Toggle sidebar

Front-page Tempcast 

AI Recomping Controversy, PC Games on Switch and more (ft. Scarlet) - Tempcast #99

tempcast-169-logo-png.297330

​
If you expect a detailed discussion regarding the topics listed in the title you are going to be disappointed. This weeks episode is a vaguely threaded conversation about some of the prominent topics circulating in the past month or so, with some of our hosts forgetting that we don't record video and others showing off their absurd gaming collection, all neatly tied together by our special guest chief editor Scarlet. Join us on this purely improvisational episode as we try to make sense of what's going on.


Theme by VinsCool and AlanJohn​

The podcast is also available on Spotify!
 
  • Like
Reactions: raxadian, SylverReZ, bonkmaykr and 1 other person
Click to expand...
Dark_Ansem said:
The problem with AI decomps is that AI will hallucinate what does what, so at the end...
Click to expand...
for that sort of things, I am pretty sure the AI creates deterministic tools that will not hallucinate.
But I could be wrong? maybe someone who has done this before can confirm?
 
Costello said:
for that sort of things, I am pretty sure the AI creates deterministic tools that will not hallucinate.
But I could be wrong? maybe someone who has done this before can confirm?
Click to expand...

If your goal is just to create PC port, sure. The AI can do that. It will probably eventually create a functionally-equivalent but nonsensical codebase full of fakematches (things that compile to the same binary, but are not anywhere close to what was in the original code) and without naming stuff properly.

Said codebase is effectively useless to humans afterwards (and possibly also AI, lacking context from the previous runs), because trying to understand what it does requires you to repeat the reverse engineering effort.

I mean just look at this AI Decomp of Banjo Kazooie. Does any of this code actually tell you anything about how the game works? https://gitlab.com/banjo.decomp/banjo-kazooie/-/blob/master/src/core2/code_C4B0.c

Compare with a decomp done properly:
https://github.com/zeldaret/oot/blob/main/src/code/z_collision_check.c
https://github.com/SMGCommunity/Petari/blob/master/src/Game/AreaObj/CollisionArea.cpp
 
Last edited by skawo,
  • Like
Reactions: PalindromicBreadLoaf, lauty29, Chrono_Nightmare and 11 others
skawo said:
If your goal is just to create PC port, sure. The AI can do that. It will probably eventually create a functionally-equivalent but nonsensical codebase full of fakematches (things that compile to the same binary, but are not anywhere close to what was in the original code) and without naming stuff properly.

Said codebase is effectively useless to humans afterwards (and possibly also AI, lacking context from the previous runs), because trying to understand what it does requires you to repeat the reverse engineering effort.

I mean just look at this AI Decomp of Banjo Kazooie. Does any of this code actually tell you anything about how the game works? https://gitlab.com/banjo.decomp/banjo-kazooie/-/blob/master/src/core2/code_C4B0.c

Compare with a decomp done properly:
https://github.com/zeldaret/oot/blob/main/src/code/z_collision_check.c
https://github.com/SMGCommunity/Petari/blob/master/src/Game/AreaObj/CollisionArea.cpp
Click to expand...
And I thought I was writing messy code. Jesus Christ that is awful.
 
  • Like
Reactions: FateForWindows and bonkmaykr
skawo said:
I mean just look at this AI Decomp of Banjo Kazooie. Does any of this code actually tell you anything about how the game works? https://gitlab.com/banjo.decomp/banjo-kazooie/-/blob/master/src/core2/code_C4B0.c
Click to expand...
Not humanly... but I'm sure AI can "decrypt" the code.

If I were to write my code using StarFox Adventure's Dino Language, AI will easily get everything but to a human that doesn't know what "Saurian" is it'll be nonsense. There's absolutely no way that code was human written, and maybe is on purpose to effectively keep humans from reading it.
 
skawo said:
I mean just look at this AI Decomp of Banjo Kazooie. Does any of this code actually tell you anything about how the game works? https://gitlab.com/banjo.decomp/banjo-kazooie/-/blob/master/src/core2/code_C4B0.c
Click to expand...
Jesus. I can't even call this a decompilation. Most of the work reconstructing the code was not even done. This looks like someone just threw the entire Ghidra output at Claude and said "make it perfect💫 and robust🏔️ and super fast🚀".

And this is why people who exclusively rely on AI without a prior understanding of what to do without it are doomed to fail. I have yet to see a large-scale AI codebase that does not have code smells or major architectural fuck-ups. I don't mind if it's used conservatively by experienced programmers, this was obviously not.
 
  • Like
  • Love
Reactions: lightwo, FateForWindows, supershadow64ds and 1 other person
Costello said:
for that sort of things, I am pretty sure the AI creates deterministic tools that will not hallucinate.
But I could be wrong? maybe someone who has done this before can confirm?
Click to expand...
It will generate whatever ”source” translates to the same machine code after compilation which may or may not (likely not) match the original source due to how compilers work (you can code something 8 different ways and once the compiler is done with it you end up with the same binary anyway, you just used a different convention or approach). This has no impact on the end result since the end result is a binary.

EDIT: Nevermind, somebody already responded in a more verbose way. :lol:
 
  • Like
Reactions: FateForWindows and bonkmaykr
Whatever degenerative AI is doing today, it will be doing 1000x better two years from now. This type of work is perfect for AI because it is tedious, technical, and does not require creativity. And even today, whatever spaghetti-code that was created in that AI Decomp of Banjo Kazooie can be cleaned up and commented by other more powerful AIs today. Bad AI examples should not reinfornce a believe that all AI is bad and will continue to be. I fcking hate what these companies are doing to us, the absolute corruption involved, and how there are no meaningful safeguards being put into place to protect humanity from the tech. But it is amazing what it can do when trained properly and utilized responsibly.

Just to be clear, some of the very worst moments over the next 20 years for large swaths of humans will be a result of advances in AI. All the other threats I have seen come and go in my 50+ annual trips around the Sun are not even worth mentioning by comparison. AI is being added to every type of large scale weapon. It's is able today to reach hospitals, water treatment plants, military bases and everywhere else were humans are dumb enough to connect critical infrastructure to the whole rest of the world. But unlike Russian or North Korean ransomeware clowns, AI is a million times faster, more adaptable, and being used by all states right now in preperation for attacks and counter attacks. But also by criminals gangs and terrorists.
 
  • Haha
  • Like
  • Wow
Reactions: zfreeman, Dark_Ansem, Lop1 and 2 others
  • Haha
  • Like
Reactions: Lop1, FateForWindows, supershadow64ds and 3 others
Jayro said:
I don't think there should be any controversy, AI has accelerated recomps and porting games to the MiSTer FPGA at an alarming rate. I'm excited to see what else comes down the decomp/recomp pipeline.
Click to expand...

I'd rather have lower RAM and SSD prices myself.
 
Jayro said:
Sorry buttercup, but we can't put pandora back in the box. We need ai models to shift over to HBM modules so they leave PC RAM alone, and then RAM will be cheap again.
Click to expand...
Memory companies have all the incentive in the world to build out production capacity, the bigger problem here is that the memory market is cyclical and, up until 5 minutes ago in historical terms, sales were high whenever standards changed (in-between DDR/GDDR versions) and moderate when they stayed the same. Now all memory that’s manufactured is either already spoken for or sold instantly which is *unusual* for that market. This causes a cooling effect in regards to otherwise obvious and surefire investment as there’s no foreseeable timeline for how long the market will need to stabilise (if it will stabilise at all - this may be the new normal for all we know and the usual cyclical pattern will flatten out long-term). On one hand the more you produce the more you sell, on the other you don’t want to build new fabs that may end up being dormant in the future. They’re cautious to avoid over expansion which causes shortages. Market testing phases like this are a pain in the ass for all parties involved. I assure you, if memory companies had the capacity to make *more money* without the prospect of long term loss, they’d jump right to it.
 

Site & Scene News

New Hot Discussed User Submitted

Go to forum More news

Popular threads in this forum