I wasted 10-15 minutes trying to get ChatGPT to write a simple algorithm in Clojure and it fails miserably each time, from hallucinating functions to even hallucinating arguments to stdlib functions.
I've expressed the requirements in short bullet points, provided 5 test cases, instructed it how the stdlib functions it misuses actually work, and it still fucks up.
Meanwhile in 15 seconds I wrote a solution in three lines if code that passes all tests. Thanks for lying to me Bot.info. This thing sucks.
>inb4 buy ChatGPT Plus
No thanks. I'll wait for GPT-5 instead so I am not disappointed a second time.
Why did girlish number flop?
It was too girlish.
Try a different language
Give me another language with namespaced keywords. Neither Ruby nor Lisp offer this, they have keywords in a single shared namespace.
projection
cope
nakadashi inside her
>clojure
nobody cares neet
Clojure? The boots, boys, the boots?
they probably didnt train it on that dead language
more cope from a jobless undergroid
I'm reasoning with a language model, but I guess it was overtrained on some homosexual language like Python which is why Bot.info and HN rave over it being able to stumble its way into a half-functional, 10-line selenium script.
I've also tried getting it to answer questions about "popular tools" like Docker yet it will fail. I tell it to load a BuildX image and it insists on pushing to a registry first then pulling in image immediately after.
But yes, I'm aware this thing lacks any reasoning capabilities given that you can claim it is wrong for any reason and it will spit out the same "I am sorry" template, only to repeat the exact same thing as before. It gives me a strong resemblance of ELIZA, which does similar template pasting and can fool people for a while until they notice the reptitive patterns in its output
Yup. In the end I decided to implement something menial myself. I first tried using it to output some GitHub actions boilerplate with BuildX and it never gave me a solution; 5 minutes of Google did.
>inb4 some retard says "nooo its the prompts being bad!"
>spit out the same "I am sorry" template, only to repeat the exact same thing as before
This right here, it gives me broken shit and I tell it it's broken, then it just makes up some fucking function name and inserts that in there like some kind of "exercise is left to the reader" level of fuck you.
It's almost useless. OpenAI is talking about it like they created AGI or close to it. This meme technology will be completely forgotten in a few years. It slows you down more than it helps you. Unironically, porn might be its best use.
you're trying to reason with a probabilistic madlibs machine
of course it will fail
It really fucking sucks dick at anything remotely complicated. I spent a full fucking day trying to get it to build me a mongodb system that could use certificates I generated from C, and I ended up doing 95% of the work myself. It probably doesn't help that it was using old mongodb config names and shit, but it's nowhere near as useful as everyone is making it out to be.
Hopefully someone makes some decent performance and ACTUALLY FUCKING OPEN alternative that I can feed my own projects into.
Its alright for some things. You just have to learn its advantages and disadvantages
try racket
It's only decent for basic js and python shit. It's absolutely fucking dogshit at making database designs too.
>ChatGPT training dataset didn't include data about fossilized crap called clojure and therefore it can't use its dialect
what did you expect retard, AI to suddenly break laws of physics and write you accurate algorithm for dialect it does not know? NOOOOOOOOOOOOO MUH CHATGPT BAD, IT CANT USE QUANTUM NONLOCALITY TO SUPERLUMINALLY LEARN AND USE STUFF IT HAS NO PHYSICAL ACCESS TO
braindead midwits like you is why AI threads should've been on Bot.info
are you ok
I can make ChatGPT eventually ignore its filters. I just can't get it to do anything remotely useful for work.
I'll probably wait for GPT-5 since it's coming out by the end of this year anyway.
It definitely influences its output. It didnt bother with null checks or considering empty seqs until I added tests for those in a revised prompt.
>I can make ChatGPT eventually ignore its filters. I just can't get it to do anything remotely useful for work.
stop using the chat interface, use the platform payground. you can alter the system prompt so it stops doing that.
Several factors at play here, from my perspective. Here are some tips that can help you reflect on your performance:
1) You can't proompt. You suck at proompting and should just git gud. Simply put, it's a skill issue on your end
2) You're using the wrong model. Instead of using the free ChatGPT, consider using GPT-4. All the recent hype you've seen was about GPT-4, not ChatGPT. ChatGPT is GPT-3, or GPT-3.5 as the kids call it. GPT-4 is considerably more advanced, and expecting GPT-3.5 to perform at GPT-4 levels is unreasonable.
These are some tips that could help you get better results from chat-based AI assistants. I hope this helps, and if you have any questions, feel free to ask. Good luck with your project!
yeah give me your prompt ill put it in gpt4
how the fuck am I supposed to know what you're trying to accomplish turbotard?
>hurr durr I put "solve all my problems" into ChatGPT and it failed. AI is just vaporware
hey dumbass I was agreeing with you and telling op to post his prompt
maybe have gpt4 teach you some context
>write me a response to the Bot.info technology post complaining about chatGPT's performance. Write in the style of an AI-optimistic Bot.info Bot.info poster, and be sure to include the word "proompt" as a misspelling of "prompt"
>provided 5 test cases
I don't think that helps at all because ChatGPT can't run code, maybe Copilot 2 will let it run code locally
the model you tested is not only a general language model trained on mostly non vode databut it's also outdated. why not give an actual programming trained model a try like github copilot and then tremble at the implocations of the fact that those godlike results are achieved in these super early pre alpha stages of this groundbreaking new tech
you dont think we will accelerate towards diminishing returns?
I hope you mean Copilot 2, not Copilot. I tried Copilot several months ago and it feels like ChatGPT without the interaction as far as emitting code goes.
Is there any free demo or will I have to pay money to get baited by impressionable homosexuals from Bot.info?
Why do you fags keep saying "hallucinating" if the "AI" is just doing her job, that is, larping. As long as she's larping, then she's working as intended.