@eniko the main gist was that as an illustrator, animator, etc. you shouldn't give the client what they asked for. Instead it was important to give your interpretation of what the customer expected/needed.

A mouse was a good example. Don't draw a realistic mouse. Draw an *impression* of a mouse that conveys the client's needs. Scary shape if for an exterminator, cute if for a children's book. And for a commercial...? 50's housewife scared of nothing.

AI just doesn't get the artist's eye.

@justafrog @eniko To be fair, this was true of the chatbots of 20 years ago, too (people claiming "we've done it, we've passed the turing test!" and then a real turing test was performed and it was marked failed because it couldn't answer "how are you feeling today?")

@david_chisnall @eniko Interesting, that’s also how rollback netcode for online games works. When you don’t have the other players inputs for a given frame due to network conditions you just assume they’re holding the same inputs as last frame and a lot of the time that turns out to be correct

@david_chisnall @eniko My grandfather was a meteorologist, and he often quoted that fact. In meteorology, it even has a name: “persistence of type”.

@david_chisnall @eniko this tradeoff is why I think the best use in the near future for LLMs will turn out to be voice controls a la "star trek: the next generation".

The captain can ask for simple, commonplace high level commands and be confident the computer will do it mostly correctly. And that's a huge time saver. But for anything that requires any amount of precision, or technical knowledge, you better believe he's asking Geordi La Forge (or maybe a real AGI like Data).

@karadoc And yet a regular Google search easily found this line. LLMs do not search for information.

@karadoc @eniko

Awesome! I'm laughing but can't help feeling a little creeped out, too. Thank you.

@karadoc it also loves to claim that you can do things that you actually can't, like connect an Xbox controller to a Switch without an adapter.

I accidentally ran into this nonsense when looking that question up, as expected it turns out it's not outside the shit AI-generated articles that just claim it's possible because, well, the algorithm was told to write a guide on how to do the impossible thing, and of course the thing it assumes people want to hear is how to do that, reality be damned

@eniko Once people are presented with 100% LLM-generated games, they'll like them just as much as they like the books that produces.

@eniko There was a scifi writing prompt on tumblr where they invent a transporter with lossy compression and people become more and more like each other over the decades due to average regression. I keep telling people hyping genAI to go read it. Of course the ingest-millions-of-examples-and-extract-the-most-predominant-features-machine can make DOOM-looking sorta-ok-looking screenshots! That's literally the easyest ask! Now try get it to make something niche like Microsoft Flight Simulator

@eniko This so much. Every time they're pushing a new iteration of a product, when they're showcasing its output with HAND CURATED best case picks, it's blatantly, obviously dogshit if you pay any attention.

It will always come short. With exponentially larger and larger indiscriminately pilfered datasets they define the exact contour of falling short with ever increasing precision.

@eniko and they first need to make the game to get the footage to train their IA on it.

@eniko stable diffusion has limitless potential for games! you just have to:
1) make a normal game
2) film millions of hours of said game

@eniko absolutely. Also like most ML things I don't even find the tech interesting. It's just a big optimizer and if you add some billion parameters to a model, that might just cover like any surface level data variation by sheer size alone. And it immediately breaks down with something like games or videos because the additional temporal dimension just would add impossible memory requirements. So tired of this wave of blackbox brute-force "programming" approaches these last years

@eniko My favorite part about the Minecraft one is that they were crowing so hard about it being "playable". But "I can technically input controls and get a response" has never been what "playable" means in gaming. If a game from a normal publisher accidentally forgot the whole environment and erased all your work when you turned around, literally every review for it would have the word "unplayable" in it.

@eniko I'm surprised that AI doesnt hallucinate twitch chat with random nonsense, considering if was probably trained on twitch VODs.

@eniko what makes the whole "AI will always be bad" much worse is the environmental effect it is having, just to provide something that approximates an approximation poorly based on stolen data

@PixelPodium @eniko It's fine, they just need to screw up once in the right way, like crowdstrike.

But considering it is AI, they will screw up a lot in smaller ways until they add up.