improving AI. He highlights differences between AI model training analogies and human learning and emotions, offering the speculation—drawn from a neuroscience case study—that the human value function is modulated by emotions hardcoded by evolution, which may be part of what makes human decision-making so robust. The conversation explores the concept of “straight-shot superintelligence” pursued by Sutskever’s new company, Safe Superintelligence Inc. (SSI), focusing on creating AI systems that learn continually from deployment rather than delivering finished intelligent agents. The discussion touches on economic impact, model generalization, the importance of gradual deployment, and how SSI plans to balance research and deployment while addressing safety and societal integration of powerful AI systems.

Highly recommended to tune into this podcast. Curious to learn what you take out of it!

[1] Dwarkesh Podcast, “Ilya Sutskever — We’re moving from the age of scaling to the age of research,” Nov 25, 2025 (full transcript with timestamps; the quoted pre-training/AGI passage is in the 00:46:47 section): https://www.dwarkesh.com/p/ilya-sutskever-2

• AI
• Innovation