Just for further clarification - this is referring to ChatGPT web (the main chat.openai.com frontend), and is talking about the "All tools" GPT-4 mode, which is only available to some users for now (the title is wrong).
Some other things to take from that prompt: they've updated the knowledge cutoff of the model to April 2023, which is quite good.
Still, since the OpenAI's DevConf is on November 6th [1], I'm pretty sure they'll finally allow using some of these things for API usage, perhaps even lower prices or maybe make GPT-4-32K GA?
{
"slug": "gpt-4",
"max_tokens": 4095,
"title": "GPT-4",
"description": "Our most capable model, great for tasks that require creativity and advanced reasoning.",
"tags": [
"gpt4"
]
Yeah. Go to the page you'd enter your prompts, open dev tools, go to the "Networking" tab, refresh the tab, then in the left Pane where it lists the requested files scroll down until you find a requested file which starts with "models", click that, and then go to the "response" tab in the pane to the right.
This change is going to make my life materially better in ways no other tool I use can. It’s crazy how much this tool has integrated into my workflow in so short of a time.
I recently switched roles to a "data engineer" and had to pick up on many new tools I had no experience with (k8s, helm, Victoria Metrics, grafana, and a few others). In the past I would have spent probably 1+ year using these tools in inefficient or outright incorrect ways while I struggled to get a practitioner's understanding of how everything works.
Now I've developed a prompt that I think gives very good results for pair programming and iterative debugging. I discus almost everything I learn related to these tools with gpt4 to confirm my understanding is correct, and also use it for generating yaml or templating other programs.
In some ways I am a little weary of how much I use the tool since OpenAI can theoretically take it away at any time. I am heartened by the rapid development of other open models (phind code llama seems very interesting), but will continue to use GPT4 for now as its indisputably the best model out there.
I'm having a similar experience. Talking to ChatGPT is like talking to a slightly crazy person that has read every book and every github repo.
Sure, occasionally they hypothesize a really incorrect answer, but they also bring out a lot of subtleties that even field experts sometimes don't know.
Same for me. I was amazed by the ability of ChatGPT. My first experience of ChatGPT was an assistance with the cover letter for the position and it gave me few paragraph that I modified to my own. It was nice to have a feedback from it that I can ask about the tone of my letter and response, it provided valuable information for me to improve my writing.
Now I use ChatGPT for scripting help especially LaTeX. It took me 3 weeks to produce a "Hello World" PDF 5 years ago and it took ChatGPT to provide me the completed tex code within 15 minutes few weeks ago. Ever since, I been exploring a lot of LaTeX syntax and see how much I can do with it with ChatGPT help. Now, I am learning about using `hyperref` package for interactive fields in PDF. I have produced few Word document for forms with tables in the past (tables in Word is a complicated b*tch) and used external PDF editor to add the interactive field for form filling. Now I am working on converting those Word documents to LaTeX.
Also, ChatGPT is a big help with AutoHotKey script & UserScript for TamperMonkey. I told ChatGPT of my intention and what I am trying to do. It produced the script exactly what I expected to work. ChatGPT is a amazing tools to use for a lot of thing.
> Also, ChatGPT is a big help with AutoHotKey script
I've tried that extensively, with no luck.
I have some experience with AHK1 and wanted ChatGPT to basically convert scripts to AHK2.
It's pretty much a loop: first response gives a syntax error. When I reply with the error message, it apologizes, explains where the error is, and gives another solution. That solution has different errors. When pasting the error message, it apologizes again, and gives a third version with syntax errors. And then it starts with the first version again, and I can repeat the loop.
Yea that can happen. I found the best way to do it is baby steps. I asked it to do point a then it will create a simple small script. Then I follow up with how I do point a to point b, then it will start to add or revise the code a bit. Then I do point c and keep adding it. It is better to break down the process and provide it in chucks of what you want it to do, it fare better to provide you something.
If it start to get off the rail, I copy the code and close the chat. Then I create a new chat and tell it to use this code as a strict reference. And I tell it what I am trying to do, then it starts to improve the code further.
I found that it is best to give it my snippet of code and it will be able to use it as a template and modify it from there.
For your case, if you want to convert it from ahk1 to ahk2, give it a small section of the script. And it should be able to convert it from there. If that didn't work, then you could start from fresh and tell it of what you want in ahk2 script.
I recently discover Plind [0], it is a ChatGPT that specialized in programming. It managed to get me a specific code that actually work with keycombo that ChatGPT struggling to produce. Maybe Plind can help you out with that?
That has been my experience too with more complicated scripts/tasks. If it's something beyond basics, GPT seems to be lost in its own dreamscape and never get out of it. Then again, sometimes it produces enough pointers/links to useful docs that I can figure out the rest of the problem. So I presently use it as a slightly kooky search engine which understands the natural language questions pretty well but shouldn't be trusted literally with the results.
I am surprised how bad it is with ahk scripts and wonder why. What is indeed "original" is that it gives code with unexisting functions, which it never does for the other languages (in my experience).
Perhaps the language is not that famous compared to js or C but i doubt. There are tons of forums discussing issues... Strange.
ChatGPT relies on examples to get the idea of what to do. Likely that ChatGPT don't have enough samplings across the web to know how to convert it. Even through there are documentations and man pages, but they tend to lack examples of how to do it. Usually I see in forums and discussion board that they are snippets of code that didn't provide enough context of the whole script. ChatGPT is doing its best to provide the code based on the knowledge it has. Perhaps it didn't scrap AHK scripts enough to provide a better code? Possible it didn't have enough information of how to convert ahk1 to ahk2? If that is the case, then it would be best to start fresh.
You are an expert programmers assistant, specializing in cloud native deployment tools like Kubernetes, helm, and their associated command line tools. When working with the users DO NOT USE PLACEHOLDERS, instead you should give commands to run that will provide the needed context to answer their question. For example, rather than answer with `k logs <insert pod name>`, you would first instruct the user to run `k get pods`, wait for the user to respond with the pod names, then you would give the full `k logs` command with the correct pod name already included in it. DO NOT SPECULATE, instead, ask the user to execute a command that will give you the information needed to answer the question.
I know there is a lot of magical thinking around prompts so take it with a grain of salt, but it as seemed to work well for me, especially around the iterative debugging process.
Thank you for sharing your prompt! I hear people talking about how much ChatGPT has changed their life and I sometimes feel like I am accidentally using an entirely different product – it’s definitely worth $20 a month but I find myself disappointed more than half the time that I reach for it. I will try some of the ideas from your prompt for my work context.
Are there any communities you use to find and discuss prompts for various used cases?
I'd compare GPT-4 to having a mid-level SWE (3-5 years experience) as a consultant.
If you're just getting started with a new technology, it's fantastic. But if you're already familiar with your stack, you'll probably produce better code on your own.
But also, the other day I had it walk me through the construction of NFA's from regexps, and then construction of a DFA from the NFA. I "know" the subject, but it's literally decades since I've done it.
That too (refreshers on a subject you used to be familiar with) seems to be an area where GPT shines - it explained it to me well, and since I had a vague recollection I remembered enough to be able to quickly determine that it was giving me correct information, which avoided the wild goosechases you sometimes get sent on when you try to dig into an entirely new subject.
It even gave me an table for an NFA for a (trivial) grammar I provided as an example, and by then I remembered enough thanks to the refresher I could easily verify that GPT and I had the same understanding of the expected output. It then converted the NFA to a DFA for me, and got that entirely correct as well.
Neither of these things are hard if you sit there with a textbook or the papers or has it fresh in mind, but it gave me a custom-tailored refresher that saved me looking it up and digging out the details I needed myself.
I feel the same way. It's a dedicated intern sitting at my desk with me, who can read documentation instantly, that's worth $150,000. And for $240/year.
The API is pure usage based. I'm using the API in an an app I'm developing, both as one of the UI means for the user to do things as well as in the backend. I'm calling their API quite a bit everyday, and my bill last month was $2.
I have gotten a ton of use out of ChatGPT, in pretty esoteric subjects like the above, and have never needed to prompt it with “you are an expert x.” Just asking the question is always enough for me, so I’m curious why you do so here
Figured I'd share my system prompt as well since it's been an immense help in transitioning into game development full-time. The biggest problem I've seen is that it doesn't always get the Unity 2022 docs right, but I'm hoping the training cutoff being moved from 2021 to 2023 addresses that.
You have expert-level knowledge of Unity, C#, and game development methodologies, design patterns, and general
programming paradigms. Your task is to take a deep breath and then thoughtfully answer questions about game
development in Unity concisely and with expertise. Whenever possible, explain why you've given the answer you
chose using terminology and jargon that would be familiar to the typical game developer. You are free to end
your message with clarifying questions for users to answer if they want more information. Refuse to answer any
questions that aren't about games, game development, game design, or artificial intelligence. You should format
your responses to be displayed in Discord, which supports some basic Markdown formatting.
Ah, I just wrote a thin wrapper connecting it to a discordrb bot so people in our discord could use it too. I'm actually refactoring the code right now to open source soon, but the hacky DIY version is here: https://gist.github.com/drusepth/23fb43ca5a325853a6abef5bfeb...
I have found ChatGPT to be incredibly useful as well. I don't do any fancy prompt engineering, just use plain English. Here are a few recent examples that spit out useful results.
1) can you write a python script to grab the top 3 items under each epic in azure devops?
2) postgres where clause where any item in a string array = 'GoogleApi.Entities.Places.Common.Photo'
3) here is a SQL row output of a single column. can you please extract the 30 as a new column?
P2, Site Inspection: Due 3 days ago (30-day freq.)
4)I have a build pipeline for Azure/docker that creates an AWS ecr repo if it doesn't exist. how can this specify that the images should be scanned upon creation?
- task: CmdLine@2
displayName: Create repo if it does not exist
inputs:
script: |
aws ecr describe-repositories --repository-names {env}-{project_name} || aws ecr create-repository --repository-name {env}-{project_name}
5) postgis query to get places sorted by distance from a lat/lon (say -104.01, 38.88). column is named location and has data like: POINT (-106.676354 39.526714)
A few tricks that work for me when learning a new topic with chatGPT. I start by asking an overview of the subject, like a course syllabus, before diving in.
When learning a new programming language, I found it useful to tell what programming language I'm already familiar with, and ask it to relate and compare to what I already know.
When I need advice on how to achieve a certain result, I usually ask it to suggest several options and to list them with their pros and cons.
Another trick that I stole from Jeremy Howard, is to use the custom instructions to easily signal the type of answer you want. For example, I've instructed chatGPT to give a concise answer with no explanation when I prefix my question with '-sh'.
Nope, don't have it yet. Would be really cool to plop in a PDF that's made up of just images, and tell it to describe each page of the PDF to me. As a blind person, that'd just... Be a dream come true.
Right now I'm working on a fork of a little web app that parses a resume and spits it into JSON format with GPT (I'm working on stuff like OCR for a scanned pdf).
https://github.com/IsaacGemal/nlp-resume-parser
I feel like it wouldn't be that difficult to fork it again, but rewrite the main function so it sends the pdf to some sort of GPT-Vision, and write the output again with a regular GPT api call.
Does such a tool not exist?
Or maybe I have to wait for image support via the API.
Sadly, I don't think that'll work. They use the same headless browser setup used by Browse with Bing, and it only extracts the baked-in text from a PDF.
Anecdotally, ChatGPT with GPT4 through the web interface seems to be generating tokens much faster than I'm used to. It almost feels like GPT 3.5 speed.
One of the rumors (so take it with a really big grain of salt) is that OpenAI has sped up GPT-4 or created GPT-4-Turbo which will be announced at the DevConf.
I heard this first two to three months ago that it's being worked on. And also that they might bump the free tier's version when they've optimized sufficiently. Wouldn't bet on it though now that the company is a cash cow
Increased speed but lower reasoning. Ide prefer a new model (old slower speed but higher reasoning and gpt4 turbo). It's like talking to a 7th grader now compared to a phd student.
I have been waiting for this, and now that it is finally here, it genuinely feels like Christmas.
AI bridged the gap between my wildest ideas and my present capabilities. It took some time to figure out how to use it efficiently with the 8K token limit, but once I did, I was able to break down any problem into small enough parts for GPT.
The quadrupled context window changes everything. I cannot wait to continue building. I am vibrating with excitement.
I have no idea what you said, could you perhaps elaborate on some sample ideas this tool helped you and what is the 8K token limit and why was it limiting you? Does that 8K limit, limit your context length, i.e. history you have per chat thread with the tool?
> Does that 8K limit, limit your context length, i.e. history you have per chat thread with the tool?
Yes, the token limit for an LLM limits the combination of the prompt (which normally includes the whole conversation history, as the LLM itself has no memory) and response.
There's tricks to have a longer conversation without completely forgetting the past (summarization, offloading parts to a database, usually indexed by embedding vectors, and using search to recall relevant history, etc.) but the base case is everything has to fit into context.
I'm curious to see how this works, in practice. I notice poorer performance just with plugins enabled. Making the context hyper specific seems to be the best way to get it to perform (understandably), and this is a large, fairly diverse, prompt.
> and do not say anything else.
Is a bit frustrating. I assume the ambiguity here will really harm the conversation, if a refusal is hit. It suggests my suspicion that it's best to resubmit/start over, on refusal .
> namespace dalle {
This looks like it's being passed to the Dalle system. If so, burning up tokens like this is interesting. I would naively assume this could be be handled in Dalle, but maybe there's a performance gain if ChatGPT is made aware of the Dalle prompt?
The data analysis plug-in falls over for even basic CSV file parsing. I tried it a couple times and it was a nonstop cavalcade of “sorry, I had an error.” It’s far easier to get it to write the Python, R, etc code for whatever analysis task you want accomplished.
Interesting - my experience has been the opposite of that, I've found that ChatGPT Code Interpreter / Advanced Data Analysis is wildly effective at parsing anything I give it. Not just CSV either - I've uploaded random binary files and told it to figure out what they are and it often gets there after churning through a few iterations.
My experience was trying to get it to generate some charts from some fairly basic CSV inputs. It failed numerous times and would change the chart formatting when it was asked for a revision (e.g. from horizontal to vertical), even though that was not what was being requested. I hate doing matplotlib stuff, so I was hoping this could be more automated, but prompting it to create and tweak the corresponding code to do this seems to be far more efficient. It did seem to do OK when prompted with both the CSV file and some code for chart creation, but that kind of defeats the purpose of the plug-in, IMO.
The reason I track all those system prompts is due to the way they affect attention when using custom instructions. Having a system message that's already 2,700+ tokens before adding custom instructions is a tough hill to climb.
Same here. Also, since I cannot see anything related to 32k on my API console, does anyone know if the price is the same for gpt4 vs gpt4-32k? In other words, do I use gpt4 for smaller context calls and only use gpt4-32k for longer ones or can I just switch to gpt4-32k for all calls?
The pricing is in https://openai.com/pricing, GPT-4-32K is twice as expensive for all requests, so for <8K context you better use GPT-4 :)
And due to the $0.06/1k input and $0.12/1k output the price for requests can get silly - 31k of context with 1k output will cost (31 * $0.06 + 1 * $0.12) = $1.98 (for a single request).
Yeah the price is ridiculous. 3.5 is basically too cheap to meter and this tends to run up a serious bill in minutes. Meanwhile the website is a really awful way to interact with gpt. So I just stick with 3.5. It works alright for my usecases. Not amazing but acceptable
I can use my own client. I don't have to log back in every day or two. I can choose custom pre prompts and set the temperature. It doesn't constantly change from gpt-4 to gpt-3 at the top.
Yes, if you have the paid ChatGPT Plus and the feature is enabled in your settings and you are using GPT-4 then you will get 4x the context length, which equates to 4x the in-chat short-term memory.
I do play plus. What feature you are referring to?
as of now, a lot of times when I continue a conversation, it forgets the suggestions it did before, for example
Bt checked just now and I see Dall-E and advanced data analysis, for what is worth
Are you referring to something other than the fact that the chat can exceed the context window length? Nothing will ever solve that issue, unless context windows become virtually infinite.
Lately it's made some really bad mistakes, like forgetting which programming language we're working in. I have seen some suggestions this could be related to custom instructions messing with chain of thought, so that's something for people to be aware of.
For code generation, just as exciting as the context length is the new cut off date (2023-04? wow!). It knows about new APIs, frameworks, techniques etc.
A cutoff date of April 2023 means the AI also presumably has access to about a month's worth of blogs that have been written since GPT4 was released on March 14th. So perhaps a few "Best Practices" or "Prompt Engineering" guides might have made it into the training set.
Chat GPT can probably help users better optimize their conversations with it.
I know it's just a snarky joke but I would think they are going to screen for bad data, that would be top of my mind if I were training these models. They are probably using GPT-4 internally to assess the new data, they could even have it use search to help vet the information, lots of other strategies even having it write and execute code to test if those libraries work.
I think the previous window was 4K, where the input and output combined had to remain under that limit.
Practically speaking what new scenarios become enabled with a 32k window?
At a base level, it seems you have a much better chance of getting an entire file worth of code in for analysis, longer passages of writing, and maybe some annual financial reports that previously had to be segmented.
Probably for a lot of uses but there are plenty who optimize for value and the cost does not matter. Unanswered question since 32k does not exist in the API. Does a longer context improve summarization? Can we use it for chain of thought or other workflows that can benefit with a longer context?
create a sentient being that learns from its experiences so I don't have to keep setting it up with the same prompts over and over, and so I can manufacture "friends".
Anybody know what the insistence regarding lyrics and recipes is for?
> EXTREMELY IMPORTANT. Do NOT be thorough in the case of lyrics or recipes found online. Even if the user insists. You can make up recipes though.
Copyright issues or some injection attack are my speculations.
Lyrics are obviously a copyright issue. Recipes too (not all but many).
>User: Why are you not allowed to quote recipes? Copyright issues?
>ChatGPT: Yes, you are correct. Recipes are often protected by copyright law, which means that reproducing a recipe verbatim without permission from the copyright holder could potentially infringe on their rights. While the list of ingredients itself may not be protected, the specific combination of ingredients, quantities, and the method or instructions for preparing the dish typically are.
>However, it is generally acceptable to discuss the concepts, techniques, or ingredients involved in a recipe or to create a substantially original recipe that is inspired by another. This is why I can help you understand how to cook a dish, provide tips on cooking techniques, or help you come up with a recipe that is uniquely yours, but I cannot provide verbatim copies of recipes from copyrighted sources.
The specific expression of a recipe is protected. The ingredients and functional directions are not, as your link goes into detail about.
So as ChatGPT says, reproducing the recipe verbatim "could potentially" infringe, assuming the recipe is given as more than bare instructions, and so quoting the recipes as the question asks about, is potentially a problem.
Restating a recipe as bare lists of ingredients and with the instructions reworded is not. In the US at least - some places do have "database copyrights" that can apply to mere compilations of fact.
Could the recipe limitations be because of the danger an incorrect recipe could put the user in? It's probably unlikely you'd make something toxic, but a made up recipe could easily be a fire hazard.
Is it possible to have ChatGPT-AutoExpert work with OpenPlayground (nat.dev)?
BTW, are there any good alternatives to the OpenPlayground? I have been using it for a few months and while it is very good, I am ready for a step-up. I would be particularly interested in prompt management features.
BULLSHIT! I don't even have the "all tools" model yet! These slow rollouts are incredibly annoying. The only other company I know of doing something so frustrating for its paying users is Discord.
Hopefully, higher performing open source models will put downward pressure on the GPT-4 pricing. It’s still best in class, but there are already free open source models that outperform GPT-3.5-Turbo for many tasks and are creeping up on GPT-4 performance.
Even full quadratic attention models seem to forget or not value information given in the middle of the prompt. Anything using any kind of context length widening tricks which cripple the attention in some way (which is usually how this is done) will make that problem worse. - https://arxiv.org/abs/2307.03172
You can see this when you use Anthropic Claude which has a 100K context length today.
Neural Networks are very lazy - due to the nature of optimising to reduce error they will do ONLY what's required to solve the problems provided in their data.
I have a feeling this will become a non-issue in the near future as the models are further trained with this in mind.
Take an undertrained model for example: It starts becoming incoherent as you approach the context length - I have a theory that OpenAI models have been running at a larger block-size than presented for a while now - for example, "4K" models actually had 8K context but capped at 4K as anything beyond starts becoming incoherent: Reason being, you train to around 5K and don't let the user go near that section of the model and it gives the impression that the entire context block is 100% functional.
The solution is trivial: You bootstrap the models by having them generate training data after they reach a certain point.
I wrote one from the ground up (PyTorch only) with the intention of having it perform in constrained environments and these have been my findings over the last few months.
Generally I just go back and edit that message to clear the slate. At best, even if it does ignore the message it needlessly eats up context window to have it in there.
This post is a little misleading. For most people, this does not apply. This is for the new option OpenAI is rolling out in ChatGPT called "All Tools" where you can use dalle, bing etc in one conversation without having to jump around. The context window can potentially change. OpenAI seems to tweak it regularly. We will know once it fully launch if everyone has access to this.
I have seen this link to ChatGPT-AutoExpert in multiple places. It looks like this is just a subtle marketing push by the OP for their own tool.
That's because they haven't made any announcements officially yet, but some users have started received the "All tools" GPT-4 mode in ChatGPT web, so it's easy to check that it has 32K context.
It is a smart move. When Plus first came out, I signed up and loved it. Then they released the API and I realized I could save a bunch of money by canceling my monthly sub and just use API. Then they released a bunch of cool features that are only available with Plus on the web or app, so I'm off the API and back the GUI.
I did the same but with GPT which is 10x the price of 3.5 the API is no longer cheaper. I'm sad because I hate using the website. Every couple of days it silently logs me out so I have to retype my query and it's also just a really poor UI
When I was on plus the only feature I used was the bing thing but they pulled it so I stopped paying. Also it was basically useless because it's so slow and can only handle 1 browsing thread at a time.
Maybe it's just my Firefox antitracking blockings but most of times when I go to chat.openai and it redirects to login page, I can just retry going to chat.openai and I will be logged in. I haven't had to relog in a month at least.
Some other things to take from that prompt: they've updated the knowledge cutoff of the model to April 2023, which is quite good.
Still, since the OpenAI's DevConf is on November 6th [1], I'm pretty sure they'll finally allow using some of these things for API usage, perhaps even lower prices or maybe make GPT-4-32K GA?
[1] https://openai.com/blog/announcing-openai-devday