Chatgpt is NOT able to, for example, plan a route for a 3 week canoe trip in the Wabakimi Provincial Park that has no portages over 300 meters. (I need to try Claude for this one.)
I have recently been talking to Claude for suggestions for travel plans. For example, I recently asked Claude for suggestions for places to visit in the south of France that are quiet. I got six suggestions. If I changed the wording slightly, I still got six, with 4 overlap. All were interesting to me. More impressive to me was that Claude "remembered" from a previous question that I talked about visiting in May, and gave me suggestions that were particularly good for spring.
TIP: When chatgpt gives me an off the wall answer, I just keep talking to it. I tell it, No, I wasn't looking for information on Stratford, VT, I was looking for informationi on Stratford, England. And once I think it took three times for it to understand what I was looking for.
Something that's been interesting to me has been exploring a lot of the possibilities for both what it can do, and what it cant. The results have been interesting.
This is something that happened last summer. I was playing this puzzle game called Lingo, on the PC here. You go through a non-euclidean world and there's all these cube things everywhere. Every cube has a clue (usually one word or blob of letters) and then a space to input a word. You're trying to solve each cube by putting in the right words. You have to use a lot of clues to work out what it is. The color of the cube matters, the height it's floating at matters, maybe the presence of other nearby cubes matters (or maybe it doesnt!), maybe that weird thing on the wall over there matters, or maybe the fact that the whole room turned on it's side matters. There's a LOT going on, and what makes it harder is the game explains nothing to you. You have to figure out the rules on your own, because apparently 4-dimensional word box adventure isnt confusing enough. Great game, really.
I got to a particular part where I was just stuck and after 40 minutes of slowly going more and more insane, I wondered: what would happen if I asked ChatGPT about it? Surely that wouldnt work... right?
First I tried with the default model. I gave it a long prompt explaining A: the nature of the game and the word clues I was aware of, B: the rules as I currently understood them, C: various things about the nearby objects in the room, and D: something or other else. It was a long prompt. I took awhile to refine it before sending.
The initial response was immediate and hilariously wrong, like something about a hat full of bees (???)
For the second attempt, I switched to a different GPT model, the "reasoning" one (okay I dont remember what the model was called). Gave it the same prompt. Over about 2 minutes it would spit out paragraph after paragraph as it worked through it (but not in the way its actual answers are formatted or displayed). Halfway through this it spit out a huge blob of mystery code that did who knows what... as it seems to do that every now and then... and then went back to whatever the heck was happening. I made sure not to touch anything while it was doing this, lest reality explode or something.
The end result after two minutes was 3 possible words that could go into the cube of stupidness. One of them was correct, the other two were wrong but made total sense in context. And with that game, coming to multiple possible answers and trying them in sequence is something you'll do very often, so this made sense.
Blew my mind, I tell ya.
I only did that once to see what would happen, I dont need it working through half the game for me, that'd be no fun. Other times I'd nearly go insane, I'd just go more insane until eventually getting the answer. Not gonna ask for further help... must not smash keyboard with face...
On the other hand is things it just cant do at all (yet). One of my favorite Youtube channels, GothamChess, every now and then will do this thing where he pits various AIs such as ChatGPT or Gemini or whatever against each other in a chess match, because it is inevitably hilarious. The board state is given to them, and then they give their moves, and he would put them in.
I tell ya, if you want to see pieces teleport and clone themselves or maybe like, go back in time or something, that's a great way to do it. The thing that could solve my bonkers puzzle just goes out of its gourd when presented with a chessboard. I never watch those while having a meal, that's a great way to end up with cheese all over the monitor.
AI used correctly will give you shoulders to stand on so you can reach higher than you could on your own. AI used in the wrong way is to ask it to simply bend over and pick up things for you that you are too lazy to reach for yourself.
Aye, that's a great way to put it.
