One of DeepAI's renderings of “2007 pontiac solstice with 1953 corvette body rear.”

Artificial Intelligence Revisited

It’s been over a year since I talked about artificial intelligence, but that's entire generations of the technology in the past so it’s time to revisit it.

Last time the major theme was “Garbage In, Garbage Out,” or how dumping all kinds of data into AI models without the computer knowing what's fact or fiction yields terrible results.

What also hasn't changed is what AI is and what it is not. Mike Hall, host of Skeptics With a K (episode 390) summarized AI programs best as “really fancy autocompletes.” Matt Watson calls it “a collection of IF statements on a massive scale.” They take data fed into them, and respond to your prompts with what it thinks you want to hear. They do not think, create, or learn on their own. And they take an enormous amount of energy to do so.

Google’s AI currently lacks the ability to look at the data with basic context, asserting last year that thirteen U.S. Presidents attended the University of Wisconsin at Madison. You might not be surprised to see the list include some very common names, like Andrew Jackson and William Harrison, who “graduated” in 2005 and 1974, respectively. They also died in 1845 and 1841, before the University was founded in 1848.

When you see that Andrew Johnson received an astonishing fourteen degrees from the school from 1946 to 2012, you come to realize that it’s just seeing names matching our former Presidents and assuming it’s the same person, no matter how easily the human mind can still see that these are obviously different people.

Some AI errors are simple middle school math, such as when ChatGPT thought that 9.11 was larger than 9.9. These can be fixed, as math is the most “pure” in the sense that something is either right or it is wrong and computers are built to do math at their very core. Someone just flubbed a line of code. It's probably already been fixed. I hope.

Some of the errors might be misguided attempts at contextualization, such as when ChatGPT insisted there were two R’s in strawberry. Yes, there’s a double-R, but the three total. Some are because it mixes up words that have multiple definitions, like Google AI did when it said, “sauces add flavor and texture to dishes, while dressings are used to protect wounds.” These can be figured out with some more complicated, but doable improvements since Google is already pretty good at figuring out context from search terms in its flagship product.

Sometimes the AI takes jokes as facts. Google said, “Yes, John Backflip is a real person. According to a popular story, John Backflip was the first person to perform a flip in 1316, and people from all over Europe came to see him.” These will be a little harder to correct, as computers aren't good at discerning tone via text. Hell, a lot of people aren't good at discerning tone in people’s voices. This will take looking into human categorizing of the data fed into AI models, and that's quite labor intensive (read: costly).

Other times you just don't know what the hell it's even trying to do.

The source of all these symptomatic problems is that every major corporation is trying to incorporate a half-developed The Next Big Thing into everything that they do to appear more advanced than the competition. Particularly problematic is Google’s use of the tool. The examples I picked out were because they were funny. Even the underdeveloped mind of an elementary school child could figure out these results weren't right. But there are plenty of other examples where the errors aren't so glaringly obvious, and people might take the errors as fact. After all, it came from Google!

But what happens when AI makes a mistake that's less obvious, like when it has recommendations for coolant temps and radiator size for a 1969 Volkswagen Beetle. Most people probably don't know that the Beetle was air cooled and didn't have a radiator. To a car guy, this error is as obvious as JFK not graduating from UW-Madison 30 years after he was shot. But to anyone else? Not so much.

So where did the error come from? Did it mix up the Beetle with a different model? Was it mistaking the old Beetle for the New Beetle? Was it reading an article about engine swaps? These aren't as easy to fix as the previous problems, a lot of errors will require both knowledge of AI and knowledge of the subject matter.

And then there are just so many errors when it comes to AI recognizing AI. This is a problem not only for the poor souls who are accused of academic cheating, but also for the future of AI itself. There’s a problem of AI “inbreeding” where using AI generated data to train AI models results in worse and worse models until the AI collapses. If AI checks its intake of the increasingly AI-cluttered internet, and accepts AI-generated content as human-generated while falsely excluding human-generated content, inbreeding will accelerate.

We've already seen AI tools used for election shenanigans. Then there's the risk of all these AI errors and misuse eroding trust in the basic institutions of the internet. Or worse yet, being misused without ill intent by the fundamental insitutions of our society in such a way that we lose what little faith we still have in them. It is in these environments that evil people thrive and take control.

Plus we now have to worry not only about the intentions of the humans that wield this power, but also the intentions of the artificial intelligence programs themselves.

So what does this all mean for automotive history? Well, it means that I still have a hobby/side job! It’s going to take years of human labor to categorize and contextualize all of the data AI models are using to learn from. If I do a good job of keeping AI-generated slop out of The Crittenden Automotive Library, it will become a more important and trusted source of information.

Even if most of these problems were solved tomorrow, there is still so much that isn't online yet. Scanning, proofreading, and generally turning printed material into computer-readable data. Then the humor needs to be marked, so that it can be read in context and not mixed with straight facts. Lies can be valuable data, but only when they're connected to their corrections and also not taken as fact. Indeed, knowing the difference between hearsay and direct evidence will be crucial for AI to make decisions between conflicting narratives, and while wildly imperfect humans are still currently better than AI at this.

The question is, will these problems be fixed? In a profit-motive-driven society, does being right even matter?

Well, it still matters to me.

History Beyond the Bumpers

The Crittenden Automotive Library includes information from all aspects of automotive transportation and competition. This section highlights topics related to automobiles other than vehicles themselves.

Now, on to something more positive about artificial intelligence. I have ideas. I can describe the concept cars in my mind in great detail, but I draw like the average third grader.

One of my best ideas from 2009, one that a lot of people apparently thought of independently at the same time, was to rebody a Pontiac Solstice as a modern C1 Corvette. Of course all I could do was crank out 650+ words on the idea. In 2021 Abimelec Design came up with a graphic rendering of their version. I loved everything about their version except for the wheels, it was amazing seeing on the screen something I had only seen in my mind for over a decade.

With top shelf AI image tools like Midjourney looking more and more realistic, the possiblity of creating these concept cars with a few words of an idea is getting close. Even free tools are getting better, as DeepAI created this month's featured image off of the prompt “2007 pontiac solstice with 1953 corvette body rear” and it looks pretty good. The black plastic interior wouldn't be my choice, but it makes sense given the prompt. Also the tire width, tail lights, mirrors, and exhaust are updated to more modern types while the shape resembles the C1 enough to be easily recognizable.

It is, however, currently unable to recognize my idea of putting Mercury's final styling on a Ford Mustang of the time. That would have been a great Mercury Cougar, a little tan leather luxe and a big break from the black plastic and retro styling of the common Mustang of the day. It was a great style, but not for everybody. A new RWD V8 Cougar might have even saved the marque. DeepAI just cranks out different 2010 Mustangs with blurry badging. Oh well. I'll check back on that in a year. Or maybe spend a few bucks to give Midjourney a try.

629.2

The Dewey Decimal System's designation for automobiles falls within the 629.2 range. This section is about the printed materials in The Crittenden-Walczak Collection.

December was a great month for making progress in the new office. I finally reached a corner I haven't been able to since January, and changed the last daylight white ceiling light to a warm white bulb that matches the room's oak-timmed-dusty-archive future. Now that I'm able to get there I reached into a drawer this evening, and randomly pulled out these three 1967 Pontiac brochures!

There's a heck of a brochure collection in three filing cabinets still mostly blocked by boxes, and largely unsorted other than being dumped into different drawers for Ford, Chrysler, GM, and AMC. I'm sure that was something John had intended to get around to someday, but health issues prevented him from going to the basement in his last years.

Normally I'd say these would be a great scanning project, but they're already online from Dezo's Garage! I'm so happy to find another project preserving the printed materials of automotive history, and another one with a “mismatched” internet address (xr793.com) too! The Automotive History Preservation Society used to be at wildaboutcars.com, and of course The Crittenden Automotive Library is still here at CarsAndRacingStuff.com.

So another project for the upcoming years will be sorting, inventorying, and combining collections. Then I'll be able to see what I have that's not available elsewhere on the internet and get it online for us all.

Telemetry

CarsAndRacingStuff.com site statistics.

Oof, I hope that December drop was holiday-related. Having started to put together a better statistical tracker, I'll be able to tell if there are seasonal cycles once I get the data entered.

The Top 5 pages for the month of December (not counting basic index pages) were...

Article: The tricks to resetting a Dodge Grand Caravan Computer

Topic: Chevrolet C30

Topic: 1979 Ford-Lincoln-Mercury Lineup

Topic: Pontiac Laurentian

Topic: Oldsmobile Ninety-Eight

About The Crittenden Automotive Library

The Crittenden Automotive Library @ CarsAndRacingStuff.com, based in Woodstock, Illinois, is a free online collection of information relating to not only cars, trucks, and motorcycles, but also the roads they drive on, the races they compete in, cultural works based on them, government regulation of them, and the people who design, build, and drive them. We are dedicated to the preservation and free distribution of information relating to all types of cars and road-going vehicles for those seeking the greater understanding of these very important elements of modern society, how automobiles have affected how people live around the world, or for the general study of automotive history and anthropology. In addition to the historical knowledge, we preserve current events for future generations.

The Library currently consists of over 896,900 pages of books, periodicals, and documents, over 57,500 individual articles, more than 18 days of video & 24 days of audio, more than 36,100 photographs & other images.

About The Crittenden-Walczak Collection

The combined personal collections of John Walczak & Bill Crittenden provide reference materials for The Crittenden Automotive Library. The collection currently includes 1,436 different book volumes/editions, 2,495 unqiue periodical issues and 861 catalog issues, as well as booklets, brochures, comic books, hero cards, event programs, and 371 hours of video.