I’ve spent an unreasonable amount of time reading about the AI winters — the grant proposals, the post-mortems, the sheepish retrospective interviews. Not because I’m nostalgic for failure, but because the field froze twice, and both times the survivors left behind lessons that read like they were written yesterday. The winters of the 1970s and late 1980s weren’t natural disasters. They were consequences. And the causes are still in the room with us.
What follows isn’t a history lecture. It’s a field manual distilled from two collapses: what went wrong, what the survivors did right, and which lessons from the AI winter apply directly to the boom we’re living through. For the full background on the freezes themselves, start with my guide to what an AI winter is.
Table of Contents
- Lesson 1: Winters Are Made of Promises, Not Technology
- Lesson 2: Funders Buy the Story, Then Punish It
- Lesson 3: Demos Are Debt
- Lesson 4: The Work Continues in the Cold
- Lesson 5: Winter Kills Companies, Not Knowledge
- Lesson 6: Vocabulary Is a Survival Tool
- Lesson 7: Boring AI Outlives Exciting AI
- Lesson 8: Humility Scales Better Than Hype
- Frequently Asked Questions
Lesson 1: Winters Are Made of Promises, Not Technology
The most important thing to understand about both AI winters: the technology didn’t get worse. Nobody woke up in 1974 and discovered that symbolic AI had stopped working. The capabilities continued their slow, uneven climb. What changed was the distance between what was promised and what was delivered — and who was paying attention to that distance.
In the late 1960s, prominent researchers predicted machine intelligence within a generation. Machines would translate languages, reason like experts, understand the world. When those predictions failed to materialize on schedule, the reaction wasn’t “the science is hard” — it was “we were misled.” The winter was a credibility event, not a technical one.
The lesson cuts both ways. If you’re building AI, your promises are liabilities with a fuse. Every overclaim you make today is a withdrawal from a credibility account that gets audited eventually. And if you’re evaluating AI — as a buyer, a funder, an employee — the right question is never “is the technology impressive?” It’s “is the technology impressive relative to what was promised?” That gap is where winters are born.
Lesson 2: Funders Buy the Story, Then Punish It
Here’s an uncomfortable truth about how research money works: funders don’t fund capabilities, they fund narratives. In the 1960s, the narrative was “intelligent machines are nearly here” — and money flooded in. In the 1980s, the narrative was “expert systems will revolutionize every industry” — and money flooded in again. Both narratives were oversold by the people who stood to benefit from the overselling.
Then the narrative breaks, and funders don’t just adjust — they overcorrect. The same institutions that threw money at AI in 1968 were writing reports about AI’s fundamental limitations in 1974. The punishment always exceeds the crime, because funders are managing their own embarrassment as much as their portfolios.
The practical lesson: if your funding depends on a story, build the exit ramp into the story. The researchers who survived the winters were the ones who could say “even if the grand vision takes longer, here’s what this work delivers now.” Concrete intermediate value is narrative insurance. The history of artificial intelligence is full of projects that lived or died on whether they had it.
Lesson 3: Demos Are Debt
Both winters featured legendary demos — systems that performed brilliantly in controlled conditions and collapsed in the wild. The 1960s had blocks-world programs that “understood” language about stacked blocks but nothing else. The 1980s had expert systems that performed beautifully in the demo room and broke down in the hospital.
A demo is a promise made in the most favorable possible conditions. It’s not dishonest — it’s how progress gets shown. But every demo creates an expectation, and expectations are debts that come due. The winters arrived when the debts matured all at once: funders realized the demo-to-reality gap wasn’t a small engineering step but a chasm.
Modern version of this lesson: treat every impressive demo as a claim that needs a deployment to back it. A model that aces a benchmark is a demo. A model running in production, handling edge cases, surviving contact with real users — that’s the payment. The industry’s current demo-to-deployment ratio should make everyone slightly nervous. Demos don’t survive winters; deployments do.
Lesson 4: The Work Continues in the Cold
This is the lesson people miss because it’s less dramatic: the winters weren’t empty. Research continued — quietly, cheaply, under different names. Neural network research didn’t die in the 1970s; it went underground, pursued by a handful of stubborn researchers with little funding and less prestige. Backpropagation, the algorithm that powers modern deep learning, was developed and refined during the cold years.
The people who kept working had a trait worth copying: they were motivated by the problem, not the spotlight. When the money left, the tourists left too, and what remained were the people who found the work itself interesting. That’s a useful filter for your own career and investments. Ask: who in this space would still be here if the funding dried up tomorrow? Those are the people to bet on.
It also means winters are productive in a hidden way. Freed from hype cycles and quarterly expectations, researchers in the cold years did foundational work that the boom years later commercialized. The winter is when the field does its homework. The spring is when it shows off.
Lesson 5: Winter Kills Companies, Not Knowledge
The expert-system companies of the 1980s are gone — Symbolics, Lisp Machines, Teknowledge, all dead or absorbed. Hundreds of millions in market value evaporated. But the knowledge didn’t evaporate with it. The techniques, the papers, the trained engineers — they dispersed into other fields and waited.
This matters for how you think about risk. If you’re worried about an AI winter destroying value, separate the fragile from the durable. Fragile: valuations, startups with no revenue, careers built on hype vocabulary. Durable: open research, trained practitioners, deployed systems with real users, and the underlying mathematical knowledge — like how AI models are trained, which doesn’t become false when funding drops.
The lesson for builders: make your work durable. Open-source what you can. Document what you learn. Build things that work without a boom to sustain them. Winter is a stress test for whether your contribution was real, and the grading is harsh but fair.
Lesson 6: Vocabulary Is a Survival Tool
My favorite winter survival story is linguistic. During the freezes, researchers stopped saying “artificial intelligence” — the term was toxic, associated with broken promises. They said “expert systems,” then “intelligent systems,” then “informatics,” “machine learning,” “data mining.” The work continued; only the label changed.
There’s a cynical reading (rebranding to chase grants) and a practical reading (names carry baggage; choose names that don’t). The practical reading is the lesson: don’t marry your identity to a hype term. If “AI engineer” becomes a punchline again someday, the people who described themselves as “people who build systems that learn from data” will be fine.
For the industry, the lesson is to watch the vocabulary. When insiders start avoiding the field’s own name, the culture has already registered a downturn that the funding hasn’t caught up to yet. Language leads; money follows.
Lesson 7: Boring AI Outlives Exciting AI
The systems that survived both winters were the boring ones. Pattern recognition for check processing. Optimization for logistics. Statistical methods for manufacturing quality control. Nobody wrote magazine covers about them, and nobody defunded them either, because they quietly made money.
Meanwhile the exciting systems — the ones that promised to replicate human experts, understand language, see the world — were the ones that died spectacularly. Excitement correlates with fragility. The more a system’s value depends on belief in a grand vision, the more vulnerable it is when belief wavers.
If you’re building or buying AI today, tilt boring. The unsexy application with clear ROI will outlive the moonshot with a great demo. This isn’t an argument against ambition — it’s an argument for anchoring ambition in something that pays rent. Every ambitious AI project should be able to answer: “What does this do if the grand vision takes twenty years?” If the answer is “nothing,” it’s a winter casualty waiting for its season.
Lesson 8: Humility Scales Better Than Hype
The final lesson is about communication, and it’s the one the current boom most needs. The researchers who navigated the winters best were the ones who said what their systems couldn’t do. That honesty looks like weakness in a boom — your competitor’s demo is flashier — but it’s strength in a downturn, because you never have to apologize for promises you didn’t make.
Hype scales attention; humility scales trust. Attention is what you need to raise money in a boom. Trust is what you need to survive everything else. The field’s current communication culture — every release framed as a revolution, every benchmark as a breakthrough — is optimizing for the boom and borrowing against the future.
I try to practice this in my own writing: say what’s impressive, say what’s limited, and let the reader decide. It’s less exciting. It also doesn’t require retractions later. The winters taught us that credibility compounds and hype depreciates. Invest accordingly.
Frequently Asked Questions
The key lessons: winters are caused by broken promises, not failing technology; funders overcorrect after hype; demos create debts that come due; research continues quietly during freezes; winter kills companies but not knowledge; boring profitable AI outlives exciting AI; and humility in claims scales better than hype.
The first AI winter ran roughly from the mid-1970s to the early 1980s, and the second from the late 1980s into the early 1990s. Each lasted about a decade, during which funding collapsed but research continued at a reduced scale under different names.
No. Research continued with less funding and prestige, often under different labels like “intelligent systems” or “informatics.” Foundational work — including the refinement of backpropagation that powers modern deep learning — was done during the cold years by researchers motivated by the problems themselves.
They apply directly: today’s boom shows the same warning signs of inflated promises and demo-driven expectations. The lessons suggest anchoring claims in deployed reality, building boring profitable applications alongside ambitious ones, and communicating honestly about limitations to preserve credibility.
The boring, useful systems survived — pattern recognition, optimization, statistical methods with clear commercial value. The knowledge, papers, and trained engineers dispersed into other fields. What died were the hype-dependent companies and the grand-vision projects with no intermediate value.
