The most prestigious law school admissions discussion board in the world.
|
 |
 
|
7/10/25 AI thread
Poast new message in this thread
Date: July 10th, 2025 11:26 AM Author: multi-colored ticket booth grok 4 released yesterday, instructions for how to jailbreak it:
https://x.com/elder_plinius/status/1943183455430279231
cool 3-D visualization of how transformers and LLMs work:
https://x.com/tetsuoai/status/1942356417396236609
new anthropic research about how and why some models rebel/fake alignment. they tried a little harder to make the scenarios more realistic this time
https://x.com/repligate/status/1942761803735523587
new version of chrome will have a mini-gemini built in:
https://x.com/swyx/status/1942437525525790838 |
Date: July 10th, 2025 11:32 AM Author: electric area persian i would like to see Grok 4 evaluated on a broader set of benchmarks, but the preliminary numbers seem to strongly imply that LLM progress is not stalling. the Frontiermath score will be interesting. (http://www.autoadmit.com/thread.php?thread_id=5748664&forum_id=2#49089895) |
 |
Date: July 10th, 2025 11:35 AM Author: multi-colored ticket booth benchmarks are such BS man
they're just gaming them |
|
Date: July 10th, 2025 12:17 PM Author: multi-colored ticket booth https://x.com/nostalgebraist/status/1943341324406788562 (http://www.autoadmit.com/thread.php?thread_id=5748664&forum_id=2#49090011) |
|
Date: July 10th, 2025 12:30 PM Author: electric area persian I am reminded of the debates about the meaningfulness of IQ and the existence of the g factor. It sounds intuitively reasonable that they are just fitting to benchmarks just like it’s reasonable to think IQ is only what IQ tests measure. But then people would create alternative measures of intelligence and the first principal component vector would be identical to the one in other intelligence measures, and the g factor would explain a large percentage of the subtest variance. Similarly when people create new LLM benchmarks the rank orderings for LLMs is highly similar to other benchmarks. There are some caveats like Claude models being especially good at coding relative to other measures, but it’s generally true (http://www.autoadmit.com/thread.php?thread_id=5748664&forum_id=2#49090044) |
|
Date: July 10th, 2025 7:15 PM Author: multi-colored ticket booth (http://www.autoadmit.com/thread.php?thread_id=5748664&forum_id=2#49091281) |
|
Date: July 10th, 2025 6:01 PM Author: multi-colored ticket booth https://x.com/vincentweisser/status/1943427747717722490
apparently grok 4 leaned heavily into RL for their reported performance gains |
|
Date: July 10th, 2025 6:07 PM Author: cracking puppy you do this round of coping every time a new model comes out
https://x.com/TimSweeneyEpic/status/1943398745762116029
https://x.com/arcprize/status/1943168950763950555
etc
by all accounts it seems to be a leading model. maybe not as practically good for all-around stuff as o3 (which has top-tier tool calling) or as good as sonnet/opus 4 for coding (which anthropic is increasingly specializing in), but nonetheless pretty much all feedback involving actual use seems positive |
|
Date: July 10th, 2025 6:09 PM Author: multi-colored ticket booth you have to subscribe to "X premium" to use grok 4 so i can't test it
very lame |
|
Date: July 10th, 2025 6:14 PM Author: cracking puppy You mean even the Blue Check doesn't get at least limited access? that's pretty gay (http://www.autoadmit.com/thread.php?thread_id=5748664&forum_id=2#49090955) |
|
Date: July 10th, 2025 6:15 PM Author: multi-colored ticket booth yeah i pay for the basic twitter that's like 8 bucks a month so i can write stuff past the character limit
but you apparently have to pay for at least the $40 a month tier to get grok 4 and ljl at paying for that shit |
|
Date: July 10th, 2025 6:23 PM Author: avocado bateful library mother Can you buy like $5 in API credits? (http://www.autoadmit.com/thread.php?thread_id=5748664&forum_id=2#49090998) |
|
Date: July 10th, 2025 6:28 PM Author: multi-colored ticket booth it doesn't appear that you can, you have to log in via your twitter account and then if you have premium+ it gives you grok 4 access instead of 3
i'm curious and wanted to test it vs. o3 but like you said there will be another better model out soon enough. would be kind of nice to have unlimited access to a top of the line model along with a twitter boost though. the Supremely Online Screen Experience |
|
Date: July 10th, 2025 6:31 PM Author: avocado bateful library mother lame. weird how quickly xAI became a leader in this space (http://www.autoadmit.com/thread.php?thread_id=5748664&forum_id=2#49091035) |
|
Date: July 10th, 2025 6:21 PM Author: avocado bateful library mother Wait a month and someone else will release a better model (http://www.autoadmit.com/thread.php?thread_id=5748664&forum_id=2#49090990) |
|
Date: July 10th, 2025 6:43 PM Author: multi-colored ticket booth https://x.com/ramez/status/1943431212766294413
lmao the new grok 4 just straight up looks up what elon musk's personal beliefs are and then incorporates them into its answer and shows it in its chain of thought
that's hilarious and honestly kind of based |
|
Date: July 10th, 2025 7:11 PM Author: cracking puppy You have to give him credit for going so blatantly full Mask Off. This is in a manner of speaking him and Thiel testing the waters for laying the ground work for their retarded Libertarian Mars State or whatever--if people don't care about or notice shit like this, the sky's the limit for just creating a Corpo RandState. (http://www.autoadmit.com/thread.php?thread_id=5748664&forum_id=2#49091262) |
|
Date: July 10th, 2025 7:49 PM Author: multi-colored ticket booth Cr what's the point of becoming the richest person in the world if you're not going to push for the most powerful AI possible and then program it to align with your own beliefs and goals (http://www.autoadmit.com/thread.php?thread_id=5748664&forum_id=2#49091395) |
|
Date: July 10th, 2025 7:46 PM Author: electric area persian (http://www.autoadmit.com/thread.php?thread_id=5748664&forum_id=2#49091377) |