Does it matter what year I used initially? It was a quick guess, and the context should have made that clear. I always go with 2023 because this is the year most people decided LLMs were dumb (rightfully so) and didn’t try them since.

Honestly, last year was still early, it had just begun to start thinking. LLMs only really got good ~6 months ago. With the current frontier though, not happening.

In 2023 maybe. These days, it’s not happening, terminal use is baked into their training, they do not search for it. The training quality is high enough that there’s no chance this goes through either. In reality, this is just hurting the beginners that will actually believe it.

Perhaps it could trick eg. google search’s builtin one though, but I don’t think anyone trusts it anyway.

That doesn’t change the fact that there used to be a cheater every ~10 games, now I don’t think I’ve seen a single one since Vanguard. It doesn’t have to catch 100% of cheaters, if it went from catching 10% to 99.99% that’s a massive win.

What do you mean I shouldn’t have plugged Claude into the prod database?

What do you mean it doesn’t work? As much as I hate having to keep a windows partition for the occasional game of league, Vanguard has made a huge difference in Masters+ games.

I use it all the time as a kind of brainstorming tool.

“I want to do X (and details), can you tell me what tools or algorithms are available to me? List their pros and cons and give me some comparison”

Or on a somewhat recent project of mine, I has to effectively stub an entire library (but didn’t have to be done well), so I just told AI “take this page (the docs) and generate empty function stubs from it”. It doesn’t need to be high quality since it doesn’t run, it’s just to fool the dotnet engine.

I also tend to ask AI to add logging to some functions, since it’s annoying to do and impossible to mess up.