Can ChatGPT write an Eulerian Fluid Simulator in 200 lines of code?

[@patryk6769]

It doesn't get stuck, but it has a limit of tokens it can produce per user request. Sometimes I just ask it to "continue your code" to get rest of the answer.

[@aditya_a]

right i usually ask it "what's after the line ____" and it continues

[@vaan2900]

I just say 'continue'. always works

[@johnlin9665]

"Sure, I can do that!" Gets stuck... Sounds like OpenAI trained GPT to write code the same way I do.

[@TenMinutePhysics]

Thanks for the hint about the possibility to generate more code. I might do a follow-up video and discuss some more findings... and if somebody gets the flipper or tutorial 3 working, let us know in the comments

[@ThreadingReality]

Funny true story, after your eulerian vid I wanted to code it in unity, and had issues converting the Java, so I signed up to chatgpt to help expand on some of your techniques and explain or change things. You should also check out gpt4 it's generation abilities are way better for code now

[@EMAN360ELITE]

Hey @TenMinutePhysics! I converted your 2D Javascript implementation of PIC-FLIP to C++, then ChatGPT 3.5 converted it from a 2D simulation to a 3D one on pretty much the first try! I just told it I was going to feed it several methods of a 2D fluid simulator one at a time (since the whole program was too big), and asked it to convert them to a 3D simulation (I'm paraphrasing). It forgot what it was doing about 4 methods in and I had to remind it, but the 3D version it created works great!

[@tekperson]

I found this channel recently, and it's become one of my favorites. I ran the first several problems from Project Euler through chatgpt3, and the results were mixed. More complicated requests give either mixed or poor results. That's very consistent with Ten Minute Physics results. I have used chatgpt3 to help prototype a technical presentation. The results are very wordy and often a little off base - but speed up initial development. Frankly, there needs to be more transparency about where the training material comes from and what restrictions the code it was trained on has. Normally licenses propagate with code, but somehow this short circuits that. This makes using chatgpt3 for business applications a bit iffy.

[@rodai4999]

to do more extensive code, just ask chatGPT "complete the code snipped where you left off in last answer", you can do it multiple times, and manually merge the code to run! it works!

[@MrButkam]

Keep up the good work ! Thank you for your valuable videos

[@bengineer_the]

If you ask it: to continue from the line: "contents of last line" then it will produce the rest of what it thought of. :)

[@LordCogordo]

It has a limit of tokens, whenever he gets stuck just ask "keep writing" and he will

[@compphys]

Something that's not trivial and cannot be found on the net (easily, bouncing balls are all over the place) to copy/paste cannot be generated (correctly, that is). I tried to get from it a function that computes the Clebsch-Gordan coefficients, got a function that spat out -nan. It's definitively far from a silver bullet.

[@WizardOfAILand]

LOVE THE VIDEOS!!!!!!! GREAT JOB!!!!! Could you do: Moving Least Squares Material Point Method ????

[@Ghostrider-ul7xn]

ChatGPT absolutely can write complex simulations, I'm a physicist and I've done it. You just need to tell it to break the task into small steps and write fully completed blocks of codes for each steps. In other words, tell it to modularize the code and write blocks of codes for each steps. Not only will it comply, its accuracy will improve significantly.

[@RonM_1971]

Really? Can you provide us with some useful cases, as video content as well? Thanks.

[@Sgrunterundt]

The video is a year old, before GPT-4 was available.

[@plasid2]

Its possible make this using material nodes in UE5?

[@nielsdaemen]

You should try it again with the newest GPT4 version

[@ksitau]

I think you need to write "continue" to get remaining code

[@444haluk]

You shouldn't expect semantics from a syntax machine

[@johnpelitidis6297]

This scares me... who will own and control this technology?

[@zebfross]

ChatGPT is way more useful to me than Dall-e 2. Such impressive output!

[@takayamayoshikazu2782]

If chatgpt has such low processing capability, I can't use it for my engineering work...

[@samsaraAI2025]

I tried many easy things and it is not working for serious things.

[@Djellowman]

I feel like this video won't age well

[@bakmanthetitan]

Amazing video. Like the other commenter pointed out, there may be ways to get around the output limit. I would love to see a follow-up.

[@GrinderFX]

Actually there is nothing great because chatgpt got used fully wrong and it can solve the problem really good and not like shown in the video!!!

[@MarcelNKemet]

Just relax. ChatGPT is just a NRP program aka Natural Language Processor. It takes your command, encodes it and foward it to a Web Search API. It's just a lazy way to make a goggle search.

[@BumboLooks]

THIS... Why does nobody realize this? Why is everyone treating it like a sentient program? Why is everyone so stupid?

[@Compguy321]

ChatGPT-4 is better now!

[@ummerfarooq5383]

Can the paid version translate your code into faster languages?

[@KlaudiusL]

"ChatGPT won't be a competitor for a while" ... nop .. will be a replacement

[@nolram]

No. It cant innovate or create new methods as its merely a sentence completion algorithm; it doesn’t actually know anything about what its saying, it just says what is most likely.

[@KlaudiusL]

@@nolram AI is not your last phone or your car, which are "finished product", they lack some functionality .. and never will has. AI is a dynamic system, evolving permanently. Is always "half way to the end". Don't be fooled. Maybe, TODAY, it hasn't such functionalities, but take for sure it'll has.

[@nolram]

@@KlaudiusL AIs are not really dynamic. They are just equations, building connections of virtual neurons to associate patterns. They may mix, they may match, they may process in ways unknown to us, but they can never create something truly from scratch.

[@KlaudiusL]

@@nolram Neither humans. And here we are.

[@ShamJam85]

Chat GPT , as the name indicates, is meant for CHAT . Not for CFD code