Interview, Fireside Chat
Marty Chavez (Sixth Street): Finding a Single Source of AI Truth
- Marty Chavez anticipates that computational capabilities will solve the protein folding problem and enable "digital twins" for experiments that are dangerous or impossible in reality, facilitating actions based on simulation results.
- Predictions indicate that AI will identify valid drugs from vast chemical spaces and slightly improve the probability of success in Phase 2 and Phase 3 clinical trials, which currently suffer from high-frequency, multi-billion dollar failures.
- Over the next few months, specific problems are expected to transition from extremely difficult to easy due to context window breakthroughs, exemplified by the Gemini 1.5 model's capacity for one million tokens.
- AI is projected to be highly effective in rapidly changing markets, though its success is considered less certain in non-stationary environments compared to slowly evolving subjects.
- Future biological mapping is expected to mirror chip design methodologies, although the required orders of magnitude for complexity and data remain uncertain.
- Extreme confidence in simulation is predicted to become mandatory for chip fabrication and drug development as costs rise, with current fabrication costs reaching $500 million and drug development costs exceeding this by significant margins.
- Companies failing to establish a "single source of truth" for data are expected to struggle with hallucinations and non-actionable imaginings derived from incorrect information.
- Regulatory frameworks are forecasted to follow a historical pattern of being implemented promptly after major negative events, often resulting in over-regulation, with 99% of Dodd-Frank cited as ineffective red tape while CCAR is viewed as beneficial for system safety.
- Proposals for AI regulation suggest targeting control points at intersections and boundaries rather than holding creators liable for every negative outcome, aiming to prevent LLMs from becoming fully general operating systems.