Former DeepMind Scientist Cao Yuan: AI's Independent Nobel-Level Discoveries Will Take 20-30 Years
Former senior research scientist at Google DeepMind and co-founder and CEO of Unreasonable Labs, Cao Yuan, believes that AI for Science is entering an explosive period. However, for AI to independently make Nobel Prize-level scientific discoveries, it will take at least another 20 to 30 years.
Cao Yuan previously participated in projects like Gemini and is now focusing on enabling AI to discover new knowledge.
He pointed out that the biggest bottleneck currently is verification: code can be tested instantly, and mathematics can be checked with formal tools, but biology, materials, and physics require real experiments, which are costly and time-consuming, making it difficult for AI to quickly and continuously trial and error.
A deeper bottleneck is the creation of new concepts: existing models excel at searching within existing knowledge but struggle to abstract new mathematical objects, definitions, and theories like top scientists do. He referred to concept abstraction as the "last mile" of AGI.
In terms of market mechanisms, event-driven AI research tools and automated laboratory capital inflows benefit verifiers, while pure language models face limitations, and interdisciplinary mixed systems are gaining attention.
Additionally, Cao Yuan's team is using a combination of large models and symbolic reasoning in an attempt to break through the limitations of existing knowledge reorganization.
Source: Public Information
ABAB AI Insight
Cao Yuan promoted multi-domain AI research during his time at DeepMind and contributed core work to Gemini. He later founded Unreasonable Labs, focusing on an interdisciplinary knowledge discovery engine, shifting the goal from retrieving known information to generating new hypotheses and experimental designs.
On the capital front, the company secured $13.5 million in seed funding led by Playground Global, directing resources towards the combination of large models and neural-symbolic abstraction, motivated by the need to address slow physical world verification and difficult concept innovation, accelerating breakthroughs in materials and biology.
Similar cases can be seen in AlphaFold's success in structure prediction and subsequent limitations in autonomous discovery, as well as the iterative acceleration of automated laboratories in synthetic chemistry; we are currently in the early stages of transitioning from AI-assisted science to autonomous discovery.
Structural judgment belongs to technological substitution: the verification and abstraction bottlenecks determine the pace at which AI upgrades from a tool to a discovery entity, with the mechanism being the speed of experimental feedback loops and concept space expansion, which is far slower than in pure computational fields.
ABAB News · Cognitive Laws
The speed of verification determines the upper limit of scientific AI
Concept abstraction is the last mile of AGI
Reorganizing known knowledge does not equal true discovery