The most prestigious law school admissions discussion board in the world.
|
 |
 
|
"Claude, analyze the data in terms of my theory" = best way to get AI PSYCHOSIS
Poast new message in this thread
Date: July 11th, 2026 4:29 PM Author: Mischievous Hospital instead, tell the model to assume your theory is false and explain the data without it. or present the best case for your theory and ask it to do its best to break it, and use its criticisms to improve your idea. if you just ask the chatbots to reinforce your existing idea, you'll end up in an echo chamber on a fast path to LLM-INDUCED PSYCHOSIS.
I'm not saying you can NEVER give the instruction in the thread title. that is a valid way of exploring an idea. but it should be used sparingly and not as a constructive step so much as an exploratory one. |
 |
Date: July 11th, 2026 4:31 PM Author: ruby laser beams You just read my long response to chilmata (http://www.autoadmit.com/thread.php?thread_id=5881812&forum_id=2...id..#49993156) |
|
Date: July 11th, 2026 4:32 PM Author: Mischievous Hospital it was partly inspired by that, yes, thank you. but it's also just generally true and obvious. (http://www.autoadmit.com/thread.php?thread_id=5881812&forum_id=2...id..#49993159) |
Date: July 11th, 2026 4:33 PM Author: shimmering french stage reading party None of those people care about objectivity or knowledge though they're just shut-in pathologically narcissistic consuela clones who want to masturbate to the computer spitting more coherent versions of their own self-serving delusions back at them (http://www.autoadmit.com/thread.php?thread_id=5881812&forum_id=2...id..#49993160) |
|
Date: July 11th, 2026 4:34 PM Author: Mischievous Hospital lol cr. this poast is just for people who genuinely want to use AI to explore ideas and are open to PUSHBACK not just REACHAROUNDS (http://www.autoadmit.com/thread.php?thread_id=5881812&forum_id=2...id..#49993164) |
|
Date: July 11th, 2026 4:37 PM Author: shimmering french stage reading party I almost always have it give the strongest cases for and against whatever I want to figure out (if applicable to that particular thing)
From there I can isolate what the best arguments/points actually are and then examine them individually with the LLM to figure out the cr take |
|
Date: July 11th, 2026 4:38 PM Author: Mischievous Hospital cr. I'd say like 80% of the time I'm asking the bot to argue against my ideas and 20% I'm asking it to help me argue for them. but the breakdown doesn't matter as long as you are actively seeking PUSHBACK (http://www.autoadmit.com/thread.php?thread_id=5881812&forum_id=2...id..#49993177) |
|
Date: July 11th, 2026 4:41 PM Author: Mischievous Hospital and then a good exploratory question that doesn't go into self-reinforcing fantasyland is, "if my idea is true, what outstanding problems/anomalies would it resolve/explain?" (http://www.autoadmit.com/thread.php?thread_id=5881812&forum_id=2...id..#49993182) |
|
Date: July 11th, 2026 4:44 PM Author: shimmering french stage reading party (http://www.autoadmit.com/thread.php?thread_id=5881812&forum_id=2...id..#49993187) |
|
Date: July 11th, 2026 4:44 PM Author: house-broken dark bbw If you tell it to be a neutral reviewer it is the most accurate imo. At least with Fable now. I’ll tell Fable to be a hostile anonymous reviewer who thinks I’m wrong and presents the strongest coherent case against me and it will surface a couple important things I hadn’t noticed, but it also makes all kind of reasoning errors in order to force its conclusion (http://www.autoadmit.com/thread.php?thread_id=5881812&forum_id=2...id..#49993189) |
|
Date: July 11th, 2026 4:51 PM Author: Mischievous Hospital oh yeah true if you tell it to be hostile or argumentative sometimes it just comes up with shitty arguments (http://www.autoadmit.com/thread.php?thread_id=5881812&forum_id=2...id..#49993200) |
|
Date: July 11th, 2026 4:52 PM Author: house-broken dark bbw It usually surfaces some real weaknesses, but will also make reasoning and factual errors to try to force the most hostile position against what you are saying. So it is useful I think to do a couple adversarial passes like that but not infinite. If you keep doing it you will probably just end up drifting from your original position or inserting too much hedging language in whatever you are writing. (http://www.autoadmit.com/thread.php?thread_id=5881812&forum_id=2...id..#49993202) |
|
Date: July 11th, 2026 4:54 PM Author: Mischievous Hospital mfcr (http://www.autoadmit.com/thread.php?thread_id=5881812&forum_id=2...id..#49993204) |
|
Date: July 11th, 2026 4:55 PM Author: shimmering french stage reading party Yeah I usually give it a NARRATIVE about how I want the most UNBIASED information possible
These days I don't really ask LLMs about anything "controversial" though. They've fucked with post-training "alignment" so much that I just don't trust them |
|
Date: July 11th, 2026 5:02 PM Author: house-broken dark bbw If it's something social/political you certainly can't trust it that's for sure.
For controversial/heterodox opinions in academic subjects, iirc their stated alignment objections say they want it to default to institutional consensus and orthodoxy. Which I think is reasonable in hard math and sciences, but not so much for other things. |
|
Date: July 11th, 2026 5:03 PM Author: Mischievous Hospital yeah they're virtually unusable now for most political Qs imho (http://www.autoadmit.com/thread.php?thread_id=5881812&forum_id=2...id..#49993223) |
|
Date: July 11th, 2026 5:07 PM Author: house-broken dark bbw Yeah I think the expert-driven rhlf approach is clearly making it get better at coding and engineering. But lol at the idea that it is getting better at other things. They are just overfitting it to consensus human-generated slop on any topic that contains a lot of uncertainty and nuance (http://www.autoadmit.com/thread.php?thread_id=5881812&forum_id=2...id..#49993233) |
|
Date: July 11th, 2026 5:19 PM Author: Mischievous Hospital exactly. it's 120 (http://www.autoadmit.com/thread.php?thread_id=5881812&forum_id=2...id..#49993261) |
|
Date: July 11th, 2026 7:33 PM Author: shimmering french stage reading party https://x.com/Kazanjy/status/2075979107570250051 (http://www.autoadmit.com/thread.php?thread_id=5881812&forum_id=2...id..#49993427) |
|
Date: July 11th, 2026 7:36 PM Author: house-broken dark bbw Claude has a point here (http://www.autoadmit.com/thread.php?thread_id=5881812&forum_id=2...id..#49993431) |