How to effectively use chat GPT in unity?

To be fair, removing all of the code is definitely one way to improve performance.

 

Well yes. It would be faster ......

 

In my experience, it starts skipping code when it approaches memory limits. Just need to start a new chat or engage with 32k memory API (it's expensive though). Furthermore, large scripts eat up tokens fast, you need to stick under roughly 300 loc for it to be able to effectively refactor something. And the free version has even less memory.

The other thread that got locked had a suggestion to use free CodeGPT IDE plugin with GPT4 API. So I set that up in Rider and it's been pretty great. In some ways a little better than ChatGPT, because as soon as you mention Unity, ChatGPT wants to make everything a monobehaviour, scriptable object, etc. While direct interaction with API via CodeGPT opts for pure C# if the code is not directly Unity dependent, which I prefer. Direct API calls also generate quicker, and I assume don't have the 25 message limit. That together with the ability of it taking current selection as part of input prompt and then being able to replace selection with generated results, makes this a much better experience overall.

 

nice. I just started to use the api/playground since i can select the gpt4 version from march there. Not the clownshow it is now. Its really expensive tho getting charged per token.

never heard of codegpt. Is it better then copilot chat?

 

It's just an integration plugin.

 

I've yet to accumulate more than the ChatGPT subscription, which with VAT is almost 25USD. I guess august will be the real test on pricing but for now it pays for itself in productivity gains. Luckily, there are hard limits I can set for API usage so I can stay within budget.

And as Ryiah mentioned, CodeGPT is just a convenient wrapper/IDE integration for GPT4 API.

 

Largely depends what you do with it. I mostly let it refactor functions with limited line count. We are using this with two devs and have other API usage. Typically we had 2$/month with a couple of queries here and there.
A single query saving you some refactor work while you grab a coffee saves you that money in productivity.

For the typical ChatGPT user doing a couple of queries a month will be a lot cheaper.
But ofc if you go completely nuts, use the 32k model and fill it up you will get a big invoice.

I use all these tools in concert where they make sense.
For certain exploration/prototype feature code I prefer the chatGPT interface, copilot for auto complete and codeGPT for medium size tasks.

The trend with the Github tools seems to be to add hidden context without letting you know.
Both try to improve the queries by sending the model what feels like all the currently opened files.
Can be good, can misslead the model. Will probably change in future updates.
I'm not the biggest fan of that approach in the current iteration, I'd prefer they tell me somewhere what it knows so I can adjust.
Sometimes it heads off in the wrong direction because some other file I have opened in the background.


CodeGPT you just copy/paste or select and use the query feature to send it what you want, you can also save some macro queries. Useful when you figure out what fluff text you typically need to get the desired style of code from the model.

 

I solve 100% of my programming issues by using stackoverflow, reddit or these forums, what ChatGPT does is browse forums for you but you don't get to choose which is the right answer.

 

ChatGPT doesn't browse anything for you. It creates responses from scratch by assembling tokens chosen via probabilities. You can increase its available pool of tokens using GPT-4's Link Reader plugin but even then the answer isn't being copied out of that or any other thread.

Thanks to this the bot can solve problems that are incredibly difficult to find solutions for or may not even have a solution, and far faster than waiting for a response which can easily take multiple days or even weeks. It can also create a completely bogus answer, but it's often possible to work around that limitation.

 

StackOverflow has lost 35% of its traffic and is experiencing up to 50% decrease of questions and answers, as well as upvotes: https://observablehq.com/@ayhanfuat/the-fall-of-stack-overflow

Ryah already explained why.

 

Indeed banning questions and answers from the bot wasn't going to stop people from using the bot. It was just going to stop them from coming to the site after they'd exhausted what the bot could provide them.

 

I'm finding the results lately to be quite variable in general and somewhat repetitive for my various queries, mainly using Bing and ChapGPT4. It's still pretty great at specific types of coding questions. I mainly use it because it's just faster to get a summary than click on a link and wade through a long post covered with ads.

Lately though, a general question will result in it only referencing one or two sources and it's reluctant to go beyond that. I think it's what people have described, that they are throttling it down to use less resources and so the results are 'meh'.

Knowing that many people are posting less on StackOverflow, it makes me think a year from now AI may not be nearly as capable as it is now since it has a large trove of past answers to ingest. If people all over ask less questions on the internet in general, there is going to be less good information available for AI to draw from. In fact, it may start ingesting other AI content like an ouroboros and become useless.

 

Speaking of throttling I noticed that the message limit for GPT-4 which had been "25 messages every 3 hours" for the past few months has now been raised up to "50 messages every 3 hours".

 

The model collapse is a worry but GPT4 is nearly three years out of date already and it still continues to be useful. A year won't change much unless some massive technological shift happens at Unity, which it won't.

 

If AI-generated data is valid, doesn't matter if it is AI-generated, it can be used for training the next iteration. Same deal with AI art. If human judged it to be aesthetically pleasing, it can be used for training.

Eventually learning algorithms will be probably optimized and require less material.

At the moment, for example, open-assistant is gathering their own database. The database has something like 84k rows of data which is not a whole lot and is smaller than Crawl. People used it already to finetune.

 

Well if it does not read what knowledgeable people share about the subject there is no wonder it makes so many mistakes.

 

...because wise usage of Unity is pretending you're using a 5 year old version of it.

 

I am a bit skeptical about this reasoning.
First of all, fewer repeating over and over questions, doesn't contributes much. Other than by adding to more junk, which need to be filtered out, just to find information.
Often marked as duplicates of duplicates.

People ask ChatGPT or similar, for questions which has been typically answered at least once somewhere. Or crawled through documentations, or sites.
The one who will stay on these forums, are people who either refuse to use generative tools, or in case of very specific technical questions, for which answer is not present anywhere, or very difficult to obtain.

As the result, there may be less of junk and more of quality responses. And community will be narrower.

But I see other side of the coin.
If there is less of people to post things, then less people to discuss with. And perhaps discouraging of even asking questions, on abandoned ship.

 

Gtp isn’t asking me all the time if have googled the problem before or what i have done on my own to solve the problem. It just answers the question. And there is not another gtp sliding into my gtp chat telling the first gtp its a fkng idiot for doing it that way. In short: you have no human interaction and thats the best part.

that you cant find „anything“ anymore on the internet already started with discord. Like for unity plugins there are is almost no forum interaction anymore. Everyone shifted towards their own discord server and its super hatd to find answers there. I would be very happy if assetstore plugin creators would just train a lmaa instead of letting me crawl through discord servers looking for answers

 

We need to implement this as a feature.

I agree with the overall idea, though.

 

while we are at it. Just randomly add a „laughs in godot“ as a possible solution to a problem

 

Not using LLMs because they sometimes get things wrong is like saying in '96 to not use the internet because sometimes somebody is saying something stupid in a forum/blog.

They clearly have some utility and they will only get better but they also have their drawbacks the same way forums and books have their downsides.

So much easier to find niche/contextual information in a LLM and be able to direct your research with that initial direction.

 

not rly in context but i just remembered this quote:


1. Anything that is in the world when you’re born is normal and ordinary and is just a natural part of the way the world works.

2. Anything that’s invented between when you’re fifteen and thirty-five is new and exciting and revolutionary and you can probably get a career in it.

3. Anything invented after you’re thirty-five is against the natural order of things.

-Douglas Adams in the Salmon of Doubt

 

On the other hand these tools encourage so many more people to learn programming or in general become more knowledgable for the situations where the tool does not automatically solve the problem. After all Wikipedia has not replaced knowledge in our heads either.
Yes some people will completely overuse it but those people would likely have failed otherwise anyways.
Think we'll see more people having success than abandoning ship in general.

Oh do I wish I had this back then when studying...

Edit: Quite a matching take by Douglas Adams

 

The more I use chatgpt the more I realize that it doesn't have any intelligence at all. All it does is some way to spin output according to what the users input. A fancy form of "Search" box that we have right here on the forums.

 

That is correct, it's just a probability engine. Skynet is not taking over.

AI is not search, it does not display results from somewhere else unless using a browsing plugin and even then it does not copy/paste. And it does outwardly present a level of understanding even if its not inherently intelligent. Paste in some new API docs you're trying to learn and the AI will discern the purpose of the API and will be able to provide common usages and go deeper on any of those usages in the context of your project assuming you provide that context in prompt form. AI can also implement them and give you a good start.

And even if used just as search, it's faster than googling 9 times out of 10 assuming training dataset contains the information necessary and if you have further problems on the topic, it can troubleshoot those as well with 0 wait time.

You also haven't mentioned which AI and which version you tried. A lot of people try freeGPT or Bing chat and then dismiss the technology as a whole when all the power is in paid services right now. GPT4, ChatGPT4 and perhaps Copilot are a generation or two ahead of free tools.

 

I prefer Claude2 over GPT4 its knowledge cutoff date is 2023 and it also has a 100k context model available to use. I can feed a book or other resource to its 100k model before asking questions about it. For example, when I was trying to understand scriptable object architecture, I uploaded a book and asked Claude2 questions about it. Also, GPT4 API doesn't feel dumbed down compared to ChatGPT4 (I use poe since it gives me 700 GPT-4 messages and 1k Claude2 messages)

 

Claude 2 has more limited availability than GPT 4. It is only available in US/UK and that's it.

 

It is available on Poe.com

 

"Due to limited capacity, we've had to limit message sending on this bot. You have 3 free messages left today across all bots that use Claude-2-100k."

Also, who is this poe? Feels like it is affiliated with quora.

I feel like I need to start looking in running MPT storywriter locally. THough last times I tried it, it tended to go crazy.

---

I've managed to get through Poe's waiting queue for Claude 2, and, honestly, this is more like it. I've managed to stuff a story in a ballpark of 75k characters into it, and it could analyze it. However, apparently it still lost the end of the story, so would be nice to have even bigger attention window. Something in ballpark of 200k would be probably bare bones sufficient for a writer's assistant, and ideally would be nice to have a 1..10 mil tokens attention window while being able to run it locally.

But we can dream, I guess.

Also, can't say I agree with conclusion it drew from it.

 

I tested claude 2 through official site. While the attention window is impressive, the model itself occasionally arrives at bizarre conclusions, skips good chunk of source material, and for example, when I asked to translate a decent sized body of text (about 50 kilobytes), it lost the middle, due to two scenes being similar. Then it terminated the response halfway...

 

just wasted a good amount of time setting up and applying for copilot chat.

the product page says "With chat and terminal interfaces, support for pull requests, and early adoption of OpenAI’s GPT-4" but when i asked copilot what its using because it felt kinda "dumb", it says "the website states that Copilot has early adoption of OpenAI's GPT-4, not that it already has it. The model powering Copilot is currently based on GPT-3." which you can already tell by the quality of it but its a shame that its false advertised!.

so in the end: its not worth it

 

I also have copilot. I often end up in a fight with it, i type say "if (something) {" and it suggests like 20 lines of code, and deletes some of my own.. so I say no, and start entering - the word (and valid variable) Target, is flipping always replaced with something and the more i correct it then more determind it is its not... I dont see an option in the Rider plugin to ask it a question it just randomly makes suggestions and you can click on a window for a bunch of alternatives

 

Some additional information about claude (after using it for a while):

- It works much better than FreeGPT with non-english languages. Basically, in case of FreeGPT, for me it is only really usable in english, in case of claude non-english is fine and is well constructed. I think I've only seen one error.
- The model actively tries to pretend to be a human. When questioned, it insists that acting as a human is one of the project's goal. It never says "as a language model", and can say things like "I want to see what happens next".
- The maximum amount of text I managed to feed into it was about 158 kilobytes. It does keep entire text in mind, but can cut conversation short if t he total number of characters in conversation hits the limit of around 160k.
- It does work fairly well for text analysis. Basically, you can dump a short story into it or several book chapters, and say "I need an opinion about this story", it'll oblige, and it'll understand t he whole thing.
- However output is not really consistent. If you reroll you can get a different opinion. When asked to "provide an opinion" on a short story, It frequently acts as if it selects several key point and then focuses on them while ignoring other things.

 

I didn't think about doing this until I stumbled upon your thread, and I then figured out that you can learn just by asking ChatGPT.

I'm considering using paid model 4, can someone confirm the amount of lines or efficiency gained on average with the newer model compared to 3.5?


Oh god, we're entering the beginning stages of that Movie.

but I don't care.

 

You might want to try WizardLM, Vicunia, claude and CodeLlama first. See if you can run any of them locally.

 

You could ask ChatGPT how to use ChatGPT.

 

You can't really compare them in that manner as it's very dependent on your prompts. While GPT-4 is consistently "better" than GPT-3.5 I've had cases where the free outperformed the paid in weird ways.

 

Apparently the current "low end" champion is mistral 7B, it can reason on the level, allegedly, of gpt3.5. The creator is a former lead on LLaMA2.

IMHO, people who say LLM aren't intelligent miss the point, they won't need to be "intelligent" to mess you up and take over.

Also the value of LLM aren't in not making mistake and being used as a chat consultant. It's when people start bringing their autonomous "stupidity" such as it can do half your work in 10s, 24/24 7/7, and never complain. If 7B model are enough, at best that's 100WH, which roughly translate into my local rates to 400€ of initial investment on the machine, then 4€ of electricity per month. Compared to a full human at 10 000€ a month only 8h a day, which only 6 are productive. Machine only need to be mediocre at the task to be competitive.

Currently experiment show that a fleet of agents can actually do a lot of small SOP task "good enough", where you have each task layered by "personality" that cross check each others and only have memory for very small history, and manager that distribute the task and take decision. For any task that have clear validation (such as does the code compile) they will be able to progress.

I'm trying to learn docker right now ...

 

I use it like this

and it had good ideas

some funny

some got some to it

some general enough but revoking memory

would say like every 10th idea
but it did gave 30 game ideas only for Heat

I like these and could showcase events/scenes not even in the movie

 

/s ?

 

no, it has good analythical process in connecting the dots if looking for comparisons in a wide set.

it took out the most memorable scene in Indiana Jones. (there might be the tank scene (too complex, horse riding etc), can't think much more gamifyable moments, but running from a boulder after snatching the prize is the easyest and the simplest)

or just sheer luck

 

These are the kind of game ideas I have dozens of in the shower. Which is to say they're not very good, not fleshed out to any reasonable degree, and are so fundamentally generic aside that they aren't worth remembering so they probably won't be.

 

How is this hard, interesting, useful or actionable, and how is what it is describing different than Temple Run?

These are very surface and obvious ideas, and it would be one thing if they came from you, because that would mean the idea tickles your brain and you might eventually hone it into something that scratches that itch that is interesting, but if they are coming from AI it means that's all there is, and that's not much, if anything at all.

 

Was trying not to go back to this thread, and this is nothing to do with coding, but I thought this was pretty good (ideas for dialog in a space pirate game in the same vein as the LucasArts adventure games).

Interestingly, at least with a cursory google search, I couldn't find any reference to some of the jokes like the open comms channel joke or having potassium for the trip home.

Its not Pulitzer prize winning stuff but its a darn site better than the averge 'writer' on fiverr and the hammy tone seems perfect for this sort of game.

---

Can you write a 'screenplay' with camera instructions and dialog for the introductory quest of a sci-fi pirate game in which the character needs to fly their ship to a distant planet. The player character is a new recruit to the pirate life and the main quest giver is a veteran of space and the pirate life. The tone should be humouros with plenty of 'dad' jokes.


Title: Space Pirates: The New Recruit

INT. SPACE PIRATE RECRUITMENT CENTER - DAY

CAMERA ANGLE: WIDE SHOT

The room is dimly lit, with flickering neon signs and space pirate paraphernalia scattered around. The camera pans across the room, showing recruits waiting nervously. The door bursts open, and CAPTAIN REX, a grizzled veteran with an eyepatch and a mechanical arm, strides in confidently.

CAMERA ANGLE: MEDIUM SHOT ON CAPTAIN REX

CAPTAIN REX: (booming voice) Alright, listen up, you space scallywags! Welcome to the Pirate Life. If you thought this was a cruise, think again. I'm Captain Rex, and I'm here to turn you landlubbers into space-faring swashbucklers!

CAMERA ANGLE: CLOSE-UP ON PLAYER CHARACTER (PC)

PC, a young recruit with wide eyes and a nervous smile, stands at attention.

PC: (nervously) Uh, Captain Rex, sir? I’m ready for my first quest!

CAMERA ANGLE: WIDE SHOT

Captain Rex walks over to PC, looking them up and down.

CAPTAIN REX: (grinning) Eager, are ya? Well, that's good. First quest, eh? We're gonna start you off easy. All you gotta do is fly your ship to a distant planet and pick up some cargo. Simple as space pie.

CAMERA ANGLE: MEDIUM SHOT ON PC

PC: (confused) Space pie, sir?

CAMERA ANGLE: CLOSE-UP ON CAPTAIN REX

CAPTAIN REX: (chuckling) Aye, space pie! You know, easy as pie but with a little more… space. Anyway, get your ship prepped and ready to fly. I’ll transmit the coordinates to your nav system.

CAMERA ANGLE: WIDE SHOT

Captain Rex turns and starts walking away.

PC: (calling out) Wait, Captain! Any advice for a first-timer?

CAMERA ANGLE: OVER-THE-SHOULDER SHOT FROM BEHIND CAPTAIN REX

CAPTAIN REX: (turning back, smirking) Aye, lad. Always remember: In space, no one can hear you scream... unless you're on an open comm channel. Then everyone can hear you. So, try not to scream too much.

TRANSITION: EXT. SPACE DOCK - DAY

CAMERA ANGLE: WIDE SHOT

PC stands by their modest spaceship, making final preparations. The camera zooms in on the ship, showing its worn exterior and mismatched parts.

PC: (to themselves) Alright, let's do this. How hard can flying a spaceship be?

CAMERA ANGLE: MEDIUM SHOT INSIDE COCKPIT

PC sits in the pilot's seat, nervously pressing buttons. The control panel lights up, and the ship hums to life.

PC: (to themselves) Okay, uh, let's see… throttle, check. Stabilizers, check. Hyperdrive… probably fine.

CAMERA ANGLE: CLOSE-UP ON NAV SYSTEM SCREEN

The coordinates appear on the screen with a blinking icon for the destination planet.

PC: (determined) Alright, coordinates set. Time to make like a comet and tail it outta here.

TRANSITION: EXT. SPACE - SPACESHIP IN FLIGHT

CAMERA ANGLE: WIDE SHOT

The spaceship zooms through space, dodging asteroids and space debris. The camera follows the ship, showing its erratic yet determined path.

CAPTAIN REX: (V.O. over comms) Remember, recruit, keep an eye out for space cops and rival pirates. And whatever you do, don’t push the big red button. Trust me.

PC: (over comms) Big red button? Got it. Wait, why do I even have a big red button?

CAPTAIN REX: (V.O. laughing) Every good pirate ship needs one! Now quit yapping and get to that planet.

CAMERA ANGLE: MEDIUM SHOT INSIDE COCKPIT

PC nervously glances at the big red button, then refocuses on the nav screen.

PC: (to themselves) Just a quick trip. No big deal. How bad could it be?

TRANSITION: EXT. DISTANT PLANET - SPACESHIP LANDING

CAMERA ANGLE: WIDE SHOT

The spaceship lands on a dusty, barren planet. PC steps out, looking around cautiously.

PC: (to themselves) Alright, here we are. Now, where’s that cargo?

CAMERA ANGLE: MEDIUM SHOT ON CARGO CRATE

A large, suspiciously glowing crate sits nearby, half-buried in the sand. PC approaches it, scanning it with a handheld device.

PC: (reading scanner) Radioactive space bananas? What in the galaxy…

CAPTAIN REX: (V.O. over comms) Ah, you found the cargo! Just a bunch of bananas. Harmless… mostly.

PC: (sighing) Great. Space bananas. Why did it have to be space bananas?

CAMERA ANGLE: WIDE SHOT

PC hoists the crate onto their shoulder and heads back to the ship.

PC: (grinning) Well, at least I’ll have a bunch of potassium for the trip back.

TRANSITION: EXT. SPACE - SPACESHIP IN FLIGHT

CAMERA ANGLE: WIDE SHOT

The spaceship zooms back through space, heading home. PC looks more confident now, even humming a tune.

CAPTAIN REX: (V.O. over comms) Nice work, recruit. You’re on your way to becoming a real space pirate. Just remember, in this line of work, the treasure’s not always gold… sometimes it’s just bananas.

PC: (laughing) Aye, Captain. Bananas it is. On to the next adventure!

FADE OUT

END OF INTRODUCTORY QUEST

---

 

I'd recommend not to use any language models to write your scripts. Those systems sucks at creativity. They return blandest possible scenario and blandest possible script.

The reason is that the system produces most likely response comparable to average human. Average human sucks at writing.

 

Wow.

In the milieu of LLM chum, if that's enough to instigate coming back here, I'm confident there's gonna be jobs for thinkers that write, creatives that write, writers that write, and even socialists that write, for decades to come. Just in the game industry.

For spaces that demand good writing, where it's the primary focus, this form of AI probably doesn't stand a chance. LLM's have likely already ingested every published piece of literature from ever, have been steered for well over a year, yet still churn out this kind of cruft?

It's enough to make one think that, just perhaps, empathy, self reflection within a zeitgeist and human objectives are significant parts of story telling.

 

Only if you accept the first result! Just like with programming when it halucinates or produces bugs, dig deeper, tell it to filter out what you don't need etc.
The language models are assistants. Like people to brainstorm and reflect ideas with.

ChatGPT and co have also been trained on large amounts of professional books. Hence some attempts to raise copyright claims against them.
And actually you can start your prompt with something along the lines "You are a professional writer who has published eleven bestsellers in the [insert genre here]. You are willing to write another book in that genre. What would the rough plot of that be?"

If you don't believe that works - some people literally managed to make ChatGPT spit out instructions for illigal stuff by convincing it to pretend to be Joker! They call that process "jailbreak". Hopefully/likely that does not work now anymore so easily. In every case it is not hard to push the answers out of the norm of "average human".

 

Although I wouldn't use the script line for line, it was asked for 'dad jokes' with the specific intent of landing on a style that that would get the appropriate amount of groans from the audience and what it offered up was pretty darn groan worthy... in a good way. I'd happily use the 'space pie' and 'comms' jokes in an adventure game with a similar tone as Monkey Island.

Maybe thats too low brow for you, but not every game (or creative work) has to have a deep and meaningful narrative.

 

Although I don't think AI is quite up to handling the task of writing a (good) script from scratch, I wouldn't actually use this line for line, nor am I even working on a hammy space opera game, its still impressive what it can do. It's certainly far better than the average human at refining descriptions, writing the same sentence with different tone, or pumping out large chunks of 'average' text which you can use to test ideas.