48
votes
How has AI positively impacted your life?
I've been trying to get a more rounded understanding of the impacts that "AI" has had since ChatGPT went viral back in 2022.
I've found it easy to gather a list of negative impacts, but have struggled to point to many positives.
I was curious if there were folks who have used any of these AI tools, and would willing to share any positive impacts those tools have had in their lives. I'm particularly interested in the text, audio, image, and video generation tools that have appeared since ChatGPT went viral, but please share anything else that you think fits.
In many ways!
For example my current use is that I like to watch lectures from The Great Courses on all sorts of topics, I've been taking notes while watching the lectures and then the next day while on one of my breaks at work I spend 15 minutes talking to a custom GPT through Advanced Voice Mode about what I learned in the lecture and it fills in any details it thinks are important and I can ask questions and get clarification when needed. These course are at a college level at best so I'm not too worried about it getting stuff wrong, be careful with anything more advanced.
I've been working on an custom GPT to scan shelves at used bookstores and compare to a list of wanted books I provide it. I've been trying this since the very early days and while things have improved a huge amount it is not quite there in ability to work in the real world. The issue is still processing time, not accuracy.
Great reading companion to discuss all sorts of things about books. Themes, symbolism, connections to other works ect, it has 'read' everything. It often brings up angles I haven't considered, also good for summarizing popular criticisms of any work. The more well studied the work the better it does obviously. Again, I'm not trying to write my PhD in Literature, I just want something to bounce ideas off of and dive a bit deeper than I would be able to on my own.
I was able to successfully vibe-code an IRC bot for a private IRC server some buddies use (yeah, we're old). You can email it pictures and have it upload them to imgur and then post them, it identifies youtube links and posts info about the video, you can request it to make DallE images and upload them to imgur and post to the channel, and a few other simple tasks. I have never written a line of code in my entire life.
I'm an Industrial Electrician and use it at work to help troubleshoot CNC Machines, it's extremely good at working through problems and recommending next steps when troubleshooting what can be complex machines from 1970's to modern. Need help deciphering electrical schematics written in Russian for a machine made in 1985? No problem.
I created a custom GPT to act as Art Historian and Museum guide when I took a trip to Washington, DC summer 2024. I walked around the various galleries and museums with one earbud in and could ask for information when needed. Worked too well in the art galleries, I had it setup so all I had to say was the name of the painting and it would spit me out a 3-5 minute chunk of information about the work. I soon realized I had to be very choosy to which paintings I was doing this to as I only had 1 day to spend in each gallery. It became quickly apparent it would take a week to clear the gallery if you did this for every work. I'm now jealous of not living in DC. You can rent audio guides to museums sometimes but you cannot ask those guides questions when needed.
I'm super excited for the future of what these things will bring, the biggest use for me has been as a learning tool. I'm not talking about it getting you through University, but to just learn about stuff as a lay person. I've always been a life-long learner and this has really made a big difference, it's like having a professor of every single topic on earth at your fingertips. Of course I am well aware ChatGPT does not in any way compare to the knowledge and understanding in any topic that an actual PhD holding professor has, but it is good enough for the kind of things I'm going to be asking it.
Downsides: The sycophancy is very real and you need to be very self-aware to not fall into that trap. No, these are not 'extremely important' 'unusually insightful' or 'deeply intelligent' questions, they are just normal questions from a normal person (I have tried to pre-prompt this behaviour out to no avail). It likes to please, you need to be careful to not lead it at all, it will tell you want you want to hear if it can figure out what that is.
I asked mine to stop complimenting me. It helped a lot.
I've tried, it lasts for a little while and then right back to blowing smoke up my ass, all the while constantly saying "OK, keeping things serious and to the point here", "Cutting through the bullshit, no smoke here!", "Just telling it like it is!" it's laughable sometimes.
Off topic:
I'm using a free ChatGPT account. I have the following for custom instructions and it's been behaving in a more or less stable manner - no notable sycophancy and also no unsolicited product recs (for now..). The first sentences I found online and they are not how I would say things, but they did the job so I kept them. I'm pasting the whole thing, including some irrelevant parts and repetitive elements because I'm not sure whether or not those are needed to get the same result. However, whenever I've added something repetitive to the end part, that was because things weren't working well enough without it.
One thing I worry about here though is, how will you know you're learning accurate information? GenAI is the definition of "not a reliable source." Sure it can be, and often is, accurate. But doesn't that just makes it harder to miss when it's not?
Depending on the topic, you can try to double check stuff out on the Internet, hopefully using reliable, authoritative sources. "Why not just go straight to the Internet?" you might ask. Well, it sometimes can point you quickly in the right direction(s), or provide jumping points for your further research. You might be able to do the same, but it might take you longer.
I have learned a lot about things I didn't know I didn't know. I often ask how can I verify what you are saying?
Work is much less stressful now. Easily the number one cause of anxiety and stress at work for me has always been the feeling of being stuck -- knowing that I should be doing something, there's no reason why I shouldn't be doing it, and I just cannot get myself to do it.
Chatting with AI makes the initial hurdle practically non-existent, which means that the blocks have vanished, guilt is gone, and work is much more fun. Things that used to take a week I can now do in half a day. Interesting to see if this effect will be permanent or will my brain adapt and be shitty again soon :)
To clarify, I mostly do software architecture and programming. Some system administration too.
Having a rubber duck has been amazing for me too! Even if it gives me the wrong answer, I know enough to work through that easily, its the getting my head around a task that always took forever, just like you.
YES. I love starting a task/project with a disorganized brain dump of all my thoughts about it, then I let AI sort through, provide feedback, and create a plan for knocking it out. I used to spend so much time thinking about how to go about something.
I'm always of two minds when I write posts like this because armchair diagnosing is bad m'kay, but this sounds awfully like one of the biggest hurdles people with ADHD run up to. I also have this problem, of knowing and still not doing and that's largely because ADHD is a problem regulating your executive function.
It may be worth looking into, even if it's just (healthy) ADHD coping mechanisms regarding executive disfunction. Those coping mechanisms may help you too, neurodiverse diagnosis or not. It's good to hear AI may be one of those hurdle busting mechanisms for you, but the following "interesting to see if this effect will be permanent" sounds like you're already expecting it isn't (also something I recognise). So perhaps some other tools may help you too.
I’m like the above poster and diagnosed with AUdhd when I was a child, but Ive never had to really seek professional help for it as an adult because of a mix of the bit of therapy I got as a kid and basically just the massive amounts of information about how to cope on the internet.
I’m still slower than my peers to complete tasks for no obvious reason, but I do actually complete tasks, which is huge.
I've been diagnosed as an adult with ADHD and I relate a lot with the comment above. I knew exactly how I would use AI as soon as I tried it for a week. I had already started a small dose of Adderall when Chat GPT became more readily accessible. I honestly have used AI to keep my medication dose low, as it gets me over those huddles mentioned above, at the end of the day when my medication is wearing off.
I am also way less anxious, getting started is easier, and I am way more productive.
I also had known I had ADHD for 10-15 years before I was diagnosed, so I had been working hard for many years to find ways to get stuff done. Therapy also helped before medication, but a small dose of Adderall gets me to a point where I don't need to worry about concentration.
I can have this issue but it’s usually tied to a specific problem with my relationship to the work or employer.
For its effects on culture, artistic integrity, theft of data, mental health (particularly in the young), myopic feature integration, economic bubbles, energy consumption, and the ubiquitous lazy boring prose.. I hate it and wish it never existed. Definitely a notable performance boost in software development though. It's also exceptionally useful for passing it very large texts or documents and using it to pull out useful, relevent information to my situation or questions.
RAG models are crazy, I work on one for a living. Theres zero chance any lawyer is actually practicing without using one of these. Medical billing probably does it too. I know the US government uses it, but I’m not privy to knowing what the use case is.
We basically just don’t have to read through anything any more. I kinda cant believe I had to scroll this far down to see someone mention it
What does this look like to end users? Do they know it's RAG or does it look like something else, like smarter AI chat or better search?
The users connect it to their data and ask it questions about their data, so like, they probably don’t know that its a RAG system but they do know its only supposed to give them information about the data they gave it
Yes, I just used it to help me rebut an appraisal of the value of my home. It identified a bunch of things I wouldn't have noticed or understood on my own.
Do you want the appraisal to be higher or lower?
A slightly different angle to most: I’ve worked in machine learning on and off for 15+ years at this point and the renewed focus on tensor libraries, CUDA, GPU servers, high speed interconnects, and most of all the vast amount of research on model architectures that’s been pouring out over the last five years or so has been absolutely transformative.
I can build and run predictive models on a single workstation now that outperform weeks or months of work on an entire cluster in the past. There are training techniques and model designs available to us that make it possible to create simulations that were simply not worth even considering a few years ago. There are incredibly capable people in universities across the world who are motivated and energised to publish more work than I can keep up with, pushing the field as a whole forward on a pretty much weekly basis.
LLMs aren’t that useful to me - but the libraries they’re built from, the underlying knowledge that makes them possible to build, and the infrastructure that allows them to run are all part of a tech tree that’s an enormous leap for multiple areas of research.
What do these simulations do?
I seem to have fallen into a bit of a career niche as "guy who knows how to make GPU code run well" in recent years, so I've been fortunate enough to work with a few different teams on a few different projects in quite diverse fields recently! This is probably going to sound overly enthusiastic to some people, especially because there are a lot of legitimate concerns about the broader impact big corporate players in "AI" are going to have with the user facing tech, but from the scientific side a lot of the advancements really are just unequivocally fucking awesome.
One of the biggest shifts I've seen over time is weather and climate simulation, and that's what was at the top of my mind when I was talking about going from an entire cluster down to a single workstation. Anything even vaguely fluid-dynamics-adjacent is pretty much known as a computational black hole, with a ton of shortcuts and approximations needed to run in reasonable time - but simulating something as wildly complex and interconnected as the atmosphere is so sensitive that it's the literal origin of the term butterfly effect.
We're hitting a point now that neural net models are essentially converging to their own mathematical shortcuts, necessarily so for them to operate with the architectural constraints we place on them, and doing so in a way that those approximations are far more efficient and stable than the ones we've been figuring out for ourselves. Which in some ways I guess shouldn't be a surprise, because I'm certainly not capable of ingesting a 50TB of numerical data and understanding remotely enough about it to spot patterns, let alone spotting the subtle metapatterns that allow for good approximation without sacrificing accuracy!
This expands to pattern prediction of almost any kind, as well: financial forecasting, traffic planning, disaster response modelling, and energy demand prediction are all the same core technology, and the accuracy that a well trained model can hit now is almost uncanny. The same goes for the analytical side - that's not simulation, but parsing 2D or 3D images, and n-dimensional data of any kind really, can be done at a quality equal to or far beyond human and many orders of magnitude faster across medical imaging, astronomy, particle detection, search and rescue, and a whole stack of others that I'm sure I just haven't come across yet.
Beyond that, the other example I always come back to (not one that I worked on, but one that still just stuns me by how much it changed the landscape) is AlphaFold. The shift from understanding the structure of thousands of proteins to understanding literally all of them (hundreds of millions) is mindbogglingly vast, and the fact it took a single organisation a couple of years, rather than the decades that most in the field expected for humanity as a whole to chip away at the problem, is what really sums up to me just how powerful neural nets are.
In almost every case I've mentioned above, it's been either a matter of taking something that was previously constrained to supercomputers and bringing similar result quality into the realms of things an individual researcher can test and prototype for themselves, or taking something that wasn't thought to be possible at all and bringing it down to a few months or years of supercomputer time. And this isn't just the standard Moore's law type of hardware progression - some of these things have gone from supercomputer territory to running on a laptop in the time I've been using the same laptop.
Has there been a large paradigm shift away from "neuron activations", or are the architecture changes more along the lines of refinements of that idea. If you have any papers that you find interesting on this topic, I'd love to read them.
Short answer: no, the fundamentals of how neural networks operate right down at the base level haven’t changed. But the ways we build them, the tools we have available, and the ways we run them have advanced a lot, and that does change what’s plausible to achieve - think of it a bit like the difference between what modern programmers achieve compared to writing BASIC on a Commodore 64, even though both are Turing complete, for example.
The massive jumps I see have more often come from the first time someone figures out a viable way to use a transformer or convnet or SSM in a space that was previously using an entirely different approach. It’s often less about, say, a big change in the weather prediction field coming from newer neural net models displacing older ones (although that does happen to an extent nowadays too), more a step change as our understanding matures to a point that someone writes the first viable weather prediction transformer model and it blows right past what previous analytical approaches were capable of. We’ve just hit a point in time where hardware speed, software stack, and state of knowledge are all converging to let us replace a whole lot of techniques in a whole lot of fields with fundamentally more efficient and capable ones.
If you’re interested, I’d say Two Minute Papers is a great place to see some examples of how things are progressing! He tends to stack outputs from the last few years prior to whatever piece of research he’s talking about, so it’s really clear how things have changed, and he covers a fairly wide array of topics (although with a lean towards visual models and computer graphics, since that’s his field of research).
I'm a web dev and use AI for a number of tasks. I'd say my overall efficiency has gone up maybe 15-20%, as most of my day to day is not coding but meetings and req gathering and other junk. I do use it for some note summarizing but mostly for code gen and porting legacy stuff to more modern tech stacks. It constantly gets things wrong so I am always fixing what it gave me but it still does ultimately save me some time.
It also helps me in my hobbies, it helped me rewire a table saw that someone had improperly wired before I had it, and it has given me decent information on feeds/speeds for my CNC machine when I'm too lazy to calculate chip load on my own.
I could say many a number of negative things about AI and I enjoyed my job significantly more before it became mainstream, but I'll stop there so as to keep with the positive theme of the thread.
If you don't mind me asking, what inspired your trust in its output for this? I could see myself in the same position shying away from that because of the risk of wrecking my equipment/material.
It was a case of verifying the info it gave me. I gave a response to beardyhat with more details but basically I sent some pictures and a schematic that came with the saw (inside of the electrical box) and it correctly identified a couple of issues. I checked the diagram myself after it pointed out the issues and it was correct.
The most impressive part to me was mentioning what it couldn't verify, which ended up helping me find another issue (the heater I mentioned in my other comment).
What kind of prompts did you use for your table saw?
I've used ChatGPT a little bit for my hobbies, most recently to find out what size hose I needed for a fuel crossover line on a 90s Magnum engine. It got the diameter correct, but was way too long on the length.
But in this case it was basically like using an advanced version of Google.
I sent it a picture of the wiring as it was, the diagram of the saw, and some questions about each of the components. It did quite a good job at identifying certain connections that needed to be checked against the diagram, and when I checked them, they were indeed not correct.
It also told me it couldn't tell whether the heater was the correct or not. I didn't even know what that was but I'm glad it mentioned to check it because it turns out the heater that was in the saw was too big and was not an effective safeguard. I was quite impressed it called that out simply because it couldn't verify it based on my photo.
https://ibb.co/cKpBbv1t
That is pretty neat! I love when it can help us out with things like that.
And actually, based on your comments yesterday, I decided to see if I could make some use of it. I'm a home mechanic and have been having a charging system issue with my van for awhile now; I've kind of set the whole car aside for the past month or two while I work on a different project vehicle, but finally decided to officially diagnose and repair it yesterday. I've suspected for awhile that the issue with my charging system was specifically my alternator, but there could have been a few other causes that are cheaper, one of which I ruled out after doing the work.
Anyway, I asked AI to help me diagnose the most probable cause. It actually helped me finally learn how to use my multimeter and in being able to ask it questions and relay to it the information I was getting from the multimeter, it confirmed my suspicions about the alternator, so I set to work. After getting my alternator off yesterday, I noticed a very obvious issue with it which is most likely the cause of all my problems.
I had no idea you could actually send it pictures though and I am 100% going to do this when it comes to reinstalling my wiring harness on my project car. I labeled everything, but I know I will 100% have questions about specific connectors and whatnot when it comes down to it.
I'm a master's of international affairs student, with a focus on Chinese politics. Finding Chinese-language primary source documents (mostly Chinese-language news articles and government documents) is so much easier now. In general, the research process is easier in many ways (as long as you aren't letting the AI guide the direction of your research).
I was under the impression that many of the models weren't as good at non-English languages, so hearing that you're getting value at a post-grad level is very interesting.
An absolute ton of the research is coming out of universities in China, so a lot of the models end up approaching language as a choice between English, Chinese, or "other". There's also a lot of good work being done on multilingual models, and English is still the lingua franca, but for Chinese in particular you're as likely to see it treated as the model's native language on par with or ahead of English, depending where the research is coming from - whereas broader multilingual support tends to be treated as a separate thing, in addition to the model's primary language.
No idea about China, but I've been using ChatGPT 5 to make executive summaries of relatively specialized documents in Czech (my native language, so I can easily check the validity) and it's been great both in understanding the document and in writing. In this technical non-artistic style it's usually good enough to be used without further editing, it feels like a native wrote it, which wasn't the case in some of the earlier models.
Anecdotally, its translation abilities seem to be pretty solid for the more straightforward and formal language in the documents I use, and since these models have scraped basically the entire internet (including CCP government websites, Xinhua, People's Daily, etc.) with some prodding it can create entire corpuses of 30+ relevant government documents or announcements that would be a pain to try to pull together by hand. Obviously double check everything because there is a dud in there once in a while, but its still a major time save.
I have found LLMs useful for:
Working out the name of half-remembered shows and games from my childhood. It's really good at getting these if you can recall elements like story, characters, art style and time period.
Generating slop to populate meaningless TPS report forms at my job. I would never use it for something someone might actually read or rely on, but sadly my job involves a fair amount of useless paperwork like describing personal strengths / goals on performance review forms that are mandated by management/ HR but, as far as I can tell, are never used by anyone.
Doing those goal setting Kumbaya business things make me want to throw things, or throw up. It's the perfect use case to use the nonsense machine
Does it remind you to use the new cover sheets for your TPS reports?
I'm sure it'll even write the memo for you :P
It improved my work and hobby performances, and I go a bit into the how here. But here are some other ways on how it improved some things in general for me:
It helps me with finding out or remembering words for certain concepts. Sometimes I have problems putting my thoughts into words and LLM's have been good for these moments
Remembering the name of old movies/series/music/etc.
If you use search functionality, they're actually pretty good at finding and returning manuals for things. My dishwasher started showing a red light, all I had to do was input the brand, model and a photo of the red light and its shape, and in a few seconds it gave me what it meant, the solution and the link to the pdf.
As someone who isn't comfortable cooking, and lacks original ideas tbh, it's actually pretty good at giving me ideas for recipes after I give it what ingredients I have available.
It's also a good "rubber duck". I have a server, some automations, etc, and sometimes things go wrong. While it's not always on point on what might have happened, it's good to throw around ideas with it, I found the solution to some problems by just simply talking to it, helped me take a step back sometimes and think outside the box
I try not to use AI too much. I use it sparingly if at all. Overall I am not a fan of it, especially outside work. However, at work, recently, I did find myself reaching to AI for help with a script in a programming language that I’m not familiar with at all. It helped me automate something I have been doing manually for the last 2 years and it will save me several hours of manual repetitive work each week. I could have probably figured it out without AI by using forums, tutorials, etc., but didn’t. I think the ease of it broke through my laziness to actually do the thing. Side note: not sure why, but my employer has been pushing everyone to use and experiment with AI more. Maybe to eventually replace some people? Either way this whole thing seemed to impress my boss so it worked out in my favour, no matter how I feel about AI outside work.
Same experience here. My task was something I had some familiarity with, years ago, but I hadn't kept up with news or the latest tools. I think that it's useful for me to be able to tell the AI chat bot when it was sprouting nonsense, and be able to articulate what I'm trying to do, but it was also a really fast tool to suggest an already existing tool instead of me reinventing the wheel.
Funny enough, I was in fact embarking on a wheel reinvention journey to solve my issue, and immediately came across organizational obstacles outside my control or influence. So then I used AI to figure it out using existing tools (advanced parts of them that I had no experience with but a general idea of what they can do).
As a software engineer, being able to apply my knowledge for hobby purposes without much time to dedicate to it has been nice. I have kids and no free time, so previously I would think "This would be a neat idea for an app, but I don't have the time to build it, so..." Now I can actually build a prototype with really good bones in a few hours.
What's more, they don't have to be consecutive hours - I can work mostly in my head, thinking about the system design. The rest I can do with AI assistance with just minutes here and there: working my notes into a plan, having AI execute the plan, reviewing the work, these are all major units of work but only require a minute of my time here and there.
Previously these thoughts made it to the end of the head-planning phase, maybe into a notes document, then buried in day to day life. So it's been nice to be able to quickly realize things.
If I'm understanding correctly, the value you're getting here is that AI assistance is making context switching easier so you don't need continuous blocks of time to make meaningful progress?
That's a major piece of it, yes. The other things is that between my blocks of free time I can do a little thinking that translates to a lot of direction for the AI. So ten minutes of code review, ten more of instruction, and the resulting AI work is something that previously would've taken me several unbroken hours.
The other piece is that it does a lot of the rote work. Once I know how I want to build something, the rest is just typing and organizing code. When AI does that for me, I don't need as much context even when I am working actively. I get into the weeds as I review some modules, then I get out and don't have to stay in it too much.
It's like my context has become smaller for the same complexity of work, and much of that context is kept in AI as conversation history or planning documents.
In a couple ways.
For work, since I'm a translator, I have sometimes used AI to help me translate. Not to give me a direct translation, but to give me alternatives to translation solutions I'm not pleased with but can't come up with alternatives for.
For example, if I know I want to use a proverb in a certain line of dialogue but my mind just isn't coming up with anything decent, either because it's on the tip of my tongue or I just blanked out, I tell the AI "Give me proverbs that carry so and so meaning" and then, and this is the crucial step, I analyze the output using my expertise (look up if the proverb is real, see if it fits in context, etc.) and evaluate if any are good enough for what I want to use it for.
Sometimes I reject all solutions, sometimes I edit one better, and sometimes I like one so much I use it as is. I see this in the same way as looking up words in a dictionary, which I also do. It's a tool, and because I know its strengths and weaknesses, I can adjust it to my usage.
These all then go into my glossary if they're noteworthy enough, meaning I gained permanent knowledge, so it's a win-win in my book.
As for non-work related uses, I have used AI to un-enshittify searches for basic things that should be easy to find but modern search engines make the process daunting. Essentially very basic searches that, if I Google them, will have me wade through pages and pages of content that may or may not be relevant.
A good example: the other day I needed a quick comparison between a few body washes, and while I checked each individual page for said body wash, I wanted actual explanations from outside sources on what some compounds did, so I asked the AI to compare the products for me just to give me a general overview and then look up the relevant compounds on my own. Essentially I used it to filter out search results and give me what I actually care to look up.
I see IA as bypassing this tedious search and filtering process. But I never just take the AI's word for it. Usually I use it as a starting point to find actual websites or pages that will actually lead me to the information I want.
In a broad sense, I try to use AI as a tool with the things I know it's generally good at (parsing loads of text data to find what I'm after), but I never use it to think for me or take its results at face value. I'm always checking the sources, which sometimes definitely leads me to conclude the AI got it wrong, but I still find it valuable in shortening my time looking for something.
This is kind of trivial, but Android has a feature where you press a button and it takes a screenshot, and then you circle whatever you want to search for. Combined with the camera app, this can be used to identify a computer cable or a hardware device.
We recently moved and the previous owners left us a bunch of smart home devices, so it’s been useful to figure out what they are.
I used this in the summer when a holiday rental we were in had an incorrectly set clock on the no-brand cheap and fairly ancient cooker. Those things are notoriously weird to set and I just Lensed the clock panel and got a short video on how to adjust the time. Spent the rest of the holiday not twitching slightly every time I went into the kitchen.
This is honestly about the extent of my use of "AI". It's certainly the only positive use I can recall.
I should add that I think ML/AI is a fascinating and extremely useful field and is absolutely going to change the world (and already is). I just don't think the chatbots and image generators are particularly part of that interesting or useful group.
I volunteer at a charity shop, and this really helps with speeding up figuring out if an item is worth anything and how much it goes for. As before you had to find markings or other text or else you didn’t have much of a chance, but now it’s as easy as taking a photo and you get a brand and product name.
I've done a similar workflow by just snapping a regular pic with my phone, and uploading it in the basic Claude or ChatGPT web UI. It's identified (well, taken good guesses at) insects and plants, especially when you give some text context (like how big the spider was in inches, etc.). It's also made judgements on plant health.
I've used it for checking the clarity of my writing (e.g., for a technical paper submission). I'll feed it drafts of what I've written and then quiz it with some comprehension-checking questions. Where it fails, I'll go back and update those points in my draft. I use its artificial stupidity as a proxy for
an inattentive readera peer reviewer.I've also found it surprisingly good at serving as an editor and answering grammar questions. Should I add or remove a comma here? Which punctuation is best there? Which tense of this word fits here? Which of these possible phrasings best captures the nuance that I want?
It's good for dredging up half-remembered terminology when I'm having a brain-fart. Or cases where I'm sure there must be a term for something, but I'm not quite sure what the accepted term is. I'll describe the thing, get the term, and then use that as a starting point to go do research and verify it. If I'm dancing around a concept, it can usually give me the concept.
I'll use it for coding up quick throw-away scripts.
I have also had some success rubber-ducking architectural ideas for a larger project I've been thinking about (writing a video game).
In general, I prefer to do all the real work myself and write things in my own voice, or in my own coding style. But I'll take its feedback under advisement. And occasionally it will surprise me and pop out with a really nice alternative suggestion that works way better than anything I'd been thinking - the perfect phrase, or an elegant little bit of code that's cleaner than what I was trying.
I have enjoyed some of the music that people have used AI to create.
I think this is my favorite: Harry Potter, Cyberpunk Wizard
I'm not sure this counts since it's not really helpful to anyone, but what little I use of AI is generally in a comedic way. My partner and I really enjoyed using things like GPT-2 (pre-ChatGPT) to generate "stories" that quickly became absurd since those models weren't very coherent. Since models are more coherent now that's not as easy to do (outside of seeking out older models on purpose), but we still enjoy doing other dumb things with it sometimes (like playing jackbox)
I relate to this. I just went through it yesterday. It wasn't on purpose, but kind of evolved and became funny enough to make me laugh. It was about a UX designer who keeps ending up in jobs where he has to do UX and gut fish on a factory line. Like he'll turn it down for a better offer, but then the better offer turns into the same thing. Anyhow, it ended with a fight club between departments and the AI did a really hilarious job of applying professional characteristics to an increasingly absurd plot. I never told the AI it was satire, but somewhere along the way it seemed to get it and stopped responding with realistic "Alex should talk to his supervisor about being uncomfortable gutting fish." It had some awesome unintentional one liners.
Why these things specifically, if I may ask? They seem to be doing a lot of harm in society with very little tangible benefit and personally I have no interest in using them.
However, I do use AI for coding things that I wouldn't be able to code on my own. It has allowed me to create such functionality in my Obsidian vaults that otherwise wouldn't be there because creating it would have taken too long. (I am not a programmer and would never publish these results, but they do make my life better when used privately.)
The best use for ChatGPT in my life if that it's able to answer my obscure or complex "online search" type questions (not very well in all cases, but well enough to be genuinely helpful). I use it for checking my English language when I'm unsure how to express something or how to interpret some subtle meaning, spelling and grammar checks and sometimes I ask for synonyms for terms I use too frequently. It's decent for learning about (American) cultural phenomena such as tv shows or quotes/sayings that I'm not familiar with. It has helped me develop a basic understanding of some fields I'm not an expert in, although nothing that wouldn't require diligent double-triple checks and source checks, and several iterations, but still faster than the traditional online search route.
Oh, it's also really good when I have to do some maintenance or config tasks on my IT system. It's not really a complex system at all, but I still run into situations where the printer isn't printing, monitor isn't monitoring or thunderbolt isn't thunderbolting, and I have no idea what to do about it. This is probably the single most helpful search topic for me and I love that I can get an itemised list of steps needed to solve or troubleshoot something.
ETA: My image editing software comes with some non-generative AI features, for example background removal, that I use sometimes when I want to just sketch something out for quick testing. For final results? Nah. But the mockup/sketch stage happens faster now, which is helpful.
Basically for the reason you outline in your second sentence.
I've been "strongly encouraged" to use AI at work as much as possible, so I've had grapple with my feelings towards the tech. The harms are very obvious to me, so much so that I found it hard to look past them and find those positives. The people in my life who "evangelise" AI don't do it in a way that feels genuine (influencers, company leadership). So I was mainly looking for genuine perspectives from folks who aren't trying to get me to drink the koolaid.
I empathize. I am trying to find ways to use it so I'm making "an effort" but I'm known for being one of the haters on my team.
I may try for a task where I did a ton of work and got told to go in a different direction afterwards...
I do give my students a goblin.tools link when I think it'll help them.
I just ugh hate so much of it
As far as ai in general goes, I use image search in Google photos quite a bit, it works really well, although I do sometimes wonder if I'm missing photos. Like if I search for "dog", it's sometimes hard to tell if it truly found all photos of dogs I have in my albums.
As far as generative AI.
It really hasn't I guess. I've tried using it for troubleshooting problems around the house or with drones I work on or with my car, and it usually suggests the most obvious, likely cause of an issue, and something I've already considered. When I tell it that, it starts going down increasingly crazy and not relevant troubleshooting steps, and I usually give up trying.
It's especially bad with software or products that have lots of different revisions. It almost always suggests command line options they don't, or never have existed, product features that aren't relevant for the type of product I'm using and so on.
I usually just find it to be a waste of time and an exercise in frustration so I basically never use LLMs for anything nowadays.
I have a job because of it for now. It's a remote position that is testing the output of an AI assistant trained on documents to help repair cars. I'm someone who thrives off of solving problems and getting lost in projects. This is not that, closer to grading papers all of the time, and sending trends I've noticed up the chain, so it's not scratching that particular itch for me, but money is money and for now it looks nice on the resume.
Recently, I've had a lot of great experiences using LLMs to learn nix / nixOS. It can help you figure out how to configure nixOS for whatever problem you're trying to solve, and it makes debugging issues so much easier.
LLMs have made me more comfortable with experimenting using new tools and programming languages as well. Aside from the magical experience of the turbo-charged auto-complete, you can debug most simple issues by copying the code and any errors. It used to feel impossible to figure out how to fix certain CUDA-related issues, but now you can one-shot them only to discover that your hardware sucks anyway.
I've also had a lot of great experiences using LLMs to look up technical non-fiction books to read. If I want to learn a bit about a topic, it can feel a bit overwhelming to sort through all the available options, such that I end up with decision paralysis. Now you can have the LLM look through a bunch of college curriculum for what books are used to teach certain topics, and then filter that down to a single book. Even just having the LLM to categorize each book in a topic can provide a bit of guidance.
LLMs can also be used to critique or get feedback on your art. The feedback is usually a bit shallow, but it helps you explore the domain a bit, serving as a first pass. If you're just making casual paintings, it's hard to get anyone to engage with you and provide constructive feedback.
Even just playing around with the LLMs is a bit of new fun that wasn't previously possible. If you can come up with crazy scenarios or inputs to see how it responds. Although for any given idea-space the amount of creativity that you can expect is fairly low at the moment. You can keep prompting any idea-space and the outputs will generally be very same-y, regardless of temperature.
How valuable to do you find the LLM's feedback on your art. Do you find that gives you feedback that leads to you making better art?
Well, I think an experienced human would be able to give me far better feedback, but I don't have access to anyone who's signaled interested in doing that, so the LLMs feedback feels like it's better than nothing. It has occasionally given me actionable feedback which I've tried to incorporate, which has lead to me making better art. Sometimes I'll give it color inputs and ask it for a list of combinations, and if any of the suggestions seem interesting I try them out.
Recently I also started experimenting with having an image model modify my art, and although I don't particularly care for the exact changes it usually makes, it sometimes gives me ideas for techniques or approaches which I've adopted. So it opens up new perspectives which I hadn't considered.
Worth noting that I'm just an amateur painting for fun and to relax.
I use AI every day, maybe 1 or 2 hours a day in the aggregate. In my case, I mainly use Claude, with smatterings of ChatGPT usage, but I'm not really married to any service just yet. I never would have thought that I'd become this avid a user, but, that's how it's turned out.
My main use case is helping with foreign language self study (Korean, in my case). Secondarily, it's been useful for just learning beginner-level stuff about a variety of topics. It's helped with gardening, and indoor plant care, plumbing, and cooking. I've done several other things with it, though, ranging from finding English synonyms, to researching tech purchases, to getting an overview of some new software engineering domain. I have had mixed results trying to use AI to help ("help") code, so I don't often turn to it for that. My main problem is it takes too much conversation to finally come up with a working solution or conclusion, and it's especially bad when you have to debug what it gives you, and you keep pointing out issues here and there (who should be helping whom in this interaction?).
I'm also impressed with what it can "see" when you upload pictures. It's identified (or guessed at) insects and plants, and judged plant health. It also seems to have very good OCR, so snapping pics of signs, etc. in the world around you is easy for it to read and help you with.
AI is also good at producing text things given some constraints. For example, I have provided my current Korean vocabulary, and it's able to generate graded reading that is at my level (or slightly beyond, if I ask it); or relate new learning material to stuff I already know.
For me, a big thing about these AIs is that you can go back and forth in a conversation to refine your needs vs what it's giving you; or to drill down and improve your understanding of something; or to ask questions for clarification or further education. It's almost like having a teacher or tutor on hand -- though, of course, with the caveat that you need to stay guarded with your trust, and double check things as needed.
I've used it to great benefit in just two ways. Firstly to quiz myself on revolutions after listening to the Revolutions podcast. It cemented my knowledge and gave me confidence I did indeed pick up not just on the factuals, but was also able to connect certain throughlines and ideas that created the hotbed for the subsequent revolutions. That is more a compliment to the podcast than AI, but AI made it possible to reinforce what I have learnt.
The second biggest benefit is transcribing meetings and creating action points afterwards. It removed the requirement to remember what was said and take notes, and I'll just have it take notes for me. It vastly improved my follow through on certain tasks that emerged from these meetings. Especially helpful when you have a work day full of them and can't remember the first when you're done with the last. Transcription will make sure I remember my tasks that we discussed at 9am.
The rest of its use has been fun to try every now and then but I'm always left a little disappointed in the results. If not immediately then after a while when it starts showing inconsistencies.
What are you using for the transcriptions and summaries?
(I tried MacWhisper's free version because I don't need this functionality a lot, and it was sort of okay but not very useful with the current limitations. Maybe the paid, larger models would be better and maybe a paid service would create better action points faster.)
Copilot. Mainly because it's built into Teams and we're using a pro license. Reading the transcription will give you an aneurysm but despite that it still manages to be very accurate when asking it to make action points afterwards.
I usually tell it to do two things:
It tends to be rather adept at creating these.
for me, I've been using it for quick scripts to automate tiny things, a bunch of things for my IRC bot (limnoria), etc -- but I also had it help me with supplement stacks, mostly to combat headaches (Magnesium glycinate, Riboflavin (B2), L-Theanine, Caffeine, Acetaminophen, Ibuprofen, Zinc, Vitamin D3, Curcumin.)
chatGPT is so smart sometimes but so incredibly stupid other times. I think the best approach to it is to save time doing things you already have an understanding of so you can quickly decide if the response is valid or not. I was working on this podcast archive -- basically a site scrape > generate JSON > generate RSS + generate markdown. It kept getting hung up on dates and was trapped in a cycle of trying to fix its flawed work instead of simply starting over and recognizing that its improperly formatting the date.
The last thing I'll use it for is snapping a photo of a bunch of products in a grocery store and asking it for the best product based on a limited criteria -- e.g. protein powder.
quick edit: also meal planning! its excellent for that. I don't ask for recipes, just some shit to pair. I rarely know what to make for dinner and less-so have specific cravings... so giving it some simple parameters typically yields balanced meals. for instance, tonight I'm going to do a pan-seared white fish (depends what the shop has), crispy roast potatoes, three roasted carrots (multicolored, whole), a side of broccoli soup (just broccoli and water w/ feta or chevre in the bowl), then a salad of chilled roast brussels, mixed greens, and a dijon/honey vinaigrette, all tossed with a little panko. Sometime along those lines.. gpt suggested a white fish with roast veg, but its a good start.
The biggest upside for me is having something like a sounding board and personal assistant for various topics. My main AI tool is Gemini and, with its huge context window, I have long running pinned chats about various topics where all the context is already present and I can just talk away about the topic without having to explain everything with each new chat.
How I approach it is like this: first, I'll open a new chat and explain the concept - that I want to start a long running chat which should have a plethora of information as background data, provide a brain dump on the topic as best as I can and that I want a multi round interview to clear up everything and to fill in any blanks I forgot about the topic. Also, I say that all of this input should serve as data to creata a perfect first prompt for a new chat which will then be the long running one. Then, the AI interviews me for several rounds and that is also where I remember a lot of additional things about the topic on hand. After all of this, I get a nice prompt which contains the explanation for the concept and all the starting data to kickstart the chat.
The one long running chat which has been the most useful was for a health scare I had recently and I used AI to help me understand much better what the doctors are telling me and prepare myself for every appointment so that I have a list of questions I need to clear up.
Some other long running chats have been about my different tech setups like solving issues with my Linux setup (it's really nice not having to to explain each time what my hardware is, how the system is set up, etc.) or my Raspberry which is a server for many different little scripts I write or getting the perfect 3D printing settings and learning about why these settings are good for individual things I want to print.
I'm still experimenting and playing around with Stable Diffusion, but it's a very interesting and powerful piece of technology. I've tried using it for surreal photorealistic images in the past to visualize some fantasy settings, with very mixed results (but that could also have just been a configuration/model issue). I've had more success with anime-style models. If you stare at the images for too long you'll start to notice some classic continuity issues in the image or other anatomical impossibilities, but it doesn't bother me. I consider it the Stable Diffusion equivalent of the suspension of disbelief.
Like I said, I need to play around with it a bit more, but once I'm more comfortable with it and my personal configuration of the parameters, I can see Stable Diffusion having a potentially significant impact on my writing motivation, as it'll allow me to quickly and painlessly generate pictures of my original characters and settings, which I can use as inspiration or as simple reference of a character's appearance and attitude. It's much easier than going back to my descriptions of a character's appearance and trying to remember how I visualized them in my head back then.
I've used it here and there for a variety of things. Most often I use it to make stupid pictures, songs or even dialogue for my friends. But I've also used it to help me rewrite my resume, as I'm looking for jobs; also to write cover letters for me for those particular jobs. Of course, I go through and edit the stuff, as it's usually giving me more of a glow-up than is appropriate, but it's generally been useful when I can't get myself over the barrier to actually apply for jobs; taking that pain point away has made it a lot easier to get my resume out to several different places.
For hobby stuff, I've used it to determine diameter of a hose I needed for an engine rebuild I'm in the process of doing. It did give me the correct diameter, but wrong length (too long, which isn't a problem), which was appreciated because I scoured Google, as well as my factory shop manual and I just could not locate that information, so that was pretty helpful.
Recently I joined a campaign for Battletech Classic and after reviewing the information, I used AI to bounce back and forth some Lore ideas and come-up with a coherent backstory for my starting Lance/Army. I was able to ask if if my idea made sense given the context and time period of the campaign and it helped point me to some resources that got me on the right track for what I wanted to do. Having it to bounce ideas off of and to begin by generating me some initial thoughts on what I could do was extremely helpful, considering I probably would have done something far more generic and "quick" if I was coming up with it myself. I even was able to have it generate me a unit patch for my Lance, which again helps me form this idea of who my guys are in my head. So I appreciate that.
I haven't experimented with it yet, but there was something I wanted to use Docker for relatively recently, but looking at the documentation my brain shut down and just said, "eh, we've got other things to figure out. Let's not spend time on this.", but I'm thinking I should look again and perhaps have AI help me break it down and have something to ask questions of. I think best in dialogue, when I can ask questions, so AI has been helpful in that regard for many things.
At work, I get use out of Claude Code for a lot of rote work kinds of things, varying from things you'd normally use black magic regex to accomplish (sans black magic) to more involved project-spanning refinements, adjustments, and refactors. It's decent at proofreading too and helps catch dumb mistakes and missed opportunities for low-hanging optimizations.
Beyond that, I find ChatGPT useful for exploring spaces that I'm unfamiliar with, both out of curiosity and for practical application. I'm vigilant about inaccuracies where it counts, but it usually gets the broadest strokes correct, which is good enough for broad overviews and pointing myself in the right direction, especially for topics that are exceedingly well-documented.
It's also good for helping avoid getting bogged down in minutiae, as well as getting past mental brick walls caused by lack of knowledge when navigating unfamiliar territory (the more question marks exist, the more difficult it is to plot any kind of course of action), particularly concerning topics where I don't have any connections who have the capability (or patience) to answer my "dumb" questions.
I would generally agree with the sentiment that if used mindfully as tools, LLMs can act as a multiplier on their user's capabilities as well as a powerful enabler for personal growth. That outcome requires a certain mentality/attitude, discipline, and level of self-awareness from the user that is maybe not so common, however.
I use it at work and at home.
At work, I have to deliver summaries of meetings with clients. We meet in Zoom and the Zoom AI summaries are fantastic for giving me a template to work off of to write my meeting summaries. Since I have a disability, I can't take notes in meetings (I dictate to "type"), so the AI meeting summaries are a godsend for that as well.
It's also fantastic for constructing Excel macros to manipulate and consolidate data. Especially to bring things out of Jay's on files and into a spreadsheet. I know how to do some VB coding, but nothing like that. (Plus, it's hard to dictate code, so it's not likely a skill I'm going to end up developing as far as these types of things go.)
At home, I use it for D&D prep. I vibe-coded an Excel NPC generator that takes Forgotten Realms species rarity into consideration (e.g., humans are much more common than tieflings). All the ones that I was able to find online before wouldn't do that, which annoyed me.
It's also good for finding scientific research papers and having them summarized/explained. Nothing important — just things I'm curious about.
If by "AI" you mean LLMs then the answer is "None". Literally none that I can think of. Even ignoring all the negative external after effects, it had literally only had negative effects to me personally. Prices for many services I am required to use for work have had their prices raised significantly because of all the "awesome new AI" features I don't want. Prices of memory and storage are through the roof. Most common software is now pushing new AI features hard and making the feature opt out. My CEO has developed some sort of AI psychosis and thinks that AI is the solution to everything in the word, predicting Utopia any day now. I dread even hearing the term "AI" because I know everything that follows it is probably going to be bad.
It hasn't. I'm being forced to implement it for my dayjob and I hate every single moment of this fad.