Lisp Machine Crash and the Second AI Winter
Dedicated Lisp machine market collapses, AI hardware companies fail, starting the second AI Winter.
Overview
In 1987, as general-purpose workstations and personal computers improved in performance, the market for expensive dedicated Lisp machines rapidly shrank. AI hardware companies like Symbolics and Lisp Machines Inc. failed or pivoted. Meanwhile, expert system limitations became apparent: high maintenance costs, knowledge acquisition bottlenecks, and difficulty handling uncertainty and common sense reasoning. Many corporate AI projects failed to achieve expected returns, leading to investment withdrawal. This became known as the second AI Winter, lasting until the mid-1990s. During this period, "artificial intelligence" carried negative connotations in both academic and commercial contexts, and many researchers adopted terms like "machine learning" and "neural networks" to describe their work.
Study Notes
Background, mechanics, impact, and key takeaways in one read
Background & Context
The second AI winter lasted roughly from 1987 to the mid-1990s. Its immediate trigger was the collapse of the specialized Lisp-machine market: as general-purpose workstations and personal computers became dramatically faster and cheaper, dedicated hardware companies such as Symbolics, Lisp Machines Inc., and Thinking Machines lost their customers and folded or were acquired. The deeper cause was the disappointment of expert-system promises: knowledge-acquisition bottlenecks, high maintenance costs, and poor handling of commonsense and uncertainty meant many corporate AI projects failed to deliver expected returns. The Fifth Generation project's overruns and delays deepened the gloom. Around 1987 governments and companies sharply cut AI funding, and "artificial intelligence" became a word to avoid in business and academia, with research migrating to more pragmatic labels such as machine learning.
Key Points
- 1From about 1987 the specialized Lisp-machine market collapsed and firms like Symbolics folded or pivoted.
- 2Expert systems were hard to acquire knowledge for, expensive to maintain, and brittle; corporate projects underdelivered.
- 3Japan's Fifth Generation overruns deepened global skepticism about AI promises.
- 4Government and corporate funding retreated sharply, and "AI" became almost a negative label in industry.
- 5The research community pivoted to machine learning, statistical methods, and connectionism, setting up the 1990s revival.
How It Works
The winter had no single technical cause; it was the intersection of several technical-economic curves. First, hardware: general-purpose processors improved exponentially, so by the late 1980s a workstation could run the symbolic programs that once needed a dedicated Lisp machine, erasing the hardware premium. Second, knowledge: expert-system performance grew roughly linearly with the number of rules while acquisition and maintenance costs grew superlinearly with domain complexity, so the return-on-investment curve turned downward after a few hundred rules. Third, representation: deterministic rules could not naturally express probability and utility, MYCIN's certainty factors were a stopgap, and the lack of commonsense and large knowledge bases made systems fail quickly in open environments. Fourth, Japan's FGCS concentrated huge resources on logic programming, and its delays externally confirmed that intelligence could not simply be engineered on schedule.
Impact & Legacy
The second winter reshaped the field. In the short term it destroyed a generation of AI companies and funding channels, marginalized the Lisp and Lisp-machine industry, and drove many researchers out or into euphemism. In the long term it was a healthy re-centering: statistical learning (SVMs, ensembles), probabilistic graphical models, connectionism, and reinforcement learning matured quietly during the winter, and the later successes from Deep Blue to deep learning were built on that pivot. It also changed narrative strategy: breakthroughs were increasingly announced as concrete capabilities (speech recognition, image classification, question answering) rather than "general intelligence", lowering expectations and raising credibility. Today's periodic fears of an AI winter almost all echo this one: overpromising, vanishing hardware premia, and underestimated knowledge-acquisition costs.
Key Takeaways
- Technology adoption follows cost curves: exponentially improving general technology erodes specialized premiums.
- The scaling bottleneck of symbolic systems is knowledge acquisition, not the inference engine.
- A winter forcibly realigns expectations with reality, destroying old industries while seeding new approaches.
- Announcing concrete capabilities instead of grand narratives is a survival strategy the field learned the hard way.
Glossary
A dedicated computer designed to run Lisp efficiently, with hardware tagged architecture and GC support; it died as general workstations rose.
A period of sharply contracting funding, reputation, and research activity in AI, typically triggered by unmet expectations and commercial failure.
The cost and difficulty of manually eliciting expert knowledge into rules, growing superlinearly with domain complexity; the core bottleneck of expert systems.
Related Links
Related Events
LISP Language Born
John McCarthy creates LISP, becoming the dominant programming language for AI research for decades.
Lighthill Report and the First AI Winter
British scientist James Lighthill publishes a critical report, leading to massive funding cuts for AI research and starting the first AI Winter.
XCON: Commercial Success of Expert Systems
XCON (R1) saves DEC millions of dollars annually, marking a huge commercial success for expert systems.