Superbreakthroughs: How to engineer a brighter future with AI & science
Presenting two essays, ”What do we mean by AI for Science?” and ”Superbreakthroughs: Four vignettes of how AI will transform the way we live” that examine the present and future promises of AI & science.
The conversation around AI and science has been dizzying. Last month, OpenAI announced an AI-led solution to the Navier-Stokes existence and smoothness problem, one of the Millennium Prize Problems long held as a pinnacle of mathematical challenge. Anthropic revealed their new life sciences laboratory along with their discovery of a novel enzyme system which they compared to CRISPR, the Nobel-winning and field-defining genetic editing mechanism.
But as swiftly as those announcements came, swiftly too did the backlash arrive. Mathematicians quickly decried a perception of academic misconduct by OpenAI (which the company denied), and 25 Fields Medalists referred to this as a “severe misalignment” that could be a “general threat to intellectual work.” There’s already a Wikipedia page on it. Similarly, biology researchers noted that Anthropic’s announced enzyme system had been previously discovered though unpublished, and others questioned the scientific significance of the claim, if they didn’t understand the mechanism.
All of this has been a reminder to me that scientific research exists in a complicated ecosystem of scientists, institutions, training, competition, prestige, legacy, identity, and ego—and as far as I know, AI hasn’t automated all of those yet.
I should say that I am incredibly excited for the future of science and the ways AI will accelerate discoveries and usher in transformative technologies. A cure for every disease, unlimited clean energy, or room-temperature superconductors? Yeah, let’s do it. But science also demands a healthy dose of skepticism and fora where we can debate ideas, methods, and conclusions. So what do we think the future of AI for science will hold?
This recent research piece by economists at Google DeepMind dives into it well: we are already seeing AI change the way scientists work and what they think about the future of science. Their survey of US and UK scientists show that 46.9% believe that AI has contributed a moderate or significant increase to their scientific output in the last three years, and 55.6% believe AI will have a further >10% increase to their scientific outputs in the next three years. But, there is also a sense of how science and the scientific ecosystem are changing—49% of respondents said AI increased their preference for incremental, safer questions rather than riskier endeavors.
To understand where science is going, we must examine what AI makes easier and what it makes harder. In our first essay, What do we mean when we say AI for science?, we walk through the scientific discovery process to see how AI tools—which can include large language models, specialized scientific models, autonomous lab equipment, and more—are transforming science. But science exists against a nuanced backdrop of science policy, for example, how do we verify results, or use new information to make decisions, or allocate science funding, or train junior scientists? As AI speeds up certain parts of the cycle like hypothesis generation, data collection, and coding, we must make concomitant efforts in strengthening verification, communication, and data governance, to ensure that we’re asking the right questions and we’re creating a scientific flywheel for the benefit of humankind.
To understand where science is going, we must examine what AI makes easier and what it makes harder.
And new breakthroughs in science can lead to breakthroughs in technology, which we explore in our second essay, Superbreakthroughs: Four vignettes of how AI will transform the way we live, which is an optimistic telling of some really cool technologies on the horizon. The trajectory of technology is one that is intimately linked with all of human history—how we have developed tools and knowledge to make our lives better, longer, and richer. And many types of technologies get cheaper at a power-law rate known as Wright’s Law: that there exists a consistent learning curve for technologies such that the more you produce a technology, the more efficient that production becomes.[1] Succinctly, the lesson is: the more you build, the better you get at building. And AI is helping us build a lot faster—and that means improving the core technology but also recognizing that technology adoption occurs within a complex web of moving parts, including manufacturing, permitting, tax incentives, labor, community involvement, behavioral science, and more.
The more you build, the better you get at building.
The tale of solar energy is a prime example of Wright’s Law. In a single lifetime, solar has gone from a niche technology—some of you may remember it as the small row of dark rectangles powering pocket calculators—to the cheapest and fastest-growing source of new electricity generation, and one of our biggest bets against climate change. The U.S. alone built 43 GW of solar capacity last year—that could power about 200 trillion calculators.
The optimism around AI & science is not so much a belief that it will magically pop out a technological solution to many of society’s challenges, but that we can decompose problems and solutions into stages that get tackled in turn. That AI tools will organize and accelerate trials, iteration times, and traditional process bottlenecks, so we can learn more and learn faster to figure out what works and what moves us forward. And if we do that thoughtfully and methodically, thousands of times over for the problems we seek to solve, well the world is going to look a lot different in a handful of years. That’s what we want to drive home in our essay, Superbreakthroughs, where we look at four examples of energy technology.
If you ask a physicist (or an occultist), they’ll tell you that you can describe everything around us in terms of energy—movement is energy (E=1/2mv2), matter is energy (E=mc2), existence is energy (H=Ea†a). So a fascinating investigation is: what is the future of energy? Both in how we produce energy and how we use it and control it. In Superbreakthroughs we explore a pathway to geothermal energy, or how to generate 24/7 clean energy using the heat from the core of the Earth. We also discuss carbon monitoring and marine carbon dioxide removal, and how models of the ocean and the atmosphere can help inform some of the world’s most ambitious attempts to remove carbon from the atmosphere. Finally, we explore electrochemical manufacturing, or how to move atoms around to construct the molecules and chemicals we need for everyday life.
How do we get there? Four policy recommendations for AI & science
- Treat data and provenance as infrastructure. Domain-specific, interoperable data standards will set the foundation for shared and verifiable AI model inputs and outputs to construct a sustainable flywheel for AI & science.
- Build new institutions for communication, verification, and replication. Traditional journals and peer review will quickly become insufficient for this new wave of scientific discovery. We need to invest in a new system of communication that builds trust in technology and directs us to the most interesting and pressing questions.
- Fund validation through to pilots. Make it easier to turn new lab discoveries into technology that works for people.
- Guard against the concentration of capital and the erosion of small science. Traditional scientific discoveries may become tightly coupled to access to compute and large models.
There’s a lot to be excited about in the coming future of AI, science, and technology, but how do we get there? In our essay, we describe four policy recommendations, summarized above, for how we can build the foundations for a sustainable and inclusive science and technology policy ecosystem to serve as the practical engine to make these futures possible.
Read the full essays here:
