How long do you think it will take till we can train a GPT that can create SC maps for us using natural language?

I have so many ideas that I'd like to see come to life that just need a generic shell to exist in, that I can then tweak.

I figure that's probably a pretty niche setup that you would have to program yourself or hold your breath a looong time for. I've toyed with asking ChatGPT for suggestions on map concepts and it has been interesting, but I haven't walked away with anything particularly useful yet.

I would say the tech is already there. It's "just" a question of training (fine-tuning) a large language model (LLM) for this specific use case. Just in quotes because I doubt there is enough crawlable data publicly available to pull it off. But I'm not an AI researcher, so take that last sentence with a grain of salt.

And since SC mapping is so niche I doubt that OpenAI or anyone licensing their GPT model would even attempt that. If it is to happen it will likely be on the sc community to train an open source LLM like BERT.

So in conclusion I would assume never.

Post has been edited 1 time(s), last time on Nov 19 2023, 7:42 am by NudeRaider.

You don't think AI will ever be advanced enough that we could tell it to learn how to use a tool, and then make something with said tool? I think we're already pretty close. It's just that currently, we need to do a lot of manual work teaching it how to use a tool, but I don't see any reason it can't eventually figure out how to teach itself something. They already taught themselves how to play starcraft.

Not sure if you actually read my post or just the last 3 words... Spoiler:
Were my first words.

That's not how fine-tuning a LLM works. Basically you give it a massive amount of data and it "magically" makes sense of it and then is able to apply the knowledge.
What you're describing is the classic approach where we tried to program smart (and numerous) enough rules into it so it feels intelligent.

ChatGPT is a real PITA. I used to take submissions for short stories for a mod I work on but then people started flooding it with painfully obvious chatGPT submissions so I had to shut it down.

ChatGPT tricks people who don't know what good writing is, thinking it is generating good writing, but it's not. It creates the most bland, cliche, generic stories EVER.

I am an AI researcher. Generally speaking Nude is right, this is not going to happen unless someone from SEN or maybe Korea reeeeeeeeaaaaaaally takes an interest.

Edit: The reason - ChatGPT basically regurgitates statistically relevant information it has memorized, it doesn't produce new or novel things, except insofar as combinations of existing things it knows about may be novel. So you can ask it to write a Harry Potter Star Wars crossover because it's been trained on millions of lines of Harry Potter and Star Wars fanfic and can draw on both those things already.

The first problem with having ChatGPT create a map for you - how do you reformulate the problem into something a General Purpose Transformer architecture can understand? Triggers use a language based format, and are a good candidate to feed into something like this, but how do you encode terrain, building placement, and a sequence of abstract game based events into it? If you could overcome this hurdle the problem is possible (see: AlphaStar playing SC2 for example), but when you look at existing creativity models (ChatGPT, StableDiffusion, etc) the lack of artistic vision really creates more problems than it solves, once you want to go beyond basic regurgitation.


What you're describing is general purpose AI, which is quite a different level than existing technologies like ChatGPT. Nominally speaking this technology is at least decades away, though I can't claim to have read every research paper in existence, and given the break-neck speed at which the field is progressing, it could be closer than expected. Perhaps human intelligence isn't actually that high of a bar, for instance (See: Westworld).

Post has been edited 4 time(s), last time on Nov 20 2023, 11:48 pm by Vrael.

It's interesting that you mention that. I too have looked at what AI produces, especially their mistakes and it does remind me a lot of flaws of human minds, especially how fuzzy memory can be, or how blind people can be to facts if they do not fit their world view. Conspiracy theories could be our equivalent to AI hallucinations. So I too, couldn't help but wonder if the human speech center isn't also just a statistical machine that produces words based upon what seems to fit best next.

It's a bit mind blowing when your whole life you basically compared your brain to a biological computer where bits are represented by neurons firing, misremembering was attributed to degradation of biological "components" when in fact it seems to be a fundamental part of how it works.



I believe we currently are in the exponential phase of a S-curve development and are going to plateau soon enough. As for general AI I share your assessment, that it could be decades away and only really gaining traction once we have a better understanding of generative AI.

I went to a colloquium yesterday at my university on "neural networks through a geometric point of view" given by a guy from MIT. It was pretty interesting.

The jist of it is that there are some hidden convexity properties in some commonly used AI algorithms when viewed in a certain way. It lets them basically run the algorithms faster.

So yea research is progressing steadily, but it's mostly an academic venture though. The main tools of all academics nowadays is to find novel ways to bridge one field to another so you can use techniques from one in another. IMO, to get some "sci-fi level" AI we need completely original and novel ways to breakthrough some problems. That's not to say we can't refine stuff to a certain degree.

These are just my feelings, but like, to give a dumb way to explain this: Probably to break into this "sci-fi level" AI we need to at the bare minimum relate the problem to something stupidly impossible like the Riemann Hypothesis or Navier Stokes or whatever; which are like totally unpenetrable problems. And honestly to create something that can reallllly mimic human creativity we would need to go even farther than that.

Tldr; we need to progress math further first, but the ceiling of math knowledge is imo still like we're babies and really pathetic. That's all just my feeling though.

no

What a basic low-level AI answer. Are you sure you're even human Corbo? Or are you just three llamas in a trench coat maybe?

But I'm not trying to have it come up with a unique and creative starcraft map. I feel like should be able to upload it 100's of examples of (say) "Turret Defense" maps to train it to understand what a turret defense map is. Then I can ask it to make me a generic one.

Why not simply build off an existing "Turret Defense" map then, if all you want is a template to start from?


The critical piece here in my opinion is that a starcraft map requires 100% precision in terms of its function - you can't randomly place a location in a non-walkable area and then expect the trigger its attached to to actually make something happen when a marine walks into it. So even if someone trained a StarEditGPT you would end up spending more time debugging these kinds of "creative errors" than you would just making the map from scratch. The second, but probably solvable problem, is actually encoding all the pieces of a starcraft map into something that a transformer architecture could actually learn on. ChatGPT/Dictionary words are often run through a One Hot Encoding for instance - you have a vector the size of your vocabulary (say, 10,000 words) and there is a single 1 in the location that corresponds to the current text item. More sophisticated approaches use a latent space encoding. To encode a starcraft map, you'd need to come up with some kind of vector encoding for the locations, for the triggers, for the terrain, for the interactions between them, and it gets worse because the triggers arent going to be a fixed size, etc. If you managed to solve this, the third problem is data size. Transformers take tons more data to train than their convolutional counterparts - 100's of Turret Defense maps isn't going to cut it, you would want something like 10,000 examples of each category - Turret Defense, LOTR, Zoids, Campaign, whatever, (100,000 of each would be better, and you may even want millions). If you input the entire corpus of existing maps, the model still might not converge. However, if someone were to solve these issues, or if someone comes up with a better architecture than the Transformer, or some small-data technique or something, hey maybe it could be done. I'm not going to be working on this though

Why not, you seem to know what needs to be done. You're the one! You can't just turn your back. WE NEED YOU!

Ok fine, I'll start on that after I finish Tenebrous