Can AI Make Scientific Breakthroughs?
The article explores whether AI can achieve genuine scientific breakthroughs, moving beyond pattern recognition to hypothesis generation and discovery. It examines current AI capabilities in science, limitations in causal reasoning and creativity, and the potential for AI to act as a collaborator rather than a replacement for human researchers.
Background
- The article is from the Cosmos Institute, a think tank founded in 2024 by tech investors and AI researchers (including people from Stripe and DeepMind) focused on AI's philosophy and culture.
- It critically examines claims that large language models (LLMs) can make fundamental scientific discoveries, arguing instead for a "mechanistic" view where AI is a tool within existing scientific processes.
- Key reference: Karl Popper (science progresses through falsifiable conjectures) and David Deutsch (breakthroughs require explanatory theories, not pattern-matching).
- "The Bitter Lesson" is a 2019 essay by AI researcher Rich Sutton arguing that general methods like search and learning (scaling with compute) always beat human-crafted knowledge — a view the article pushes back against.
- Why it matters: AI labs (DeepMind, OpenAI, Anthropic) increasingly claim their models can accelerate or automate science; skeptics argue this confuses pattern recognition with understanding.