Home > Programs > Zenith Grants > Zenith Grant Awardee

Zenith Grant Awardee

Gerardo Adesso

University of Nottingham

Co-Investigators

Paul Knott; Ludovico Lami

Project Title

The ultimate brain: Hallmarks and limitations of intelligence in general probabilistic theories

Project Summary

What is intelligence? In judging whether someone is more or less intelligent – relatively to some specified task, e.g. finding the exit of a maze – we are unlikely to answer "they are intelligent if their neurons are connected in such and such way". Rather, we judge them by their ability to output the right sequence of moves to win. The details of how their brain reaches the solution do not make any difference. We signify this by saying that intelligence, with its main ingredients (memory, processing, learning) is a substrate-independent phenomenon. Here, we postulate that intelligence is even theory-independent: classical systems can exhibit intelligence, as can quantum systems, and even systems abiding by laws beyond those of quantum theory. Using mathematical tools from general probabilistic theories, we will then identify the fundamental laws that enable intelligence to exist, while placing quantitative limits on the ability of intelligent systems to perform tasks like solving mazes, tic-tac-toe and Go. We will then address fascinating questions such as: How many resources does Nature need to evolve an intelligent agent like the human brain? Can even cleverer brains exist in a universe where communication is more nonlocal than in quantum mechanics?

Technical Abstract

This project aims to provide a rigorous framework to understand which laws of physics enable intelligence to exist, which resources sustain it, and what are its fundamental limitations in different – classical, quantum, and post-quantum – physical theories. Our primal tool will be the formalism of general probabilistic theories (GPTs), within which we can model information processing in abstract systems, including but not limited to classical and quantum theory. We will use this to formulate an axiomatic framework for intelligence. We will also benchmark intelligence quantitatively in specific tasks and games, both when an optimal strategy is known and when one first needs to learn it (e.g. by machine learning). Within the GPT formalism, we will investigate the ultimate performance of intelligent agents on purely a priori grounds. We will assess whether physics with stronger nonlocal correlations than quantum mechanics could allow for more integrated and more powerful brains to exist, and we will estimate the computational resources required for intelligent agents to evolve in quantum and post-quantum physical theories. The proposal relies on foundational tools (GPTs and fundamentals of information theory) and exploits them unconventionally to lay a rigorous topical groundwork for intelligence in the physical world.

Skip to content