A post in Epoch AI’s Gradient Updates newsletter highlights a common futurist claim: once AI research is automated, superhuman AI agents could quickly produce radical technologies — molecular nanotechnology, Dyson swarms, or near-light-speed spacecraft. Ajeya Cotra, for instance, argues that a large population of superhuman AI agents could invent technologies much faster than humans and within a year or two produce truly science-fiction-level capabilities.
The authors identify two parts to this claim. First, if AI research becomes fully automated, AI systems could improve very rapidly and surpass current human experts. There is empirical and theoretical work suggesting this is plausible — bottlenecks do not clearly appear strong enough to prevent rapid improvement.
Second, many assume that with enough hyper-proficient AIs, building radically futuristic technologies would then take just a few years. Put another way: the main constraint is a shortage of extremely skilled workers. The authors question this: is lacking workers really the primary bottleneck? If AI cannot do essentially anything, the time to develop any advanced technology will depend heavily on how hard the technology itself is to develop. Making a nuclear weapon is far harder than making an axe, for instance. Yet debates about Dyson spheres, nanotech, or super-bioweapons often fail to rigorously examine those difficulties.
This omission is the missing half of AI-futurism debates: we’ve thought a lot about how capable AI might become, but much less about how difficult it will be to engineer specific technologies.
A proposed research agenda: three steps
The post recommends a focused research direction that follows three steps:
-
Pick a concrete technology and define it precisely.
- Example: self-replicating interstellar probes — machines that land in a new star system, assemble copies of themselves from local materials, and launch those copies onward at a large fraction of light speed.
-
Specify assumptions about AI capabilities.
- For example, assume AIs can act as “drop-in remote worker replacements” (i.e., substitute for human workers remotely) and require the runtime compute of an H100 GPU, or explicitly set the number of AI workers, how fast they can be run, and what physical actuators they can access.
-
Estimate how long these AIs would take to develop the technology, or what resources they would need.
The authors emphasize that they do not expect everyone to work on this, but they do think far too few people are doing it now. If an intelligence explosion is plausible — and possibly within a few years — then more concrete analyses are warranted.
Objections and responses
Objection 1: Haven’t people already done this?
There is relevant prior work. Many analysts frame AI’s technological effects using economic metrics (for example, how AGI might affect GDP growth). If AGI accelerated growth tenfold, that could compress a century of progress into a decade. But GDP is a noisy proxy for certain crucial questions: would a hundredfold increase in GDP lead to interstellar colonization, brain uploading, or easily developed molecular nanotech or bioweapons? Not necessarily.
Some researchers point to biological growth rates (e.g., fruit fly populations doubling in days) to argue that physical systems can change quickly; this serves as an existence proof but is a weak analogy for AI. The closest body of work is exploratory engineering (a.k.a. scientific roadmapping), which constructs detailed engineering pathways for specific future technologies — nanotech, neural recording, positional chemistry, brain emulation. Armstrong and Sandberg’s analysis of replicating probes, which used a 1980s probe design and a planetary disassembly procedure, argued that humanity could plausibly launch a project to colonize the reachable universe within decades.
The difference the authors propose is to add explicit AI assumptions to exploratory engineering: ask how much sooner nanotechnology could arrive with an army of AIs that match human abilities and work at human speed, or work a hundred times faster. One public example is Damon Binder’s work on post-AGI energy production, which designed a minimal solar power system that an economy with abundant robot labor could build and double on the scale of weeks. The authors envision similar analyses that vary AI assumptions to see how conclusions change.
Objection 2: The future is too uncertain to make these estimates useful
A common critique is that exploratory engineering is the “trick that never works.” The authors disagree, pointing to historical successes where theory preceded practice: Konstantin Tsiolkovsky in 1903 derived rocket motion equations and argued that liquid fuels could enable spaceflight decades before the first rocket crossed the Kármán line, and Arthur C. Clarke proposed geostationary communication satellites about 19 years before they were realized.
They do not claim certainty about the future, but argue that demonstrating a plausible engineering pathway given sufficient resources is easier and still valuable even if precise timelines remain uncertain.
Objection 3: Modeling superintelligence is too hard
Some argue that without knowing what superintelligence will look like, we can’t draw useful conclusions. The authors reply that we can still gain insights by bounding capabilities. For example, if you assume a billion AIs each at least as capable as top human experts (what the AI Futures Project calls Top-human-Expert-Dominating AI, or TED-AI), and you can show those AIs could build molecular nanotechnology within a few years, then it’s reasonable to conclude that even smarter AIs could as well. Conversely, showing such AIs probably could not do it highlights crucial capability cruxes and focuses debate on concrete skills.
Existing influential analyses about explosive economic growth used similar simplifying assumptions; they did not perfectly model superintelligence but still yielded useful perspectives on potential rapid growth.
Conclusion and call to action
The post’s main point is that more researchers should perform exploratory engineering analyses for specific, high-stakes technologies while explicitly modeling AI assumptions. If you are interested in doing this kind of work, the authors invite you to contact them at js@epoch.ai. The general recommendation is straightforward: combine detailed engineering roadmaps with clear, varied assumptions about future AI capabilities to better understand which futuristic technologies are plausible within different timelines and resource constraints.
