Once or twice a week
I forgot how the conversation went, but one day, a conversation I had with someone about comprehensibility (which was often an issue) compelled me to talk to an AI, a talk which I remember from the fact the AI did now have such issues as the complaining humans had.
Yeah I’ve run into this a bit. People say it “doesn’t understand” things, but when I ask for a definition of “understand” I usually just get downvotes.
Never
It’s simple: I don’t.
If by conversation you mean asking for a word by describing it conceptually because I can’t remember, every day. If you mean telling it about my day and hobbies, never.
That is basically the best use of LLMs.
A few of the most useful conversations I’ve had with ChatGPT:
I don’t unless a website requires that I talk to one as a poor excuse for customer service.
So, less than once a year.
I just type “Speak to a human” until it relents. Usually takes 3-4 times. Kind of the chatbot equivalent of mashing 0 on telephone IVRs. The only question of its that I answer, after it agrees to get a human, is when it asks what I need support with since that gets forwarded to the tech.
I’ve done it once or twice in the early days to see what was up, never since then.
Maybe 1-3 times a day. I find that the newest version of ChatGPT (4o) typically returns answers that are faster and better quality than a search engine inquiry, especially for inquiries that have a bit more conceptualization required or are more bespoke (i.e give me recipes to use up these 3 ingredients etc) so it has replaced search engines for me in those cases.
Daily.
💯
I’ve attempted to use it to program an android app.
2 weeks of effort… It’ll finally build without issue, but still won’t run.
Never
Not as much as I did at the beginning, but I mainly chalk that up to learning more about its limitations and getting better at detecting its bullshit. I no longer go to it for designing because it doesn’t do it well at the scale i need. Now it’s mainly used to refractor already working code, to remember what a kind of feature is called, and to catch random bugs that usually end up being typos that are hard to see visually. Past that, i only use it for code generation a line at a time with copilot, or sometimes a function at a time if the function is super simple but tedious to type, and even then i only accept the suggestion that i was already thinking of typing.
Basically it’s become fancy autocomplete, but that’s still saved me a tremendous amount of time.
The closest I come to chatting is asking github co-pilot to explain syntax when I’m learning a new language. I just needed to contribute a class library to an existing C# API, hadn’t done OOP in 15 years, and had never touched dotNet.
About as often as I have a conversation with my dishwasher: never.
Jeez…how do you think your dishwasher feels about that? Monster!!
It varies. Sometimes several times a day, sometimes none for a week or two. I’d say about half of those conversations are about software design.
