The scenario this manual has to take seriously
Every page in this pillar reasons about a fast-moving field, and the forecast and dashboard pages treat continued rapid progress as the working assumption. This page is the counterweight: what if AGI does not arrive on schedule? A serious manual plans for the distribution of outcomes, not the modal headline, and the distribution includes a stall. This is scenario analysis about a historically documented phenomenon, not a claim that progress has stopped.
What the last two winters actually looked like
The field has been here before, twice.
The first winter (roughly 1974–1980) followed a direct institutional kill-shot. Sir James Lighthill’s 1973 report Artificial Intelligence: A General Survey, commissioned by the UK Science Research Council, concluded that the grand promises of AI research — machine translation, robotics, general problem solving — had failed to deliver at anything like their projected scale, and recommended redirecting funding toward narrower work. Within a few years, major government funders pulled back on both sides of the Atlantic, and graduate programs contracted sharply.
The second winter (roughly 1987–1993) killed a different victim: expert systems. These rule-based commercial products had absorbed enormous corporate and government investment during the 1980s, then collapsed under their own brittleness — systems that could not learn, could not handle exceptions, and were expensive to maintain. The market rout took funding for the whole symbolic-AI paradigm down with it. The National Academies’ history of federal computing research, Funding a Revolution (1999), documents how DARPA’s retrenchment and the disappearance of the Lisp machine vendors turned a subfield’s failure into a discipline-wide drought (National Academies Press). The fuller narrative histories are Crevier’s AI: The Tumultuous History of the Search for Artificial Intelligence (1993) and Nilsson’s The Quest for Artificial Intelligence (2009) — both worth reading for the pattern, not just the facts.
The pattern: each winter was not a technical dead end but a credibility-and-funding collapse. The science kept moving in the shadows; the money, talent, and public patience did not. Winter ended when a low-cost demonstration — chess, then ImageNet — made the field investable again, almost overnight.
The case for a false dawn
The strongest version of the stall case has three legs.
Data and algorithmic limits. Epoch’s analysis of human-generated text estimated that the stock of high-quality public training data would be exhausted within roughly the decade if then-current trends held (Villalobos et al., 2022), forcing reliance on synthetic data and other workarounds whose long-run value is genuinely uncertain. Separately, the field has repeatedly mistaken benchmark saturation for general competence — a warning the manual’s emergent-capabilities and Turing-test pages develop in detail. Benchmarks can saturate while real-world reliability, the thing economies actually pay for, lags badly.
Jaggedness may be the ceiling, not the floor. The persistent finding that models are superhuman on some dimensions and inexplicably weak on adjacent ones has a pessimistic reading: that scaling a next-token predictor buys breadth without buying the kind of robust, compositional reasoning that scaling-law optimists project. Chollet’s case that contemporary benchmarks overstate intelligence because they allow memorization rather than requiring skill acquisition under novelty is the canonical statement of this skepticism (Chollet, 2019).
Overhang cuts both ways. Even if the gains are real, they may already be extracted: the “capability overhang” argument implies today’s models underperform their training because products, evaluations, and organizations haven’t caught up. If so, further scaling yields diminishing visible returns — which markets read as a plateau, whatever the underlying cause. There is even a documented phenomenon running against scale: inverse scaling, where larger models score worse on some tasks, such as certain types of truthful or sensitive reasoning (McKenzie et al., 2023).
The case against a winter
Winter advocates have to explain why this cycle rhymes with the last two, and the differences are large.
The money is real this time. Previous winters followed hype without revenue. The current cycle has deployment revenue measured in tens of billions of dollars annually and private generative-AI investment that keeps setting records — the Stanford AI Index 2025 reports $33.9 billion in private generative-AI investment in 2024 alone, up sharply year over year. Funding on that scale does not evaporate because benchmarks disappoint; it evaporates when balance sheets break.
The hardware base is enormous. Prior winters could starve the field by cutting a few grants. Today the field rides on a capital-intensive semiconductor, cloud, and energy buildout whose investors expect decades of returns. That is a floor under continued progress — and, more darkly, a constituency that will keep pushing even when prudence says stop.
Widespread integration creates its own demand. Hundreds of millions of people and a large share of firms now depend on AI systems for ordinary work. Utility, unlike novelty, does not go out of fashion. The burden of proof sits on the winter case: it must show not merely that progress slows, but that it slows enough, for long enough, to repeat a funding collapse against an installed base this size.
What a winter would mean
Assume the stall happens. What follows is genuinely double-edged.
Talent dispersal. A winter pushes frontier researchers toward hedge funds, biotech, and whatever the next hot field is. That is bad for safety in one way — alignment expertise, which is thin and concentrated at a handful of labs, scatters — and possibly good in another: dangerous capability work slows while some of the same brains work on narrower, more auditable problems. The net effect is unknowable in advance, but the dispersion of alignment talent is a real cost that winter optimists tend to skip.
Hardware overhang. Paused software progress does not melt down the chips. A world that stopped training frontier models tomorrow would still hold an enormous stock of deployed capability, much of it under-incentivized to maintain safety staffing. Winter thins the defenses around existing systems; it does not thin the systems.
The misreading risk. This is the worst part. A genuine multi-year plateau would be politically decoded as “the risk passed” — a conclusion that would be actively wrong. Progress paused is not progress refuted. The underlying drivers (compute, algorithmic efficiency, commercial incentive) would still be compounding in the background, and the restart, when it came, would likely be faster and less supervised than the run-up. The manual’s treatment of benefits and opportunity costs cuts the same way: even a stalled transition has real gains, and a world that declares victory early loses both the gains and the vigilance.
The misreading risk deserves emphasis: the most dangerous winter is the one that convinces the public the danger is over. Historical winters ended in restarts. Nothing in a stall guarantees the safety work survives to the next thaw.
Plan for the distribution, not the headline
The honest posture for a manual is the same one this page opened with: do not bet the analysis on continued exponential progress, and do not bet it on a stall either. Prepare for capability arriving on schedule, arriving late, arriving unevenly, and arriving amid the institutional failures covered elsewhere on this site — because the same preparations (robust institutions, preserved knowledge, distributed response capacity) are the ones that pay off across the whole distribution. A winter would change the timing. It would not change what a well-prepared society looks like.