Ask HN: How do you use AI to get things done faster?

I use it alot to help me write functions and small python scripts.

It helped me build a script that takes a wakeup word "GPT" or "Hey GPT" and then grabs the next words and sends it to the the GPT api and responds back with TTS.

Also, helped built some scripts to take notes from web pages, youtube videos, a mini chat window, FTP images from Dall-e and Pexels to my website.

had it write me script that can call GPT api to generate a list into a sqlite database then takes that list and calls GPT api with another prompt.

it helped me build a script that pulls my email from an imap account.

tons of bootstrap/css snippets for website widgets.

It helped me build a poormans vector database. I took a list of keywords got embeddings then saved it to a column in sqlite and ran it through a sentence similarity function to find the id of the closest match then had it mark the one as ignore.

.htaccess expressions for taking url slugs and redirecting to a posts.php page which is a real pain in the butt no way I could have done that without ChatGPT.

had it generate a regular expression cheatsheet.

Mostly to get around bad docs. I need to implement things using libraries, services, and APIs that aren't well documented, but the source code is public, and people have written code against them, and ChatGPT has read all of the things that I would read (and more) to get to the point where I can use it, so I just ask it.

The 2021 knowledge limit is pretty annoying, since libraries change so often, but it's still very useful.

It's ALSO very useful for asking stupid questions that I don't want to waste someone's time with. Like for instance, in bash, when you redirect stdout to a file with > filename it works, and when you redirect stderr to stdout with 2>&1 it works, but when you try to redirect stderr to stdout and stdout to a file, it only works when you do it in this order:

command > filename 2>&1

it doesn't work if you do

command 2>&1 > filename

which feels more natural to me, so I asked ChatGPT why that is, and it explained that you have to consider it from a filehandle perspective and that if you look in /proc/pid/fd you can see that 2>&1 is really redirecting to the terminal, and > is redirecting from the terminal to a file.

I would have had to find someone deeply steeped in unix/linux fundamentals to explain that to me, or I could just ask ChatGPT. I've done the same thing again and again - how are HSMs really different than TPMs? How are heat pumps different than ACs?

I'll read a reference to something, and immediately go to ChatGPT to learn more - "Can you give me a brief summary of the writings and opinions of Cicero?" and then I can spend 20-30 minutes learning more about stoicism, epicureanism, and whatever else I'm curious about. It's like being able to interview wikipedia.

Oh yeah, unless I know I need to interact with a website or view pictures, if I want to know the answer to a thing, I just go to ChatGPT now and talk to it. Way faster and the signal to noise ratio is incredible.

I use it (LLMs / GPT-4) as a sort of "realtime intern" to offload the more commoditised tasks that I don't really need to do - and that I don't contribute unique value to by doing myself. So some boilerplate CSS layout code, sketching out unit tests, initial stubs of API client code. Lots of "filling in the gaps" - which frees me up to spend more time thinking about higher level architecture, structure, business problems, abstraction/refactoring, planning, design etc.

Due to the way my brain seems to work, it also keeps me from getting stuck on small distracting problems which would previously create some resistance or procrastination - causing a break in my flow of work. It essentially keeps me in a productive flow state for longer periods than I could sustain without it. Any problem that is not in my "critical path" that I want to be focused on in the moment, and that could be easily solved by an LLM, gets "outsourced" as such.

This is in addition to it replacing about 80% of my software development / devops related Google searches. Because I work across quite a wide range of disciplines, I'm often looking for quick answers to questions about some technology stack that I'm not using daily. It's perfect for that. And I have enough familiarity with what I'm working with to sense-check/QA the responses.

I believe you do need some subject matter knowledge and experience to get the best out of LLMs though. I think many people are verbatim copy/pasting code out and complaining when it doesn't work. I very rarely find I waste any time debugging or correcting problems - because I either spot them and correct them in real time - still saving me a lot of time regardless - or I structure my prompts in a way that avoids these problems in the first place - by breaking the request down in to granular enough parts that I can pretty much predict how accurate the response will be (most of the time; very).

And in the scenarios where there is a bit of back and forth, trying different ideas and debugging in realtime - this is almost always a much faster (net) process than if I had done the same iteration myself.

As a point on usage and confidentiality, I don't use integrated coding assistants like Copilot - everything I do is sandboxed - so nothing confidential goes into the LLM. Specific details in my prompts are "anonymised" as I enter them (as in, I self-censor) - so I get the benefit of a lot of assistance from LLMs but with no sharing of any information that I would deem confidential. I plan to experiment with tighter integration into my workflow (eg. Copilot type assistance) with a private LLM instance at some point, but I'm comfortable with the balance of productivity and confidentiality at this point.

I do also have a Hammerspoon shortcut that will take the currently highlighted text in any app, and send it directly into OpenAI's endpoint. So I can highlight a mixture of comments and/or code in my IDE and immediately send them to GPT-4 and have that highlighted text replaced (or appended to) by the response. This gives me contextual assistance without having a constant live feed into a proprietary LLM ala Copilot.

Mainly cost. I also don't mind the insane performance of bare metal :)

https://world.hey.com/dhh/we-have-left-the-cloud-251760fb