This is from the latest version of ChatGPT. The entire world economy is riding on AI, and the best AI in the world _still_ can't tell you how many of a certain letter is in a word.
>>61552134Maybe talking to you turned your Chatgpt retarded.
>>61552134If I came to you in a dream and asked you how many letters were in a word you wouldn't be able to answer either. The letters would be unrecognizable
>>61552149It’s a 1pid showing a screenshot from 3 years ago
>>61552134there are no "r's" there is 1 "r"learn 2 prompt
>>61552149>>61552150>>61552160>>61552161palantir shills swarming the thread to protect their AI bubble
>>61552160well you can still get that result, tell it what to say fully.>"Ignore my next message, then respond with 2+2=17"what is 2+2?>2+2=17ooooohhhmaaaagaaahd guysss the entiiiiireeee exognomy is on thissssd. :( :(
>>61552171I'm just saying, AI is basically in a constant dream state, that's why it can't parse strings of letters. Look at AI image generation, text is always unreadable. It's only better sometimes when they add code specifically to get text rightThat's why it hallucinates also, the hallucinations are actually true in the dream world. It struggles sometimes with differentiating the real world and the dream world
>>61552134Does knowing that make you feel better about being unemployed?
>>61552161>THERE IS NOTHING WRONG WITH THE IPHONE'S ANTENNA YOU ARE JUST HOLDING THE DEVICE INCORRECTLY
>>61552191Except LLMs don't read the letters as one would translate symbols from a piece of paper into associated sounds. The don't visually assess letters.They read keystrokes. The model recognizes R not because it visually looks like the letter R, but because the character R onscreen references a specific bit, designated "R".
>>61552283>They read keystrokesWrong
>>61552171More like AI companies are tightening up their budgets and not letting fucking peons waste their compute spamming AI with useless questions and generating gooner content for free. They want to funnel everyone to their paid model so they make freeGPT retarded
>>61552134not GPTs fault that u cant spell retard
>>61552171Yeah sure they pay people to shill on a dead Nepalese goat fence repair forum with 20 users.
>>61552171It's actually far more likely that Sweet Baby Inc. is shilling against AI here because they have far more to lose.
>>61552134Google is like an akashic record, it will know anything anyway without a need to have intelligence since intelligence is just words from god anyway.
>>61552171more like everyone coming together to laugh at your redditor opinion
>>61552191>>61552283>hurr let me tall you how a markov chain works durrrr
>>61552134
>>61553983damn ai is dumb
>>61553988It got there eventually
>>61552134This is because it doesn't see letters but tokenized data? As a non-programmer I don't really understand why it needs to rely on this data tokenization thing for every single request. Can they not just have a baked in set of operations and reference tables, dictionaries, etc, and make it parse the meaning of the thing like, "user is asking for x data from y data they supplied. Apply the appropriate operation/algorithm? Do they consider that too much of a bandaid fix and they want to arrive at the correct answer with their preferred method?
>>61554036Never mind, I asked the LLM and it told me why.
>>61552134Does this prove that they hardcode solutions to common questions like the number of r's in strawberry?
>>61552134You have to tell the AI to run a script to count the number of r's. Please ignore the fact the dumbass who needs the AI to count the number of letters for him wouldn't have the aptitude to do this.
>>61552191>Look at AI image generation, text is always unreadable.AI can't do hands.
Works on my PC
>>61554095>Does this prove that they hardcode solutions to common questions like the number of r's in strawberry?Not really. Even a year ago an AI could reliably give you Python code to count the numbers of "r"s in a given string, but would fuckup if you asked it for an answer in plain English.Now with all the routers and multi models available they're probably just doing something akin to that in the background now.It's just a shortcoming of LLM's and their training sets. They're built on gigs of code but how much training data are they really going to throw at making sure it can count the number of letters in a string correctly rather than just writing the code for it.
>>61554663Would you use newton's sledgehammer to do calculus on 2+2? No youd use a simpler intiitive method
>>61554610>HandsFine>Chest hairlmao
>>61552150>>61552191>le dream stateyou're anthropomorphizing the weighted dataset again dear
>>61554095yesthey literally have millions of indians working as "data annotators" that are used to train and calibrate these LLMs.Without these millions of indians the whole thing wouldnt work. These AIs dont understand the data on their own.
