With copilot included in Professional-grade Office 365 and some politician claiming that their government should use AI to be more efficient. I am curious on whether some of you did use “AI” to get some productive things done. Or if it’s still mostly a toy for you.
I find that the very best use case of LLMs are in the name it self–language. I can check my text is grammatically correct or not for example
Its great for reading docs, I don’t have to search through to find out how to get a user’s playlists using the Spotify api
I’ve used it productively this week by…
- Summarising and finding relevant parts of Microsoft Teams meetings.
- Finding relevant parts of the labyrinthine policies I have to comply with.
- Quickly finding out what’s going on with corporate events in the market.
- Generating SQL code instead of starting with a blank document (I can never remember the exact way to declare various structures)
- At home, feeding private documents into Ollama for insight and producing compliance reports.
- Instructions for stepping through flashing some temperature sensors.
I use it as a glorified Google search since Google search is absolute dogshit these days. But that’s about it. ChatGPT is one of the most over hyped bullshit I’ve ever seen in my life.
Absolutely agree!! LLMs are good for quick “shallow” search for me (stuff I would find on google in a few minutes). Bad for “deeper” learning (because it’s not capable of doing it). It’s overhyped.
You shouldn’t use it for search like that. They (Gemini and ChatGPT) love to be confidently incorrect. Their perfect grammar trick you into believing their answers, even when they are wildly inaccurate.
I use FastGPT on Kagi and it lists the sources for its conclusions, so it’s like a better aimed search
I use GPT in the sense of “I need to solve X problem, are there established algorithms for this?” which usually gives me a good starting point for actual searching.
Most recent use-case was judging the similarity of two strings: I had never heard of “Levenschtein distance” before, but once I had that keyword it was easy to work from there.
Also: cmake and bash boilerplate
Describing a concept and getting the term is awesome with an LLM.
I’ve found documentation and discussions of various strategies I’m considering in tech work.
I describe my idea, the LLM gives me the existing term for that strategy, and then I can find discussion, guides, and theory about that. Keeps me from reinventing the wheel.
As if Google is any better
I think I’m going to disagree with the accuracy statement.
Yes - AIs can be famously inaccurate. But so can web pages - even reputable ones. In fact, any single source of information is insufficient to be relied upon, if accuracy is important. And today, deliberate disinformation on the internet is massive - it’s something we don’t even know the scale of because the tools to check it may be compromised. </tinfoilhat>
It takes a lot of cross-referencing to be certain of something, and most of us don’t bother if the first answer from either method ‘feels right’.
AI does get shown off when it’s stupidly wrong, which is to be expected, but the world doesn’t care when it’s correct time and again. And each iteration gets better at self-checking facts.
I have it provide me with its sources
Copilot is actually linked directly into their search engine and it provides the links it pulls its data from. But you’re correct, ChatGPT is not hooked into the live internet and should not be used for such things. I’m not sure if Gemini is or not since I haven’t used it or looked into it much, so I can’t comment on it.
That’s not correct. ChatGPT is hooked into the live web.
It seems like exactly the moment google’s successor showed up, google has a stroke. it’s awful these days
Used it as a toy for the longest time but by now I had to do a lot of coding and I was actually able to make good use of code completion AI.
Saved me about a quarter of my time. Definitely worth something. (FYI I use supermaven free tier).
Also I’m using ChatGPT to ask dumb questions because that way I don’t have to constantly interrupt other people. And also as a starting point to research something. I usually start with ChatGPT, then Google specific jargon and depending on the depth of the topic I will read either studies, articles or forum threads afterwards.
It did take me a long time to figure out which AI and when to use it, so mandating this onto the entire government is a gong show more than anything.
No AI is not useless, but it’s always a very specific use case.
If you’re interested, I suggest using the free ChatGPT version to ask dumb questions together with Google to get a feel for what you get. Then you can better decide if it’s worth it for you.
The amount of shit we have to clean up from devs using AI generated code nowadays is insane. Some people are using it to generate the bulk of their code and the output can be trash tier.
I was supposed to have a nice long weekend to rest and I spent most of it cleaning up after clients who pushed AI generated code into production that froze all data processing. Even after we found the problem and fixed it for them, the data didn’t catch up until yesterday afternoon. The entire holiday I had to spend with a laptop a few feet away on a Teams call because a dev used AI-gen code.
I am not saying that it isn’t helpful to your situation. What I am saying is that a growing number of outfits are starting to depend on “devs” who’s code is mostly LLM generated, and they push it without understanding what it does or how it interacts with the environment they are deploying it in.
Yeah. I think AI literacy is a real thing and should be taken seriously. Before generating everyone should internalize the boundaries and limitations of any model used.
If you have a hammer, everything’s a nail. And that reflex exists with AI as well, so everyone who uses is has to be careful in regards to that.
Same.
Gen AI is very helpful when you’re working on modifying or filling holes in code.
Search (with some double checking), rough drafting and frameworks, templates, and scripting.
I have been this week, for the first time.
I’m using Hugo to design a new website and Gemini has been useful in find the actual useful documentation that I need. Much faster and more accurate than trawling the official pages, and does a better job of providing relevent examples. It’s also really good at sensing what I’m actually asking, even if I’m clumsy at the phrasing.
And for those who continue to say AI isn’t really useful for learning - another thing I’ve been using it for. “write perl to convert a string to only container lowercase, converting any non-alpha chars to dashes” - I’ve learned how to do stuff like that over and over again, but the exact syntax falls out of my head after a few months of not doing it. AI is good at providing a quick recollect. I’ve already learned perl properly (including from paper books - yes, I first wrote perl a quarter of a century ago) - and forgotten it so many times. AI doesn’t prevent me learning, just makes it faster.
I don’t use AI for productive work, for the same reasons I don’t stir my soup with a dishrag.
Pretty good for recipes, tho’.
When I had a mold problem it was affecting my mind. I couldn’t think straight or focus, so I had ChatGPT make me a step by step plan for dealing with it, and it had it break each step down into nested sub-steps until no step was more than five minutes of effort, then I had it format the plan to copy-paste into workflowy.
It was really helpful. I could have made that plan myself, except that I was fucked up.
I use it all the time, and not just for myself or for work. Yesterday I fed my son’s study guide into ChatGPT and had it create a CSV file with flash cards for Anki. It’s great at any kind of transformation / summarizing or picking out specific information.
When school sends me overly verbose messages about everything that’s going on I can feed the message into ChatGPT and have it create an ical file that has events for the important stuff that happens in school in in the coming week.
I used it to write a greeting card for my dad on his birthday (“I’m giving him X, these are his interests, give me ten suggestions for greeting cards”).
I have it explain the reasons behind news stories (by searching for previous information and relating it to the news story). I ask tons of questions about anything I wonder about in the world such as chemical processes, the differences between oil frying and air frying, finding scientific papers about specific things, how to factory reset my Bose headphones… the list goes on.
how to factory reset my Bose headphones
I don’t think it can get the information for this with 100% accuracy unless the process is same for all Bose headphones. How did it go?
and have it create an ical file that has events for the important stuff that happens in school in in the coming week.
How did this go? It can hallucinate stuff even when you post static data to it, last time I tried.
I don’t think it can get the information for this with 100% accuracy unless the process is same for all Bose headphones. How did it go?
Why not? I told it the model (Bose 700). It searched the web for information for that model, found an article that described how to do it, and provided me with the key points without having to scroll past tons of ads and noisy language. Of course it sometimes gives me the wrong info (usually because the sources are incorrect), but I’ll notice soon enough.
How did this go? It can hallucinate stuff even when you post static data to it, last time I tried.
It went perfectly. Again, there are certainly times when it makes errors / hallucinates, but I can fix those manually. In my example of producing flash cards for my son, we obviously had to proofread the cards but that’s much faster than writing all the cards by hand. One out of the 20 flash cards had a nonsensical question/answer so we just removed it.
I have used it as a nicer version of web search, mostly for “How do I write code using this library I’m not yet familiar with?” It provides passable tutorials when the library’s documentation is sparse (I get it) or poorly written (they tried 🤷♂️).
I use it in two ways.
ChatGPT as an interactive search. Last one was about EU GDPR compliance checklist to give a quick answer on what areas need to be looked at. I use it like once a week for work.
Productive in othen ways I use it once a month for recipes. Recipes are probably my favourite since I can say “Write it using grams and ml” and "give me some options to replace eggs and it writes out a legit recipe based on these millions of annoying blogs recipes.
Jetbrains AI auto complete for programming which is getting better slowly and I’m getting the hang of using it. It’s really good for cases where I have a common thing that I don’t remember the syntax of and I just type a name of a variable like “cspHeaderValue” and it will format thing that’s very annoying to look up based on what I some values I wrote above.
I’m not a 10x engineer now for it, it’s more like +10% overall and really depends on the task. I can see it go up to around +50% but an AI plateau might come before then.
I would say that I have used an LLM for productive tasks unrelated to work. I run a superhero RPG weekly, and have been using Egyptian & north African myths as the origin for my monsters of the week. The LLM has taken my research and the monster-creating phase of my prep from being multiple hours to sometimes under one hour - I do confirm everything the LLM tells me with either Wikipedia or a museum. I can also use an LLM to generate exemplary images of the various monsters for my players, as a visual aid.
That means I have more time to focus on the “game” elements - like specific combats, available maps, and the like. I appreciate the acceleration it provides by being a combined natural-language search engine and summary tool. Frankly, if Ask Jeeves (aka ask(dot)com) was still good at parsing questions and providing clear results, I would be just as happy using it.
Absolutely. I’ve used it to write basic scripts that I didn’t feel like spending time on. I’ve also used it to write cover letters. I always make sure to peruse through it to see what it did and make sure it works or sounds right.
My physics professor has us compare our answers to physics problems with a LLM’s output. Somehow, the AI is even worse at physics then I am, it once simplified (4pi2) to 4.
Are you in my class?
