Eventually, that becomes interesting. Before we get to that stage, I think we pass through (almost) everyone following almost the exact same advice as each other.

Such people are extremely predictable.

You may have already noticed, before LLMs, cliché? Talking points that make a group identifiable? Words and phrases that act as applause lights or cognitive stop signs? (That last sentence itself being a pair of clichés that you can use to identify where I hang out online).

Anyway, point is, LLMs will give us a memetic monoculture before they turn us all into a world of correctly personalised Whispering Earring wearers. That makes them predictable, that makes them exploitable. It'll be like playing chess against someone you know is using a specific version of Stockfish: even though it would beat you if you tried to fight the system unaided, you can win by asking your own copy of the AI to go one step further ahead, and it will be accurate precisely because it's playing against itself and reacting to its own moves.

(Of course, the fact I've said this in writing means this is in the training data; in the general case this means the LLMs will know that and account for that, but I suspect comments like this won't shift the needle all that much compared to the aggregate output of 3 billion people reacting to short-form emotional manipulation A/B slop)

▲

csh0 2 hours ago | parent | next [-]

This is already underway.

My dog often gets misidentified as a restricted breed. This used to make apartment hunting difficult because, occasionally, the property manager would visually ID the dog breed as banned, I’d have to go to the vet and get paperwork, potentially gene testing, arguing she wasn't, it was a whole thing.

But, recently, the apartment I moved into had an online portal where I had to upload a photo and it would identify the breed to determine if it was approved.

I correctly assumed the portal was using an LLM for this purpose. I wrote a script which submitted different photos of my dog to the major LLM providers until it found a photo which all the LLMs would identify as the correct breed.

I simply submitted that photo and, as expected, passed with flying colors.

	▲	ForOldHack an hour ago \| parent [-]
		I was anticipating that it would have reversed course and flagged you. Blessings on you, and your four footed friend.

▲

sebastiennight 2 hours ago | parent | prev [-]

> It'll be like playing chess against someone you know is using a specific version of Stockfish: even though it would beat you if you tried to fight the system unaided, you can win by asking your own copy of the AI to go one step further ahead, and it will be accurate precisely because it's playing against itself and reacting to its own moves.

I don't believe this is how chess works, and I don't believe this is how Stockfish works, and I don't believe this is how AI works.

Stockfish isn't winning because it's playing a better sequence of programmed steps, and having access to "the next version of Stockfish" doesn't mean it can "guess the next move" and play against that.

▲

ben_w 2 hours ago | parent [-]

I didn't say "the next version of Stockfish".

You have Stockfish version n, see board state s. I have Stockfish version n, see board state s. I want to know what you're about to do, so I put Stockfish into state s, ask it what the best move is, and I know you'll make that move because I know you'll ask Stockfish version n the same question of the same state. I now know board state s+1.

The steps are not pre-programmed, but the program itself is (modulo hardware imprecision) deterministic. If there's a RNG in there then sure, this doesn't work as easily as I wrote it; and there may be randomness in the thing that this is a metaphor for, regardless of if there's one in Stockfish or not, but that's not hard to work with when you want to win against an aggregate: we invented the field of statistics to deal with random numbers because they come up so often.

	▲	muvlon 26 minutes ago \| parent [-]
		There is deliberate randomness in stockfish. The easiest way to see this is from the fact that, when playing the white pieces, it won't play the same opening move every time. Often it's e5, but it also goes for e4 or Nf3 or something else entirely. This is by design, and very much necessary for a competitive chess engine. Otherwise, people could do basically what you say: Run an offline (as in, ahead of time, with ample compute resources) search against stockfish that finds a line where it loses, then make an engine that plays that every time. As a consequence, even if you know that your opponent is running stockfish, you can't really use that against them. Your best bet is also just running stockfish.