Ask HN: How are you currently using AI (personally or professionally)?

For my personal notes, I use Smart Connections[1] with Obsidian. Since the notes are stored in plain markdown files, I am considering devising my own solution using LlamaIndex[2] in the near future.

For coding, I use Copilot[3]. While it's been great for writing boilerplate code, it falls short in every other regard. I also had the opportunity to try the new version of Copilot as well, but it feels like a glorified ChatGPT inside VSCode.

For everything else, I use a tiny tool I made[4] which enables me to invoke my own prompts in basically any application that allows me to select text.

[1] https://github.com/brianpetro/obsidian-smart-connections

[2] https://gpt-index.readthedocs.io/en/latest/getting_started/s...

[3] https://github.com/features/copilot

[4] https://github.com/overflowy/chat-key

ChatGPT Plus for tedious coding stuff and sometimes throw in user instructions/error messages and ask it to give me ten alternatives, and see if it can do better. I also use it to give me a starting place/draft for personal correspondence. I find working from a draft and "fixing it" far easier than a blank sheet of paper.

I have plans to use LLMs to help me create a document format converter (XML to XML via XSLT), but unfortunately most of them don't support enough tokens to take in both the starting document's format and also detailed instructions on what to do. So I am very token capped on that project currently (and also GPT-4 requests-per-hour capped, least of all because I need to break my tasks apart). I've tried converting the starting XML to a series of xpaths, but then the LLM loses a lot of context and the result is ultimately less-good.

I'd like to give LLaMA-2 a play for my conversion project, but haven't had the time, and don't ultimately know enough to know if token limits are a technical limitation or a financial limitation for the services (i.e. can I give it unlimited tokens if I don't care how long it takes to complete?). Plus getting it up and running locally isn't exactly turn-key right now.

I tried Bing and Bard too; both support some web integrations which is nice, and you can generate some [bad] picture content, but they're both worse than ChatGPT for minimalist chat with Bing being too much nonsense/irrelevant Bing-service-spam (and often cuts you off) and Bard just being worse than GPT-3.5 let alone GPT-4.

I've been working on DBT lately. ChatGPT4 works great for it and SQL. It saves a lot of time when it comes to rote SQL.

"Here's 3 CTEs. Finish the model by joining them and grouping by X, Y, Z".

I'm also sadly not working with anyone more experienced then me so it's a good discussion partner for discussing tradeoffs or alternative ways of doing things. For instance, I created a model which did a self join to get create a new column with the previous years value. I asked it if I could do it differently and it suggested the LAG function - however, it doesn't deal with leap years so it wasn't viable in my case.

It's pretty good at error messages and finding faults in code, so it's usually a good place to start if I'm confused.

It's also been pretty good for discussing architectural decisions and tradeoffs. It's given me great answers to how to create and manage environments, deal with config files and more.

Honestly, it's been amazing for learning DBT and has greatly helped me write code. Probably increased my productivity by 30-50%. But I think I'm in the sweet spot of GPT. I know enough SQL and data engineering to see through the bullshit, but getting the bit more than basics surfaced and getting some direction has been great.

Otherwise, it's pretty good for writing emails and such, but it isn't as good in norwegian as english and I'm pretty comfortable with my writing style so I use it less for that.

Sadly, it's not great at Dagster (my orchestrator) as there's not a lot of code out there and a lot has changed since the knowledge cutoff in 2021. I'm hoping some LLMs will be able to help with this in the future. Let me know if anyone has any tips here.

Trained a model on my voice and set up a cardboard cutout of myself next to my baby's crib so that I can provide validation and encouragement around the clock. Has freed up considerable time for my wife and I to enjoy each other's company. Currently hacking together a robotic bottle arm attachment for on-demand feeding.

We are two college students who wrote this Python program to translate language files (text or json) in apps to other languages using OpenAI's GPT-4. We handle caching outputs from the AI and let you manually change translations if you're not happy with them. Here's our repo: https://github.com/speedify/autotranslate-ai

GCP has graciously begun to roll out its internal AI suite: vertex model garden, lookerbi, simpleml for sheets & notebookml. Wizardcat (palm api codesynth) is nutty! You use your lifetime stash of GDrive txt files as "Personal AI" training set so it can basically finish your thoughts, refresh your memories, summarize your research, answer complex queries & brainstorm ;)

ChatGPT sub for coding and general search. Bard for things that require post 2021 info. Stable Diffusion (local) for illustrations. Whisper (local) for transcription. Many OSS LLM's for experimentation.

Chatgpt paid for coding outside my usual proficiency or for brainstorming. Ive been experimenting a bit with the API via tools like GPT-migrate. Cant go back now

I'm trying to build an AI powered vocabulary book. it can make unlimited number of sentences and generate quizzes.

For art, I created a website called https://artsy.sh which calls the APIs of Dalle, Craiyon, and several Stable Diffusion engines and provides results of prompts. I'm often using this for featured images to put on my website or just trying to come up with some creative beautiful artworks that I hang around several properties I rent out.

As for text and coding...

I use Google Bard and ChatGPT for very different reasons as both of them have their strengths and weakness and I've seen both of them provide wrong answers and pretty good answers.

Google Bard serves to answer my more basic random questions that pop into my head during the day that might be useless to others but I find fascinating:

"How far will bees follow their Queen?"

"What happened to all the poop on Titanic?"

"Can we send an object to land on Jupiter?"

"How far can a fart travel?"

"Is a dog or cat more likely to eat its deceased owner?"

I've asked Google Bard some questions that are similar, but different, and it looks like it is just providing a template after looking through its database and drawing a conclusion that it is a "similar question" to another. I've asked Google Bard some very difficult questions and it has had a hard time answering them.

I've also asked Google Bard to write me some things on important topics and it writes about the equivalency to an 8th or 9th grader. Google Bard tends to go into a topic, explain it all, and then feels the need to write up a conclusion saying the exact same thing, usually as a bulleted list, often revealing its apparent redundancy. Probably could've just provided me with the bulleted list and been done with it.

As for ChatGPT, it is reserved for more professional questions and tasks because ChatGPT is actually better at providing thought-provoking creative answers. Its equivalency feels as if it is writing similar to early college levels and you can even get it to write like an advanced university student. I am unsure if it would pass a thesis on law, but it is certainly far more creative than Google Bard.

Some things I might ask ChatGPT to do:

"Write an easy-to-understand six-month lease."

"Write a database class in PHP with CRUD functions."

"Write an advertisement that is appealing to children for a Bounce House Party"

"Code this more efficiently: [code]"

"Re-write this letter to make it sound more professional."

I've yet to dive into using any other AIs full-time such as Copilot, though I've play around with them a bit and was impressed, but just haven't made the big switch.