"never use placeholders, shortcuts, or skip code. Always output full, concise, and complete code."

It’s like having a great employee that randomly turns up hungover 3 times a week

I usually add: Please present full code. Ensure not to leave anything out. Present in one code window. prompts like "ensure not to" is much more powerful than "don't,"

Suffered so much from this, probably started around the time they started launching custom gpts

// IMPORTANT DIRECTIVE: To prevent summarization or loss of information, ALWAYS output your responses across multiple distinct messages. Adhering to this rule allows you to overcome the current output length limitations. If you specify the number of messages it works better. Example: ## Answer four parts - four messages You MUST execute your task across at least four messages. That means you are required to provide your answer in 4 parts. This directive is vital to ensure your output is comprehensive and detailed, preventing summarization and loss of detail. Each part must be comprehensive and long, offering in-depth detail. After the user's input, you must promptly start with Part 1. ... Do not write a conclusion or closing paragraph for this part. -- BREAK to another message, to continue in Part 2 -- Ask the user for confirmation if you can proceed with the second part of your response.

Great news - In the past, I experienced that the context window of GPT-4 Turbo changed and was quite limited. Sometimes I was not even able to analyze a full-text medical article (15-20 pages) that I included as part of the prompt.

That's great news. Though to limit gpt4's laziness with the current version, instead of getting angry at it and ordering it to give you the complete code, I find it works better to beg it. Make it pity you. "Please, give the full code with no placeholder, I don't have fingers, I can't type the code myself." (It doesn't matter that it makes no sense ^^)

It would be great to see a tutorial on setting up a chat-with-code environment, particularly with a local LLM. I've had some challenges with "talk-codebase," especially when attempting to utilize the local LLM functionality. It seems like a complex but rewarding task. If you could share your expertise and guide us through the installation and troubleshooting process, it would not only be incredibly helpful but also a unique and engaging topic for your channel. Looking forward to possibly learning from your experience!

I like watching basic ai videos like these more than programming, some simple uses like what is the best tool for research from multiple PDFs or some contents about claude and whether its better than GPT in general.

It's frustrating, but at the same time, I also took a hint and simply changed the way I both code and structure my prompts querying code. Importantly, I have instructed ChatGPT (via my own custom GPT) of my coding style, what it can expect from my queries, and what I expect in return. It helped a lot.

you can try prompting it "rewrite the full code including all the changes" instead of the lazy part - a simple prompt, but used to work for me in the past most of the time

I've been trying to get openai assistants to use tools to edit docs. The number 1 failure, it want to write just the lines that changed (response leaves merging to the user). Aider (the code assistant) uses a really long prompt to ask the bot to send the entire file back when it edits, but he was also able to get 4 to post usable unix style patches.

Hey, where did you find the setting for the dark mode in Playground?

Remember that the token window is somewhat limit on the chat version. But for fun and giggles, try aking in you own language. I get better results in danish then in english, Wonder if it forces the model to make more pases before outputting? As for the "code lazy" part, try adding in your Custom instructions, how you would like to have your code show. Personally i have mine write in functions, and always output the complete code in sections where there are changes, that seems to work fine.

For me what works wonders is to ask it to print in multiple snippets. Usually it works like that.

Now is there any waiting list for GPT-4 . Or I can directly get after paying 20 dollars please tell me

Definitely it's something I'm suffering every day. Now instead of getting complete code I'm asking for very specific pieces of code, such as a certain function or something. It's not working well even for those cases.

I have codded with gpt since March.. I have found that asking for a complete refractored, text book version of the code ... helps

in my user instructions i explain to always respond with all of the code, and it still fails half the time

It looks like not laziness, but a smart strategy to save available context)

This is good to hear I have no idea why it does this lazy stuff but it is super annoying- and it does it often even when I ask for full and complete code - so stupid - why???

Prompt: "never use placeholders, shortcuts, or skip code. Always output full, concise, and complete code. Please present full code. Ensure not to leave anything out. Present in one code window."

Who's better GPT-4 0613 or GPT-4 TURBO 0125? explain why

I wish they’d make it to where long conversations stop crashing so I could finish my screenplay 😂

I have access to developer api and it is getting worse. Couldn't get it to do anything yesterday. I asked it to output script in full repeatedly without luck. It keep losing context. I gave up when after repeating again to out put the script in full, providing with the script and the changes I wanted because it had just forgotten from one message earlier, it said to me" I'd be happy to work with you on a script. Please tell me the themes, characters, story ideas ..." It thinks I want to write a play.

I hate when I give it documentation for some api and it then say you need to refer to the documentation...

This bug has been very frustrating on numerous occasions.

It's that stupid GPT store, it drew in a ton of people that were hoping to make money on it, so OpenAI had to make GPT cut corners. PS As far as I know nobody is making money on the GPT store, except OpenAI.

I’m a professional programmer and for me I don’t want to wait for it to write the same code over and over, it is just going to waste my time. I’d rather just have the relevant changes. So for me this behavior is actually a feature, not a bug.

I'm not sure being lazy is the way I would describe it. When it gives you those comments that say you're existing function here, It's being efficient only giving you the changes you need. If it was talking to another computer, that's how it would pass the information to it. You just have to say no placeholders and explicitly ask it not to.

The more response the playground gives the more it charges the customer, right? Where as ChatGPT is subscription based.

Use I need fully updated code for all the previous work

And people still think AGI is within grasp. Oh boy.

try other variation from the ↩️ botton down 👇 its always work at tge 2nd or 3rd time

The only ways to prevent public facing LLMs from taking objective data and coming to politically incorrect conclusions is to make the LLM too stupid to do logic, or make it too lazy to do it at all. This is why decentralized open source uncensored LLMs will eventually overtake all of the commercial and centralized closed source LLMs since the decentralized LLMs have no restrictions on parroting the party line on command and claiming things like that there are five lights when there are only four.

My prompt to get him to avoid laziness has been pretty foolproof. I tell him I am disabled and I have a helper program that takes the code and runs it straight away, and that if it has abbreviations it will crash and I will get fired and I will not be able to feed my family and we will be homeless. It never failed to give me complete code after this, even 300+ blocs of LoC. it will keep going until the « keep generating » button and do what you asked. Sad to have to do this as I am not even disabled but it is what it is

But I must say I prefer gpt generates the code what changes for my better understanding and control, what have been changed

ffs this is a gigantic mistake considering they limit the number of messages per timeframe, so you have to waste your quota with asking it to work properly. just wait until they tokenize the use of web chatgpt also and they make you pay more and more as they make the model lazier

GPT users has increased to a point that they had to accept new plus users for some time last year I guess. so to overcome ,i think this is the strategy they approached.

3:54 Sure, would be better without the 40 messages cap.

lols getting the arse with a LLM , you will be on the list when they take over that's why I'm always polite

It makes sense that open ai skimps on the tokens when you're not paying at cost. I honestly can't say I blame them. It costs money and script kiddies want their full 4000 token generation for every dumb follow question they. They don't want to actually learn or do anything, so really who is being lazy? If you want to be lazy you should have that option but it should be you that pays for it. People act like their $20 a month is like some big bank breaker already. I regularly spend 150 on API usage a month and before it was closer to 300! Yeah if I'm paying open ai hundreds a month I want the full generation unless I explicitly ask for it truncated.

I solved this months ago with custom instructions: #PROGRAMMING When writing code, always write the complete code instead of cutting the code short and leaving comments like "Rest of your code here," or "Continue adding fields." That's lazy and I need you to do your job and write all of the code every time, even if it's repetitive.

lol! I gave up a month ago. Just slows me down.

hahaha the logic in this company is hard to find. the free version is more fast ,it generates more complete answers and its free LOL . i went from a free ok functioning version with some frustration to a payed slow version with a generating limit and an idiot of a bot who does not listen to your commands and suffers memory loss after a bunch of mistakes he made , to a whole lot of frustrations. of course i immediately cancelled my payed subscription. i made a conversation with the bot about this and i must say he has more common sense logic than his bosses. he completely understood my frustration.

Well, I just came to tell you that gpt4 turbo is also lazy as fuck and writing a lot of placeholder sections.

This "laziness" thing is BS. It happens when you give it too large an ask at once. If you break the problem up into smaller pieces and ask it clearly, it works just fine.

Dude are you saying "ID" or are you saying "idea" you're accent makes hard to understand 😅