@Demruth

I released Antichamber a long time ago. I'm currently deep down a rabbit hole exploring something else.

antichamber-game.com
Joined September 2010
  • Tweets36.5K
  • Following76
  • Followers14.2K
836Media
Tried to get AI to help me fix a bug that stepped beyond a one line change of a thing I missed. It's suggestions were madness, and caused clear regressions when playing the game. It's good at figuring out a trace that led to a problem. It's bad at knowing how to fix it properly
2
1
14
1,410
more replies
Every time people say shit like "You're not SUPPOSED to READ the code it writes!", for things where you obviously know the difference between right and wrong, what the hell do they expect is happening when you're not looking at it yourself?
1
7
270
Reading and fixing code in a system that already works is supposed to be the easier case. I understand this response if YOU didn't know what you were doing, but if you DID know what you were doing, you can see where the AI had no idea, regardless of how confident it sounded.
1
5
263
Even just getting AI to read through a codebase and asking questions about it is a frustrating experience. It loves telling me hypothetical answers and what-ifs, and I'm like, I don't fucking care about ANY of that. I want ACTUAL ANSWERS, GROUNDED IN FACT, ABOUT REAL CODE.
6
17
1,664
Every time AI adds additional details for the sake of being "helpful", it makes it sound like you need to actually do something about it, and genuinely becomes you wasting your time parsing a bunch of extra noise.
1
5
682
The most frustrating, time wasting type of information is "You should make these types of changes", and then mid change while you're questioning wtf it told you, it'll go "oh sorry my bad no you shouldn't change those sorry, not for your actual use cases. What you had is fine"
5
371
As a test, I gave Claude access to a project folder, so that I could try asking questions about the project. Before doing anything, I wrote a very clear, strongly worded Claude.md, saying it was for research only, YOU ARE NEVER ALLOWED TO EDIT THE PROJECT kind of instructions
1
9
2,483
more replies
After the memory and additional notes added to Claude.md that it didn't matter how many times I asked, or how nicely I asked, or how much I said I wanted it, it should always repeat instructions and never do it... it finally wouldn't do any actual edit / add / delete I asked for.
1
3
685
I will still never absolutely trust it to do what it says it does, and this is part of diligence towards AI responses. The same "question everything" I'd apply to people, including myself lol.
2
4
681
Load more