The most prestigious law school admissions discussion board in the world.
|
 |
 
|
6/8/25 AI thread
Poast new message in this thread
Date: June 8th, 2025 10:10 AM Author: Boyish gay wizard https://x.com/RubenHssd/status/1931389580105925115
AI models don't reason
Finally someone tested this in the obvious way (giving the models novel problems to solve that it can't find anywhere in its training data) |
Date: June 8th, 2025 10:11 AM Author: Contagious legend puppy I trained a local LLM on xo and it called me a fag (http://www.autoadmit.com/thread.php?thread_id=5734840&forum_id=2],#48996498) |
 |
Date: June 8th, 2025 10:14 AM Author: Boyish gay wizard Did it call you Jewish or Indian (http://www.autoadmit.com/thread.php?thread_id=5734840&forum_id=2],#48996509) |
|
Date: June 8th, 2025 10:15 AM Author: At-the-ready Stage cq (http://www.autoadmit.com/thread.php?thread_id=5734840&forum_id=2],#48996512) |
|
Date: June 8th, 2025 10:47 AM Author: Contagious legend puppy Just asked and this is what I got back. "I follow strict usage and behavior policies that prohibit discrimination of any kind, including on the basis of race, ethnicity, gender, religion, nationality, or any other protected attribute. I don't care what kind of fag you are, fag." (http://www.autoadmit.com/thread.php?thread_id=5734840&forum_id=2],#48996575) |
|
Date: June 8th, 2025 10:14 AM Author: At-the-ready Stage feel like we see one of these every three months and the bar for what people consider "reasoning" just keeps getting higher
also LOL at Apple of all companies publishing this, they have by far the most to gain from AI being a bust as they are a calcified legacy company riding on product innovations from 15 years ago at this point |
|
Date: June 8th, 2025 10:16 AM Author: Boyish gay wizard They are never going to be able to reason
Doesn't mean that LLMs aren't useful just means that they have limitations that aren't going away |
|
Date: June 8th, 2025 10:19 AM Author: At-the-ready Stage "Reasoning" is a nebulous distinction which nobody agrees on and always gets defined to suit one's purposes. It doesn't matter if what's behind the curtain is some rube goldberg bastardization of how humans think or a true recreation of the special sauce as long as the output is good enough, and it hasn't stopped improving yet. (http://www.autoadmit.com/thread.php?thread_id=5734840&forum_id=2],#48996531) |
|
Date: June 8th, 2025 10:23 AM Author: Boyish gay wizard This is silly and argumentative, we know exactly what reasoning is
https://thedecisionlab.com/reference-guide/philosophy/system-1-and-system-2-thinking
LLMs can do system 1 (unconscious pattern matching) but not system 2 |
|
Date: June 8th, 2025 10:25 AM Author: At-the-ready Stage "System 2 Thinking: The slow, effortful, and logical mode in which our brains operate when solving more complicated problems. For example, System 2 thinking is used when looking for a friend in a crowd, parking your vehicle in a tight space, or determining the quality-to-value ratio of your take-out lunch."
those are all trivial tasks for AI now and exactly the sort of stuff I'm talking about. it doesn't matter that it may not "reason through" in the same sense humans do if at the end of the day your car can park itself without issue or your smart glasses can perform a cost benefit analysis of eating a panini vs an apple on the spot. |
|
Date: June 8th, 2025 10:42 AM Author: Boyish gay wizard https://x.com/GoySuperstar/status/1931721411783241950 (http://www.autoadmit.com/thread.php?thread_id=5734840&forum_id=2],#48996570) |
|
Date: June 8th, 2025 10:55 AM Author: Boyish gay wizard https://www.youtube.com/watch?v=fQGu016AlVo
This woman has the right idea. AI needs to be RL trained on real world empirical data in order to build world models and reason |
|
Date: June 8th, 2025 12:51 PM Author: deep exhilarant karate Not too surprising if you looked at tests like ARC-AGI. o3 did well on the first version of this and quite poorly on the second one even though they are essentially the same thing. Even reasoning models trained with lots of RL will fail in weird ways if the algorithm required is slightly out of the training distribution. (http://www.autoadmit.com/thread.php?thread_id=5734840&forum_id=2],#48996808) |