Note: This post is a summary of a speech given at CERN Sparks! Serendipity Forum in September 2021, which can be viewed here.
When people imagine a world with artificial general intelligence (AGI), robots are more likely to come to mind than to enable solutions to society’s intractable problems. But I think the latter is closer to the truth. AI is already enabling huge leaps in addressing fundamental challenges: from solving protein folding to predicting precise weather patterns, scientists are increasingly using AI to infer the rules and principles that underpin highly complex domains of the real world — domains they might never have explored without help.
Advances in AI research will increase society’s ability to address and manage climate change – not least because of its urgency but also because of its complex and multifaceted nature.
the control
Looking at the field of AI research today, there are two common categories of problems that scientists focus on: prediction and control. Prediction models attempt to learn about an area (such as weather patterns) and understand how it will evolve, while control models prompt agents to take action in that environment. Building a successful pathway to AI requires understanding and developing algorithms in both spaces, taking into account all the differences our natural and social environment throws at us, from how viruses mutate or how language evolves in usage and meaning over time to how it aids production. Energy from fusion power. Two real-world areas that scientists at DeepMind are contributing to address climate change while developing what is needed to build artificial general intelligence are weather forecasting and plasma control for fusion.
It’s almost impossible to accurately model weather patterns – it’s an example of nature’s variances to the fullest. However, causes and effects can be inferred based on vast amounts of historical data. In collaboration with the Met Office (the UK’s national meteorological service), scientists at DeepMind have developed systems that can take 20 minutes of weather data to create multiple hypotheses for radar maps and accurately forecast Heavy rain in the next 90 minutes.
Crucially, these models will help meteorologists provide forecasts that aid decisions related to emergency services, energy management and activation of flood warning systems – enabling better preparedness for and response to extreme weather events, which are becoming increasingly common around the world. Helping to predict important weather events by predicting accurate weather patterns is one example of how AI research can make a meaningful impact as it becomes more generally applicable and “smart”.
global challenges
In addition to responding to the effects of climate change, resolving its sources is of equal, if not greater, importance. Fusion, which is a single, clean, unlimited, self-sustaining source of energy, is a long way off, but it remains one of the most promising solutions in the world—I think it requires developing a general algorithm that can solve many different components simultaneously. We are already seeing progress on one of the components, which is the very difficult problem of maintaining new plasma shapes to enable better power production and to stabilize the plasma for as long as possible.
Working with world-renowned experts at the Swiss Plasma Center and the École polytechnique fédérale de Lausanne (EPFL), we are able to go beyond existing manual models, and apply deep reinforcement learning algorithms first developed for robots to control plasma. The result is a controller that can successfully handle different shapes and configurations of plasma at a rate of 10,000 interactions per second.
Without the collaboration of experts, AI researchers cannot make much progress in real-world areas. Defining the right paths forward in these areas requires partnerships across disciplines, leveraging a joint scientific approach to developing and using artificial intelligence to solve complex questions at the heart of society’s most pressing needs. This is why dreaming alongside a diverse group of natural scientists and socializing about what the world could look like with AI is so crucial.
As we develop AI, addressing global challenges such as climate change will not only bring about critical and beneficial impacts that are urgent and necessary for our world, but also advance the science of AI itself. Many other classes of general AI problems are yet to be solved – from causal to efficient learning and transfer – and as algorithms become more general, more real-world problems will be solved, incrementally contributing to a system that will one day help solve everything else, also.