10×, then 100×: the case for scaling R&D

A position I have held and argued for a long time: if we are serious about exploring the universe and becoming a multi-planetary species, we should act on multiplying the number of people doing research — fundamental and operational — by 10, then by 100. AI automating operative jobs makes this more feasible, not less urgent.

The claim

Humanity devotes roughly half of one percent of its collective time to producing new knowledge — about ten million full-time-equivalent researchers among eight billion people, some 2.5% of world GDP (the research band is the sliver you cannot see on this chart). For a species that says it wants to cure aging, understand the brain, build fusion, and settle other planets, this is not a budget; it is a rounding error. The claim is simple: the right scale for research is 10 to 100 times the current one — from ~0.1% of the workforce toward 1%, then toward 10% — and the transition AI is forcing on the labor market is precisely the opening to do it.

Two kinds of R&D, both starved

• Fundamental R&D — moving the knowledge frontier: physics, biology, mathematics, materials, the science that does not know yet what it will be used for. This is where fertilizer, semiconductors, vaccines and rocketry ultimately came from, and it is the part markets systematically under-fund because its returns are diffuse, delayed, and impossible to appropriate.
• Operational R&D — the unglamorous engineering that turns known science into working systems at scale: launch cadence, closed-loop life support, in-situ resource use, radiation shielding, orbital manufacturing, reliability engineering. Becoming multi-planetary is mostly this: not one breakthrough but millions of engineer-years of iteration. Apollo at its peak employed around 400,000 people — an operational-R&D program, and we have never run one at that intensity since.

The two are complements, not substitutes. Fundamental research without operational capacity produces papers; operational programs without a moving frontier asymptote. A 10×–100× program scales both.

Why so much more — three arguments

• Ideas are getting harder to find. The best-known empirical result on research productivity (Bloom, Jones, Van Reenen & Webb, 2020) is that maintaining a constant rate of progress requires an ever-growing number of researchers — research productivity per person falls as the easy ideas are exhausted. This is routinely read as pessimism. It is the opposite: it is a direct argument that the researcher headcount must grow by orders of magnitude simply to keep the frontier moving, let alone accelerate it.
• Space is an engineer-hours problem. Nothing in physics forbids a self-sustaining settlement off Earth. What stands between here and there is an enormous, enumerable backlog of unsolved operational problems — biospheres that close, supply chains that survive without Earth, medicine at 0.38g. Backlogs of enumerable problems respond to headcount and iteration rate. A species running this program at 0.1% of its time has chosen, implicitly, not to do it.
• The portfolio logic of discovery. Most research bets fail, and which ones pay off is unknowable in advance — so aggregate returns scale with the number of independent bets. Small research systems are not just slower; they are more conservative, because every grant must justify itself. A 100× system can afford the weird bets from which the largest returns historically came.

The AI dividend

The strongest version of the argument is new. As AI automates operative jobs — clerical work first, then much of the non-physical band that today occupies nearly a third of human time — societies face a choice about where the freed human time goes. The precedent is on the occupations chart: when machinery collapsed agriculture from ~40% of human time to ~10%, the freed time became education, services, science and retirement. That reallocation was the twentieth century. The AI reallocation will be larger and faster, and it can be spent in three ways: leisure, redistribution fights, or capability. Redirecting even a tenth of the automated time into fundamental and operational research is the 10×; history suggests the societies that choose capability set the terms for everyone else. And AI shifts the supply side too: with models as research assistants, the marginal productivity of a human researcher is rising just as the opportunity cost of becoming one falls.

What 10× and 100× would look like

• 10× — research as ~1% of human time. On the order of 100 million researchers and engineers. Mechanically: research budgets moving from ~2.5% toward ~10%+ of GDP across the OECD and China, PhD-scale training pipelines widened by an order of magnitude, and standing operational programs (fusion, biosphere, launch, longevity) run at Apollo intensity permanently rather than as one-off crash efforts.
• 100× — research as a plurality occupation. The speculative end: close to a billion people whose primary occupation is discovery and engineering, feasible only in a post-AI-automation labor market — which is exactly the labor market being built now. At that scale research is not a sector; it is what a civilization does, the way agriculture once was.
• The talent is already there. The binding constraint is not genius but access: most of the people who could do research never get the chance — born in the wrong place, funneled into operative work the economy needed yesterday. A 100× program is mostly a claim that the talent lottery should stop discarding its winners.

Objections, briefly

• “Diminishing returns — more researchers produce less each.” True, and it is the argument for scale, not against it: if per-researcher productivity falls, the only way to keep the frontier moving is more researchers. The alternative reading — do less research — concedes stagnation.
• “AI will just do the research.” Perhaps eventually, and partially. But AI-assisted research still needs humans to pose problems, run physical experiments, build hardware, and validate results — and if AI does multiply research productivity, the return on each additional human researcher rises with it. Either way, the answer is more research time, not less.
• “Quality can't scale — you'd fund noise.” Scaling badly is possible; the fix is institutional (replication funding, diverse funding mechanisms, operational programs with hard deliverables), not a lower headcount. We already run noisy, hyper-competitive science at small scale — the pathologies come from scarcity, not abundance.

The multi-planetary test

“Multi-planetary species” is a useful test precisely because it cannot be faked with announcements: either the biospheres close and the supply chains hold without Earth, or they do not. A civilization that means it will show it in its time budget — in the width of the research band on the occupations chart, the one band that has never yet had its turn to grow. That width is a political decision, and I have long argued we should make it: 10× first, 100× when the automation dividend arrives.