Ilya Sutskever
Co-founder and Chief Scientist of OpenAI (2015–2024); co-founder and CEO of Safe Superintelligence Inc. (SSI, founded 2024). Key architect of the modern deep learning paradigm: co-author of AlexNet (2012), GPT-3 (2020), and the original transformer-based scaling approaches. Left OpenAI after the November 2023 board crisis; founded SSI to pursue a “straight-shot” approach to safe superintelligence without product-market pressure.
Appearances in this wiki
| Episode | Source | Date |
|---|---|---|
| Ilya Sutskever on the Age of Research, Continual Learning, and Alignment | Dwarkesh Podcast | 2025-11-25 |
Key positions
- From the age of scaling to the age of research. Pre-training scaling has yielded most of its gains; the next breakthroughs require new ideas, not more compute applied to the same recipe. We are back to the research era “just with big computers.”
- Models generalise dramatically worse than humans. The fundamental unsolved problem: why do systems with superhuman benchmark performance still fail at tasks a five-year-old handles? Ilya’s thesis: insufficient generalisation, possibly related to missing value functions.
- Value functions and emotions. Human emotions function as a biological value function — enabling rapid, self-directed learning without external reward. AI systems lack this, which explains both sample inefficiency and poor generalisation.
- Continual Learning as the path to superintelligence. Not a fully-trained omniscient model, but a model that can learn from deployment the way a human joins a job. The “superintelligent 15-year-old” frame.
- Alignment: build AI that cares for sentient life. More tractable than human-aligned AI because the AI itself will be sentient; empathy-like properties may emerge naturally. Pairs with a capability cap on the most powerful systems.
- Research Taste as top-down belief. Directions are chosen by an aesthetic — beauty, simplicity, correct inspiration from the brain — and that conviction is what sustains a researcher when experiments contradict them.
- SSI’s straight-shot. Build superintelligence without product-market pressure, releasing gradually as capabilities are proven. Estimate: 5–20 years to human-like learning systems.