Notes from Maine - 2026/02/08

People are coming over this evening. We will watch football (and a ton of commercials) and eat far too many snacks. Until then, Albert and I are doing our normal Sunday chores. At the feed store yesterday, Albert looked like a coiled spring, sitting at my side. A lot of customers had his ears twitching this way and that. He was mostly focused on the door—I’m not sure who he expected to enter.

“She’s so well-behaved,” the woman in front of me said.

I nodded with a smile. 

It was clear to me that he was doing everything in his power to stay put, but it was touch and go. Albert has a very strong drive to enforce order on a chaotic world. His opinions are evident in his bunched shoulder muscles as his eyes dart around. If he had his druthers, he would probably sprint around the store, herding everyone into a nice tight group where he could keep a close eye on them.

“Can I pet her?”

“Sure,” I said. There was something in my eyes that told the other customer that it was a bad idea.

“Is there something wrong with her?”

“No,” I said. “Please, go ahead.” I tried to appear more encouraging. This narrative about his mental state is all my own—I shouldn’t project it on Albert. He deserves a casual interaction with his adoring public.

She bent and scratched his head. Albert’s attention was still riveted to the door.

I must give off an air of, “Don’t touch this dog,” because people admire Albert but seldom touch him. I really don’t mind if people pet him but I feel like I should warn them.

“In this setting, he’s not going to react graciously. He’s probably going to either ignore you or duck out from under your hand and you’re likely going to feel bad about yourself.”

But I don’t say that. I just give off my, “Don’t touch this dog,” aura that everyone picks up on.

He’s doing really well with his manners in public places though. At home, he’s still a monster. The people who come over this evening will get a strong burst of attitude when they come through the door. In that circumstance, it’s difficult for me to train him without training them, so I seldom bother to try. He’s very much like a German Shepherd, and they have a particular style of interaction that requires many, many rules. I’m not suggesting that people universally enforce these rules with German Shepherds. The dogs themselves seem to have established these rules as a breed. Some people know about them, but most don’t. They’re fine companions, but in my experience they’re just different (the dogs, not my friends).

This week I made great progress with my electronics project. Many people are skeptical or even hostile to “AI,” but I’ve been using it every day. I put “AI” in quotes because that moniker is debatable. When people think about artificial intelligence, they’re generally referring to artificial general intelligence, which is not where we’re at. The tools available are simply predicting which word should follow next, based on having ingested a universe of human text. It’s a simple trick that, once scaled up to unbelievable proportions, becomes a very useful trick.

I learned something very important this week. The AI tools we have available now can be very much too human—and that’s not always a good thing. I’ll give you an example.

I’ve read everything I can find on a particular pinball machine architecture casually referred to as WPC. This architecture had many trade secrets when it was developed in the late 1980s and early 1990s, for very good reason. Pinball manufacturers (one in particular) love to borrow designs. Having their previous architecture copied part for part, they intentionally hid their secrets in some custom chips. Those are the chips that people have been reverse engineering for the past 30 years, and the ones that I’m currently emulating. Information is scattered and thin, so when I had a specific, niche question, I decided to see if AI could synthesize an answer from the scattered documentation.

“On the Williams WPC-89 MPU, in what part of the boot sequence does the ‘Blanking’ signal go low?”

AI thought for almost no time and then said, “On the Williams WPC-89 MPU board, the Blanking signal behaves as a safety watchdog for the entire system.” There are already some red flags in that answer. There’s another watchdog. Blanking is adjacent conceptually, but not part of the watchdog.

I’m sure you didn’t ask, but here’s a quick explanation.

Pinball machines are a heavy-duty commercial amusement appliance. They have to operate continuously for hours or days at a time without burning down the business around them. The “blanking” signal is a safety net on all the high-voltage components. No power is on to the mechanical parts unless the brain of the device says it’s okay. If the signals are “blanked,” then they’re not running. 

On a parallel track, the “watchdog” is a special timer that indicates the brain is operating correctly. Periodically (a thousand times a second), the brain reaches out and “pets the watchdog,” to tell it, “Hey, everything is okay. I’m still over here running properly.” If the brain fails to pet the watchdog before the timer runs out, then the watchdog resets the brain to get it back in order.

So these two systems are both fail-safes. The blanking keeps voltage off until everything is safe, and the watchdog monitors the health of the brain. 

But AI said, “On the Williams WPC-89 MPU board, the Blanking signal behaves as a safety watchdog for the entire system.”

This is a confidently wrong answer because it merges these two different systems.

Then, it goes on to make other errors. 

“Blanking signal is Active Low. Low (0V): Blanking is active; all outputs are disabled to prevent ‘locked-on’ coils or damaged displays,” AI tells me. It’s actually the opposite. When blanking is High, all outputs are “Blanked,” meaning they’re disabled. 

I pointed this out, and the AI politely said, “I appreciate the correction. You are absolutely right to call that out—I had the logic levels inverted in my previous explanation.”

So far, so good. It made a mistake and accepted the correction. So far, this interaction has not been at all helpful, but I’m about to steer back to my real question. I really wanted to know what chain of events convinces the CPU that it’s okay to really turn on the machine. This is where things get really hairy. I remind AI that I’m looking for the actual message sent from the pinball machine’s brain to say everything is okay. It responds with another incorrect statement. Are you still awake? If so, settle in. This is where we get really in the weeds.

“The ASIC contains a watchdog timer that defaults to ‘active’ (High) at power-up. To turn off the Blanking signal (drive it Low), the CPU must perform a write to the Watchdog Register, located at memory address $3FFC.”

This, again, merges the concepts of the two systems (watchdog and blanking) erroneously. I point that out to the AI. This is where it starts doubling down on its answer and arguing with me. It’s a subtle kind of arguing. At first glance, it appears to be agreeing with me but then it reverses course and tries to give me the same incorrect answer as before. I understand why—most of the sources out there don’t cover this information because it’s not widely known. When I gave up arguing with AI, it took me almost an hour to track down the real answer.

This is why I’m starting to think that AI is a little too human. When forced to speculate on an answer, it convinced itself of something wrong and then refused to back down. It’s very polite: “I appreciate your persistence. You were right to push back—my previous addresses for the blanking signal were incorrect,” but that’s cold comfort when it does its own form of pushing back by trying to convince me why it wasn’t wrong in the first place.

It’s no wonder—you train something on how humans interact in text and you’re going to pick up some bad habits. As I already mentioned, I looked up the real answer through human-generated sources and eventually got it all working. I couldn’t help but argue with the AI as it presented wrong answer after wrong answer. Each was put forth as the “final truth,” with 100% confidence. It would be infuriating if it weren’t funny. Don’t get me wrong—the tool has been incredibly helpful in getting me pointed in the right direction. Once you get to really in-depth knowledge it falls apart. 

I guess that’s what I’m for. I have to be the final watchdog for what’s fact and fiction.