What Did We Learn, Palmer?

I don’t think the people who are building LLMs (Large Language Models) are all that smart.

I’ll start off by noticing that they have no idea what to even call the things. As usual. I’ve been using Chat GPT for a while now. I just call it Chad, because you have to call it something less goofy than Chat GPT, or any of the other attempts to label the things. LLMs? Generative AI? Neural Language Model? Ugh. But everything in the internet age is named by the same sorts of people, and badly. Is there an uglier and less informative possible word for this essay than a blog? But we’re stuck with it because a bunch of people who attend marriage ceremonies held in Klingon said so.

So around the Cottage we call it Chad because it’s close enough to Chat, and it’s very like Brad Pitt’s character in Burn After Reading. It’s amusing in a child-like way, helpful in mostly unhelpful ways, energetic and somewhat obsequious. It can also get you shot in a closet or killed with an ax in the middle of the street if you listen to people who listen to it uncritically: (Some swearing)

That’s a minor quibble, of course. I use Chad often for a lot of things, although I gather from seeing what my nieces use it for on their phones that my Chad workload is a solid minority, as usual. I’ve never described my cat to Chad, and had it turn it into a person, for instance.

So it’s great for silly business. Fine. I find it’s superb for reading the internet for me, and vomiting what I’m actually looking for instead of having to wade through oceans of Search Engine Optimized drivel to find the part number on a washing machine or something similar. It’s not the birth of artificial intelligence, but it is the death of Google, so there’s that.

But back to the point, if I have one. I’ve always suspected that the wrong people are working on all the various versions of Chad, and that’s why its such a muddle trying to figure out where it’s going. These companies have hired anyone who’s good at math, statistics, and computer languages to work on the things. These people aren’t smart. They’re like the IT Crowd version of Chad Feldheimer. Being good at heavy lifting and running in place doesn’t make you smart when you’re looking at “…names and shit and these other files are just, like numbers arrayed. Numbers and dates, and numbers and numbers, and dates and numbers, and I think that’s the shit, man. The raw intelligence!”

Internet math geeks are like plumbers. They’re not dumb, but they have a very narrow worldview. If you ask a plumber to design a house, it will have two bedrooms and fourteen bathrooms. It’s human nature. Ask math geeks what to call a collection of essays, and they’ll call it a weblog. Ask them how to make a machine intelligent, and you get numbers and dates, and numbers and numbers, and dates and numbers, and I think that’s the shit, man.

Now back when I was in high school, and Noah was still looking for his bevel square, they used to acknowledge that verbal skills were more important for high-level thinking than math skills. That’s right, you’d take the SAT, and your final score was your verbal score X 2, plus your math score. It was how a stoner who skipped the better part of the last three years of high school classes became a National Merit scholar while the kids who got an 800 on math grumbled and wondered. The math wizards are the same people who became computer programmers, and claim on Reddit that speling doesn’t matter. Of course, a spelling error in their computer program occasionally destroys an entire company in an afternoon, but C’est la vie! They’ll say it’s all part of their “journey” on their next job application.

So the founders of these companies are ardent bullshitters and sophists, but that’s not the same as being verbally brilliant. Shamelessness is more useful when you’re pitching some vaporware to angel investors. And the glorified statisticians they eventually hire are always busy looking at the digital world as one big spreadsheet.

Now on top of what little I know about the truly technical matters involved, I heard an interrumor that at this point, the people that built the LLMs no longer understand how it works. This attracted my attention, because it made perfect sense to me why that would happen. To stretch my simile, you don’t hire plumbers to pick out your curtains, either, because the catalogs don’t have any porcelain drapes in them. And the math wizards are looking for mathematical drapes in Chad’s output.

So, who to ask about my suspicions? Why not Chad hisself?

Uh huh.

So there you go. We asked Chad some questions about how Chad works, and got some Chad answers. So let’s sum it all up, roll up the ball of wax, and survey the whole megillah: