The Second AI Winter
🥶 The Second AI Winter
🖥️ The Lisp Machine Crash
Companies like Symbolics and Lisp Machines Inc. (LMI) had built entire businesses around specialized workstations engineered to run LISP fast — hardware tuned for exactly the language the expert-systems boom depended on. For a few years in the early 1980s, that specialization was a genuine performance advantage.
It didn't last. General-purpose workstations from companies like Sun Microsystems and Apollo, built on conventional processors, improved rapidly — and paired with better LISP compilers, they closed the performance gap while costing a fraction of the price. By the mid-1980s, buying a purpose-built LISP machine no longer made economic sense when a cheaper general-purpose workstation could run the same software nearly as well.
Symbolics stock peaked in 1987 and then collapsed almost immediately as the specialized-hardware market it depended on evaporated. Across the LISP machine industry, companies that had raised significant investment on the strength of the expert-systems boom were bankrupt, acquired, or wound down within just a few years — an abrupt, market-driven crash rather than a slow fade.
⚙️ Why Expert Systems Stopped Scaling
The narrowing trick that revived the field in Chapter 6 had a hidden cost that only showed up once systems grew large. A handful of rules is easy to reason about; thousands of rules start interacting with each other in ways no one predicted when writing any single rule in isolation. Adding a new rule to fix one case could silently break the reasoning behind an unrelated one.
Even XCON — the very system Chapter 6 held up as proof that expert systems delivered real commercial value — wasn't immune. Its rule base reportedly grew from a few thousand rules to somewhere around ten thousand over the following decade, and keeping it consistent and correct became a substantial, ongoing engineering effort in its own right. Many companies that built smaller in-house expert systems found the long-term maintenance cost eventually outweighed the benefit, and quietly abandoned them.
📢 A Term Coined Before the Crash
Here's a detail worth sitting with: the phrase "AI winter" itself — introduced back in Chapter 5 to describe the Lighthill-era collapse — wasn't actually coined until 1984, at an AAAI (American Association for Artificial Intelligence) conference panel featuring Marvin Minsky and Roger Schank. They used it as a warning about an approaching pullback they saw coming in the then-current expert-systems hype. The term predates the very winter this chapter describes — a rare case of researchers naming a boom-bust pattern accurately enough to predict its next occurrence a few years in advance.
💸 DARPA Pulls Back
Government funding followed the same trajectory as private investment. DARPA's Strategic Computing Initiative, launched in the early 1980s partly in response to Japan's Fifth Generation project (Chapter 6), had promised ambitious near-term military and industrial applications of AI. When those promises didn't materialize on schedule, funding was sharply reduced in the late 1980s — echoing, on a smaller scale, the exact funding-versus-promises dynamic that produced the Lighthill Report in Chapter 5.
📜 Two Winters, Compared
| First AI Winter (Chapter 5) | Second AI Winter (This Chapter) |
|---|---|
| A research-community reckoning — the Lighthill Report assessed unmet promises | A market correction — commercial hype met real maintenance costs and cheaper competing hardware |
| Hit government-funded academic research directly | Hit a private industry built on venture capital and corporate contracts |
| Triggered by a technical limit — combinatorial explosion in symbolic search | Triggered by a practical limit at scale — rule interaction complexity — plus being economically outcompeted |
Two winters now, separated by roughly a decade, each following the same underlying shape: bold promises attract funding and investment, the field delivers real but narrower results than promised, and the gap between promise and delivery eventually triggers a collapse. Minsky and Schank were able to name the pattern accurately enough to warn about it in advance. The next chapter asks the harder question this course has been building toward: what exactly is the repeating mechanism behind that pattern, and does recognizing it in advance actually help anyone avoid it?
🤔 Questions to Sit With
The Lisp machine crash was primarily an economic story — cheaper general-purpose hardware caught up — rather than a story about AI research failing on its own terms. Does that make this winter feel less serious than the first one, or does a purely market-driven collapse carry its own kind of warning?
Minsky and Schank named "AI winter" as a warning in 1984, and the winter they warned about arrived a few years later anyway. What does it tell you that accurately predicting a hype cycle didn't seem to prevent it?
Rule-based systems became hard to maintain once they scaled past a few thousand rules, because new rules could interact unpredictably with old ones. Can you think of other kinds of complex systems — software or otherwise — where growth itself, rather than any single flaw, becomes the main source of fragility?
🎯 What's Next
Next chapter: Lessons From Two Winters — the final chapter of this course, naming the recurring hype-cycle pattern directly and bridging to Course 3's statistical, data-driven era.