By Marc J. Elzenbeck
Hard to believe at this remove, but blues music was once considered a fringe, largely forgotten art form. It was also commercially dead. Then the comedy duo of John Belushi and Dan Ackroyd formed “The Blues Brothers” in 1978, and their album “Briefcase Full of Blues” is credited with almost single-handedly re-vivifying the genre. Even harder to believe, their eponymous blockbuster movie that came out in June of 1980 smoothly introduced the concept of AI into everyday life. It also beat late tech-themed titles like “Blade Runner,” “War Games,” and “The Terminator” to the punch. How did it do that?
The windup: Elwood Blues and his brother “Joliet” Jake run a red light in front of a suburban Chicago
police cruiser parked out of view. The cruiser hits the siren, Elwood pulls over after passing by the Nelson
Funeral Home just east of O’Hare Airport. Elwood remarks, “Man, I haven’t been pulled over for 6 months. I bet those cops have got SCMODS.” Jake: “SCMODS??” Elwood: “State/County/Municipal Data System.”
We cut to the police cruiser, where they do indeed have what Elwood dreads. We see a dash-mounted keyboard between the two officers. Almost the same as the machines look now. The following message types itself out on its terminal:
“Blues, Elwood
Illinois License: B263-1655-2187
Currently Under Suspension
Warrants Outstanding: Parking 116
Moving Violations: 56
Arrest Driver … Impound Vehicle”
One of the officers walks up to Elwood’s window and courteously informs him that his license is suspended, and requests that he step out of the vehicle. Elwood glances at Jake, starts up the Bluesmobile, a battered police surplus Dodge, and an epic car chase ensues. Watching it in the theater, the scene made a deep impression on me.
As an infamous local hot-foot, the last thing I wanted was for cops to have SCMODS, or anything at all like it. Since I was also fooling around with computers, I kept thinking about it. Keeping all that mainframe-stored data straight was one thing, and kind of boring. But at what point did a computer decide, “Arrest Driver. Impound Vehicle?” Clearly, this was some form of expert system. In fact, SCMODS was a superb example of a Finite State Machine, part of what we’ve come to jumble into the catch-all brand ”AI.”
Would 116 parking tickets be enough to get arrested, or just the vehicle impounded? Fifty-six moving violations would be plenty for both, but where were the thresholds? What were the weightings? Some algorithms might have it say, “Give Driver Warning. Use Discretion. Driver Legally Blind. Escort Driver to Bat Cave.” My parents’ rather unsporting threshold was one speeding ticket, then no more driving privileges. So going 89 in a 20 MPH school zone would qualify both for my parents and State penalty points to suspend my license.
I kept thinking about it. Nine years after “The Blues Brothers” helped me find the music of Pinetop Perkins and Elmore James, I was jet-setting through the financial centers of Europe, a young information technology engineer at a start-up. I simply called up CEOs and Managing Directors at big banks and insurance companies to tell them I was making and selling systems to automate risk analysis. So they could make more money with less effort. They listened.
My systems looked a lot like SCMODs. AI’s huge advantage is speed. You can scan a whole lot of data, sort and boil it down to associate past good decisions with a limited number of outcomes. That’s a finite state machine. If you’ve got clean data – and that’s the big “if” – you can flag exceptions and route the tougher cases to where they should go: to humans with the contextual experience to know what they’re doing. It’s automation. It’s what we do: figure out how to do something faster and better.
The AI priesthood class constantly rambles on about large language models (LLMs), neural networks, machine learning, internet scraping, hidden Markov this and generative AI that. Gobbledy … fricking … gook. It’s database look-ups. It’s mostly SCMODS, a collection system that kept track of your parking tickets and moving violations hooked to weights, rules, and some rudimentary math. Social credit score? Surveillance plus math.
Speaking of LLMs, where AI gets into a great deal of trouble is when you ask it questions. Especially open-ended questions. Ask a 5 year-old about the Moon or Mars and you’ll get a lot of variability. Language is words, which are signifiers of meaning, which in turn spring from cognition and context. AI does not do well with strings of logic, cogito ergo sums.
As we all know, one word can simultaneously be an insult, a compliment, both, or a nebulous place-holder. (Kek.) So, it is no surprise that OpenAI’s best current version, GPT-4.0, when tested against 4,326 factual questions with only one clearly correct answer in subjects like science, art, and sports, scored 38.2%. Anthropic’s top model, Claude, got only 28.9% correct because its algorithms are set to be less overconfident than OpenAI’s.
The mistakes in these test batteries don’t even repeat in the same way from one try to the next. Mixing LLM into regular old Boolean search functions is making the error rate skyrocket. Try querying Google with “How many furlongs from San Francisco to NYC?” The closest hit you’ll get is “Flight time from San Francisco to New York – FlightSphere.”
Managers are always looking for easy answers. For magic beans and silver bullets. The mayor of Chicago might want to hook SCMODs up to harpoon-equipped robots with which to affix the fleeing Elwoods and Jakes of the world. Sometimes they would get that old Dodge, and sometimes they’ll snag a Tesla.
