从 1980 年代专家系统到 LLM：一位工程师对 AI 的反思

In the early 1980s, while working for ETA Systems, I got picked to be our representative to Control Data (our parent company)’s working group on AI. At the time, it was all Lisp and Scheme-based “expert systems,” and my assessment was that the hard part was teaching the systems and the danger of the system making connections that were not factually-based, but which sounded very plausible.

I never really got any deep understanding, as I saw it as an intractable problem, for a number of reasons, not least of which was liability. “If you build an expert system, and it mis-diagnoses someone’s cancer and they die, are you prepared for the reputation hit, and the years of lawsuits?” The systems of the time could still suggest interesting avenues for research, but it would take a human to do the research and see if the machine’s suggestion has any basis in reality, or was just a case of correlation without causation.

This was not seen as a productive direction, as most researchers had plenty of bad ideas, and having a computer generate even more bad ideas just added noise without adding signal. Plus Control Data started having financial problems, so the whole thing was dropped. “Any useful results are still a couple decades off, and we don’t have the resources to spare for this research today.” It reminded me of the people saying that fusion power was going to save us; it’s just another decade of research away (still is, four decades later).

At the same time, a friend was playing with Markov Chain text generators to make plausible sounding “lorem ipsum” text. The results were interesting, and plausible-sounding (just like today’s LLMs), but there was no fact-checking. It was a fun toy for distracting the computer-naïve though, much like Eliza.

Today, we have much bigger models, so the text is even more plausible sounding, but still without fact-checking. And we have the companies run by greedheads who don’t care about whether the LLM is right as long as they’re making money, and shrug off any liability, but the warning from IBM in 1979 is every bit as true today as it was then. “A computer can never be held accountable. Therefore, a computer must never make a management decision.”

Anyway, I’m thinking it’s another case of a waste of research money, but with the greedheads wasting much more money, energy, and resources than would have been possible in the Reagan years. And we’ve got LLMs polluting the web with their slop, so a guy can’t even find out easily that there’s no such thing as alcohol-free port without having to wade through dozens of pages of slop.

Sigh.

Discuss... Reply to this in the fediverse: @[email protected]