Andrea
Procopio.
I'm 21, and I'm a first year PhD student at MIT with Prof. Dimitris Bertsimas, and a Research Fellow at Harvard with Prof. Tomer Ullman.
My research focuses on AI systems for open ended scientific discovery that can continually learn and improve from experience, using new knowledge, and becoming themselves better at performing the process of discovery. Similar to what humans do, and possibly in a more holistic way.
Previously, I studied CS and Economics at Bocconi University and spent my final year doing research at Harvard on computer vision and cognitive science. I also co-founded BAINSA, Italy’s largest undergraduate-run AI research lab, and EuroTech Federation, a community of young founders in Europe.
Research Thesis.
What makes a 5th year phd student produce better research —in quality, success rate, and time— than the 1st year version of himself? Or again, what makes an experienced artisan create more intricate and tasteful artworks than his younger self?
I think a fundamental part of the answer lies in the ability to extract and consolidate experience into reusable knowledge: what worked, what didn't, which trajectories of reasoning and experimental directions proved promising, and why.
My hypothesis is that the process of discovery itself can be modeled, and that some of its underlying principles are domain-agnostic. If that's true, we could build AI systems that continually improve their ability to discover in any domain, rather than approaching every new problem from scratch, and hyperspecialize models to solve them (e.g., AlphaGo, TTT Discover).
Questions I'm Thinking About
-
Learning from Experience: How can research agents extract and consolidate generalizable knowledge from their reasoning, experiments, successes, and failures, in a way that persists across different runs of the same family of problems?
-
Learning to Discover: How can agents improve their research judgment and “taste” over time, learning which questions to investigate, which hypotheses to pursue, and which experiments are most informative?
-
Continual Self-Improvement: How can research systems modify their own selves, using accumulated experience to improve the quality, success rate, and efficiency of future discoveries?
-
Domain-General Discovery: To what extent are the mechanisms underlying successful discovery shared across different scientific disciplines, and how can knowledge acquired (and the process itself) in one domain transfer to another?
What Motivates Me?
Throughout history, much of human progress has been driven by the hominid or civilization’s ability to discover and apply new knowledge.
A scientific breakthrough can advance a particular field. But a breakthrough in how we make breakthroughs could accelerate progress across many fields, at the same time, and potentially producing positive externalities as a collateral of cross domain interactions. And if the systems making those discoveries can themselves learn and improve from experience, the slope itself of that progress increases over time.
I believe that making the process of discovery cheaper, faster, and increasingly domain-agnostic is the best moonshot to expand humanity’s capabilities, understanding of our outer and inner world, and more concretely raise the median quality of life. I also think it’s just super cool and exciting