The Next Fifteen Years

A forecast built from first principles
Section future / 08-method / base-rates / winters.md

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#

CycleRough datesWhat was promisedWhat actually boundWhat killed the money
1st winter~1973–80General problem-solving, machine translation, battlefield automationCombinatorial explosion; weak compute; brittle symbolic systemsLighthill (UK), Mansfield Amendment / DARPA pullback (US): no path from demo to deployable system
2nd winter~1987–93Expert systems everywhere; Japan’s Fifth GenerationKnowledge-engineering cost; narrow transfer; hardware (Lisp machines) overtakenExpert-system ROI failed at scale; maintenance cost > value; specialized hardware market collapsed
Minor dips~1990s–2000s; 2010s agent hypeAGI-adjacent branding; “AI winter over” narrativesSame: capability below commercial threshold for the sold use caseProduct-market mismatch, not a field-wide funding freeze

The shared structure is not “people got bored of AI.” It is:

  1. Capability plateaued below the threshold of a paying use case (or the use case required integration the demos skipped)
  2. Funding had been justified on crossing that threshold soon
  3. 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:

What did not end the winters#

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 historyPresent in mid-2020s?
No revenue attached to deployed capabilityNo - material revenue, usage, task performance
Capability stuck below sold thresholdMixed - strong on verifiable tasks; weak on open-ended autonomy sold as near
Funding justified only on next threshold crossingPartially - train-run scale and AGI timelines still do this for a share of capital
Specialized stack with no residual valuePartially - accelerators depreciate fast; power/shells do not → Capital
Government single-buyer cancelsLow - 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:

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 patternClass matchNot a winter
Equity drawdown, no credit event, train runs continuePrice moveDo not credit class 4 or 2
Credit event, secondary labs exit, inference keeps growingClass 2 / row 6Capability line intact
Seat revenue collapses, outcome revenue holdsRevenue quality cullTechnology not falsified
Incident freezes enterprise buy, research continuesAdoption freeze (Game 2)Looks like winter in stats only
Capability plateaus and paying use cases fail and funding freezesClass 4 full winterRequires 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

View markdown source

select · Enter open · Esc close