Oyster Sauce wrote: ↑
March 19th, 2024, 02:00
Cheats regularly advance and become more difficult to detect, guys watching replays to find evidence of cheating have peaked and will never get better.
Who has peaked, the interns looking at the evidence they were sent? You don't know what you're talking about. I mentioned RTS, there's no one good who never watched their own replays to learn from them. if you check it and find e.g a cheater clicking on units
outside the fog of war which is impossible without cheats, that is a dead giveaway. that's scouring a replay hard to find evidence, to you, I guess, clear sign they'll never improve.
There were more obvious cheats than maphack, too. There are two layers of detectability, it doesn't matter if something is undetected or undetectable when most cheaters are stupid. In any case, you had an offline record of the online game that took place and that could serve as evidence of cheating if need be (as opposed to dev logging everything you do) and it could be given to a person to look at it. Much better than the accepted malware and spyware of today, I'll take a few cheaters over that any day.
Human intuition is pretty good, much more trustworthy and far more reliable than anything based on neural nets—like LLMs—because they have fundamental flaws that have been known for devades and cannot be solved and are the cause of many of their issues, e.g.: if A equals B, then B equals A; obvious to a human, yet anything based on neural nets can not and will not ever comprehend it.
You really think LLMs that can't read the time on a watch with 100% reliability are the way forward with something way more complicated? If "technology is the way forward," then it's certainly not in neural nets, deep learning, or LLMs, what nonsense, and it's obvious to many who work on AI software. Unless you just think openAI is 7 trillion short.
Rand wrote: ↑
March 19th, 2024, 02:11
That's why AI anti-cheat has me feeling hopeful. AIs can already see shit humans at their best can't, and only improve with more data and examples.
Like I said before, what makes me happy is the idea that they can't even get away with new accounts after a ban, since the AI will quickly recognize them from their idiosyncratic playstyle, even with hacks.
More data and scaling hasn't improved the core problems one bit, why do you think LLMs never ended up replacing radiologists? Know what else they do? They see shit humans at their worst can't and make mistakes no human ever would.
More data and scaling has brought us little other than multimodal hallucinations and spam, not to mention automated propaganda since the main interest in them is state interests. Everyone who actually works on AI software knows it's nothing but hype at this point. When reliability is lost from telling the time of statistically less common positions on a clock, how will LLMs ever be useful at detecting cheaters? It's pure fantasy because no matter how much data you throw at them, it's still next-token prediction; no understanding. It won't ever be useful for anticheat, it would be awful, not to mention the cost of training and running one for anticheat.
Just hire some real employees. low tech is the way to go. Like others pointed out, if they can get people to police their games for "hate speech" there is no reason they couldn't do better to stop cheating, but they want an easy, passive solution that avoids the hard work of human input.