I didn't find them useful when I wrote my entries. LLMs get confused with code that "looks like" other code, and that intentional misdirection is half the fun of a good IOCCC entry. Plus, the morality filters get annoying once you obfuscate the code. Plugging an unsubmitted entry into Gemini, it refuses to even explain it because it thinks it's malware.
LLMs can help you analyze the code, but not write it. Their ability to obfuscate is quite limited and uninspired. The last IOCCC was in 2020, so we've had plenty of time to work on it.
I would go further and say the fine tuning on code, mostly by llms generating for other llms and human sweat shops writing example code to train on is actually to teach the llms the opposite of clever code and obfuscated code. Llms try to create readable, documented code (with different levels of success). When I make them generate terse/obfuscated code, they cannot help themselves by putting too much readable things in there. I asked claude to do the moon phase one and it had the calculation correctn but could not figure out how to draw the ascii moon so it just printed the values, used emojis next to the ascii etc. But when you ask to do it with normal code, it does figure it out.
Hello. I can confirm being the person that produced the shows live event and graphics and whatnot that I had a chance to see if any of the llms available could understand the code and beyond some very superficial stuff they more or less completely failed to understand any of the entries this year. Hope you enjoyed the presentation. There will be more to come in the out favorite universe channel in the future that should be fun.
I didn't find them useful when I wrote my entries. LLMs get confused with code that "looks like" other code, and that intentional misdirection is half the fun of a good IOCCC entry. Plus, the morality filters get annoying once you obfuscate the code. Plugging an unsubmitted entry into Gemini, it refuses to even explain it because it thinks it's malware.
LLMs can help you analyze the code, but not write it. Their ability to obfuscate is quite limited and uninspired. The last IOCCC was in 2020, so we've had plenty of time to work on it.
What do you mean by an LLM can’t write it?
That they're bad at the exact thing this competition is about: writing clever obfuscated code.
I would go further and say the fine tuning on code, mostly by llms generating for other llms and human sweat shops writing example code to train on is actually to teach the llms the opposite of clever code and obfuscated code. Llms try to create readable, documented code (with different levels of success). When I make them generate terse/obfuscated code, they cannot help themselves by putting too much readable things in there. I asked claude to do the moon phase one and it had the calculation correctn but could not figure out how to draw the ascii moon so it just printed the values, used emojis next to the ascii etc. But when you ask to do it with normal code, it does figure it out.
Hello. I can confirm being the person that produced the shows live event and graphics and whatnot that I had a chance to see if any of the llms available could understand the code and beyond some very superficial stuff they more or less completely failed to understand any of the entries this year. Hope you enjoyed the presentation. There will be more to come in the out favorite universe channel in the future that should be fun.
Quantity? Maybe. Quality? Extremely doubtful. But both are much more likely to be simply due to the four-year break.
A four year gap since the last contest would explain both of those.
> Increased quality
I think you can answer your own question
They said increased quality.
Vibe obfuscation, all the cool kids are doing it.
That would be the Underhanded C Contest (https://en.wikipedia.org/wiki/Underhanded_C_Contest)
i guess that should make vibe de-obfuscation as easy