I promise this question is asked in good faith. I do not currently see the point of generative AI and I want to understand why there’s hype. There are ethical concerns but we’ll ignore ethics for the question.
In creative works like writing or art, it feels soulless and poor quality. In programming at best it’s a shortcut to avoid deeper learning, at worst it spits out garbage code that you spend more time debugging than if you had just written it by yourself.
When I see AI ads directed towards individuals the selling point is convenience. But I would feel robbed of the human experience using AI in place of human interaction.
So what’s the point of it all?
People keep meaning different things when they say “Generative AI”. Do you mean the tech in general, or the corporate AI that companies overhype and try to sell to everyone?
The tech itself is pretty cool. GenAI is already being used for quick subtitling and translating any form of media quickly. Image AI is really good at upscaling low-res images and making them clearer by filling in the gaps. Chatbots are fallible but they’re still really good for specific things like generating testing data or quickly helping you in basic tasks that might have you searching for 5 minutes. AI is huge in video games for upscaling tech like DLSS which can boost performance by running the game at a low resolution then upscaling it, the result is genuinely great. It’s also used to de-noise raytracing and show cleaner reflections.
Also people are missing the point on why AI is being invested in so much. No, I don’t think “AGI” is coming any time soon, but the reason they’re sucking in so much money is because of what it could be in 5 years. Saying AI is a waste of effort is like saying 3D video games are a waste of time because they looked bad in 1995. It will improve.
AI is huge in video games for upscaling tech like DLSS which can boost performance by running the game at a low resolution then upscaling it, the result is genuinely great
frame gen is blurry af and eats shit on any fast motion. rendering games at 640x480 and then scaling them to sensible resolutions is horrible artistic practice.
rendering games at 640x480 and then scaling them to sensible resolutions is horrible artistic practice.
Is that a reason a lot of pixel art games are looking like shit? I remember the era of 320x240 and 640x480 and the modern pixel art are looking noticeably worse.
that’s probably more to do with a lack of dithering and not using tubes anymore. lots of those older games looked better on crt than they do on digital
a good example is dracula’s eyes in symphony of the night, on crt the red bleeds over giving a really good red eyes effect
on lcd they are just single red pixels and look awful
Quite possibly, old games also look worse on emus (and don’t even let me start about those remasters, i got incredibly hyped for incoming Suikoden 1+2 on PC but my eyes fucking bleed).
In the context of programming:
- Good for boilerplate code and variables naming when what you want is for the model to regurgitate things it has seen before.
- Short pieces of code where it’s much faster to verify that the code is correct than to write the code yourself.
- Sometimes, I know how to do something but I’ll wait for Copilot to give me a suggestion, and if it looks like what I had in mind, it gives me extra confidence in the correctness of my solution. If it looks different, then it’s a sign that I might want to rethink it.
- It sometimes gives me suggestions for APIs that I’m not familiar with, prompting me to look them up and learn something new (assuming they exist).
There’s also some very cool applications to game AI that I’ve seen, but this is still in the research realm and much more niche.
I treat it as a newish employee. I don’t let it do important tasks without supervision, but it does help building something rough that I can work on.
shitposting.
Need some weidly specific imagery about whatever you’re going on about? It got you covered
Here’s some uses:
- skin cancer diagnoses with llms has a high success rate with a low cost. This is something that was starting to exist with older ai models, but llms do improve the success rate. source
- VLC recently unveiled a new feature of using ai to generate subtitles, i haven’t used it but if it delivers then it’s pretty nice
- for code generation, I agree it’s more harmful than useful for generating full programs or functions, but i find it quite useful as a predictive text generator, it saves a few keystrokes. Not a game changer but nice. It’s also pretty useful at generating test data so long as it’s hard to create but easy (for a human) to validate.
Money. It’s always about money. But more seriously, I also wonder what’s the point since all my interactions with GenAI have been disappointment after disappointment. But I read Dev saying that it’s great at creating drafts
I wrote guidelines for my small business. Then I uploaded the file to chatgpt and asked it to review it.
It made legitimately good suggestions and rewrote the documents using better sounding English.
Because of chatgpt I will be introducing more wellness and development programs.
Additionally, I need med images for my website. So instead of using stock photos, I was able to use midjourney to generate a whole bunch of images in the same style that fit the theme of my business. It looks much better.
deleted by creator
I use it to help with programming and writing. Not as a way to have something so it for me but as something that can show me how to do something I am stuck on or give me ideas when Im drawing a blank.
Kinda like an interactive rubber duck. Its solutions arent always right or accurate but it does help me get past things I struggle with.
deleted by creator
I use it to sort days and create tables which is really helpful. And the other thing that really helped me and I would have never tried to figure out on my own:
I work with the open source GIS software qgis. I’m not a cartographer or a programmer but a designer. I had a world map and wanted to create geojson files for each country. So I asked chatgpt if there was a way to automate this within qgis and sure thing it recommend to create a Python script that could run in the software, to do just that and after a few tweaks it did work. that saved me a lot of time and annoyances. Would it be good to know Python? Sure but I know my brain has a really hard time with code and script. It never clicked and likely never will. So I’m very happy with this use case. Creative work could be supported in a drafting phase but I’m not so sure about this.
I use it for coding, mostly as a time saver. Generally as I’m typing, it will give a suggestion that’s functionally the same as what I was going to type anyway so I hit tab and go to the next line. It’s able to do this accurately for around 80% of the total lines that I’m writing and going from writing full lines to writing 0-3 characters + tab on most of those lines makes a massive speed difference. It’s especially great for writing one off scripts when I’m doing something that’s not even a coding project, but there’s some tedious file juggling involved. Writing a script completely by hand for that often would take slightly longer than just doing the task manually, and as I said, it’s a one-off. But writing the script with copilot often takes as little as 10% of the time which is really nice.
Even in cases where I don’t already know how to solve a problem (particularly a problem involving specific integrations) it can often be faster to ask it how to solve the problem and then look up the specific functions, endpoints, etc it uses in the docs rather than trying to find those doc entries directly with a search. And if it hallucinates a function that doesn’t exist in the docs then I tell it that and it often successfully corrects itself. When it fails more than once I’ve generally found that there’s a high probability that the SDK/API/etc I’m looking at doesn’t have anything that does what I need so it’s time for me to start rethinking my approach
Outside of coding, I also use stable diffusion to generate images of D&D characters I’m creating instead of image searching and settling for something kind of close to what I was picturing.
I also regularly use SD when I stumble upon some art I’d like to use as a desktop wallpaper, but can’t find at high enough resolution. I just upscale it and proceed. Sometimes I’ll have something at the wrong aspect ratio and use generative fill to extend the edges of the image to the desired aspect ratio, those parts of the image are nothing special, but the important part is the original image and I just need some filler to prevent it from abruptly ending before the edges of the screen.
One last case is if I need to put together a tediously long document, I generally find that having it generate a first draft with the right structure and then iterating a bunch on that comes more easily than starting with an empty page.
As a dm I do this too for npcs and sometimes for pcs if they have a specific concept. Otherwise it takes hours of searching and you always have to settle, nothing is quite right. I don’t see an issue with it, no hobby dm can spend money on an artist creating every npc and it’s not feasible to expect it. This way I get high quality art if I’m willing to put the effort in to sorting and correcting images through repainting. Especially since I run a campaign set in feudal Japan. Orcs and kobolds in japanese traditional garb is not very common! Ai even let me run an encounter of a fashion show by providing the players a bank of images to curate a collection from and compete with other collections I made using ai.
I still spend money on patreon for monster tokens and maps and I doubt that would change even if ai could make good ones. I pay for their creativity and concepts not just the art.
For coding it works really well if you give it examples like “i have code that looked like this … And i made it to look like this … If i give you another piece of code that’s similar to the first can you convert it to the second for me”. Been great to reduce the amount of boring grunt work so I can focus on the more fun stuff
In C#, when programming save/load in video games, it can be super tedious. I am self taught and i didnt have the best resources, so the only way i could find to ensure its saving the correct variables was to manually input every single variable into a text file. I dont care if its plaintext, if people want to edit their save then more power to them. The issue is that there are potentially tens of hundreds of different variables that need to be saved for the gamestate to be accurately recreated.
So its really nice that i can just copy/paste my classes into gpt and give it the syntax for a single variable to be saved, then have it do the rest. I do have to browse through and ensure its actually getting all the variables, but it turns a potentially mindnumbing 4 hour long process into maybe a 20 minute one thats relatively engaging.
Also if you know a better way lmk. I read that you can simply hash the object into a text file and then unhash it, but afaik unhashing something is next to impossible and i could never figure it out anyways.
You could encrypt and decrypt it with keys.
Or you can do something simple like scramble the letters like a cypher, still able to edit manually but it wouldn’t be as readable and obvious what everything does.
Or you can can encode it, same issue as the last but they’ll have to know what it was encoded with to decode it before editing.
Or you can just turn it into bytes so the file is more awkward to work with.
You could probably mix a bunch of these together if you care enough. U don’t think any are THE standard and foolproof but they’re options
The goal isnt to encrypt the data, i dont care if its plaintext. The goal is to find a way to save an object in c# without having to save each individual variable.
Oh, in that case serialise it into json. Just use the json serialiser in system.text. it can turn any object in c# into a json object and you can deserialise them back into objects too.
Sorry i misinterpreted what you were asking for.
Yeah, that sounds a lot easier. Thanks
Best use is to ask it questions that you’re not sure how to ask. Sometimes you come across a problem that you’re not really even sure how to phrase, which makes Googling difficult. LLM’s at least would give you a better sense of what to Google
Idea generation.
E.g., I asked an LLM client for interactive lessons for teaching 4th graders about aerodynamics, esp related to how birds fly. It came back with 98% amazing suggestions that I had to modify only slightly.
A work colleague asked an LLM client for wedding vow ideas to break through writer’s block. The vows they ended up using were 100% theirs, but the AI spit out something on paper to get them started.
Those are just ideas that were previously “generated” by humans though, that the LLM learned
Those are just ideas that were previously “generated” by humans though, that the LLM learned
That’s not how modern generative AI works. It isn’t sifting through its training dataset to find something that matches your query like some kind of search engine. It’s taking your prompt and passing it through its massive statistical model to come to a result that meets your demand.
I feel like “passing it through a statistical model”, while absolutely true on a technical implementation level, doesn’t get to the heart of what it is doing so that people understand. It’s using the math terms, potentially deliberately to obfuscate and make it seem either simpler than it is. It’s like reducing it to “it just predicts the next word”. Technically true, but I could implement a black box next word predictor by sticking a real person in the black box and ask them to predict the next word, and it’d still meet that description.
The statistical model seems to be building some sort of conceptual grid of word relationships that approximates something very much like actually understanding what the words mean, and how the words are used semantically, with some random noise thrown into the mix at just the right amounts to generate some surprises that look very much like creativity.
Decades before LLMs were a thing, the Zompist wrote a nice essay on the Chinese room thought experiment that I think provides some useful conceptual models: http://zompist.com/searle.html
Searle’s own proposed rule (“Take a squiggle-squiggle sign from basket number one…”) depends for its effectiveness on xenophobia. Apparently computers are as baffled at Chinese characters as most Westerners are; the implication is that all they can do is shuffle them around as wholes, or put them in boxes, or replace one with another, or at best chop them up into smaller squiggles. But pointers change everything. Shouldn’t Searle’s confidence be shaken if he encountered this rule?
If you see 马, write down horse.
If the man in the CR encountered enough such rules, could it really be maintained that he didn’t understand any Chinese?
Now, this particular rule still is, in a sense, “symbol manipulation”; it’s exchanging a Chinese symbol for an English one. But it suggests the power of pointers, which allow the computer to switch levels. It can move from analyzing Chinese brushstrokes to analyzing English words… or to anything else the programmer specifies: a manual on horse training, perhaps.
Searle is arguing from a false picture of what computers do. Computers aren’t restricted to turning 马 into “horse”; they can also relate “horse” to pictures of horses, or a database of facts about horses, or code to allow a robot to ride a horse. We may or may not be willing to describe this as semantics, but it sure as hell isn’t “syntax”.
