TABLE OF LINKS
3 The Seven Qualification Tests for an AI Scientist
- Selection Criteria
- The Heliocentric Model Test
- The Motion Laws Test
- The Vibrating Strings Test
- The Maxwell’s Equations Test
- The Initial Value Problem Test
- The Huffman Coding Test
- The Sorting Algorithm Test
- Can an AI possibly conquer these tests?
- Why do we need these tests?
5 Conclusions and Future Work and References
2 Related Work
2 Related Work
The idea of automating scientific research activities dates back to the early days of computer science. An article on Science in 2009 [15] provides a great overview on the early explorations. Also in 2009 a “Robot Scientist” named Adam was released [16]. The authors developed specialized hardware for conducting basic experiments, such as tracking yeast growth with varying gene deletions and metabolites. This was paired with logic programming software for selecting experiments. The software keeps track of various hypotheses and chooses experiments likely to refute many of them at once. These experiments are automatically performed, and their results guide the next experiment’s selection. Adam effectively identified the functions of multiple genes, requiring fewer experiments compared to other experiment-selection methods like costbased choices. [17] presents a research that utilizes special hardwares to automatically learn the effects of different drugs upon the distribution of different proteins within mammalian cells. Very recently a breakthrough was brought by DeepMind [18], in which the authors created a large language model that learned geometry on one billion generated problems, in order to discover geometry properties, and train itself to prove these properties. The model was tested on 30 IMO geometry and got 25 of them correct, which outperforms the majority of IMO participants.
This is the first time a neural network model learns to master a discipline of science on its own, and it will not be surprising if the same methodology can be extended to other disciplines such as number theory and combinatorics. Our goal is to let AI make scientific discoveries on its own. There are two routes towards this goal. The first is to build an AI agent that can make novel and impactful scientific discoveries that have not been made before. This is our ultimate goal. But it is very challenging because the AI agent needs to be better than the best human expert in a field. The alternative route is to build an AI agent that can make some of the most important scientific discoveries in the history, without reading human knowledge that may contain key information to these discoveries. We believe this is an easier route because some of such discoveries can be inferred from abundant data and a scientific methodology. Comparing with the first goal which is analogical to building an AI agent that can beat the best GO player in the world, the second goal is like building an AI agent that can beat an amateur GO player. We believe the second goal is a good starting point for building an AI scientist, which should eventually evolve into someone who can make new and important scientific discoveries.
Authors:
- Xiaoxin Yin
This paper is