ChatGPT was lobotomized for coding and GitHub Copilot is broken... so I made my own!

Patreon (and Discord)

Substack (Free)

GitHub (Open Source)

AI Channel

Systems Thinking Channel

Mythic Archetypes Channel

Pragmatic Progressive Channel

Sacred Masculinity Channel

[frankebert]

Thank you for helping us all learn and helping us maintain the freedom to use the tools we need to create!

[Steve-xhby]

Thanks for posting this! I use ChatGPT for tons of writing refinement, and I noticed that its writing fell off in quality since sometime in May (with one of the updates). I don't know what they did, but it tends to repeat certain adjectives, verbs, or nouns over and over when it writes now. They broke it for writing also.

[landmonitor-lsd]

Helpful and straightforward - thank you! I don’t understand code like you do but I did notice it seemed to be less effective at making changes to complex Deforum video prompts than it was in April

[scotter]

Btw: this is beautiful and i am feeling much gratitude as your code may have gotten me past a blocker!

[patrickjreid]

Its so true! I explain what I want it to do then it goes all weird, I reexplain and it gives me the same output. It normally doesn't give me good code until I ask it, "Why do you think this code works?". That is my new best prompt. Once I ask that it seems to start thinking again.

[zzzzzzz]

the larger issue is that of centralized control of GPT . whether lobotomized or downscaled it is simply unknown what changes have been made . any setup still requiring a GPT API is not a solution, we need to switch to a local model and have the structure of interaction with the LLMs be agnostic to changing and updates .

[gileneusz]

lobotomized lol, maybe they did it on purpose to sell more "sophisticated" ChatGPT Business, which will fix the issue they just created by themselves, Of course they will never admit that 😆

[hermestrismegistus]

Can anyone give some context? In what ways is it worse for coding now than before? I've only recently started using it for coding so I don't know how it was before.

[platotle]

It definitely seems to have gotten stupider. For some reason I'm getting way better results with API (using the gpt-4 chatbot on my free trial of Poe, which I would switch to if the interface wasn't so horrible). Can you tell why API seems to be smarter? It would be nice to see you run test comparisons on it.

[MyCodingDiary]

Wow, this video is amazing!😲 Your programming skills are truly impressive.

[steve_jabz]

What do you mean by ChatGPT not handling coding that well? I've had the opposite experience and my friends have been saying the same thing.
Seems to have gotten way better in the last month or so. Infact it was working so well I rewrote my gpt-4 plugin 2 days ago so that it just prompts ChatGPT 3.5-turbo to act like it's the plugin system since it seemed like it could handle it and it worked perfectly.
How long ago are you talking?

[wanfuse]

This is great!
Thanks!

Note: I figured out that if you limit your question to chatGPT to 6 questions, and wait two hours between ( terribly inconvenient as it may be, that it actually will remember a much longer history) forgetting and reloading after a rate limit message causes it only reload a percentage of the history, chatGPT really needs a reference system, separate as it may be, I don’t really understand with 20 billion why such a system isn’t made, integration of chain of thoughts, multiple agents independent analyzing expert agents without knowledge at each recursion of other agents evaluations ( may need same problem, question or task to be presented in different sessions), training of new version on sample reasoning where each faulty reasoning output analyzed for why it is wrong and fed back into the model. Having interaction box, box with API reference documentation, pre and post processing, training of model to recognize common compression methods, such as RLE, and Huffman, language short hand, vowel and abbreviation, fraction representation of numbers, look up table window for variable substitutions with short non user friendly versions for substitution using intermediate system. Separate system that passes api’s documentation for lookup by chatgpt that is open source and editable by individual trusted creators for each api, ability to quickly add one’s own plugins, that meet the needs and security of individual using it, but that might be accessible only by the user creating them, eliminating third party untrusted sources for privacy and intellectual property concerns. These are a few of the changes I would bake.

[someguycalledcerberus]

So maybe I'm a bit dumb, but what does the scratchpad do? If I understand correctly the system message primes GPT to see the contents of the scratchpad file as python code input. And the expected output is also a python code. Do you write commented pseudo-code in scratchpad along with real code you already have and then GPT returns a functional code?

So for example if I wanted to ask the bot to print numbers 1 to 10 and wanted to use scratchpad to do this then I would edit scratchpad to look like this?
# Loop through numbers 1 to 10, ten including, and print them all out.

[KunjaBihariKrishna]

I've only been using my own chatbot. I briefly tried GPT4, but as long as I can remember, chatgpt is now so bad with code that it's very first response will contain errors. And it then corrects its own errors in the subsequent response. It's a bit too absurd to deal with

[cwspirols]

More on this please. Can you show an example of using this Chatbot?

[gileneusz]

btw the issue with GPT-4 now is that they stripped the chat version to 4k tokens, instead of 8k tokens before, this is why coding is so bad, because of context limitation it "forgets" what was written before and it's doing "stupid mistakes". This is easily fixed by using API, which gives you back the 8k context, BUT IT IS PAID and expensive 😬

[JeremyPickett]

I *have not tried this at all*, but there are some things that could be really intriguing. Like learning Lisp via python examples? Decoding IDA output into something that doesnt take Peter szor or eugene kaspersky (or russinovich) level chops. It is irritating that things like shell code are so aggressively filtered. I get ya dont want tp make it easy for the skiddoes, but when youve got a couple hundred sigs/rules/payterns you need NOW, decoding shell into constiuent parts so you can actually Find the immutable parts to anchor against would be a huge force multiplier. Or just keep us locked in the dungeon with (i kid you not) infinite coffee and nacho cheese fountains. No chips, coffee and nacho cheese. Thanks hurricane electric, my wife couldnt be around me for a week! 😎

[lerpmmo]

does anyone know if there are any tools that can scan an entire codebase and answer questions? i dont have gpt4 api access, was wondering if there are any free open source alternatives??

[codonomicus]

Tried this and it works just fine. Thanks, Dave, helps a lot! I dumbed mine down to 3.5turbo because I'm a cheapskate, but my coding needs are also dumbed down, so it performed just fine for me, just like the free chatgpt used to do before it was lobotomized.

[PaddySlattery]

My chatgpt is working fine but I'm here for you lowkey taking over!