There was a moment, roughly between 1983 and 1987, when a serious AI researcher worked at a Symbolics. Not on one. At one. The machine was the workstation, the operating system, the editor, the debugger, the version control, the inspector, and the language all at once, and the language was Lisp from the microcode up. The LM-2 in 1981 was the first commercial repackaging of MIT's CADR design, and by the time the 3600 line shipped you could buy a desk-sized Lisp environment with over five hundred thousand lines of system code written in the same language you were writing your own code in. Down to the kernel. That is still an extraordinary thing to think about.

Symbolics was a spinoff from the MIT AI Lab, one of two companies that came out of a hacker schism around the turn of the decade, the other being Lisp Machines Inc. Symbolics got most of the hackers and most of the money. For a few years the bet looked obviously right. The Defense Advanced Research Projects Agency had decided that expert systems were a route to military advantage, and the Strategic Defense Initiative meant DARPA was buying. Symbolics machines went into labs and skunkworks and Star Wars feasibility studies. Revenues climbed. The company signed long real estate leases in California on the assumption the run would continue.

It did not. The mechanism is worth being specific about, because the standard summary, "AI winter killed it", obscures more than it explains. Two things happened in parallel. The first was that expert systems, the thing Symbolics had been sold as the platform for, turned out to scale badly. Rules accumulated, interactions between rules produced surprises, knowledge engineers were expensive, and the systems could not learn. The second was that Sun Microsystems and the rest of the RISC workstation industry caught up on raw compute and then surpassed Symbolics on price-performance, while compiler people figured out how to make Lisp run respectably on general-purpose silicon. A purpose-built Lisp machine had been the only way to run Lisp fast; once it wasn't, the entire commercial premise dissolved.

DARPA noticed before the company did. Funding cuts started biting in 1987. Russell Noftsker, the co-founder running operations, ended up in a power struggle with Brian Sear, the CEO the board had hired in 1986, over whether to pivot to selling Lisp software that ran on Sun boxes or double down on Symbolics' superior hardware. The board fired both of them. Sales fell off a cliff, the California leases turned into millstones, and Symbolics filed for bankruptcy in early 1993. A small remnant company kept maintaining Open Genera, a virtual machine version of the OS, for the customers who could not afford to migrate off it. By 1995 the Lisp machine era was over.

The instructive bit is that no single decision killed Symbolics. The technology was genuinely impressive. The customers were real. The leadership had founder energy and MIT pedigree. What killed it was the assumption that a specialised hardware platform could remain the cheapest way to run a particular kind of software, in a decade when general-purpose silicon was doubling in capability roughly every eighteen months. The platform was a moat that was also a clock. Every Moore's Law tick narrowed it.

You can see the same shape now in the argument over whether GPU accelerators stay a category, or whether the workload that currently justifies them gets absorbed by whatever comes next. Specialised hardware is enormously valuable while the gap is wide, and a liability when it narrows. The companies that bet on the gap staying wide are the ones who sign the California leases. There is no indication this lesson generalises, and every indication it does.

Sources: