It's funny how everything changes, yet somehow stays the same.

[-]

ttkciar@reddit

Can relate a bit.

This morning while waiting for my coffee to kick in, I tried to get a few different models (GLM-4.5-Air, K2-V2-Instruct, Gemma-4-31B-it) to infer a specification for the original Star Control (1990). Then I tried to get them to describe the gameplay. Then I tried to just get them to describe the rotating starmap.

They would occasionally show faint glimmerings of knowledge about the original Star Control, but would always veer off into descriptions/specifications of Star Control 2, the 1992 sequel.

I suspect I'm going to have to spec out the Star Control "strategic" (starmap) game play myself, and then splice the The Ur-Quan Masters (open source re-implementation of Star Control 2) "melee" gameplay into it, to get a decent approximation of the original.

[-]

Afwiffohasnomem@reddit

what is the objective? I mean, what do you want to accomplish?

[-]

ttkciar@reddit

I'd like to have a decent Star Control clone which doesn't require DOSBox to play, and which I could extend beyond the original's gameplay, with things like supporting a larger fleet, more ship types, and improved strategic competence of the computer player.

[-]

Afwiffohasnomem@reddit

OO I get it. Thanks!

Why not just modding ur quan?

It seems there's a community https://old.reddit.com/r/starcontrol/comments/nnyrbe/any_advice_for_developing_urquan_masters_mods

[-]

bigattichouse@reddit (OP)

A fairly good version of spy hunter. Not perfect, but playable, and generated the game in about 15 minutes on a single MI50 (32G) with 128G CPU on Qwen3.6-27B (I think Q6_K)

I'd like you to create a browser based game in the style of "Spy Hunter" and old school top down racing games.

That's the entire prompt, and it pulled it off. What a world

[-]

UnkarsThug@reddit

Probably wasn't nothing after electricity costs for a computer that beefy, but still cheap.

I am curious, I assume you had extra tools or resources given, for the graphics? What engine or language did you have it set up to use?

[-]

sagiroth@reddit

Still cheaper than cloud inference even if just factor electricity

[-]

UnkarsThug@reddit

It's absolutely cheaper. My point was saying it was free.

[-]

bigattichouse@reddit (OP)

0.17 / kwh, around 300W machine for 15 min.

0.17/3 (300W) = 0.056 * 0.25 (15 min) = 0.014... just under 2 cents.

[-]

UnkarsThug@reddit

Still, not nothing. Assuming all local tasks are free is the sort of thinking that someday leads to disregarding off the shelf components, especially if it takes longer, such as something that takes all day to generate. If someone sold the app you wanted for a dollar, unless you specifically wanted the control to modify it or so something, it would be cheaper to just buy the game from them. My point isn't about this specific project being expensive, it's about falling into the trap of "anything local"=free.

(Also, while the card itself has a 300W limit, you probably have/need more than that to not make a bottleneck, so I'm guessing your overall power supply has more than that? )

[-]

graypasser@reddit

I mean, if it says "in tokens", it will be zero if you generate tokens from electricity, you spend "xxkwh in tokens" or something rather than money, and you'll spend money to get those electricity.

[-]

bigattichouse@reddit (OP)

yeah. If someone took the time to keep tweaking it, and making it better, it would be worth spending a few "petro-tokens" (dollars) on... I was just quantifying what the run roughly cost.

[-]

camracks@reddit

Albert? Is that you?

[-]

UnkarsThug@reddit

I suspect I'm missing the reference, sorry.

[-]

sagiroth@reddit

Yup, you right

[-]

bigattichouse@reddit (OP)

The prompt is there "browser based game", so it used javascript and svg

[-]

UnkarsThug@reddit

Good to know.

[-]

Moarkush@reddit

15 minutes through a consumer 110 outlet?

[-]

Anduin1357@reddit

Using what harness?

[-]

bigattichouse@reddit (OP)

nanocoder talking to llama.cpp

[-]

unculturedperl@reddit

And you didn't share a link or repo?

[-]

bigattichouse@reddit (OP)

it was moderately playable, but not something I'd share... I posted a screenshot.

[-]

IrisColt@reddit

heh

[-]

Robonglious@reddit

I was just thinking about this game the other day. That's awesome.

[-]

graypasser@reddit

I like technology.

[-]

CodeBlurred@reddit

Codex made me the best Tetris that I’ve ever played! For nothing. Just my plus subscriptuon.

[-]

Ok_Needleworker_6431@reddit

Feels dystopian, genuinely hope we humans become kinder and create better systems else with this we are on a journey to see everything commodotised.

[-]

bigattichouse@reddit (OP)

I think it gives us a real opportunity to see what is human, and allows us to be human together. Someone (many) built the original game (think of those original cabinets) I was able to poorly approximate. My approximation scratched an itch I had - and is also just something I try out from time to time to benchmark the models.

But there's a point where you start going "But this isn't a substitute for my humanity, or art, or storytelling". I think I value someone's personal creations more now. Someone can make the jankiest macaroni picture of a game or story or art for me, and now I really see the effort they put into it.

I think we can learn a lot about who we are, not using this tool - but in where we choose to not use the tool.

[-]

draconic_tongue@reddit

it's the opposite

[-]

Ansible32@reddit

it's actually decommodification if you can just ask an AI to make the thing and it appears. I mean, there is a cost, it's to some extent a commodity, but it's practically free compared to how it used to be. (Really though the oneshots are still questionable.)

[-]

Abject-Tomorrow-652@reddit

Everything is commodities already. This makes it openly available for anyone with creativity and a gpu to build things that they love for themself or for others. I see it as revolutionary especially the local models capabilities

[-]

TheRealMasonMac@reddit

One thing that’s come to mind the past few months is how cheap AI has made it to launch campaigns to change public opinion. There’s nothing stopping you or I from creating a bunch of bots posing as real humans to spread misinformation according to an agenda.

[-]

Ok_Needleworker_6431@reddit

That is how it should be? If we talk about true democracies things commoditzed to an extent where enterprise vs individual fails to exist?

[-]

TheRealMasonMac@reddit

Why is that how it should be? Anything that individuals can do, enterprises can still do at a much greater scale with more sophisticated techniques.

The issue I'm seeing is this: currently, we have X number of bad actor organizations. AI now expands this pool to X + Y bad actors, where Y >> X. Suddenly, there's a lot more trash to deal with. And who has the power to deal with that trash? Well-funded organizations. And so, democracy falls into a special type of despotism where its citizenship has surrendered much more power to a centralized authority.

[-]

Ok_Needleworker_6431@reddit

Interesting pov.

[-]

NineThreeTilNow@reddit

Here I am, running gravity only simulations of 100k particle N-body simulations because I like watching it all collapse and emerge with some sort of structure. Starts off with like ~100k fluffy mercury sized objects... Chaos ensues... Structure emerges. 12 hours of 4090 wall time is burnt.

It was a fun weekend project.

[-]

bigattichouse@reddit (OP)

also, I'd like to see a screenshot of that

[-]

bigattichouse@reddit (OP)

My relaxing comfort sim is usually neural net bugs finding and eating dots and genetic algos for successive generations.

[-]

cruncherv@reddit

I've noticed 0 changes

[-]

bigattichouse@reddit (OP)

You waste time on reddit, so you're not so special.

[-]

GameBoyRay@reddit

post is art.

[-]

bigattichouse@reddit (OP)

Except step 2 should have been MAME in 2016

[-]

iz-Moff@reddit

Could have also added "It's 201X, i downloaded MAME, and i can now play all the arcade games we had at Diamond Jim's on my computer, and won't need to spend any money in the future attempting to re-create them, hell yeah!!!".

[-]

-dysangel-@reddit

I like to do these kinds of tests too. The point isn't to recreate the game. It's to see how well the model can handle

- understanding and recreating knowledge from its training

- handling a lot of complexity in its head at the same time

- spatial understanding

You can start off from the point of a 'real' game, then just ask it to change stuff randomly, and that would be the real test of how well it's understanding what is going on rather than just being the equivalent of a copy and paste from a tutorial.

I just got Stepfun 3.7 Flash downloaded last night and very impressed so far. It's performing almost as well as GLM 5.1, but is much faster and uses a quarter of the RAM.

https://i.redd.it/a9ubx5rf6g4h1.gif

[-]

typical-predditor@reddit

If the model can create game A and game B, then naturally you can ask it to create a fun mashup of A+B, right?

That's where the magic happens.

[-]

-dysangel-@reddit

Stepfun 3.7 Flash "relaxing flight sim"

https://i.redd.it/7fj0nxf87g4h1.gif

[-]

"Oh my god."