\
  The most prestigious law school admissions discussion board in the world.
BackRefresh Options Favorite

AI totally didn’t “solve” that math problem that broke a genius’s intuit

And showed that two subjects humans thought weren’t re...
Walnut passionate house dog poop
  05/26/26
yes
Boyish Idiotic Cruise Ship Tank
  05/26/26
It’s Pure Coincidence. See from the outside it just LO...
Walnut passionate house dog poop
  05/26/26
this is like showing a caveman a working tv and it thinking ...
Boyish Idiotic Cruise Ship Tank
  05/26/26
The caveman clutching the club thinks just because they can&...
Walnut passionate house dog poop
  05/26/26
...
jade goyim
  05/26/26
Not even people who make LLMs use the whole "it just pr...
bronze locus
  05/26/26
People at the big labs and top researchers might not, but to...
Walnut passionate house dog poop
  05/26/26
...
bronze locus
  05/26/26
...
jade goyim
  05/26/26


Poast new message in this thread



Reply Favorite

Date: May 26th, 2026 1:09 AM
Author: Walnut passionate house dog poop

And showed that two subjects humans thought weren’t related are actually just different languages for describing the same thing. See actually what happened was the machine jusf statistically predicted the next token in a way that happened to be correct. It totally has no “understanding” of the math.

(http://www.autoadmit.com/thread.php?thread_id=5869116&forum_id=2most#49901513)



Reply Favorite

Date: May 26th, 2026 1:13 AM
Author: Boyish Idiotic Cruise Ship Tank

yes

(http://www.autoadmit.com/thread.php?thread_id=5869116&forum_id=2most#49901516)



Reply Favorite

Date: May 26th, 2026 1:19 AM
Author: Walnut passionate house dog poop

It’s Pure Coincidence. See from the outside it just LOOKS like the machine converged on a genius answer but actually it was just a freak statistical accident. If you multiply probabilities across infinite time monkeys on a typewriter could probably do the same.

(http://www.autoadmit.com/thread.php?thread_id=5869116&forum_id=2most#49901522)



Reply Favorite

Date: May 26th, 2026 1:22 AM
Author: Boyish Idiotic Cruise Ship Tank

this is like showing a caveman a working tv and it thinking the tv is a real person or time traveller or some shit

(http://www.autoadmit.com/thread.php?thread_id=5869116&forum_id=2most#49901523)



Reply Favorite

Date: May 26th, 2026 1:32 AM
Author: Walnut passionate house dog poop

The caveman clutching the club thinks just because they can’t see the “person” inside the machine that it had to have come from statistical magic. The reality is that if “predicting the next token" results in a counterexample to an 80 year old conjecture that contradicts the intuition of Paul Erdos, then "token prediction" is functionally indistinguishable from "reasoning."

(http://www.autoadmit.com/thread.php?thread_id=5869116&forum_id=2most#49901537)



Reply Favorite

Date: May 26th, 2026 2:09 AM
Author: jade goyim



(http://www.autoadmit.com/thread.php?thread_id=5869116&forum_id=2most#49901569)



Reply Favorite

Date: May 26th, 2026 1:31 AM
Author: bronze locus

Not even people who make LLMs use the whole "it just predicts the next word" thing anymore

(http://www.autoadmit.com/thread.php?thread_id=5869116&forum_id=2most#49901535)



Reply Favorite

Date: May 26th, 2026 1:33 AM
Author: Walnut passionate house dog poop

People at the big labs and top researchers might not, but tons of software engineers and Reddit midwits still do. It isn’t even totally wrong at a certain level of description anyway just reductive in a way that strips away almost everything actually important

(http://www.autoadmit.com/thread.php?thread_id=5869116&forum_id=2most#49901541)



Reply Favorite

Date: May 26th, 2026 1:35 AM
Author: bronze locus



(http://www.autoadmit.com/thread.php?thread_id=5869116&forum_id=2most#49901545)



Reply Favorite

Date: May 26th, 2026 1:16 AM
Author: jade goyim



(http://www.autoadmit.com/thread.php?thread_id=5869116&forum_id=2most#49901521)