When I search this topic online, I always find either wrong information or advertising lies. So what is actually something that LLMs can do very well, as in being actually useful and not just outputing a nonsensical word salad that sounds coherent.
Results
So basically from what I’ve read, most people use it for natural language processing problems.
Example: turn this infodump into a bullet point list, or turn this bullet point list into a coherent text, help me with rephrasing this text, word association, etc.
Other people use it for simple questions that it can answer with a database of verified sources.
Also, a few people use it as struggle duck, basically helping alleviate writers block.
Thanks guys.
Not exactly sure this is the “right way” to use them, but I use one as an autocomplete helper in my IDE. I don’t ask it to code anything, just use it as autocomplete.
Majority of the time, it works well, especially in common languages like Python.
Brainstorming. ChatGPT and co. are slightly better rubber ducks. Which helps to sort my thoughts and evaluate ideas.
Also when researching a new topic I barely know anything about, it helps to get useful pointers and keywords for further research and reading. It’s like an interactive Wikipedia in that regard.
Very basic and non-creative source code operations. Eg. “convert this representation of data to that representation of data based on the template”
I find it’s really good for asking extremely specific code questions
kill time
Overcoming writers block or whatever you want to call it
Like writing an obit or thank you message that doesn’t sound stupid. I just need a sentence down to work from, even though it doesn’t make it to final draft.
Or I needed to come up with activities to teach 4th graders about aerodynamics for a STEM outreach thing. None of the output from LLM was usable as it was spit out but was enough for me to kickstart real ideas
This is a great use i use it for similar purpose it’s great brainstorming ideas. Even if it’s ideas are bullshit cause it made it up it can spark an idea in me that’s not.
This sort of applies to dev work too, especially if you have ADHD. I overcome blockage by rubber ducking, but sometimes my ADHD gets strong enough that I can’t, for the life of me, sit down to write some trivial code that might as well be a typing exercise. I simply get Cursor to generate the stuff, proofread it, and now that it’s suddenly a bug smashing session instead of typing out some class or component or whatever, I overcome my blockage and can even flow. Speaking as someone that often gets blocked for weeks to months at a time, LLMs have saved me from crashing into deadlines more than a few times.
That’s about where I land. I’ve used it the other way, too, to help tighten up a good short story I’d written where my tone and tense was all over the place.
I’ve used LLMs to write automated tests for my code, too. They’re not hard to write, just super tedious.
Same. It’s gets me started on things, even if I use very little or even non of its actual output.
Yes, it’s like the rubberducking technique, with a rubber duck that actually responds.
Sometimes even just trying to articulate a question is a good first step for finding the solution. A LLM can help with this process.
Just rewrote my corporate IT policies. I feed it all the old policies and a huge essay of criteria, styles, business goals etc. then created a bunch of new policies. I have chatgpt interview me about the new policies, I don’t trust what it outputs until I review it in detail and I ask it things like
What do other similar themed policies have that I don’t? How is the policy going to be hard to enforce? What are my obligations annually, quarterly and so on?
What forms should I have in place to capture information ( i.e. consultant onboarding).
I can do it all myself but it would be slower and more likely to have consistency and grammatical errors.
I use it to review my meeting notes.
- “Based on the following daily notes, what should I follow-up on in my next meeting with #SomeTeamTag?”
- “Based on the following daily notes, what has the #SomeTeamTag accomplished the past month?”
- etc.
I’m not counting on it to not miss anything, but it jogs my memory, it does often pull out things I completely forgot about, and it lets me get away with being super lazy. Whoops, 5 minutes before a meeting I forgot about? Suddenly I can follow up on things that were talked about last meeting. Or, for sprint retrospectives, give feedback that is accurate.
To add: I’ve also started using AI to “talk to podcast guests.” You can use Whisper to transcribe a podcast, then give the transcript to AI to ask questions. I find the Modern Wisdom Podcast is great for this.
I record meetings of my building’s board of management, nothing secret there, very mundane. I run it through Whisper and give the transcript to ChatGPT. It condenses everything into accurate minutes, resolutions and action items. Saves me a shit ton of work, finished in seconds. I’m never going back!
I guess my notes are unstructured, as in they’re what I type as I’m in the meeting. I’m a “more is better” sort of note taker, so it’s definitely faster to let AI pull things out.
Infosec … I guess people will have to evaluate that for themselves. Certainly, for my use case there’s no concern.
Philosophy.
Ask it to act as Socrates, pick a topic and it will help you with introspection.
This is good for examining your biases.
e.g. I want to examine the role of government employees.
e.g. when is it ok to give up on an idea?A fringe case I’ve found ChatGPT very useful is to learn more about information that is plentiful but buried in dead threads in various old school web forums and thus very hard to Google. Like other people’s experiences from homebrewing. Then I ask it for sources and most often it is accurate to the claims of other homebrewers that also can be correct or less correct.
As a developer, I use LLMs as sort of a search engine, I ask things like how to use a certain function, or how to fix a build error. I try to avoid asking for code because often the generated code doesn’t work or uses made up or deprecated functions.
As a teacher, I use it to generate data for exercises, they’re especially useful for populating databases and generating text files in a certain format that need to be parsed. I tried asking for ideas for new exercises but they always suck.
I use it to help me come up with better wording for things. A few examples:
-
Writing annual goals for my team. I had an outline of what I wanted my goals to be, but wanted to get well written detail about what it looks like to meet or exceed expectations on each goal and to create some variations based on a couple of different job types.
-
Brainstorming interview questions. I can use the job description and other information to come up with a starting list of questions and then challenge the LLM to describe how the question is useful. I rarely use the results as-is, but it helps me to think through my interview plan better than just using a list of generic questions.
-
Converting a stream of thought bullet list into a well written communication.
-
I ask it increasingly absurd riddles and laugh when it hallucinates and tells me something even more absurd.
JIRA queries, rules, automations, etc. Suggestions for how to make my rage-fueled communications sound more reasonable and professional Meeting Summaries. Not having to take notes is HUGE.
Meeting notes are the ideal use case for AI, in the sense that everyone thinks someone needs to write them but almost nobody ever goes back and actually reads them.
But when I got curious and read the AI generated ones (the ones from Zoom at least)… According to the AI I had agreed on an action that hadn’t been even discussed in the meeting and we apparently spent half of the meeting discussing weather conditions in the various locations (AI seems to have a hard time telling the difference between initial greetings or jokes and the actual discussion, but in this one it became weirdly fixated with those initial 5 minutes)
This is one area where, at least for me, CoPilot is very good. In most other areas, CoPilot is not very good.
