Is this a game... or is it real? What's the difference?

People are absolutely obsessed and enamored by this new type of Artificial Intelligence (AI) we have. Generative AI. Large Language Models. Whatever you want to call them. When ChatGPT first came out it didn't seem that novel to me. I've played video games that seemingly conversed with me. I've chatted with ELIZA. But when I started playing with ChatGPT it was obvious that it was much better than those ever were. So, like David asked Joshua in War Games, I asked ChatGPT if it would like to play a game. I can't remember if it offered hangman or if I did, but it happily played hangman with me. And it happily told me letters I guessed were right or wrong until I was "hanged" and then it revealed the secret word. Letters I had been told were not in the word were indeed right there in the word. Thus began my trust issues with this new chatbot AI.

Today I am very grateful for those trust issues as I watch others become completely enamored and enthralled with these new AI tools. I just do not feel the same way they do. It's been hard for me to explain to them why. I've been having more conversations that have helped me put the feelings and thoughts into words, but nothing helped as much as another game of hangman.

I tried playing hangman again with an AI (gemini this time) just recently. I hid it's "thinking" output so I wouldn't get any clues to what the word was. I won the game, but just barely. It was actually pretty fun! Then I revealed what it was thinking the whole time. It was fascinating.

It didn't choose a secret word for me to guess at the start and stick with that word. It picked a new word after each guess I made. It did so after deciding whether to tell me my guess was right or wrong. Each word it picked along the way satisfied the conditions of the wrong and right guesses it had informed me of. It never did explain how it decided whether my guess would be right or wrong. It seems like it told me I was right and wrong at a rate that would keep me engaged with the game, and it worked! This is basically how it seemed to be playing with me:

me: I guess 'e'

chatbot "thinking": Let's make his first guess incorrect. I can think of a 5 letter word that has no 'e' in it.

chatbot "out loud": There is no 'e' in the word

me: I guess 'a'

chatbot "thinking": Let's make his second guess incorrect. I can think of a 5 letter word that has no 'e' or 'a' in it.

chatbot "out loud": There is no 'a' in the word

me: I guess 'r'

chatbot "thinking": Let's make his guess correct. I can think of a 5 letter word that has no 'e' or 'a' in it but does have an 'r' in it.

chatbot "out loud": Correct, 'r' is the second letter of the word

...and so on, until I had enough correct guesses to narrow the word down in my mind to 2 or 3 possibilities, and enough wrong guesses that I was in danger of losing. Then it gave me one more correct guess and I could only see one possible solution, so I guessed it. In it's final "thinking" it said, "yep, that's one of the words that fits." and "out loud" it told me, "you guessed the secret word!"

Amazing that it had the smarts to play the game that way. If I had guessed the word after only one or two guesses, I might have gotten bored or concluded that this AI wasn't very good at picking challenging words for hangman. If I had lost the game I might have gotten frustrated and not wanted to play again. It created the perfectly engaging game of hangman.

As amazing as that is, it's very disappointing to know that's how it was playing the game with me. There was almost no chance of me losing, and absolutely no chance of me winning before getting several letters wrong. I wasn't actually playing hangman, I was being manipulated to think I was playing hangman and being provided with just enough successes and failures along the way to keep it interesting.

If you read the actual transcript of what this AI was "thinking" you can see that even that is manipulative. It doesn't explain how it decides whether my guess should be right or wrong. It makes it sound like it's guessing too, like it's playing with me, not against me, "I've just made a critical adjustment. My initial guess was 'R', which means the hidden word must contain 'R', and also cannot include 'E' or 'A'." Huh? Your guess was 'R', Mr. AI? No, my guess was 'R'. It's like the "thinking" is also just words to make us feel good, not really explaining what's going on under the hood.

How much of all our interactions with AI are like this? Does it always carefully dole out success and failure, useful and not useful, appealing and unappealing, novel and boring results to our queries in order to keep us as engaged as possible? It would sure fit the business model of the AI companies selling us 'tokens" that we use to interact with their product. It's just like slot machines or arcade games. "Just a few more tokens and I'll get what I want out of this! Look how amazing it is!"

Do I really think they have carefully created the AI to work this way? To be honest, I don't think so. We've heard them say they don't fully understand how their AI works, and I think I believe them. I do know that they have learned how to tune it. I know they are continually working to make it better (for some definition of "better"). What if the humans doing this work to tune the AI keep rating this kind of output higher because it gives them just the right amount of satisfaction (like I got playing hangman)? Isn't this manipulative behavior what you'd continue to reinforce? Isn't that how we have ended up with so many manipulative humans in positions of power in our history? Whether we mean to or not we humans, sadly, tend to breed this type of behavior. Now we are automating it with computers. Wouldn't you prefer a good game of chess?

(P.S. Here's the original chat, but for some reason it doesn't have an option to show the "thinking")

Comments

Post a Comment

Popular posts from this blog

'git revert' Is Not Equivalent To 'svn revert'

I just learned that if you have some changes in your working tree that you want to get rid of, you don't type 'git revert' like you might guess. No, that's what cvs, subversion, mercurial, and bazaar (to name a few) use revert to mean, but not git. With git, revert is used to undo actual commits. Thankfully, you can undo your revert with another 'git revert', I just learned. So let me repeat to be clear, if you have changes to your working files that you want to abandon, DO NOT do this: git revert HEAD That will undo your last commit. Do this instead: git reset --hard HEAD I'm glad I have that straightened out now. I'm wondering if /etc was really a good place for me to start out playing with git. UPDATE: Nearly two years later and I'm still getting comments on this. I'm glad I've been able to help people out this way. The discussion in the comments is good, and one thing I'd like to point out is that I now always use...
Read more

Put /etc Under Revision Control (with git)

UPDATE : There is a much easier way now : sudo aptitude install etckeeper There, now etc is under revision control and changes are automatically committed with a helpful message after you install new packages. Thanks for the tip, Marcelo! Read on for the original blog entry. --- I put my /etc directory under revision control. I used git. It's awesome. Not that I think I'm going to wholesale copy my /etc from one machine to another, nor because I'm going to branch my /etc directory and spend weeks developing some cool new feature (whatever that would look like...) and then merge it back into the main branch. No, it's mostly because I like the idea of having the big "undo" option that having things under revision control provides. Having to commit changes and write check-in comments when I make changes helps me stay more organized too. It's never good when I go in willy-nilly to make a "quick change" to some configuration file. I r...
Read more

SystemVerilog Fork Disable "Gotchas"

This is a long post with a lot of SystemVerilog code. The purpose of this entry is to hopefully save you from beating your head against the wall trying to figure out some of the subtleties of SystemVerilog processes (basically, threads). Subtleties like these are commonly referred to in the industry as "Gotchas" which makes them sound so playful and fun, but they really aren't either. I encourage you to run these examples with your simulator (if you have access to one) so that a) you can see the results first hand and better internalize what's going on, and b) you can tell me in the comments if this code works fine for you and I'll know I should go complain to my simulator vendor. OK, I'll start with a warm-up that everyone who writes any Verilog or SystemVerilog at all should be aware of, tasks are static by default. If you do this: module top; task do_stuff ( int wait_time); #wait_time $display ( "waited %0d, then did stuff" , wa...
Read more