Funding Blue-Sky Research to Survive the Trough

DoesDoes a Research Field Sleep? shows AI research surviving two winters because specific foundational work — Werbos's backpropagation thesis, Pearl's Bayesian networks, LeCun's early convolutional networks, Vapnik's statistical learning theory — kept going while public attention and applied funding collapsed. The obvious lesson to draw is that state-sponsored blue-sky research is what survives a Trough that commercially-funded work doesn't. That lesson is too simple, and the same history that supports it also refutes it: the funding cuts that triggered both AI winters were themselves government decisions. State money isn't a reliable exemption from the Hype Cycle. One agency's own history shows both sides of this at once.

DARPA's Two Faces

TheThe Advanced Research Projects Agency, founded in 1958, funded the packet-switching research that became the ARPANET starting in the mid-1960s — J.C.R. Licklider's early vision, Bob Taylor's 1966 project, Larry Roberts's network design, the first host-to-host connection in 1969¹. There was no product, no near-term application, and no public narrative demanding results; it took decades for that research to become anything resembling the internet the term now describes. That's blue-sky funding working as advertised: a specific programme, evaluated by a programme manager against technical milestones, on a horizon long enough that no Hype Cycle had time to form around it at all.

The same agency's Speech Understanding Research funding was cut in the mid-1970s, directly contributing to the first AI winter Does a Research Field Sleep? describes. Same agency, same broad funding mandate, opposite outcome. Whatever explains the difference, it isn't "government-funded" versus "not."

What Actually Made the Difference

TheThe Lighthill Report is the sharpest illustration of the actual mechanism. James Lighthill's 1973 review, commissioned by the UK's Science Research Council specifically to inform funding decisions, concluded that "in no part of the field have the discoveries made so far produced the major impact that was then promised"² — and that single report drove a near-decade collapse in UK academic AI funding. This is a Trough-of-Disillusionment correction happening at the level of national science policy: one committee's revised model of the field's utility, produced under exactly the kind of small-sample, hype-inflated extrapolation the second-order framing describes, propagating into a funding decision that affected an entire country's research base at once.

Japan's Fifth Generation Computer Systems project shows the same mechanism with an even more explicit publicity input. MITI committed roughly ¥54 billion over 1982–1992, framed from the outset as a national-prestige project — establishing Japan as a technologically advanced, internationally contributing nation, not just building a computer³. Tying a research programme's justification to national competitive standing makes public sentiment about that standing the publicity input driving the whole project's Hype Cycle. When the parallel-inference-machine approach was overtaken by conventional CPU improvements, the correction was as total as the ambition had been.

Contrast this with the ARPA/DARPA funding model's better-documented mode: mission-oriented programmes with a specific technical goal, a programme manager given real authority to run it, and continuation decisions made against that programme's own technical milestones rather than the political mood about the field as a whole⁴. Whether a given pot of research money survives a Trough doesn't depend on whether the government or a company wrote the cheque. It depends on whether that money's continuation is coupled to the same hyped narrative currently cycling through the field, or evaluated against something else entirely.

The Deeper Lesson: Diversity of Evaluation, Not Source of Funding

OnceOnce the mechanism is stated that way, it's a specific case of a much more general principle. Harry Markowitz's foundational result in portfolio theory is that combining assets whose risks aren't correlated reduces the portfolio's overall risk below any individual asset's own risk, precisely because a shock that hits one holding doesn't automatically hit an uncorrelated one⁵. A field's research funding is a portfolio in exactly this sense: Werbos's Harvard thesis, Bell Labs' internal signal-processing work, Pearl's book, Vapnik's statistical learning programme, and DARPA's SUR funding were not one asset, but the crash that hit SUR didn't automatically hit the others, because their continuation wasn't tied to the same evaluative narrative SUR's was tied to. Scott Page's independent result from collective problem-solving is the same principle from a different angle: groups that bring genuinely different evaluative perspectives to a hard problem systematically outperform groups of highly capable people who all evaluate it the same way⁶. Connecting Markowitz's uncorrelated-risk result and Page's diversity-of-perspective result to research-funding design specifically is this page's own extension — neither author makes that application themselves.

Bell Labs is the cleanest illustration that the mechanism, not the state/private label, is what's doing the work. Its research budget came from AT&T's regulated telephone monopoly, insulated from quarterly market discipline in a way ordinary commercial R&D isn't, and LeCun's early convolutional-network work was pursued there through the second AI winter on exactly that insulation⁷. That's not state funding at all — it's a different structural trick producing the identical decoupling from the field's own publicity cycle. Vannevar Bush's 1945 case for federal basic-research funding, which created the US National Science Foundation, and Mariana Mazzucato's more recent argument that state investment underwrites exactly the high-risk, long-horizon work private capital won't⁸⁹ are both real and don't need correcting. What this page adds is the condition Lighthill and Fifth Generation show that argument leaves out: state funding only supplies the insulation those arguments promise if its own continuation criteria are kept separate from whatever public or political narrative is currently cycling through the field it funds. A single large state programme whose survival depends on one report, one election cycle, or one national-prestige narrative is a single point of failure wearing a government logo, not a diversified portfolio.

A Practical Design Principle

PutPut in the same terms as Managing the Trough's advice to builders: a funding body should treat a Trough-driven funding review the same way a builder should treat its own Peak-driven overclaiming — as a correction to a model that was never a reliable measure of the underlying research's value in the first place, not as a verdict to act on immediately. Concretely, a resilient research ecosystem needs several independent funding lines evaluated against different criteria and different time horizons, not one well-funded programme whose survival depends on a single narrative remaining favourable — whether that money is a government agency's, a monopoly's regulated profits, or, as it was for CNNs, SVMs, and Bayesian networks alike, a specific researcher simply continuing the work regardless of who was or wasn't paying attention.

Related Topics

• Does a Research Field Sleep? Collective Consolidation and the AI Winters — the historical pattern this page draws its funding lesson from.
• The Hype Cycle as a Second-Order Cybernetic System — the model behind reading the Lighthill Report and the Fifth Generation project as Trough-correction events.
• Managing the Trough: For Builders and For Learners — the same self-model-correction framing applied to a builder's own publicity, rather than a funding body's.
• Placing a Technology on the Half-Life Scale — the substitute-pressure and fan-in signals this page's funding-diversity argument is a portfolio-level version of.

References