Apparently there's a billboard in San Francisco advertising a new AI service called "ChatTJB" Here's the catch, AI in this case stands for Average Individual.

"You ask a question, and a person named Tucker reads it, thinks about it, and writes back when awake and motivated."

That sounds legit. Let Tucker give all the responses. Maybe people are just getting sick of AI chatbots, because AI is being so heavily overused these days. You get on a call to a customer service line, and instead of getting a person, you get an AI chat bot.

Sometimes, I run AI chatbots locally because they can give some interesting feedback about things that I'm thinking about or things that I'm writing about. AI is good as an echo chamber. Once you push your own thoughts to it, the chat bot can rehash things and give you an interesting response that you might not have thought of otherwise.

What can be annoying though is when the AI chatbot starts repeating things. 😖️ Then it can start feeling like Eliza. Of course there's things you can do about that. There's parameters you can adjust in a language model to make the responses more varied and interesting. One parameter you can adjust is the temperature. By giving a higher temperature value to the language model, you can get more varied and interesting output from it. That can be useful if your language model starts getting too repetitive in its output.

Sometimes you want the opposite though. Sometimes you want the language model to be more predictable rather than more varied. You could provide a lower temperature, but another thing you could do is to set the random seed to a fixed value. Here's a thing that's true about computers in general, not just language models ' computers are very good at doing the exact same thing over and over again, but they are very bad at doing random, creative things. There's a strong tendency for them to be repetitive and consistent.

That's to your advantage in most cases, where you want a computer to be as predictable as possible. Like when you are debugging an application, you'll want to make sure you can reproduce the issue at hand. If the program doesn't behave the same way each time you run it, it will be difficult to fix the thing. Fixing stuff in programs generally entails reproducing whatever it is that went wrong in the first place. If you can't reproduce the problem, you can't easily fix it. At least, your likelihood of fixing it drops pretty low.

Now consider Tucker with ChatTJB. He's what we would call "wetware", because a biological component of the system. His brain powers ChatTJB. What if you don't like his responses? As a programmer, I'll think, hit the debugger and find out what's wrong. Since Tucker is providing all the responses, you might say that errors there would be considered "PEBKAC" errors really.

Anytime a person is part of the loop, that always complicates things in the process. You can't really debug a person. If a person goes wrong somewhere in the mix, the best you can do is remove that person from the mix.

Working a job, now, there's a thing. You work along side other people day by day, and they are prone to doing what they will. Some people might be too loud, some might be quiet, and some might be noisy. Us people are a messy, squishy sort.

Sometimes, on my drive to work, I drive past an accident. Some squished car, bent in places where a car isn't usually squished. That's always a tragic situation. Most likely, someone got hurt, and maybe our country has one less citizen. All these people acting on the road, driving their cards, with hands on the steering wheels, moving the wheels one way or another, then controlling the brake and accelerators with their feet. What could go wrong?

Plenty, plenty could go wrong. I used to drink my coffee at work after I get in. These days, I drink all my coffee at home. The reason being, if I drink my coffee at home, I'm much more alert while driving. Thus, highly recommended to have coffee at home before driving to work to stay safe while driving. Plus, coffee is the fuel of programmers to get stuff done.

If I, as a driver, make a mistake while driving, naturally I'll want to take appropriate steps to fix it. You can't fire up ChatGPT and just figure out what's wrong with yourself. Or rather, you could, but you're not likely to get the correct answer. Just like when you are sick or something then you open up Google to search for your symptoms, sure you might find something there but chances are you'll find an answer that doesn't relate to reality at all.

When I write a program, I'll fire up a debugger to figure out what could be wrong with the program. If only there were a debugger for people, then you could attach the debugger to a person to figure out what's wrong with them. Great if we could fix whatever is wrong with people instead of locking them up when they do wrong. Jails are full of people who made mistakes, some worse than others I suppose.

Now again about ChatTJB, relating all this to Mr. Tucker. If you were to fire up your ChatTJB client and have a conversation with him, probably you're not going to get the answers you want. If you were to dig deeper into the program, and attach your hypothetical debugger to Tucker, you'll probably find that he just doesn't care about what you wrote. Probably, he didn't even read your prompt and he just pasted some stock response to answer you that he came up with ahead of time.