Hard Takeoff, Soft Takeoff, or No Takeoff: Pick Your Heresy

The recursive self-improvement debate has been running for about thirty years. Both sides are still at it, with positions that have barely moved despite an enormous amount of new evidence. It remains one of the few unresolved empirical questions in the field, and not for lack of experiments. The trouble is that people read the same experiments in completely different ways.

The basic question: if an AI system can improve its own intelligence, how fast does that improvement happen? Fast enough to matter catastrophically? Or slow enough that human institutions can adapt?

The Hard-Takeoff Case

Eliezer Yudkowsky has been making the hard-takeoff argument since the late 1990s, when he was writing long essays on a mailing list that most people had never heard of. His core intuition is that intelligence improvement is not linear. Small improvements in intelligence generate larger improvements in the ability to make further improvements, so the feedback loop compounds. What starts as a modest advance becomes, within hours or days, something qualitatively different from anything that has existed before.

He coined the term "FOOM" for this scenario: a hard takeoff so fast it looks like an explosion. The word captures the idea that there wouldn't be a gradual transition or any period of warning. The system crosses some threshold, and then it's done. The world after bears no recognizable continuity with the world before.

The argument isn't crazy. Compound growth does produce nonlinear outcomes, and small edges in capability can translate to large edges in output. There's no obvious physical reason why intelligence improvement couldn't be self-reinforcing to a degree that produces rapid, discontinuous change.

The Soft-Takeoff Case

Robin Hanson, the economist, has been the most persistent skeptic of the FOOM hypothesis. His counterargument is empirical: we don't observe this pattern in the historical record. Other important technological developments, such as steam power, computing, and communications, have shown fast progress but not FOOM-fast progress. They hit diminishing returns, encountered physical bottlenecks, and required complementary inputs that took time to develop.

The bottleneck he emphasizes is that intelligence improvement is not constrained only by intelligence. A more intelligent AI can design better algorithms, but it still needs training data, compute, energy, and engineering work to implement them, and none of that scales with intelligence alone. So the compound feedback loop has friction, possibly a lot of it.

Yudkowsky's response is that Hanson underestimates the magnitude of the intelligence advantage once it crosses a certain threshold. A system smarter than all humans combined could work around most of those bottlenecks in ways we can't anticipate. The bottlenecks we can see are the ones that exist at human intelligence levels, and a sufficiently superhuman system might just solve them.

What the Scaling Era Settled, and What It Didn't

GPT-4 to GPT-5, Claude 2 to Claude 3 to Claude 4: each generation is better than the last on benchmarks, and capabilities that seemed years away appeared within months of each other. The rate of progress over 2020-2025 was faster than most people expected, even the people paid to track it.

Does this vindicate the hard-takeoff case? Not really. The progress, while fast, has been gradual and demanded enormous engineering effort and capital. The companies spending the most are also spending on things that don't scale with model capability: data centers, energy infrastructure, human researchers. If self-improvement were exponential and unconstrained, you'd expect the pace to accelerate much faster than it has. Will MacAskill's framing treats those non-scaling inputs as feedback loops of their own: an industrial explosion that kicks in only once AI can automate the factories and chip lines, arriving after the software loop rather than alongside it.

There is a counterpoint to the counterpoint. Current AI systems can't improve themselves. They're trained by humans on human-curated data using human-designed architectures. The recursive self-improvement scenario requires something that doesn't yet exist: a system capable of meaningfully improving its own training process, not just its outputs. We don't know what the scaling curves look like once that threshold is crossed, because we haven't crossed it. The AI 2027 scenario is one detailed guess at what the months right after look like.

Nobody Knows

The answer is that nobody knows. The hard-takeoff scenario is possible, and the arguments for it haven't been refuted so much as contested. The soft-takeoff scenario is also possible, and the empirical evidence tilts its way, for now, for current systems.

The uncertainty is asymmetric, though. If the soft-takeoff people are wrong, the world has time to adapt. If the hard-takeoff people are wrong, we wasted some caution. If the hard-takeoff people are right and the soft-takeoff crowd bet against them, we don't get to be wrong twice.

Yudkowsky has made this asymmetry argument for decades, and it isn't obviously wrong. The problem is that acting on it implies constraints on AI development that no government or company has shown any willingness to accept, while the case for those constraints requires levels of certainty the evidence doesn't yet provide.

The Speed of Takeoff Decides Which Safety Plans Survive

The speed of takeoff determines what kinds of safety measures are even feasible. Slow takeoff gives you time for governance, for iteration, for interpretability research to catch up. Hard takeoff means none of that works. You either solve alignment before the first powerful recursive self-improver, or you don't solve it at all.

If Yudkowsky is right, almost everything the mainstream AI safety community is doing is running the wrong race. If Hanson is right, the slower pace means that something like normal human institutions might actually work.

We'll find out. Probably sooner than either side predicted.