@@solarpellets Its not really that bad at it though. it gives very human-like responses, its just that its not very good when it comes to getting information correct. But the way it speaks can be very human-like, the thing that gives it away is that it generally just has no clue about what its actually talking about. Although the super interesting thing about ChatGPT which is what makes it exciting technologically is that it is far better at random tasks than it SHOULD be. The fact it can even pretend to be competent to any degree at random tasks is shocking when it wasnt trained to do that stuff at all. Thats why there is a lot of current AI research around LLMs, because they are better at random tasks than they SHOULD be given they werent trained for them. But ofc being "better than they should be" doesnt mean they're actually at all competent, just that they SHOULD be even LESS competent than they actually are.
"There is a significant overlap between the smartest bears and the dumbest people" - I don't know who said this, but it's in reference to bear-proof designs
Oh, I get it. Irony. Because the people blindly lauding AI aren't the only morons here. Who'd ever guess it'd be a bad idea to use a screwdriver as a catheter?
Nobody asked, but I solved the crossword the rest of the way properly. Before I give the answers, I'll mention that roughly half of ChatGPT's answers were right - It ultimately solved *12 of the 25* clues correctly. It skipped two clues, one of which had confused it, probably, because it required a two-word solution, which made it a non-standard clue in its presentation (two numbers between parentheses instead of one). But again, it did a relatively good job otherwise, especially considering it understood everything from only an image and a mention that the image was a crossword, generated its answers in seconds, and allowed the user to correct and more easily solve the crossword the rest of the way from those answers. This demonstrates the actual, wonderfully versatile use of AI to be a human-assistance tool, not a catch-all do-everything-for-us tool (not yet, anyway). Across 1 - prompt 4 - cabin 7 - canton 8 - blowup 9 - etna 10 - turnpike 12 - cotoneaster 17 - coolness 19 - digs 20 - iguana 21 - raised 22 - media 23 - adages Down 1 - plastic 2 - outcast 3 - penitence 4 - colon 5 - bowline 6 - nipper 11 - reassured 13 - orotund 14 - tidying 15 - regress 16 - schism 18 - ninja
Not to be too supportive of AI (because I sure do hate the way they're misused), but it does admit it only gave "potential" answers. I wonder if it would have been able to get the other half, if one were to fill in the ones it (seemed to) get correct and ask it to try again. I doubt most practiced crossword puzzlers would be able to fully solve a crossword just by listing out answers and not seeing how they interact with each other.
I once asked chatgpt to arrange kids in a seating arrangement based off some rules like "sam can't sit next to Ricky" "Jeff needs to be 2 seats away from sam" etc. What is a correct seating arrangement. Only for Chatgpt to completely mess it all up. Upon asking chatgpt for its inccorrect reasoning, it made up its own headcannon saying stuff like "Stacy and Jeff are dating so they should sit next to each other" ai is a language tool, nothing more. Edit: ChatGPT in its current state is just a language algorithm. Not to say all AI is. EditEdit: The rules weren't too restrictive, you could've easily came up with multiple seating arrangements that followed the required rules. But when pressed on the issue, it would make up its own rules which was outright incorrect. You could try this with ChatGPT and use a color arrangement instead with rules and I'd assume it would get it wrong most of the time still.
it's actually really good at coding though. 60% of autocomplete suggestions from Github Copilot are exactly what I wanted to type (or something better) and I just need to press tab.
Is the GPT-3.5 or 4? Well, either way, try testing it with modern 7b LLMs. Also, "ai is a language tool, nothing more" is evidently false considering how useful it is in programming, and "ai" is used in everything... generative AI is its own branch.
It's still amazing to me that it understood the problem and gave a solution that LOOKS like an actual solution to this crossword puzzle at first glance.
As someone who went through the boy scout program and enjoyed almost nothing, I did have to learn about practical knot tying, so when chatGPT triggered those memories by claiming that 5-down was either sheepshank or clovehitch, but was clearly wrong considering the length of the string allowed, i immediately knew (despite not having seen the clue) that the actual answer was bowline, which would have fit nicely with some of the other answers that chatgpt had made. So there's that.
@JBB685 isn't all bad. I did complete the program and earned eagle scout, which is claimed to be a good measure of leadership ability. So, combined with my 7 years of college and bachelor's degree in comp science, I was able to land a decent job as a ... checks the papers ... grocery store night stocker? Wait that can't be right. And what's this about advanced placement through primary school laying the groundwork for social anxiety disorder and low self esteem? 🤔 huh. Well, I can tie a square lashing to hold two sticks together, so that's good.
for some reason, after chatgpt was loading the answer my video skipped to the end where Evan said "see you next week" and I thought that was literally the whole video.
Actually it would be awesome Sudoku is a mathematical sum 0, which computers are experts at . The thing here is how crosswords are so open ended with what could fit. It's not (a line must sum 45) is (a lane has to fit letters and meanings), so retroactive value is lost, as fulfilling the crossing of letters might be more important than meaning of words, unlike numbers that just use one value for all crossing
ChatGPT can't even validate its very answers which is simply checking that squares are properly filled and letters match. What gives you half the faith that it would be eligible for solving a medium Sudoku?
As soon as that first word phased through the the wall i knoew this was going to be an absolutely golden video I mean all of CBs videos are golden. When he uploads.
I understand some people are scared AI will take over the world and enslave us, but I’m more worried about people overusing AI for everything and taking it as all-knowing when it's severely flawed
why not both, it has the potential to cause disasters from overuse in areas were the specific AI tool being used is completely inappropriate like we already use half the tools we already use and the ability to cause an increasing wealth inequality to spiral even more vastly out of control while the planet burns and everyone says they care while our societies continually prove that we don't or at the very least the people who need to care don't, while billionaires tell us guys its a "utopia there will be no jobs!" while experts say that's a dystopia your brain doesn't like having meaningful work. and how I love this sort of video and find it fun to play around with AI and do occasionally have actual uses for it, I still see it as a credible threat to the way are economy functions, or disfunctions (which is an anagram of functions)
AI doesn't use letters. Words are fed to the AI as a series of tokens, or parts of words. So it really struggles when it has to think about the letters in words.
Explains why there was once that moment that was going around where Google Bard didn't know how many times the letter e appeared in ketchup. It doesn't "see" the letter e or the word ketchup in the same way we do.
> GPT-3 has a vocabulary of 50,257 tokens, and each token is represented by a vector of 12888 elements. these tokens can be whole words, common prefixes or suffixes, letters, etc
@@populer208 its true that it has tokens for the letters a-z, but it doesnt use them to spell out words. If you put in pickle, for example, it doesnt use tokens for each letter, in fact it uses one token for "p", and one token for "ickle". It outputs the same way too, since it was trainined on the same kind of input. You search up the chat gpt tokenizer and see how it breaks down any input.
Reminds me of ChatGPT when playing chess, even the most recent iteration, which is supposed to be better at it. Still starts hallucinating and cheating within like 20-30 moves.
Yeah, it's still messing up after the opening, but it's better. What's interesting is that the newer version, while not yet in a usable state, plays in a way that is more human-like than regular chess bots. I suspect that it has been specifically trained to be better at chess because of the "ChatGPT Chess" meme.
@@NoNamer123456789 It's main fault relative to humanlike play is it, despite being able to play 20 lines of theory, it still does not fundamentally *understand* how the pieces move.
My man tried to make a low effort video an just happened to actually accidentally give himself more cause it built the expectation of a complex follow up video
If you combine this with a tool that automatically tries to fill in the answers ChatGPT produces and then asks back to ChatGPT if a clue doesn't fit, this can definitely solve Crosswords easily. And the GPT manager should only be what, a 30 line python script?
In my experience, ChatGPT is really good for giving you ideas or concepts, but is terrible with implementation. For example, I wanted to improve some code, so I asked ChatGPT how I could do that. It said to parallelize the processes -- Excelent suggestion! But when I asked for an example, nothing worked properly. I went to RU-vid to learn parallelization and then implimented it, and it did drastically improved the code. Here, it would probably be excelent at telling you "words with __ letters that fits this description", but further implementation of those words into the puzzle is where it breaks
I should’ve known from the title that this video wasn’t “Getting ChatGPT to solve a crossword” but rather a simple demonstration and answer to a highly asked question
Tried this once… same result. Solution: use the coder part of gpt, ask it to solve the problems, and write code to double check the length. Tell it to repeat the code for everything until it is correct and double check if the word makes sense at the end with another prompt to double check the answers. Got closer to 60% of the answers then
Remember. LLM give the most plausible answer, not the correct one. How much that overlaps depends on the training data, the engineering, and the question asked.
At first I didn't see this was a Day Off video and thought it was a full fledged CodeBullet video. Imagine my surprise as the video was just like "Nah ChatGPT can't do it" *end video*
people saying that this proves AI will never go anywhere have yet to consider that this is a general purpose algorithm that, using just an image of the crossword, which it SOMEHOW manages to interperet, gets a couple of words right. this is really impressive.