Interview, Fireside Chat
Ilya Sutskever – We're moving from the age of scaling to the age of research
- AI's economic impact is projected to be widespread and strong, driven by powerful economic forces, with a predicted transition from an "age of scaling" to an "age of research" as compute costs shift from a primary bottleneck to a resource sufficient for demonstrating new ideas.
- A disconnect between high evaluation performance and low real-world utility is expected to persist, potentially caused by RL training environments that prioritize metrics over general utility or create single-minded models, and pre-training is anticipated to lack the generalized understanding achieved through human evolution or lifetime learning.
- SSI plans a "straight shot" to superintelligence but acknowledges the plan may adapt if timelines extend or gradual public impact proves more valuable, noting that the organization does not require absolute maximal compute to validate its research ideas.
- Economic forecasts suggest that if AI achieves human-like learning efficiency, the next 5 to 20 years will see rapid growth, while even if companies stall on this specific capability, they are expected to generate significant revenue despite potential profit challenges.
- Market competition is predicted to prevent any single entity from monopolizing the benefits of continuous learning or specialized agents, leading to differentiated economic niches and price reductions similar to previous AI adoption cycles.
- Future AI models are expected to diverge in approach rather than remain uniform, with differentiation emerging through RL and post-training on diverse datasets, potentially incentivized by adversarial setups like debate or self-play.
- Safety dynamics will shift as AI capabilities increase, with governments and the public demanding action, all companies adopting paranoid safety approaches, and frontier labs beginning to collaborate, potentially necessitating a "cap" on the most powerful superintelligences.
- Long-term stability is posited to be more achievable through human-AI neural links than by allowing AI to act solely for individuals, while the aesthetic principles of beauty and simplicity are expected to guide research through periods of contradictory experimental data.
- Current models are predicted to be trained on identical pre-training data, limiting diversity, whereas future developments in competitive programming may not translate to improved judgment or taste for general codebases.