now that you read off the ChatGPT Hallucinations in your video about battling bugs next time It writes in the style of Code Bullet it will add more bug fighting. Which you will then read on camera and solidify it more as something Code Bullet would say until scripts by Code Bullet from ChatGPT is a full-on stories about the "BUG WARS". haha.
challenge: your next video on human benchmark cannot use AI programming tools, like chatGPT (basically no chatGPT or alternatives for programming, but you can use public/free libraries)
Very rarely it goes the other way, just write some code in two minutes and it just works. Then you spend an hour looking for the bug that are definitely there anyway only to not find any and wasting the time anyway.
Reminds me of when I tried to program a Sudoku bot. That piece of crap library had no problem deciphering a handwritten text, but you give it a single, perfectly captured number it completely craps its pants.
Sometimes watching Code Bullet reminds me of something my high school basic programming teacher said: "A computer is only as smart as the person using it." Then he told us that there were some dumb computers in here.
I love how this could have been done so much easier by just using a library that reads the contents of the number's HTML element right out of your browser, like selenium for python.
I came to the comments to make the same point. I do fear that in a world where people are thinking of resorting to ChatGPT to solve programming problems, we are going to have more cases where people don’t stop to think about how to make a solution /easy/ and instead say “f’ck it, it’s not my problem as long as it works” I seriously worry how code is going to be maintained in the future when this becomes the norm. The only way code will be able to be maintained is by getting more and more advanced AI to debug the things previous AI had generated and humans will have no idea what the hell is actually making things work anymore.
Even with screenshot, just using the characters that are used instead of tesseract would probably be faster. You just need a key value pair table, and sort by X then Y value. (So the numbers at the top all come first, and the ones at the bottom come later.)
Reading directly from the DOM would surely be easier. But in this kinda of content I feel like that would be more cheating than what's happening here. Because then you'll literally be reading the number from HTML, setting a value field and .click()'ing buttons
Absolutely love how when CB gets so excited the chair becomes an enemy. That’s true accomplishment. When you’re feeling so good.. you just NEED to prove it by beating on an inanimate object.
@@Connor-yn8pz The bell curve displayed is skewed (to the left). In order to be in the 50th percentile, you need to be at the point where the left half area = the right half area. Without a y-axis to see how large those small data points (both left and right of the large peak) actually are, we can't really tell if the ones on the right are the same height as the ones on the left. I suspect the displayed graph has a smoothing function applied to it to make it look nicer.
@@zecuse i wasnt arguing with the 50th percentile. Its still a normal distribution if a skewed one. And by your own wording hed be in the higher %tile because of skew. The smoothing function is definitely there if the number is righr
@@Connor-yn8pz He's still not guaranteed to be above 50% though. We don't know how high the right half of the graph actually is and therefore how much area it actually has. That's also what I was talking about regarding the smoothing. At best, each point over there seems to be the same height and the smoothing shrunk all of them down. We don't know if the valleys in the left side are the same size as the entire right side. We don't know how any of the valleys compare to the outlier peak on the left. The smoothing function has hidden this.
EasyOCR is another option for Image Text Extraction. Based on my experience, EasyOCR demonstrates higher accuracy in recognizing numbers/digits. However, it comes with a notable drawback, as it lacks support for a wide range of languages.
tessaract works for numbers fine if you change the config, for example "pytesseract.image_to_string(text_in_tile, config='--psm 6')" config 6 if for single number, without it it couldnt read it for me but with it it can easily with close to 100% accuracy.
really both options are overcomplicated. just use selenium and extract the number from html. Or if you're not married to the option of using python, just use JS
You inspired me to try to do the ones you hadn't done in the previous video. When I tried doing this one I ran into all the same problems at the start. The issue is that tesseract is optimised for text at about 25 pixels tall. The screenshot being taken makes it huge and it has a hard time reading it, thats why zooming out and taking a screenshot of more of the screen made it better. I did run into issues such as it recognising 7 as / or 5 as s but using psm --6 reduces the frequency of this. I had the same problem at level 18 and kinda gave up after a couple of hours because im not that good at coding. I am very happy that you posted this video and i hope you stick to uploading once a week.
You can do it with find a picture in another, then find all the numbers and sort them by X then Y value, so the numbers on the first line are all before the 2nd line.
@@satibel I ended up zooming out more and using isdecimal() to get rid of errors, spaces, and enters. It works really well now and I got to level 36. I am using the method you described to do the chimp test though.
gotta say, your voice us the No.1 reason why i watch every video of you that youtube suggest to me, doesn't matter what the video is about, i just love it. p.s. your humour is a close second place ;)
asking chat gpt to write a script to get a number for a screenshot when you could just get that number straight from the html. i love programmers!! i love that these are the people that i have to work with!!
Hey bullet, glad to see you alive! This is one of the best promotions I've seen. Usually I only watch them to support the channel I'm watching, but this time I watched it while enjoying it.
I gotta say... Code Bullet goes above and beyond in his videos! Who needs to use debugging consoles to scrape the page's layout and directly set the input value to the text you copied before, when you can just install and use four libraries to take a screenshot, apply a little post-processing to it, spend an hour to try your program, pray that it works without changing anything (aka "debugging"), and ultimately recognize the text from the pixel values so that after a small delay you send key press events corresponding to the characters in the text you extracted to complete the original task at hand? (I'd be lying if I said I didn't reread that to make sure it was actually supposed to end in a question mark. I'm actually still not 100% sure it is supposed to end in a question mark. Oh well.) In all seriousness though, this was an entertaining video and I enjoy seeing Code Bullet go through his projects, even if at times he is screaming in pain internally. I both 100% feel you and also can't help but laugh XD
Wow, this shows the dangers of using a sledgehammer for every problem. I looked at the page this test is on and it shows the numbers in plaintext on the page itself. So the whole thing could have just been a simple copy/paste job instead of all this complex screencapping to read images to get text from and such.
Eh, I could see it done easily in 3 or less lines of javascript if someone really wanted to minimize number of lines of code. But I resist the notion that the journey of getting to even a less concise working answer wouldn't have been as funny. There's plenty of room to stumble, especially for someone who hasn't done something like it before. Though the biggest difference would have been the absence of trying to bend the blackbox of whatever library was trying to read text from images to his will, which, admittedly, was a large portion of the video.
Even though pi doesn't change, I bet people that memorize it would have a good chance at using their method to do well at this. I was working at a private school several years ago and one of the middle grade classes had a contest. The winner was a girl who memorized a few pages of ~16pt handwritten font. That's impressive as hell to me. Another student had about a full page, and the others were pretty far behind at a half page or less. I have no idea how long they had to memorize though.
Soooo....this one finally motivated me to have a shot at it myself. After some back and forth I actually got a good result. First tried it purely in JS, which didnt work because the input field needs an actual keyboard input to enable the submit button. Then I built a websocketserver in C# to receive the current number from JS and type it in. Worked fine, but a bit slow and I wanted it to be in one language, so I tried the screenshot+OCR method in C#. Got it to somewhat reliably work up until a single digit would show up on a new line. Back to the JS+server, optimized that a little and left it running when I left for a couple of hours. Last thing I might try again at some point is skipping the annoying and increasingly long timer. Too much for me as a non-web developer. The result: the time you get to memorize 216 digits is about 174 seconds and it takes about 5 hours to get there (typing the numbers not included). 😄
😂 I like that analogy, because of how his code started controlling the mouse and was almost out of control as I picture the first person making fire having that same “oh f’ck!” moment straight after the jubilation ;)
Thank you for your patriotism sir. Good on ya mate. As much as I have hated all of Australia since the Cricket semi in 1999 - you do have a lot to be proud of you bastards.
~ 6:00 This is the exact excitement I get when my code works too. And just like me, outta context we're both excited over the silliest thing to other ppl: "FOUR! WE GOT A FOUR!" (relatable 100)
I did the same and hit the 100.00% with 88 digits lol. Fun little project I'll use for my thesis. Also did smth for reaction time which was bottlenecked by my ping and typing which got to ~616 wpm
You need to tell tesseract you are looking for numbers only, then it works much better. It is expecting text and seems to prefer a letter over a number with default settings. Also it seems that it prefers text a certain distance from the top and bottom of the screenshot if it's a single line.
So recently, Infinite craft has been blowing up. Maybe you could make an AI to randomly mix things? Maybe it randomly drags stuff into each other and keeps getting more stuff. this could be very cool. you could probably get world record for having most first discovered tho after enough time. This is just an idea though. your choice. Also could be on your second channel. love ur vids