Tag Archives: AI

AI testing: The Bad…Time loops, tech support quirks, and drift

The PolyBlog
April 18 2026

By now, most people have seen some form of AI crop up in their tools. The most obvious one is Google’s search engine, which provides results from its AI mode first in the list. You can go pretty far with that prompt, even asking for image creation, although that’s a terrible place to create images (full imaging tools aren’t really available in AI search engine mode).

In my case, I’ve used it for some research here and there, often against a framework I had in mind. More recently, I’ve had it helping me “test” some frameworks. I design a framework for something I’m building or writing, I outline it and paste the outline into AI, and ask it to challenge the framework from the perspective of say gender equity, under-represented groups, or literacy levels. Something more than a grammar check, something less than a full AI partner. When it’s done, I decide if I want to change anything in my approach.

But I’ve discovered some recurring oddities. Not necessarily bugs, just aspects of LLM-based tools that attempt to translate what I’ve said into something concrete.

Time Loops

About three months ago, I was testing Google’s tools to create an image. I eventually moved to ChatGPT to do the same. And both tools had the same problem.

I input a bunch of prompts. Created some sample images. Iterated a few things. All good. Then I told it to “tweak the image” in a certain way, and it said, “Okay, here you go.” But it was the same as the previous image. There was no “change” or iteration.

Okay, I thought, random glitch. Please regenerate the image with the following changes. Enter, whirr, ding. Same image. Huh?

I would then tell the AI that it gave me the same image again. Apologies, whirling indicator, bam! New image, same as the old. No matter what I did, it would not give me anything else.

It felt like a giant glitch. Or Groundhog Day. No matter what I did, same result. I couldn’t get out of the loop.

At the time, I had NO idea what was happening. Was it me? Was it the AI? Was it my browser?

I now realize it’s essentially a memory issue. Each chat in certain tools has an amount of “context” memory built into it. Once that’s full, loops start happening. Things bog down. In some tools, it will say, “Hey, I need to compact, okay?” and it will crunch your chat and go, “all ready!”. Except you have no control over what it ditched. Images perhaps? Instructions you definitely needed it to remember? Gone. In other tools, it compacts without even telling you.

The AI experts advise that where you had it generate a lot of “assets” (pictures, documents, etc.), it’s better to start your next phase with a clean prompt. You can cheat, though … if you ask an AI tool for a “handover” note, it will generate one you can prompt into the next chat, while it quietly fades into an ignored chat window. Waiting to see if you ever come back.

Google AI mode and ChatGPT seem terrible for this. I hit a lot of loop walls quickly. Gemini wasn’t so bad, but I think that was one of the ones that just compacted on its own. I actually prompted it a few times to save just to be safe. Claude, by contrast, doesn’t seem to have ANY of that happening. It hasn’t got stuck in a loop, and I haven’t seen it compressing/compacting/deleting anything yet.

PolyWogg 0, AI -25.

Technical support

One use case people recommend for AI is technical support. I’ve had four experiences using AI as technical support, and it has done a couple of things okay-to-well, and bombed on others.

The first bomb was on support in a program called mIRC. The IRC part of that is for Internet Relay Chat, and mIRC has been my go-to tool for online chatting since the late ’90s, when I used to be really into it. I have a couple of specific uses for it now, and I installed a couple of plugins recently to automate some stuff. Great, except they didn’t work QUITE the way I wanted, and the default display was in 9-point font. So, I asked ChatGPT how to tweak the mIRC settings for what I wanted.

One of the first things I told it was that I was using version 7.8.3. It has changed interfaces over the years, as well as command structures, so old commands won’t work; just like the voicemail messages say, “Please listen closely to the following options as our menu items have changed.” Okay, ChatGPT said, in its oh-so-confident way, that setting the display font to 16 points was super easy. It gave me a simple command, I entered it, and Bam! Error message. mIRC had no idea what that command was.

I told ChatGPT, it said, “Oh, right, sorry, yes, it’s done THIS way.” Another command, same error. “Oops, let’s do it through the menus, guaranteed to work. Click on DCC / Options / Display / Fonts”. Except there is no DISPLAY option under options. The menus have changed. Took me a while to find where fonts were. Made the change. No help really from chat, I just found the setting. Great.

Except no change. It would change the font for the chat window, but not the popup windows that I needed to tweak. Back to ChatGPT. Reminding it that I was in 7.8.3. Oops, it told me, the instructions were for version 4.3 or something archaic. What? Why? I specifically told you NOT to show me guesses, and to ONLY show me solutions that were validated for 7.8.3. It politely informed me that it hadn’t guessed; it had “INFERRED”.

And thus began my long descent into a deep rabbit hole with AI along for the ride, digging small tunnels ahead of me.

I knew the change could be done, that it wasn’t rocket science, and that I wouldn’t figure it out on my own. I knew just enough to know that either the default font or the plugin font was set too low. No other way for it to be wrong. I knew, therefore, Dr. Watson, that I could either fix the original setting, find a way to override the setting automatically, find a way to change it manually after the fact, or ignore it completely. As time wore on, that last option grew increasingly attractive.

To be fair, mIRC isn’t exactly a commercial application like Microsoft Word. It doesn’t have millions of users. And a user plugin within mIRC? That has even less information about it.

Yet each time I asked a question, the AI tool would say, “Oh, I know how to do that!” Except it never did. It couldn’t find where the default font was set, although I later figured out that it wouldn’t matter, it was the plugin font that was the problem. And it couldn’t figure out how to change fonts AT ALL. Nor could I. I opened EVERY file that came with the plugin. Lots of stuff for settings in the pop-up window, but nowhere where it had a font setting. It seems to be hardcoded in the plugin, alas.

I was undaunted. I knew that if I couldn’t do the first two options, I could at least set it after it loaded. Because I could go into the menu, choose Options / Preferences / Fonts / Font choice. Or something equivalent. It took about 5 clicks to get to where I wanted to change the font. But then if another window opened, I had to edit that one too — another 5 clicks.

None of the options AI suggested worked. Auto-load commands, mIRC scripts — none of them worked — and mostly ended up with the AI tool telling me, “Oh, it would have worked if you were using an older version.” WHICH I TOLD IT NOT TO DO! Grumble, grumble.

I found a workaround — I forced the font menu onto the taskbar manually and then told it to stay there forever; now when the pop-up shows up in 9-point font, I can click the taskbar, the menu opens, I change the font to 16 to 20 points, and it’s done. Super easy, two clicks.

PolyWogg 1, AI -25.

Drifting back to shore

This is a newer version of the loop problem. At least, it seems like it is the same sort of error.

I was trying to get Claude to do an image for me. I wanted to create a badge, with an embroidered edge. All of the AI tools take different approaches to images; some work in specific types of image scenarios, others in different scenarios, and others? Well, some don’t work at all.

Claude NAILED the first part of the badge problem. It gave me a perfect ring on the first try, which none of the other tools did (it uses SVG vectors to handle the geometry, hence why it was so accurate). But when it tried to do the embroidery, it failed completely. Nothing it did looked like embroidery.

I scrapped that idea, moved on. About 40 minutes later, out of nowhere, its attempts at embroidery showed up again in the margins. I was like, “Huh? Did I paste an old prompt?”. So I asked it why it included embroidery in that version. It told me because I asked for it earlier, and the algorithm forgot that I said no to it, so it went back and did it again. It had “drifted” back to the earlier setup. A little weird, so I had it add a prompt component that said very clearly, NO EMBROIDERY ELEMENTS. About 20 minutes later, working much further down in the model, the embroidery attempt came back. I checked the prompt; it clearly said no embroidery. So I asked again, “How?”.

This was a second type of drift. It had analyzed the prompt. And because I had asked for embroidery before (positive inclusion) and now was excluding it (negative inclusion), the fact that I had mentioned it at all was interpreted as positive inclusion. It ignored the “NO” part. I suddenly felt like I was working at Foreign Affairs back in the old days of TELEXes where you couldn’t afford for a word to be missed so you would type NO/NO to make sure one of the “NOs” made it through. I didn’t try that with Claude, because it was now a VERY long chat, Claude was getting on in digital minutes/years, and showing signs of confusion. I reset and started with a new chat, no mention of embroidery. It never showed up again.

I couldn’t find a way around it, other than using new chats. Not sure that’s a win.

PolyWogg 0, AI -2.

That’s the bad news. I was going to write about the tips it gave me for GIMP, but that’s a mixed bag, not all bad. And what really excites me is all the good things it’s done for me. That’s the next post. 🙂

Where is my dumb robot?

The PolyBlog
November 11 2016

I confess I have a bit of a techie mentality. I paid for part of my university life through working tech support areas and software installation services at universities, as well as helping teach a few practical sessions with professionals upgrading their computer skills. And when I started working for government, a lot of what garnered early positive feedback was my computer skills. I’ve done programming too. But where I stop being a techie usually is when it moves from software to hardware. There I’m relatively lost. Yet when people talk about Artificial Intelligence, better use of data, and all those wonderful things that are more software-oriented, they omit the part that I think is really possible in the short-run. The physical hardware with some basic programming.

People are all excited in the industry about “smart cars”, but long before I get a smart car, can someone tell me why I don’t have a dumb robot yet?

I don’t mean those simple robots that are merely self-propelled vacuum cleaners nor the ones for kids that roll like BB-8 or respond to a couple of voice commands or are in the shape of a pet. I’m talking about a dumb, simple, repetitive-task performing robot.

There’s an article over at the Harvard Business Review blogs by Andrew Ng called What Artificial Intelligence Can and Can’t Do Right Now (link may expire) and I love it for the way it approaches what AI can do by comparing it to the way humans process things. Basically, the argument is that if our brain can figure out what to do in less than a second, then the number of variables are relatively small, there are discrete choices and outcomes, and thus you can automate the task to a machine. Basically machine and supervised learning to teach a machine how to do it.

What do I want in a dumb robot? Someone who can do things for me during the day that I don’t need to do myself. Let’s walk through a typical day and the things that I should be able to have already…

I start my day with my alarm clock beeping at me. No real need to automate that, the alarm clock does exactly what it should do, a tried and true technology. But what if I roll over, turn off the alarm, and accidentally fall back asleep. I don’t mean I hit snooze, I mean I turned off the alarm. Now there is no backup. No mental nudge to say stay awake. What if my dumb robot (DR), let’s call it Jeeves, what if Jeeves was programmed that unless I override his programming the night before had access to my calendar and saw that it was 8:00 and I have a work meeting at 9:00, but I was still in bed. Could Jeeves beep at me? Or even in a nice voice (maybe reminiscent of my mom calling me when I was a kid to get my butt out of bed) saying “Paul, are you up yet?”. Maybe more insistent if I don’t answer. The backup to my own false sense of infallibility.

But let’s say I get up on time and I’m heading for the shower. Do I want Jeeves to turn on the shower for me and have it pre-heated to the right temperature before I come in? Nothing particularly challenging about that. Movement to a preset location, turning a knob to a specific point, good to go. Not much of a time-saver, most people wouldn’t bother. But you could have Jeeves do it.

Now, showering, brushing your teeth, voiding, those are tasks you’re going to perform yourself. But if you had a slight disability, are there basic things Jeeves could do to hold an arm out to assist with transitions? Hand you a towel? Monitor you in case you fall and call someone if you do? Could Jeeves even assist with bathing for those who need it? That’s probably a bridge too far right now, but not an impassable chasm.

But as you finish up in the bathroom, could Jeeves make you breakfast? Your bowls, utensils, cereal, juice, glasses are all pretty much going to be in the same place every day, so automating the robot to fill a bowl with cereal and a glass with juice shouldn’t be that difficult. You just need some flexibility to identify to Jeeves what your bowls and glasses look like, the layout of your kitchen, etc. although scanning/mapping software would do that for it pretty easily. A more advanced version might even be able to crack open a couple of eggs, butter bread or toast, make you a fried egg sandwich so breakfast is ready whenever you are.

Once everything is over, presumably Jeeves could clean up and put dishes in the dishwasher, etc. Could maybe clean them, and put them back in exactly the same spot as the day before, but perhaps not.

When I go to work each morning, there are basically six things I take with me. My tablet, my work blackberry, my personal phone, usually a book that I’m reading, my notebook, and my work pass. There are some other things in my bag, etc., but those six are pretty standard. I might or might not wear a coat depending on the day, different shoes, mitts, hats, always my car keys, but those are contextual. And once in a while, I forget something. Like my work pass. Why? Because I stopped somewhere on the way home, put it in the pocket of my jacket, got home, hung up the jacket, and forgot to put the pass on the shelf by the front door where I’ll see it. No biggie, but why am I using mental energy to remember to put it specifically in the same spot or remembering the next day? What if each of those six items had a small RFID tag on it that Jeeves would monitor. And if they weren’t all in my bag as I go to head out in the morning, Jeeves would say, “Excuse me, Paul, I don’t believe you have your work pass with you.” My first reaction will be, “What? No, of course I do, it’s right here in my … umm, why isn’t my pass in my bag? Oh right, it’s over here. Thanks Jeeves/memory jogger.” Is that a big deal? Of course not, but I bet I would program it to scan for the RFID’s when I’m leaving for the days when my brain is focused on the seven things my son, wife and I are talking about as we scramble to get out the door. Heck, sometimes it’s as simple as something got placed on top of my pass and I can’t physically see it on my shelf, and so I head out thinking I have everything.

Here’s where some of us will diverge. Lots of people would love to take the robot to work. That’s a bridge too far for me. If work wants to automate tasks, great, I shouldn’t bring my own “robot” to work to help me do my job. If so, why not just hire the robot?

But while I’m at work, could Jeeves vacuum the house? Clean a toilet? Wash pre-sorted laundry? Hang it on a line to dry or throw in the dryer and check if it is dry when done? Cut the grass? Shovel snow…oh, that would be sweet.

Could Jeeves be programmed with a more sophisticated kitchen module that would allow it to chop vegetables? Basically act as a sous-chef? Maybe even, with remote activation, throw a pizza in or a pre-assembled casserole so it’s ready when we all get home? I hesitate to go so far as having a full cookbook with multiple ingredients, but that is only an RFID tag on a standard sized container away from doable. Could he open the door and receive a package from UPS or FedEx? Could he collect the mail from a central box?

After supper, can it also double as a stand-in for a playmate for someone who is single or whose friends are busy that night? Get your mind out of the gutter. I mean rather than playing a board game or card game against a computer screen, could it roll dice, charge you rent in Monopoly, learn to throw and catch a frisbee? Or a baseball? Could it be programmed with multiple pitching styles to act like an automated pitcher that adjusts to your level and technique so you don’t have to hit balls by yourself and chase them? Could it act as pitcher with five little scouts running around it that chase balls and bring them back? Could it play basic tennis? Those are more about the design of the robot’s arms/movements than about technique for hitting or throwing a ball, so yes, they all could be done.

As I’m getting ready for bed, Jeeves could turn off all the lights downstairs (heck, an app can do that now). Jeeves could also monitor the location and charging status of my e-devices, and if they are not on the charger, go and get them and put them on charge. Or double check my to do list verbally with me to see if there is anything to adjust, delete, add. A personal secretary app, not unlike some of the functions Siri does now. But more interactive, following me around while I do other things.

And all of those things are doable. A dumb robot, personal assistant, digital butler, e-handmaiden, non-sentient slave. An article I read some time ago talked about the issue of android rights, similar basically to the idea that was raised in the first season of Star Trek: The Next Generation. Is an high-AI android property like a toaster? Or as the AI evolves, does it meet any criteria for self-awareness or even sentience? Except it missed the point.

Developers are looking for smart androids. People are looking for dumb robots.

If you had a Jeeves, what would you want it to do that you hate doing yourself?

And where the hell is our Jeeves?