Interview, Fireside Chat
AI progress is about to rapidly accelerate in 2025 – Sholto Douglas & Trenton Bricken
- AI is expected to eventually surpass human researchers in the speed of deploying and accelerating context, acting as a "fantastic co-pilot" that enables coding at significantly faster rates.
- Future progress in model capability will increasingly rely on synthetic data generated by AI, while maintaining large-scale training runs remains essential to understand emergent properties.
- The field will likely operate under conditions of imperfect information, necessitating "ruthless prioritization" to navigate a "graveyard" of failed initial runs where performance lines deviate from smooth theoretical curves.
- Research teams face a strategic allocation decision between compute for new training experiments versus scaling the best-performing models, with the elasticity of progress quantified at 0.5 (e.g., 10x more H100 compute expected to yield roughly 5x faster Gemini program progress).
- The most effective researchers will be those who rapidly expand their toolboxes by combining reinforcement learning, optimization theory, and systems engineering, rather than relying on a single academic background.
- Architectural discovery will continue via "greedy evolutionary optimization" over possible architectures, potentially resulting in solutions that appear more "brain-like," though the best ideas will often need to attack problems directly without attachment to familiar solutions.
- The ability to iterate quickly remains the primary differentiator for top researchers, with cycle time separating leaders at smaller scales and a need to balance big frontier runs with experimental regimes.
- Teams will need to apply a "simplicity bias" to navigate broken theoretical understandings, constantly guessing model behaviors and interpreting why ideas fail at small scales without guarantees that trends will hold for new architectures.