How Many Burgers Will The Robots Eat, Part Two

"Art is easy. Fast food is hard."

apparently this is what one of my friends said to another one at a temporary concours urinal, and then the second friend fired the first one

Part One is here.

There’s something to be said for productive ignorance. The always-remarkable Curtis Yarvin just wrote a piece titled “Remote Hands Are The Future Of Physical Work” in which he argues that someday, thanks to technology, various construction tasks and whatnot will be done by… enormous humaniform robots. As a sci-fi idea, it’s first-rate; so much so that it appears with tiresome regularity everywhere from the original Shogun Warriors/Gundam to, ah, well before that in various Hebrew tales of the Golem and whatnot. Yarvin enthuses that

Caterpillar’s biggest general-purpose bot is two hundred feet tall and if you give it an elephant, a blast furnace, and the right knives and stuff, it can kill and clean and cook and serve the thing almost just as fast as your country cousin can process a rabbit.

Unfortunately for humanity’s plans to fight “kaiju” monsters and whatnot, the so-called square-cube law makes human-looking creatures remarkably implausible, if not downright impossible, much past the size of a human. This is true whether we are talking mechanics or biology; I’m not even sure the “Engineers” of the Alien franchise wouldn’t have immediately stroked out when they stood up. The original one in the 1979 was about 15 feet tall and the one in Prometheus is nine feet tall.

Don’t be too depressed, because this is also why you don’t spend much of your day in pitched battle with ten-foot-long cockroaches. They don’t scale, you see. As an insect grows, the required exoskeltal material goes from chitin to… titanium… to carbon fiber… to the kind of imaginary materials of which Iain Banks makes such gleeful use in his Culture books.

I mention all of the above for two reasons. The first is to calm down any Starship Troopers nightmares you might be having. The second is because Yarvin makes a similar, but far more plausible, claim elsewhere in the same article:

Making art is easier than it looks, but manipulating objects is harder.

For humans, writing a great pop song is harder than folding a shirt. For computers, folding a shirt is harder. Generative AI can amaze us with its creativity, but it cannot help us in driving a car, folding a shirt, or planting a tree.

Managing moving objects in the real world seems to always ramify into a set of obscure corner cases in which there is not enough training data. Generative AI works where errors—the wrong number of fingers in a painting, math errors in prose—do not matter. If it was as bad for a text generator to make a logical mistake as for a self-driving car to crash, no one would use text generators.

For the record, I strongly, strongly, strongly object to the use of “AI” to describe any of the bullshit you see floating around social media. It is the modern equivalent of Aztec superstition — or, if you prefer, the idea that Jesus can take the wheel, something I personally believe but which in modern society is considered a mark of idiocy — and it is one of the most dangerous things happening in society today.

(Which is impressive, because arguably there has never before been a time in history where so many different destructive forces were working at the same time to reshape human interactions.)

There is no more “intelligence” in the “AI art generator” or “ChatbotGPT” than there was in a “Centipede” arcade game. Less, perhaps, because the “AI” in Centipede was king of its chosen domain, effortlessly capable of crushing any human opponent at any time it “wanted” to. The only reason “Centipede” was fun was because the computer opponent is hobbled to some percentage at all times. Otherwise you’d put in your quarter and be immediately pinned between a bouncing spider and falling flea. By contrast, ChatbotGPT can barely string words to the level of Brad Brownell or Brett Berk, although it tends to make fewer grammatical errors and/or commit fewer outright inanities.

The media loves to make hysterical claims about ChatbotGPT — it can solve college exams! it can write functional malware from scratch! — but the very nature of these claims tends to undermine the idea that this is “artificial intelligence”. As does the original meaning of “GPT”, which is

General

Predictive

Text

All the early models did was “read” a bunch of stuff, note which words tended to follow each other, and then do the same thing from a “seed” prompt. This didn’t feel important or disruptive (read: angel investment worthy) enough, so now GPT rather hilariously stands for

Generative

Pre-

(trained)

Transformer.

This accomplishes the core “woo” mission of making the product sound horribly obscure and futuristic. The bar for this kind of thing is set really low, at the “Washington Post reporter” level. If an angel investor sees something in the mainstream media about how artificially intelligent a product is, he will invest. Thus the name change. “General Predictive Text” sounds like a feature in Microsoft Office. “Generative Pre-Trained Transformer” sounds like Optimus Prime — and who wouldn’t want to invest in Optimus Prime?

GPT works because most communication follows general templates. For instance, in the case of my friend Rodney the phrase “fine-ass” is invariably followed by “bitch”, and the phrase “quiet as” always prefaces “kept”. With these two rules, you could program RodneyGPT in a few lines of Perl, and the reporter from WIRED would be very impressed. “Why, it captures the heart of the Black experience!”

The same is true for the various “AI” sketches and drawing programs. The only difference is that they manipulate images in predictable ways rather than words — this is no harder or easier for a computer, because it all resolves down to data and equations whether they are “vectors” of text or “vectors” of images. Thus, the oft-discussed “AI image” of “salmon in a river”:

I want to believe that is real, but I would be equally satisfied if someone Photoshopped it, because that would be an instance of human creativity. An “AI” is too stupid to have context for “salmon in a river” unless you provide it, but a human being is capable of generating humor from a knowing mismatch of context, as in the following joke that was very popular in my youth:

A young man walks into a bar, orders five shots of whisky, and quickly downs them all.

The bartender says, "Whoa there buddy. That's a lot of shots. What's going on?"

The man replies, "First blowjob today."

The bartender says, "Well congratulations! I'll tell you what, have another drink. It's on me."

"No thanks." says the man, "If the first five didn't get the taste out of my mouth, I doubt the sixth is going to do much."

This joke becomes less funny every year because society is progressively normalizing the idea of men sucking each other’s dicks. Or, perhaps, re-normalizing it, as the activity was hugely popular prior to about 1850 in the Anglo world, and remained so among the upper classes. The explicit prohibition of homosexuality for the middle and lower classes is a recent thing that just happens to coincide with the greatest technological progress in human history. As soon as homosexuality became more accepted, the rate of scientific discoveries and engineering implementations slowed. You can literally chart it, although in intellectual honesty I suspect it’s one of these. Or maybe not. Had Alan Turing been permitted to live his truest self, surely the Enigma would remain un-cracked, the same way I burned about five novel’s worth of creativity managing a stable of married girlfriends! Anyway, the joke is only funny if you automatically assume a man is the recipient of the blowjob in question. Once that context is gone, it’s just another story. As a Primecitizen of 2023 you intrinsically understand that. A computer never will.

I would strongly disagree with Yarvin’s assertion that “art is easy” for AI, but folding clothes is hard. There is no evidence that any computer in history has ever created “art”. At best, it is creating “craft”. The whole purpose of art is for the artist to send a human message to the viewer. That’s why some of Jenny Holzer’s ridiculous sentences qualify as “art” but nobody has ever thought of a Canon digital camera as an “artist” even though it is literally creating a digital image based on what it “sees”.

Folding clothes, however, is awfully hard for a computer, because computers — this should sound familiar — have no ability to understand context. A clothes-folding robot is easy fooled by: changing the fabric, moving the shirt around a bit, turning the shirt upside down. Now, you can code for these things, but code you must. Whereas even the least intelligent GAP employee can understand on Day One how to fold a shirt that is “upside down” in front of her.

There’s a bit of 80/20 principle to this. In another life, I single-handedly wrote the Web sales interface for a partnership between UPS and IBM. The idea was that IBM salespeople would visit UPS locations, evaluate the needs, then file the order via my website. I was paid for this via an authentic International Business Machines green-line check that I “blew up” on a copier and framed, because I was very proud of the fact that I had gone from sitting unemployed in a county jail on an assault charge to buying a new-build home using an IBM check for the down payment, all in something like 14 months.

I wrote the whole site and order-processing mechanism in one long thirty-hour code session fueled by Mountain Dew and Anita Baker’s best two albums on repeat. Rolled it out the next day, in production — something nobody does anymore. It was a big hit and it met almost everyone’s needs… but not everyone’s needs. I spent the next two years writing functions to handle the “corner cases” that had been easily and transparently handled by IBM’s inside sales staff prior to my website’s appearance. Eventually, “my people” at IBM moved on to other gigs, and the site was deactivated in favor of a massive e-commerce website running on a copper-core zServer or something like that, programmed by three dozen overseas resources in a year-long effort at eight-figure cost. Which got them to where I’d been on Day Two, so they ended up forming a whole department to maintain the site after the fact.

My bitterness and/or self-ego-stroking about this story aside, there’s an important lesson to it. The only things that keep humans in the labor game are:

0. The “20” part of 80/20;

1. consequences of error.

Yarvin notes, rightly so, that self-driving cars are held to higher standards than “AI Art Generators” for narcissistic Instagrammers. Can we collapse both 0) and 1) into a single idea? Yes we can: context. 80/20 corner cases are a result of not understanding context, as are consequences of error. There’s a reason that a “full self driving” Tesla will cheerfully ram you into a stopped tractor-trailer at 75mph; it doesn’t have any context for anything. It can’t “know” that this should be avoided. Yet even an 80-IQ low-function human being can usually avoid driving directly into a stopped truck…

Note that Magnus’s error here wasn’t entirely a context error. He knows that it’s a bad idea to run into a truck. He just made a mistake handling the car. By contrast, an “AI Porsh drifter” probably wouldn’t let the slide get out of control — but it might not “know” that it’s bad to hit a truck. Or it might not recognize the truck as such.

(Alternately, you could argue that Magnus did make a context error, in that he prioritized “getting the shot” over the obvious risk.)

Let’s leave the “Urban Outlaw” be and return to something he is too rich to ever worry about: making hamburgers in a fast-food restaurant. In theory you could create a just-in-time system where a completely automated McDonald’s takes in precisely-packaged “precursor” items, prepares them to known standards, then manipulates them into customer hands. It doesn’t sound as hard as, say, making a microprocessor. Yet “chip fabs” are far more automated than McDonald’s — and it’s not necessarily because their products have a much higher profit margin. The context inside a clean-room chip fab is several orders of complexity beneath that of a McDonald’s.

I suppose that with several million dollars’ worth of automation you could provide a limited McMenu… until something gets stuck, or a machine falls out of tolerance range, or the delivery truck accidentally backs in at the wrong angle. My experience in fast food suggests that something along these lines will happen on a daily basis.

Therefore, you will always need a context interpreter on hand for these failures. And here’s the worst part, at least from McDonald’s perspective: that person has to be far more skilled than any Current Year McCook. He has to know all the machines intimately. Has to be able to diagnose any problem and repair it immediately — because unlike a current McDonald’s, which can keep serving burgers when the fry machine is out of order, this robot restaurant will be riddled with single points of failure. And, of course, he will have to address customer complaints, because the rowdier sections of American humanity will burn the place to the ground otherwise.

It will be a job not unlike being a combat lieutenant; long periods of tedium punctuated by utter terror in which only your personal decision-making ability can prevent total chaos. What do you think it needs to pay?

My first guess: $100k. Or more, depending on how much repairing you have to do. And you need three of these people. Four, really, because you have to account for vacations and schedule difficulties. They need to be as smart as junior sysadmins, a job which currently pays $80-120k and never involves cleaning a toilet or being punched in the face. (Well, not usually.) So you might need to pay them more than a junior sysadmin makes.

You can run a four-person “Fight For $15” staff for the same kind of money, which is all you get at a lot of McDonald’s nowadays. So the cost savings seem a bit… tenuous, particularly once you factor in the price of the machines.

Now here’s the good part: In my proposed scenario, McDonald’s would all of a sudden become a middle-class job! If you’re willing to rent in the Midwest, you can just support a family on $100k a year — which is a sentence that seemed more sci-fi than the Shogun Warrior robots as recently as 2019. And that change in staffing would eliminate a lot of jobs that currently go to ahem, undocumented citizens. Which is not a good thing for the people currently working them, but perhaps it would slightly decrease the incentives for potential future undocumented citizens to endure horrifying personal risks on the way into our tottering empire.

Here’s what I think. I think that American corporations are far too in love with the idea of robot labor for them to be even remotely deterred by any of the above. So instead of adjusting the situation to fit reality, they will adjust reality to fit the situation. The obvious thing to do is to reduce fast-food restaurants to the level of an “automat” pre-packaged dispensary from the Sixties or Seventies. The food can be prepared centrally; you’ll still need people but the economies of scale are much greater. And then you automate the delivery, with one fellow in a van driving around solving problems in dilatory fashion like today’s ATM repairman.

(That’s ATM repairman, not a2m repairman, which is a different kind of job that mostly appears in fan fiction.)

We’ve already bifurcated fast food into Poor Folks (McD’s, Wendy’s, Rally’s, et al) and Real People (Five Guys In-N-Out, Chik-Fil-A, Starbucks) categories. You automate the former, make the food pre-packaged trash, keep the prices where they are. What are the customers gonna do? Pick up The James Beard Cookbook and start making healthy meals on the 17th floor of a Baltimore project or back lot of a trailer park? They’re sure as hell not gonna go to Five Guys, where one meal costs more than a fifteenth of any monthly welfare or basic income benefit in this country.

This solution will combine the worst aspects of modern automation and the worst aspects of how we treat people in 2023. Expect it to spread like wildfire at some point. Too attractive a concept not to, really. Is there a consolation on offer? Only this: history tells us that oppressed communities are often hotbeds of unique artistic expression. The more you crank up the pressure cooker on everyday people, the more interesting the results. So maybe we’ll get great art out of the process. Or maybe we’ll get a revolution. Ever notice how the two often go hand in hand? If you did, that’s a uniquely human observation. I have a suggestion for a password to fool the AI watch bots, when we meet to act on it: salmon in the river.