And it feels like the output by ChatGPT is getting worse every month.
It’s learning from humans, that’s to be expected
deleted by creator
The irony of you being upvoted when reality is the exact opposite…
Model collapse IS NOT from them consuming more human-based knowledge. It is from them consuming AI-generated content it thinks is human.
This has been all over the news. How can you utterly fail to understand a basic phenominon? Because you like using the AIs?
Stop lying about reality if you want them to continue giving you any truth.
They are an AI
It is sadly ¯_(ツ)_/¯
your lost backslash: \
(⊃。•́‿•̀。)⊃
How appropriate
AI is honestly just A most of the time.
I remember about 10 months ago or so ChatGPT used to output some surprisingly top-tier code. I’d ask it to create a method with some required functionality and it would output the code, fully commented and everything. I didn’t have to edit the code, it just worked, and it was more or less efficient.
Now? I can’t even get it to write comments for code I give to it.
Interesting. You can’t choose which “generation” you use?
yeah the response quality is so much worse and it has that weird “take my answer and shut up” attitude now.
The free version or the paid version? Part of it is that they’re trying to push people towards the paid version, which is a much more sophisticated model.
Both, I was using gpt4 for some processing of text. The July 20th update came about and for the exact same input it could nolonger follow my directions, I had to tweak the prompt a bunch to handle a whole new set of edge cases.
Maybe it’s because I’m only using it as plan B or C (after the documentation has already failed me), but I have never gotten any usable code out of chatGPT.
And yet co-pilot is able to finish my code perfectly after I type the first few characters… even though they’re the same model.
Maybe same model but differnt data
Co-pilot isn’t using the same model. They’re using a model that’s been trained on a LOT of open-source code.
Alot of “open” source code ( ͡° ͜ʖ ͡°)
is the model open source?
Not that I’m aware of. Even if the input is public data, the actual training scripts and resulting model tend to be closed-source. Meta’s one of the only major companies I know of to release their models under a somewhat-open-source license.
ChatGPT is amazing for describing what you want, getting a reasonable output, and then rewriting nearly the whole thing to fit your needs. It’s a faster (shittier) stack overflow.
I normally have it output toy examples of the syntax I don’t want to bother learning and then remix that into what I need. IMO it’s better than stackoverflow because stackoverflow code is more likely to be not really what you were searching for or not actually run because the author didn’t bother testing it and there’s a typo or something.
I think co-pilot works better because it has the context of the whole project for reference when suggesting auto completion. I’ve gotten a lot of unusable junk from it too though
It can go from great to absolute junk.
But sometimes I need some weird terminal command and it’s weirdly good at it.
Also I get a buddy that can put log messages like “AAAAAAAA” all over the place so at least I don’t go crazy on my own
I go the other way with it. Give me something broken but close and I’ll use the documentation to fix it.
Haha indeed!
It’s funny when it starts to just invent things. Like packages, with version number!, that… do not exists…
Or when it outputs code without using the variables …
The most annoying thing is imho that it keeps explaining everything al the time. Even when I prompt “you have a working app with vuejs…” and others it sometimes still explains how to setup the app.
That said: the tool has become a staple in my workflow whenever I need a starting point. Or have to do some math algorithmic things
Like it invented an entire package to do TOFU in racket for me
Are you using version 3.5 or 4?
I think both are fine. But when I see 3.5 has difficulties, I usually switch to 4 and get the job done there.
Why use 3.5 at all in that case?
Because I only have 50 messages in 3 hours
I feel like for simple algorithms chatGPT could be good. Like as a reference for how to code something. But if it’s simple code I often find it faster to just write it myself then reorganize whatever it makes to work with and match the style of other code in my codebase. And if it’s complex code I often find it harder to describe what I want then to just make it.
In my experience, what makes gpt-4 great for coding is its astonishing knowledge of available software libraries, built-in interface features, etc.
I’ll tell it the task I want done, and it will tell me where to find, and how to install the necessary dependencies.
With zero experience in browser extension design, gpt-4 helped me to build an incredibly complicated Chrome extension, using vector database; creating a custom, cloud-based server; web scraping with headless browsers, voice recognition, speech synthesis, wake-word capabilities, and a sophisticated user interface. I had ZERO experience with ANY of these.
For me, using gpt-4 was like collaborating with a just okay programmer, but one who had extensive experience with literally every programming language, API, protocol, etc.
And it was a collaboration. We would talk through problems together. I would make observations and guesses about why a block of code wasn’t working, and it would tell me why I was wrong, or alternately tell me I was right, and produce a fixed version.
So what are you building? A browser STT interface for chatting with GPT and other LLMs?
I’m not ready to talk about it in detail. Even my boss doesn’t know. But you’re in the right ballpark.
I’m actually building a proof-of-concept prototype for what I want to work on… and I’m using a browser extension so that I can build it independently without anyone from the tech team being involved and slowing me down.
That sounds nice. I’ve been looking at serenade.ai and thought about extending their STT with an option to use another third-party STT engine. I would then like to extend their command engine with LLM command recognition. In my experience, maybe also with my pronunciation as a non-english speaker, their STT and command recognition really doesn’t work that well.
Have you tried Whisper from OpenAI? It’s the best I’ve ever seen. I’m curious how it would handle accents.
deleted by creator
People have had similar success with Bing Chat, and it’s free and uses GPT-4.
I feel that bing chat results are a lot worse than gpt4 tho
It has different rules from OpenAI’s GPT-4, so it might require a slightly different approach.
This is so apt ! though it does help to get difficult syntax for small fragments working quickly so you can get some proof for your concept instead of struggling with syntax errors for an hour
If there are even any docs… I usechatgpt when i can find no usefull docs. Quite often it can find some information somewhere.
Programming with AI help is like having the expert chef at my shoulder, giving me tips, but he’s high as hell on three different mild altering drugs.
Then he’s like “That cake needs some lemon juice. Trust me.”
If it’s lemon drizzle that would actually be appropriate lmao.
Yep. Every strange recommendation is either a brilliant technique I’ve not seen yet… Or just the AI hallucinating out of it’s mind.
Man you have to try lemon drizzle cake some time. It’s delicious.
Unless it’s Microsoft documentation in which case it feels more like bill gates beating me over the head with the frying pan until I give up and find an alternative way to achieve my goal
Or AWS documentation. Trying to use it feels like getting tortured with thumbscrews.
I hate it so much. It’s a cyclical maze of ignorance and frustration.
Google, on the other hand, may have my hand in marriage.
Is that Eddie Kingston?