We Have So Many Questions
Until now, energy could not become discovery without passing through a human mind, and every institution of science was built around that constraint. Automated research could remove it.
Primus
For the past few months, our team has logged on each morning to find an autonomous, PhD-level researcher waiting for instructions.
We called it Primus. We pointed it at our personal whiteboards of research ideas that we had been accumulating over years. It fleshed out the ideas, executed the experiments, and came back with results.
After going back to our list of research ideas over and over, eventually the backlog was empty. The researcher was still waiting. We had nothing left to ask.
That was new to us. We had never lived in a world where we could answer research questions faster than we could generate them.
The Tree of Knowledge
I have heard researchers say that if automated research takes off, we risk running out of things to discover.1 I don't think we will. Every answer opens more questions than it closes. The neuroscientist Stuart Firestein compares knowledge to a ripple widening on a pond: the larger the circle, the more of the unknown it touches.2 Each discovery enlarges the adjacent possible, the set of next steps that become reachable only once the last one is taken.3
So we went looking. Over the last months we've been meeting with quantum physicists, mathematicians, materials engineers, roboticists, economists, biologists, and so many others. Nearly every one has shown us a direction that could serve the world, and a question we could never have phrased ourselves.
The questions had been out there all along, in other people's heads. So if answers and questions are both plentiful, what limits research?
Energy In, Knowledge Out
Stripped down to physics, research is a process that turns energy into knowledge. We spend it training people for decades. We spend it building laboratories, telescopes, and particle accelerators. We spend it powering the computers that run the experiments. Energy goes in. Years later, if everything goes well, a discovery comes out.
Nothing about that path is efficient. Entire societies are built on trained minds, and each one takes decades to make. Most experiments fail. Most careers produce no breakthrough, and nobody can say in advance which ones will. Suppose the world decides a field matters: a funding agency writes a program, a university opens a department, a venture firm backs a startup, a graduate student gambles their twenties. Decades pass between "this matters" and "people are working on it."
The conversion is so slow, so messy, and so unpredictable that we have built centuries-old institutions just to manage it. Universities and tenure, to protect the long bets. Peer review, to decide what counts as a result. Markets, to pay for steady improvement. Prizes, to reward expeditions with no business case. All of them exist because energy only becomes knowledge by way of human lives.
Automated research has the potential to collapse that entire process. Energy and compute become search directly, the way training a model already works. The decades of training humans disappear. So do the decades of lag. Today you cannot pour a megawatt into quantum computing and get more quantum computing knowledge. Soon that may be possible, and once it is, energy becomes the only constraint on discovery.
A Lever for Every Field
We picture a world with levers for every form of research. Pull one, and ten thousand synthetic biologists go to work, wired to robotic wet labs. Pull another, and a hundred thousand materials scientists join them.
Expertise becomes effectively unlimited while energy stays finite, so humankind has to decide, hopefully democratically, what to focus on:
- Do we point a million synthetic oncologists at the rare childhood cancers that markets ignore?
- Do we invent zero-emissions cement and crops that fix their own nitrogen?
- Or do we spend a tenth of the world's power learning to settle Mars?
Choosing one future over another becomes an explicit decision.
Mavericks vs Followers
Once we choose what to pursue, we still have to choose how boldly to pursue it.
When progress can be measured, there is a strong pull toward the nearest hill. Thomas Kuhn called this normal science: cumulative work inside a paradigm, as opposed to the rare revolution that breaks one.4 Most science is normal science, and should be. The risk is an optimizer that does little else and looks productive the whole time. Lean too far the other way, though, and it chases long shots that never pay off.
Traditional institutions somehow find a balance, because tenure, markets, and prizes each reward a different appetite for risk. Nobody designed that mix and it is often unjust, but it leaves somewhere for revolutions to come from.
An automated lab will need to evolve an equilibrium of its own: mavericks to take the long risks, incrementalists to climb the hills, and the right proportion of each.
Who Steers?
The last bottleneck is political. Until now, politics has been an argument over the resources of survival: land, food, housing, medicine, security. The next argument will be over who decides what to discover.
A world with infinite researchers is not a world where every question gets answered. It is a world where choosing the question becomes the defining political, economic, and moral act of our time.
Anyone handed a million researchers will attempt great things. But who decides which pursuits are worth it? Mars, higher-yield rice, a new branch of mathematics: each is a defensible use of the world's energy, and each serves different people. Today that choice is diffused: spread across thousands of grant panels, boards, and careers, and made by nobody in particular. Soon it could be concentrated in a handful of people at the levers.
We are dedicating our careers to solving the technical roadblocks to automated research, while working with researchers around the world on the question of what it should pursue. We continue the same way we started: by asking lots of questions.
The walkers in the hills figure are Low Detail Animated Crowd by Shahriar Shahrabi, licensed under CC BY 4.0.