Ask HN: How are you using ChatGPT for yourself?

39 points by jeanlucas ↗ HN
I'll start: it became my notes app, because I can have insights on the notes I'm taking. Not an ideal use case, but it is fun for side projects ideas.

Example, I was fiddling with a PGN project idea while waiting for the Christmas dinner: https://chat.openai.com/share/eb975eaa-1c37-413f-a6a0-b17eb2f1bdf3

81 comments

[ 3.0 ms ] story [ 121 ms ] thread
I made a custom GPT and uploaded my books to it, and now use it for:

* A story bible (asking questions about characters, timelines, plot points, etc that I might have forgotten)

* A brainstorming tool that uses existing themes and characters for new ideas

* A rudimentary developmental editor or critic (does not replace my human dev editor, however)

* A character and environment art generator (this one is hit and miss with characters especially)

Tip for anyone doing this: make a separate GPT for each series. When I tried to upload books in two distinct series, even clearly labelled in the data, it started getting them confused. Works better to split them up.

Do you have any links on how to make your own? Interested in the same use case.
I'll write up a quick post on my blog for this with screenshots, but I hope the following can get you started:

So assuming you want the super low barrier to entry ChatGPT way, you'll need GPT Plus ($10/month). From there, you will go to this page: https://chat.openai.com/gpts/mine

This will list your custom GPTs. You will click on "Create a GPT". The GPT Editor will ask you for some information about your GPT. Keep in mind that GPT Plus comes with an hourly rate limit, and the creation conversation itself uses your GPT 4.0 quota... So you can _also_ just go to the "Configure" tab on that creation page and put in your prompts directly (this does not use your GPT 4.0 quota). Here's mine, for example:

"Description: A sci-fi romance writing assistant."

"Instructions: 'Assistant-Name' is a specialized writing assistant and professional developmental editor for the 'Series-Name' series, focusing on elements like steamy romance, dark themes, space exploration, power plays, sensuality, futuristic technologies, and aliens. This GPT will provide insights and suggestions aligned with these themes, ensuring that discussions and brainstorming sessions enhance the unique character of the series. It will avoid veering into unrelated genres or themes. The assistant is designed to recall specific details from the series, assist in developing plotlines and characters, and provide creative ideas that align with the established universe of 'Series-Name.' It will adopt a tone that resonates with the series' themes, being insightful, a bit edgy, and deeply engaged with the elements of dark sci-fi romance."

On that same "Configuration" tab, you can upload your knowledge base. In my case, that is each book in the series in .docx or PDF format. Tick the box that says "DALL-E Image Generation" and optionally "Web browsing" if desired.

Under the "Additional Settings" (hidden by default) config options, I unticked the "Use conversation data in your GPT to improve our models" checkbox. Just personal preference.

You can try out the assistant before saving.

I just bounce random showerthoughts I periodically have about various topics and it offers different perspectives I did not previously consider and helps connect dots in my head.

It's excellent at Socratic dialogue.

I sometimes use it to solve actual problems I'm having otherwise, usually software design related, and more recently involving Stable Diffusion since its April 2023 update.

May I ask you to elaborate on how ChatGPT is involved with Stable Diffusion for your use case?
You can ask it to generate images now, might be Pro only - it's not quite as good as Midjourney etc out of the box but works.
Asking it about more technical details about how SD works for some specific task. Nothing to do with its ability to generate images itself. Although with newer breakthroughs like Latent Consistency Models, it's aging like milk already.
Essentially automating project boilerplate that is custom enough to need attention, but not quite custom enough where it's interesting. Some examples include creating dockerfiles, various database models and data parsers, openapi specs, etc.
Nice !

The christmas card it generated wasn't even that awkward but very usable

Can it make Christmas cards by including some of your own pictures?
Love the Christmas card generator! That was awesome.
used it yesterday to learn how to calculate the linear regression constants all the way from multivariable calc down to converting it into the standard statistical form. it hallucinated a few times but i knew enough to ignore that and learn some techniques. all up a fun exercise that i surely would have been stuck on if not for the few insights.
I haven't used it in a while now, but I had found it productive for a few things:

* writing blog posts and outline for blog posts

* writing SEO descriptions

Generally though, its output was just a starting point, and I'd have to make a lot of edits of my own to actually include something of a writer's voice, remove any errors, etc.

I also found it fun to use ChatGPT to generate descriptive 'image alt text' for scenes and then plug that text into ComfyUI and A1111 to generate images. There were some cases where I was generating patterns and would use it to give me long lists of ideas for things, like I'll ask it "give me 20 ideas for christmas-themed patterns" and it spits out a bunch of stuff like "merry reindeer." Some of it isn't exactly usable, but it's also easier to have ChatGPT generate 100+ ideas than to do that yourself, which becomes pretty tedious past ten (for me).

I'm looking into using it for language learning, although I trust it less than something like Duolingo, which I already don't trust. I'm wondering how I can 'chat' with it in a different language and make sure that it correctly corrects my grammar and is also using proper grammar itself. Not sure how useful that could be, but if anyone else has had success with this, I'm interested in hearing how.

I ask it to write regexes, and I've had it write code that follows known algorithms.

In general, I tend to ask Chat GPT questions when: 1: I don't know what to search for, and 2: When I need it to aggregate and interpret context.

I personally am not using Chat GPT daily. I find that I still prefer searching, because often I'm either trying to find a specific website; or the human discussion provides a lot more insight than ChatGPT can give me.

A few use cases...

-I use it to clean up emails for work (paste in email, ask it to formalize it or make it more clear)

-My wife used it yesterday to create a rhyming scavenger hunt for the kids

-create Chrome extensions (tip, make sure it uses manifest version 3)

-I use it to create themed Teams video backgrounds

-Used it to create survey questions for Work From Home feedback for my staff.

Lots of other random uses as well. I've been a paying member since the start and it still blows my mind everytime I use it.

Care to elaborate on the extensions? I just started building an extension this past week and have been heavily relying on ChatGPT for help getting going. Just curious on your idea.

My first extension is effectively a real time cheater for Wordle in the browser. I will have it published by the end of the month and fingers crossed I can make this cross platform as well as iOS app. If nothing else, I just want to learn more about building browser extensions.

I'd say it's incredibly easy to build a basic extension, and if you know a little programming (at least to know what to say to ChatGPT) then you can get more complicated with your extensions.

I mostly get my results by simply conversing with it like I would a person. Tell it what you are seeing and what is wrong (ex "The font displayed on the screen should be larger and centered on the screen."). You can also ask it to start logging errors to the console, then simply copy and pasting the error and ask GPT what is wrong and how to improve it. Another tip is to provide some custom instructions in your profile and ask it to always provide the full code when updating particular sections. Without this, it will just give you the small snippet you need to update and you have to figure out where that goes. Probably not a bad idea to tell it to always use manifest V3 when creating extensions (there are some structural changes which would brick your extension when the migration occurs in June 2024. When you are done, you can also ask it for instructions on publishing your extension in addition to creating your icon set with Dall-e for example. If you want to charge for your extension, I think there is a tool called ExtensionPay which provides an API. ChatGPT might be able to incorporate that code (I haven't tried that yet though).

My email is in my profile if you want to chat more.

I am not using it, never have. I use open models.
Which are the best you've used?
This is my favorite model as of now - https://huggingface.co/Open-Orca/Mistral-7B-OpenOrca

It's not top on the Open LLM leaderboards but it's worked well for me. Haven't had a chance to look at Mixtral but this is the one I'd try first - https://huggingface.co/Open-Orca/Mixtral-SlimOrca-8x7B.

This was a favorite of mine some months ago but it has been surpassed by other Mistral finetunes.

To me the best one right now is neuralhermes 2.5, check out how I was able to teach it basic function calling in the system prompt: https://huggingface.co/mlabonne/NeuralHermes-2.5-Mistral-7B/...

Each model varies and it's highly dependent on the use case. It's best to try a few to see what works best. Definitely isn't a universal answer.
I found it has replaced a large percentage of google searches for me - particularly programming related. I can literally put the same keywords in ChatGPT that I used to put into google, and I'll get answers from various programming language docs, with examples I can follow-up on, with no SEO spam / ads.
I wonder how long it will take for advertising to get injected into LLM replies?
Wow. That's actually a great idea. Although, I know you didn't intend it as such. Haha.

But, now that you mention it, that's pretty much the future.

It will be better at serving ads than Google ever was.

So that's how Google fell...

....So that's how Google fell... their enemies that stood before them

With the power of its Gemini - Google was able to inject product ranking placement tokens into any prompt sent through its tentacles to the deep blue datacenter where, beneath the North Sea sits a kraken of subliminal advertising.

The new concept is to put secret subliminal ads in the responses AI makes across the board in order to secretly drive sentiment while the users believe they are educating themselves, they are secretly indoctrinating themselves to be the clickslaves Google sees them to be.

--

I just installed this last night on my laptop:

https://github.com/lllyasviel/Fooocus

Highly recommend:

>"Looking up from the deck of golden gate bridge at the towers and metal work, the towers rise and arch back in an ominous and foreboding manner. more artistic, like an alphonse mucha propaganda poster - slightly fish-eye feeling" -- https://i.imgur.com/vyNg79f.jpg

the local UI and 1.27.0.0.1 - https://i.imgur.com/wRwghuN.jpg

Why would anyone use a LLM with ads injected when there are alternatives?
"I would like to roast a leg of lamb. Kindly provide instructions"

You'll need:

> A 2kg leg of lamb from Bob & Mary's Finest Cuts [Maps Link]

> A bottle of Yarden Merlot [Buy Online] or [Maps link]

> Four organic red onions from [Maps link]

> ...

> [Click here] to order all the ingredients and have them by 6pm, after checkout I'll provide the cooking instructions.

Thank you for using ChefGPT Free!

Already have all the required ingredients? Upgrade to ChefGPT Plus and get the recipe in seconds!

Where's the story? This recipe has no pathos! Tell us about your childhood.
I sometimes use it for questions about configuration around tooling, but I have pretty mixed success. I don’t know if it’s because I use the free version, or if it’s the training material for the subject matter, but it is wrong about 50% of the time unless I know how to ask it the right way.
The quality of the insight is mixed. For example, it was quite capable of explaining subtle points about PyTorch API usage, but for Spring Boot it was mostly giving me quite generic advice.

  - board game rule clarifications
  - bash scripting
  - home improvement questions
  - cooking questions
  - third rail issues and opinion questions
I use gpt APIs for a tool I am building.

link: https://ropuz.com

and using chatGPT webapp for code and almost everything.

Generally I use it to try and waste as much of Open AI and Microsoft's money as I possibly can. I try and get it to use the absolute maximum number of tokens as I possibly can for free doing pointless tasks, I use the image generation to produce the weirdest art I can without triggering the censorship and when I'm bored I just tell it as many lies as possible and make it treat them as truth.

I don't know if this is feeding into the training data or not but I hope that it is.

Sounds like they're living rent-free in your life.
Is this not acting in bad faith? What is driving you to do this, is it not better to simply let it be and focus on something else?
Censoring their models was the first act of bad faith.

Sounds like OP is just running tit-for-tat.

tbh I always wondered why do they call themselves 'openai' since models are not open. But does that mean people should do things like what OP says..
It should not have surprised anybody that publicly run models would be censored. OpenAI has a mission that they are not keeping secret at all.

Apart from that, censoring what one's web service can be used for is a very common CYA strategy to avoid damage to the brand and not causing governments to push through overtly harsh laws.

Differently from social networks, model censoring does not even remotely restrict anyones constitutional rights. And who really wants to run big uncensored models can just use open source ones (which are approaching GPT 3.5 performance) and run them on 2nd hand GPUs from ebay since for LLMs bulk VRAM is more important than raw processing speed.

Wow! This feels very...weird, especially since, on top of wasting your own time and their resources, you go out of your way to actively ruin this service for others as well.
Sound like you're having fun :)
I've been regularly using ChatGPT for translating technical documents. It does a great job, definitely better than Google Translate.
Don't use it.

I tried using it for coding but it made up APIs constantly or had subtle errors. I don't know what people mean when they say it works for them. I can see it help as an icebreaker though. It just helps me get in the mood sometimes.

Other factual stuff ex nutrition questions, health stuff, historic events etc it's quite useless.

I am an LLM skeptic. I think there're a lot of people who're claiming that they find it valuable just like crypto was claimed to be valuable

People's brains work differently. I enjoy starting from an empty file, it was never an issue for me. But I know quite a few people who struggle until they have some kind of draft in front of them. I also know enough people with programming jobs that struggle with writing code, remembering all the keywords and everything. They're OK at reading it though.

Personally I'm in the same camp as you - I don't use it. It wasn't able to really help me do anything I do. But I can appreciate that it works for people different from me, good for them.

> I tried using it for coding but it made up APIs constantly or had subtle errors.

This has been my experience so far as well. I also find it very tiring to have to review all the iterations of the code.

I do find GitHub Copilot quite valuable at work, especially when it recommends one liners (they are easy to review, so I just feel like I'm writing faster).

To be fair to ChatGPT, on one of my recent experiment it actually said something along the lines of "you need to integrate with the library/API yourself", and it only started spewing garbage after I insisted to write the code. I see a lot of people online complaining that GPT is getting lazy, but I would rather have it lazy than wrong.

I sometimes use it when I'm lacking inspiration to craft polite LinkedIn or email response.
No, I guess I'm a luddite that way.

But I may get into it from a SaaS perspective eg. get into the hype since there are customers.

Pretty similar to how a lot of people in IT are using it:

- Generating boilerplate code for a process - saving time and then just customizing basic code after feeding it an idea and some basic requirements

- Troubleshooting - especially on systems I'm not familiar with - this often gets me 95% of the way to a solution. Obviously not just copying & pasting scripts it spits out, but it's been very good at figuring out some obscure linux issues, oracle config issues, even 3rd party integration software.

- Meal prepping - give it a list of ingredients and preferences and have it spit out a plan on how to portion it out, make each different meal roughly equal nutrition-wise, plan out when and how to cook each piece on a day to make it the most efficient, and even generate a shopping list. Output it into a nice table and the hardest part is done.

As a junior developer to sketch scripts or simple methods. If often knows things that I don't know, because it read all the manuals and all of SO. It stops to be useful after the second or third iteration on the code when it becomes apparent that it knows the syntax and the examples but it doesn't understand the purpose.
I’ve had a big need for abstract algebra in a side project and ChatGPT 4 has been The Best tutor (on top of several books and YouTube videos). Gives examples on the spot for your specific needs in any way you want. Gives broad overviews of a subject so you can make deeper dives elsewhere (Google, Math Stackexchange). Relates concepts to other parts of math. Implements math in my programming language of choice.

I’m sure it’s hallucinated a few times, but I’m at such an early level that it’s typically fine. I learn either way in a way that’s set me way ahead of where I’d have been without it

I use it to improve work emails. Often time my messages read like a series of bullet points strung together, and ChatGPT will rewrite to make it flow a little more naturally.

At work we get pressured to write comments backing up our ratings on performance appraisals. Not just the overall performance, but aspects such as if the employee communicates effectively, works well with others, etc.. ChatGPT is good at generating those without sounding too repetitive.

I've also occasionally used it to generate a meal plan following a my diet.

Whenever I have to write boilerplate code in a language or library I don't know well.

For example, "Give me a quick Python script that opens a window with PyQt5", or "Split a string on spaces in Emacs Lisp".

It's pretty much flawless for basic stuff like that.

This. Also writing boilerplate in a language I _do_ known.
heh - my final question in a months long panel interview for a position at google ~2007

Interviewer: "Quick, how to do a global search and replace in VI!"

ME, blanking: "...I'd.... google it..."

Interviewer: "best answer!"

(wound up not getting job for not having a degree, and Sergey was still personally signing off on all hires, according to them... (after they told me they were sending me an offer in the morning))

Just having a memory for a sec - some other notable interview moments:

Google: "What kind of company do you think Google is? (2007)

Me: "People think you're a search engine, but youre really just an advertising company."

And at twitter: "What kind of company do you think twitter is?" (~2011?)

Me: "I think you are a global sentiment engine"

(Didn't get the job allegedly because "I didnt read enough industry news" (seriously made what I read a big deal in the interview - and I fumbled trying to think of all the various things I read - this interview seemed to be looking for what physical publications I read, like DataCenter rags and such (I was a pretty successful datacenter design-team consultant at them time))

Using both ChatGPT and some open models to handle or test some personal automations like todo list and calendar management and email sorting.

Also used for boilerplate code stuff.

No longer. Too lazy and censored. Stopped performing even mundane tasks. Cancelled my sub and have been using bard and jetbrains AI for code. Haven’t looked back.