Computers Built for Thinking

The world it was born into

Lisp, designed by John McCarthy at MIT in 1958, was the language of artificial intelligence research for decades. It treated programs as data, made it easy to manipulate symbols and lists, managed memory automatically and encouraged interactive, exploratory programming. These qualities suited AI research, where programmers often did not know in advance what their programs should do.

They also made Lisp slow on ordinary computers. Automatic memory management, checking the types of values at run time and the heavy use of lists all cost time and memory on hardware designed for numerical calculation. By the mid-1970s AI researchers at MIT were running large Lisp programs on shared time-sharing machines and finding them painfully slow.

The bet

Richard Greenblatt and Tom Knight at the MIT AI Lab proposed a radical solution: build a personal computer designed specifically to run Lisp. Work on the first machine, CONS, began around 1974; a second generation, the CADR, was built in the late 1970s and used in the lab. The hardware supported Lisp’s needs directly, with tags identifying the type of each value and support for memory management. Each researcher would have a whole machine, an unusual luxury at the time.

Around 1979–80, two companies were founded to sell commercial versions: Lisp Machines Inc. (LMI) and Symbolics. Xerox, whose research centre had developed its own Lisp environment, and later Texas Instruments joined the market. The bet was that AI and advanced software development would grow into a large market that needed the best possible Lisp environment, and that custom hardware would provide it.

Why it made sense

Lisp machines were extraordinary. Their environments offered high-resolution graphical displays, windows, a mouse, networking and powerful development tools: programmers could inspect any object, change code while it ran and debug in ways most developers would not see for another decade or more. The whole operating system was written in Lisp and could be examined and modified by the user.

The market looked promising. The early 1980s brought an AI boom. Expert systems, programs encoding the rules of human specialists, attracted investment. Japan’s Fifth Generation Computer project, launched in 1982, alarmed American and European governments, which responded with large programmes of their own, including the American Strategic Computing Initiative of 1983. Lisp machine makers sold to universities, defence contractors, oil companies and AI start-ups. Symbolics, the largest, became a significant company. In March 1985 it registered symbolics.com, the first .com domain name.

How it lost

Two things happened at once. First, general-purpose workstations from Sun, Apollo and others, built on fast commodity microprocessors, improved rapidly, riding the industry-wide gains described in the compute essay. Companies such as Lucid and Franz built high-quality Lisp implementations for these machines. By the mid-1980s a standard workstation could run Lisp acceptably for many purposes at a fraction of the cost of a Lisp machine, and it could also run everything else a company needed.

Second, the AI boom ended. Expert systems proved more limited and harder to maintain than promised. Funding and corporate enthusiasm cooled in what became known as the second AI winter. Around 1987 the market for specialised AI hardware collapsed, by many accounts almost overnight. Symbolics struggled on and filed for bankruptcy protection in 1993.

Explaining both the same way

In 1990–91 Richard Gabriel, a Lisp pioneer, circulated an essay containing a section called “The Rise of Worse Is Better”. He contrasted two design philosophies. The “MIT approach”, which he associated with Lisp, held that a design must be correct, consistent and complete, even if that made it complex to implement. The “New Jersey approach”, associated with Unix and C at Bell Labs, favoured simplicity of implementation, even at some cost to correctness and completeness. Gabriel argued, with evident ambivalence, that the second approach had better survival characteristics: simpler systems spread faster, run on more machines and improve over time.

That is a symmetrical explanation, and its strength is that it calls neither side foolish. Lisp machines were better at what they did. Workstations running Unix were good enough at it and better at fitting into the rest of computing. Where general-purpose hardware improved relentlessly and cheaply, the general system’s advantages compounded and the specialised system’s advantages eroded. Christensen’s innovator’s dilemma describes the same shape from the other side: the cheaper, simpler product, at first inadequate, improves until it is enough.

What the loss teaches

The central lesson is about where to put value when the platform beneath you is being commoditised. The Lisp machine companies’ deepest asset was their software environment, not their hardware. Their fate might have been different had they moved that environment onto general-purpose machines quickly, though with lower margins and fierce competition. Holding on to the hardware meant holding on to the part of the business the market was making obsolete.

The second lesson concerns booms. Lisp machines rose and fell with expectations for AI. A company tied to one market’s mood is exposed to its swings, a warning with an obvious contemporary resonance.

Finally, little of value was lost. Integrated development environments, interactive debugging, garbage collection and dynamic languages, all developed or refined in the Lisp machine world, became standard. The machines died. Their way of programming won.

Sources and limitations

The market collapse of 1987 is widely described but poorly quantified in public sources; revenue figures for Symbolics and others are not given here because reliable figures were not found. Technical descriptions draw on the 1980 paper by Greenblatt and colleagues and on Steele and Gabriel’s history of Lisp. Gabriel’s essay is available on his own website. The interpretation that a faster move to software might have helped is the author’s, and it is uncertain: software companies selling Lisp on workstations also struggled once the AI winter set in.