Base Rates - AI Winters, Failure Archaeology (class 4)#
Contents
The winter analogy is the most-abused reference class in AI forecasting. Used loosely, it says "hype always dies." Used carefully, it asks what specifically killed funding last time, and whether those kill conditions are present now.
What ended each cycle#
| Cycle | Rough dates | What was promised | What actually bound | What killed the money |
|---|---|---|---|---|
| 1st winter | ~1973–80 | General problem-solving, machine translation, battlefield automation | Combinatorial explosion; weak compute; brittle symbolic systems | Lighthill (UK), Mansfield Amendment / DARPA pullback (US): no path from demo to deployable system |
| 2nd winter | ~1987–93 | Expert systems everywhere; Japan’s Fifth Generation | Knowledge-engineering cost; narrow transfer; hardware (Lisp machines) overtaken | Expert-system ROI failed at scale; maintenance cost > value; specialized hardware market collapsed |
| Minor dips | ~1990s–2000s; 2010s agent hype | AGI-adjacent branding; “AI winter over” narratives | Same: capability below commercial threshold for the sold use case | Product-market mismatch, not a field-wide funding freeze |
The shared structure is not “people got bored of AI.” It is:
- Capability plateaued below the threshold of a paying use case (or the use case required integration the demos skipped)
- Funding had been justified on crossing that threshold soon
- When the miss became undeniable, the money left - often via government program cancellation or enterprise budget cycles, not via a philosophical reassessment
Secondary kill factors that recur:
- Specialized stack stranded (Lisp machines; any hardware that only makes sense if the software thesis wins)
- Evaluation was theatrical - demos in constrained environments that did not survive contact with open-world cost
- Substitute technology won the budget (conventional software, simpler stats, better process design)
What did not end the winters#
- Proof that intelligence is impossible
- Exhaustion of interesting research questions
- Permanent public rejection of the idea
Research continued through every winter at lower amplitude. Winters are commercial and procurement events, not ontological ones. That is why the right analogy for 2027–29 is closer to capex boom corrections than to 1974 - if revenue is real.
Diagnostic for this cycle#
| Kill condition from history | Present in mid-2020s? |
|---|---|
| No revenue attached to deployed capability | No - material revenue, usage, task performance |
| Capability stuck below sold threshold | Mixed - strong on verifiable tasks; weak on open-ended autonomy sold as near |
| Funding justified only on next threshold crossing | Partially - train-run scale and AGI timelines still do this for a share of capital |
| Specialized stack with no residual value | Partially - accelerators depreciate fast; power/shells do not → Capital |
| Government single-buyer cancels | Low - civilian commercial demand dominates; sovereign is additive |
By the historical test, a full capability winter is the wrong base rate. The matching pattern is: correction that culls over-build and secondary labs while leaving deployed capability and the research line intact. That is why Compute and Part V row 6 model a capital-markets event, not a field death.
How a winter-shaped event could still arrive#
Failure archaeology also names the paths that would make this cycle rhyme with 1987:
- Revenue quality collapses - seat-priced option-value subscriptions churn when CFOs cut “AI” as a line item; outcome-priced revenue does not. → B5
- Agent demos never clear unsupervised production reliability while capex assumed they would - threshold miss with modern branding
- Energy/politics freeze deployment capacity so that model progress cannot convert to revenue at the assumed rate → Energy
- A salient incident freezes enterprise adoption for a cycle even though capability is real → Game 2
P(full winter \| historical structure) is low given revenue; P(sharp capital correction with consolidation) is the reference-class central case; P(adoption freeze after incident) is a separate mechanism that can look like a winter in the statistics without being one in the research pipeline.
Adjustment: revenue quality matters more than revenue quantity. Seat-priced subscriptions bought on option value behave differently in a downturn than usage priced against a measured outcome.
Scoring a miss cleanly#
| Observed pattern | Class match | Not a winter |
|---|---|---|
| Equity drawdown, no credit event, train runs continue | Price move | Do not credit class 4 or 2 |
| Credit event, secondary labs exit, inference keeps growing | Class 2 / row 6 | Capability line intact |
| Seat revenue collapses, outcome revenue holds | Revenue quality cull | Technology not falsified |
| Incident freezes enterprise buy, research continues | Adoption freeze (Game 2) | Looks like winter in stats only |
| Capability plateaus and paying use cases fail and funding freezes | Class 4 full winter | Requires all three |
The public will call several of the top rows a "winter." This page's job is to refuse that label unless the bottom row's conjunction is met.
Outcome revenue is the winter vaccine. Seat-priced option-value AI churns in a CFO cut; outcome-priced revenue with liability attached does not. B5 is therefore also a winter-risk indicator, not only a reliability indicator.
Related: Cycles · Capital · Compute · Part V row 6 · Uncertainty 6