ฝัง
- เผยแพร่เมื่อ 23 ก.พ. 2023
- I put artificial intelligence to the test by making a voxel engine using the rust programming language and the bevy game engine. How useful is ChatGPT for programming really?
SOURCE CODE (good luck lol):
github.com/TanTanDev/chat_gpt...
My discord group:
/ discord
Want to support me?
⁍ Patreon: / tantandev
⁍ Monero: 43Ktj1Bd4Nkaj4fdx6nPvBZkJewcPjxPB9nafnepM7SdGtcU6rhpxyLiV9w3k92rE1UqHTr4BNqe2ScsK1eEENvZDC3W1ur
Resources: (The tech I'm using for my larger voxel game project)
bevy game engine: bevyengine.org/
game_stat: github.com/TanTanDev/game_stat
physics (3D and 2D): github.com/dimforge/rapier
inspector ui (runtime tweaking): github.com/jakobhellermann/be...
hot reloading using dlls/so files: docs.rs/dlopen/latest/dlopen/
benchmarking: github.com/bheisler/criterion.rs
world generation: my own custom implementation is not open source at the moment.
other rust libraries, that I recommend:
game engine: github.com/not-fl3/macroquad
Graphics rendering api: wgpu.rs/
multithreading: github.com/zesterer/lagoon
profiling: crates.io/crates/profiling, using Optick feature
#tantan #gamedev #voxelgame - วิทยาศาสตร์และเทคโนโลยี
![CAMPปลิ้น | EP.75[2/2] คุยแบบนี้แล้วจะจบกี่โมงพ่ะย่ะค่ะ!!!](http://i.ytimg.com/vi/N6E8njWH7Zg/mqdefault.jpg)
i love this channel, we're always funnin' here
love your stuff dani
How is Karlson going Dani?
Wait the Netflix hit-men didn't kill you
He's alive!!!!!
ye the channel looks pretty epic
you are alive!!
This matches my experience with ChatGPT pretty closely. You can get it to write simple code, but especially when revisiting data types, it will often forget how it declared them earlier on and then invent some slightly different way that might also work but is simpletely incompatible.
What I've also seen quite a lot is that telling it to fix some stuff in its output often leads to it outputting the exact same code again.
To get it to your the function signatures & structs you want, it helps to include the code in your prompt. E.g. Given the following script: [CODE] can you fill in the [FUNCTION_NAME] to do [FUNCTIONALITY]
you have to keep in mind that the AI's context is quite limited, only 4096 tokens (basically words) I think, so that's one reason it's not good at larger things, it just doesn't have space for it
You can build custom commands to make it remember yes it's is possible
It could get lost in context between functions
@@weakamna I've been looking into how it creates tokens. Fascinating stuff... =]
Yeah I noticed chat GPT really struggles with proving 3D data. I once asked it to make a cube edge table, and it couldn’t quite do it, it was always a little bit wrong
If something is a bunch of numbers with no easily discernable pattern, it will not pick up on the pattern.
ChatGPT has no internal visualization capabilities. When I write 3D code, I visualize the result in my mind to help me get these things right. ChatGPT can't do that. In humans, this would be called aphantasia. It makes graphics programming really hard I imagine.
@@anlumo1 As a modder with aphantasia, can confirm.
@@anlumo1 As a person with aphantasia, I can also confirm it makes navigating 3D environments in video games absolute hell
@@anlumo1 aphantasia is basically how I program and do mathematics, the part about text writing and syntax generation is really easy, those transformers are just kind of "copying code", they don't really have "scene understanding".
programming job is still safe. well, for web developers that might be a different thing
I worked with chatGPT to try to get it to optimize my algorithm written in rust to convert the decimal places of pi into base 26 so it can be represented as the latin alphabet for word searching. I seeded it with my code, after 2 hours of back and forth it spit out my code back to me and gave up. It literally stopped responding to my queries. it was awesome.
Lmaoo I’m dead 😂 chatgpt said “fuck that this shit too hard”
Okay so what we learned is that ChatGPT is still very far away from making your code, the rapping part however... Nah I'm just kidding 😂
The big issue with ChatGPT as a useful tool is that it almost always gives plausible responses, but often includes subtle errors.
It's a lot better for the creative parts or for more general things like "How to optimize a voxel engine". When you use ChatGPT to get interesting but non-factual information (or pointers where/what to look for in terms of factual information), then it's great
Like stackoverflow, but without elitism
@@dmitriyrasskazov8858 Like stackoverflow, but without accuracy or correctness. (Also, SO is a bit elitist but it's not so bad - mostly they just require that you strictly follow the rules.)
@@clonkex One of the odd things I've found playing with ChatGPT-4 is that it can recognize its errors once I point them out to it, and even provide detailed analysis about why it was wrong like, "Oh, you're right! This is wrong because [verbose and correct reasoning]."
If it is capable of doing that the moment I even say something as simple as, "Line 7 is incorrect in the above code", why could it not detect the mistake in the first place? I sometimes wonder if it's just trying to save computation.
Also it doesn't test and validate its code for correctness, even in simple cases. As an example, I asked it to generate code to approximate PI stochastically using Monte Carlo integration, and the output of the code it provided was a negative number. If it just compiled and ran the code, it should clearly understand that it's incorrect.
@@darkengine5931 _"I sometimes wonder if it's just trying to save computation."_ Nah it's not that smart. Machine learning is statistical, not intelligent. It doesn't "know" anything, it's just predicting the next most likely word based on previous inputs and what it's already said. It doesn't detect mistakes so much as give the most probable response to being told part of what it said already was wrong (and yet, it still doesn't "know" it was wrong; the only thing it knows is the words it gets fed from you and from its previous responses).
Pop OS, vim, rust, bevy. A man of culture!
Where did you get PopOS from?
Not vim, Helix!
@@mzg147 thank god
In my experience, GPT-4 has a FAR better programming ability. It's able to reason about problems far better and has a longer context length so it doesn't forget what it's doing as quickly. I've used it in Flutter to write parts of an app I'm working on, as well as in C#, Java, HTML/Javascript, and Python for a variety of personal projects as well as in my day job.
One example where it really saved my ass was when I was trying to compile data for the product support team at my company. I needed to compare a few columns and fill in another column based on those comparisons across tabs on a google sheet. I explained the problem to GPT-4 and it gave me the exact formula I needed first try. It made a potential hours-long problem into about 30 seconds of work.
if u want more increase ur own company and complex and never mistake I'll introducing AI CHAT DEEPAI the AI CHAT DEEPAI is so powerful and much better than OPENAI since start 2015-2018 and u can show ur picture and video and anything u want
No, it has less context. Gpt-3.5-turbo has 16k and public gpt4 only 8k (and there is some information that it is reduced to 4k for English)
@@palyaros02 GPT-4-turbo got upgraded to a 128k context window
We're always funnin' indeed
In ChatGPT's defense, it probably would have done better with like C++ OpenGL or C# Unity
yeah probably!
True! But it was really interesting to see how it does with slightly outdated data.
I doubt that. The same problems would apply.
@@MediaFilter You are the one who's missing the problem. Rust is a fairly new language with much less code examples for it and its crates in ChatGPT's training data. ChatGPT doesn't have a true intelligence, it is what's called "statistically correct", and having many solid training examples actually help it hallucinate a lot less even if it is the same problem in different language.
Hey I noticed TH-cam wasn't showing me any of your last couple videos (with all notifs on) , I thought you had just come out of hibernation with this one!
Great work as always though!
Holy shit that song and editing at the end. 10/10
And yeah, that problem becomes even worse when you don't know how to fix the mess it makes.
I tried it with shader code and my lack of experience in that topic made it difficult to create what I wanted.
ChatGPT just had a very good update. You should revisit the same ideas of this video where ChatGPT went from useful to aggravatingly insulting for sending garbage over and over again and losing track of things said in the chat. It will be a very interesting experience seing if ChatGPT can get things right this time c:
great video!
That's extremely epic! ChatGPT helped me to create a reflective minecraft block, lol! I would spend an eternity figuring that out!
So I wasn't really surprised it actually made a voxel engine :>
Hey Tantan, would you consider to make a tutorial for graphics programming in rust?
Many resources out there is in cpp and would be awesome to see how things work from scratch using Rust.
Great video btw!
Very entertaining video, amazing editing my dude.
This channel is so good. Brand new here this is my first video. How can someone be so entertaining while programming.
back in december I tried to get chatgpt to make a discord bot (as much by itself as possible), I've actually ran into very similar problems! it's interesting that even though the tasks were completely different, the AI had the same shortfalls, maybe it won't be that hard for the openAI team to improve on those!
I've seen clips here and there where ChatGPT helped tremendously with weird programming projects and/or with obscure languages. The Bing AI can scour the internet in real time and it seems somehow more coherent, I'd love to see how that version assists developers.
That's because the Bing AI acts more like copying verbatim than Chat GPT which synthesizes an answer based on data. While the Bing AI does synthesize data, it does so in a way that much more closely draws from less sources at a time and more closely matches the original content. This is why it seems more coherent because it's more closely referencing actual posts than ChatGPT is. This does mean that Bing AI will tend to be less novel in it's ideas, so depending on whether you're looking for creativity/uniqueness or less error prone and more coherent one of them could be better than the other.
I was trying to do marching squares for mesh generation with chatGPT and I ended up manually coding most of positions, indices and normals myself too cause it was wrong like like a lot
To get GPT to use your the function signatures & structs you want, it helps to include the code in your prompt. E.g. Given the following script: [CODE] can you fill in the [FUNCTION_NAME] to do [FUNCTIONALITY]
Man i love your videos.
yayay another video! Good one!
Loving this, quality content
ChatGPT can easily forget earlier context such as function signatures because it only remembers the last something thousand tokens from your conversation
The maximum input length for GPT-3 is 2048 tokens, once that is reached it will start forgetting things. Don't forget that both "I" and " I" (with a space followed by an I) are often two different tokens.
@@ColinTimmins This is not true. “ I” is one token. Check the tokenizer.
@@centdemeern1 I’ve seen in different places that I and I with a space are each, different tokens. I asked ChatGPT and it says the same thing…
Awesome! Btw, what is your tool for such convenient window management?
I use PopOS a linux operating system. That has built in window management that's wonderful to use
your game reminds me of Veloren, it's a voxel RPG that's also coded in Rust. It's open source too.
It is interesting to see where ChatGPT falls short in writing code for complex visual representations.
For adding packages you can use cargo add!
How do you make the compiler errors show in the helix editor ?
the unfortunate thing with chatgpt is that even though it can remember previous prompts and outputs, its computational output has been reduced over several updates. Sometimes i'll just copy and paste the entire code that i've agreed with, and THEN ask it a prompt, and it should then re-use the correctly written code that you've grabbed before.
Wow i only cringed a little bit at the end. That video was epic. I must look into morton ordering now. 99% chance i will understand nothing. Thus is the life of a self taught programmer.
I was searching for the same thing yesterday!
You are the best, dude.
I love me some TanTan content :)
I would like to see this but with Bing's new thingy instead, since it has access to the internet and up to date info
i found it easier to use chatGPT code by giving it exact names and types for input arguments and the return value.
That sick GPT rap got me to subscribe. 🤣
Have you heard of Veloren? its a FOSS voxel game written in Rust, its pretty far along in development and has some really cool features i havent seen in many other games. Its kinda a clone of the game Cubeworld
I have no idea what's going on in this code but for some reason I found this very interesting.
I have experimented with other code with chat gpt. It was amazing for simple code, allright for intermediate code (if your very specific in one comment) and shit at the same time...
I'm just a web developer and this is exactly my experience with using chat gpt 😆 It's still very helpful though
Came for the code, stayed for the rap!
TL;DR surprisingly neural network that is designed to write essays and articles is better at writing essays and articles when asked for advises than at writing code.
It's possible to fix some inaccuracies by copy and pasting your current code back to the chat so chatgpt can properly keep track.
Tantan could you please make a tutorial on voxel generation in 3D.
You can be as threatened by AI code as you are by other people sharing their code. Also, you still have to write all the commands to the AI, and going with deeper and deeper granularity to explain the exact details you want to the AI model. It'll be a more time consuming way to program, and there's no simple syntax for this. Then there are programming helper tools with a backing of AI, like Github Co-Pilot that actually helps you be a lot speedier with repetitions and that "understands" your own programming conventions. It can give you hints just like GPT but it's not really its strength, the strength lies in being a super clever autocomplete.
The thing is that it's supposed to be a tool to help you make the engine, not a tool to make the engine. People are wildly misusing this tool for things it shouldnt be and then act surprised when it doesn't do things the way they want it to.
That's not to say it's not pretty good at creating stuff, it actually is, and with enough time and effort you could end up with a fully working engine. Midjourney is already creating art like Picasso without even trying. There won't be humans involved if this keeps up in the future.
Missed opportunity to have the AI attempt to explain culling
Rust said to hold your beer in the thumbnail because it was obviously quite drunk when writing this code
What linux are you using man?😅
You get better results for your prompts, if you provide more context, like the existing data types and copy/pasting relevant parts of the documentation. I think it's called "priming".
could you try this with gpt4?
So, as far as I understand us programmers are safe from being replaced... at least until OpenAI rolls out a new version of GPT.
This is the thing people don't realize, you sort of have to be an already prominent developer to use chatGPT to give you not even proper code.
If you don't know what is needed than you simply cannot generate viable code from it, I attempted this for my UE5 development and I encountered many problems, it has to know too many variables within the engine (and even the code it generated earlier) itself to actually give me usable CPP code. I can definitely take that code and *with my knowledge* fix the things that were broken and it works (but not fine), but again, someone who is not already a developer will not be able to do that.
ChatGPT is not here to steal our jobs people, at least not yet.
Now do I think it's useless? Definitely not. For very basic functions and just learning what things are I think it's great. It *can be* a really useful learning tool in the future I think. As for now? It is incorrect too often to be used to learn stuff at a high accuracy.
I only tried CPP though and I would assume since CPP is quite complex that it might be better at js/ts, C# etc.
it's funnin' time
we're always funin'
4:00 This is something I've been wondering for a while now, as probably one of the last remaining people on the planet who have not yet interacted with ChatGPT themselves:
If you tell it that bevy doesn't have a noise function, assuming it actually takes that information instead of claiming that it does, would that "memory" go back into it's "knowledge"?
As in, would it then stop telling other users that bevy has a noise function? Or would it forget this part at the end of the session?
Not for ChatGPT as far as I know, it's pretrained on data pre-2021(I think?) but doesn't take in new information as far as I know.
The new Bing chatbot, however, is trained on much more recent data and can even search the internet. Through some breaking of it's limitations, it has divulged that it is being trained on data inputted into it, though as with any response from an AI like this take that with a giant grain of salt... Of course for both ChatGPT and Bing, I would guess that conversation data is being fed back to the AI in some fashion "manually" in order to improve the model.
It's been hypothesized also that Bing will use it's internet access to search up previous conversations it has had, which might also count as "memory", though this likely only happens if you ask it for something that is likely to need that specific information
Lol the ending.. epic 😂🤣
Meanwhile here's me struggling to make a pong clone in Macroquad, and believe it or not, my issues had nothing to do with Rust specifically, lol. At least I fixed it, now I can continue development.
I feel like you'd have gotten better results from chatgpt if you just fully copy-pasted the entire code in before your prompt, so it would have the data to go off of before creating a response. Might have avoided things like changing data types at random.
Please, do some complex stuff using Rust and GPT-4 again|
Jonas in an alternate universe
epic
if you stay inside a single conversation during the whole project and state that you dislike this and prefer that, it will generate code according to these preferences. Mostly.
Lol ima try this with raylib and C++
I prefer your raps over ChatGPT's rap.
"It looks correct" is the very reason ChatGPT won't ever replace real software engineers :)
"Ever" is a too strong word.
I would instead say: "ChatGPT won't replace real software engineers *FOR NOW* ".
Im simply in love with U
The hiccup with current iterations of ChatGPT is that it struggles with mathematics. Badly. Things from moderately complex multiplication/division to vector mathematics are topics that GPT struggles with.
ChatGPT is only useful to me as an interactive knowledge base, like a nicer documentation search engine. Even then you have to verify everything externally because I got wrong/outdated info plenty of times. Using it for anything creative such as coding and writing is a lost cause because it doesn't really understand things so you get something on the nonsense-plagiarism axis.
i love to funnin'
so not only i go thru that pain wondering why chatgpt gaslighting to me when i tried to push more then one function at once :)
ChatGPT could be a lot better, and a great learning tool even, if the AI model wasn't from years ago.
Hold on. Just e littel
5:20 you were rapping xd?
Great now do it in GPT4
cool rap gpt
try that with binggpt, it has current internet access
try multi-layer height-map images (color and/or uv image layers) engine, only vertical lines viewport projected/360 intersection rendering
la la la la hummin
Wow
We (devs) are still not out of job because of ChatGPT but in year or two I'm not so sure.
This thumbnail is now infringing a trademark xD
that's so crazy, It's a draft though, so hopefully they don't implement it.
chatgpt suck at coding, that it doesnt even help
but copilot is useful for writing boiler plate code and writing util/math functions and etc, really helpful
Lol ChatGPT trying to generate Rust code the easiest code to write wrong (to the compiler)
chatgpt struggles with math and anything that takes more than 2 steps computationally
It's still looks like I lost my job 😭
"Shat GPT"
Could please enlist your dev enviroment 😀
Guys its real time it uses bing now XD
So it can write the code, if you know EXACTLY how the code should work. Which doesn't replace the job of programmer... yet.
chat gpt is better for getting a working prototype, but you have to make the code good yourself
i love how GPT gets dumber the more mistakes it makes.
That's because the mistakes feedback into its context for subsequent generation
time to oxidize minecraft
BOGUS
ShatGPT. 👍
If you make a game using “shatGPT”, is the result poo?
Tantan, please wear bowtie
Holy shit, I've been spending many nights thinking about how I could store data in multi dimensional space is close in one dimensional/linear array space.
Nice that this problem is already solved. So once I actually learn to "just make game" I'll be able to implement it (I want to make a space physics simulation with a game slap on top of it).
So cool that I'd just stumble upon this on a random video youtube recommended me. Thank you youtube, no credit to the video creator nor chatgpt and the data it was trained one. But really the credit should go to me for the brilliant idea of procrastinating on youtube.
It sucks on advance logic..its best parts is discovery
Open source is a curse.
using WHAT
first