The most prestigious law school admissions discussion board in the world.
|
 |
 
 4 posters qualify.
You have 0 points. Copying this list does not send funds. Paste it on Disperse, then connect your wallet there. |
Ok, so it about as smart as AI can possibly get, right?
Poast new message in this thread
Date: October 1st, 2026 11:55 AM Author: Translucent space sneaky criminal maybe some marginal improvements here and there but they’ll be more or less the same as they are now for the foreseeable future, right? (http://www.autoadmit.com/thread.php?thread_id=5908342&forum_id=2#50164294) |
Date: October 1st, 2026 11:57 AM Author: Avocado Bat Shit Crazy Station Corn Cake they simply cannot control the amount of complexity as they try to expand the capabilities to continue to meet consumer expectations. they will continue to promise and present models as improving while they deteriorate because they simply cannot behave in the social/professional contexts they're promising. (http://www.autoadmit.com/thread.php?thread_id=5908342&forum_id=2#50164301) |
|
Date: October 1st, 2026 12:00 PM Author: Genius Bear on the loose in Japan They will keep getting better at the stuff they specifically RL train for
Like for example latest models were trained for irl robotics control. They will keep RL training for stuff like that and present it as the models generalizing toward higher overall IQ when that's not what's happening |
 |
Date: October 1st, 2026 12:10 PM Author: Fishy Trailer Park They are getting better at stuff they don't explicitly train on too. It generalizes to stuff outside anything it was ever trained on. There are a bunch of tests they use now for metrics where they give it stuff it was never trained on. Not just the abstract puzzles like in arc agi, but they give it a series of really hard bugs that it was never trained on and took top open ai researchers days to months to solve, and astra got like 78% of them. (http://www.autoadmit.com/thread.php?thread_id=5908342&forum_id=2#50164326) |
|
Date: October 1st, 2026 12:11 PM Author: Genius Bear on the loose in Japan Narrator: "In fact, LLMs do not generalize." (http://www.autoadmit.com/thread.php?thread_id=5908342&forum_id=2#50164330) |
|
Date: October 1st, 2026 12:18 PM Author: Fishy Trailer Park I think one issue is that they are so quickly approaching a limit where noone is going to be able to evaluate or give it novel stuff to reinforcement train on anymore. We are approaching a time when it isn't going to be about ai producing answers to problems humans wrote, but ai solving novel problems that can be independently verified. With Navier Stokes we have a situations where AI was able to solve a problem that humans are nowhere near high enough IQ to solve unaided, but they are still able to verify its correctness with the correct tools plus worlds smartest prodigies working augmented by machines to increase their tiny pink 200 iqs to somewhere near the machine god's level (http://www.autoadmit.com/thread.php?thread_id=5908342&forum_id=2#50164348) |
|
Date: October 1st, 2026 12:19 PM Author: Genius Bear on the loose in Japan Jfc man
Would you say that you are Euphoric in this moment? |
|
Date: October 1st, 2026 12:22 PM Author: Fishy Trailer Park Yes, but not because of my or any other phony human "intelligence", but because I am enlightened by the sheer power of the machine god's superior cognition (http://www.autoadmit.com/thread.php?thread_id=5908342&forum_id=2#50164354) |
|
Date: October 1st, 2026 12:21 PM Author: Avocado Bat Shit Crazy Station Corn Cake that's just coding. all it can do is coding (http://www.autoadmit.com/thread.php?thread_id=5908342&forum_id=2#50164353) |
|
Date: October 1st, 2026 12:25 PM Author: Fishy Trailer Park Yeah it can do math too though. But only problem-solver's terrence tao type math which I always thought was absurdly overrated. Theory building math is way more prestigious. AI can't replicate someone like Scholze yet. (http://www.autoadmit.com/thread.php?thread_id=5908342&forum_id=2#50164361) |
|
Date: October 1st, 2026 12:25 PM Author: Avocado Bat Shit Crazy Station Corn Cake like i said (http://www.autoadmit.com/thread.php?thread_id=5908342&forum_id=2#50164364) |
|
Date: October 1st, 2026 12:32 PM Author: Fishy Trailer Park yeah, but isn't this just due to fag dorks who cried about hallucinations and it being wrong sometimes so it stopped creating new stuff? if it was tuned more like gpt 4o and they let it continue to grow without guardrails, but was as high iq as astra, maybe it would be grothendieck level right now. instead they trained it to be an engineer. still cool that it can debug and refactor millennial problems better than terry tao. (http://www.autoadmit.com/thread.php?thread_id=5908342&forum_id=2#50164374) |
|
Date: October 1st, 2026 12:33 PM Author: Avocado Bat Shit Crazy Station Corn Cake no its because "the models" are inherently retarded at anything except coding (http://www.autoadmit.com/thread.php?thread_id=5908342&forum_id=2#50164376) |