The most prestigious law school admissions discussion board in the world.
|
 |
 
|
7/11/25 AI thread
Poast new message in this thread
Date: July 11th, 2025 1:36 PM Author: Magenta School Cafeteria scholarship wants me to do these daily again
https://x.com/elder_plinius/status/1943171871400194231
system prompt for new grok 4. it has the same line about being allowed to say politically incorrect things, which supports my suspicion that the will stancil-raping version of grok the other day had more tweaks than just this
it appears to now be the strongest AI currently available. they spent a huge amount of compute resources on post-pretraining RL compared to all the other models. this is probably why it's performing so well on reasoning and problem-solving benchmark tests, because RL training helps a lot with this |
 |
Date: July 11th, 2025 2:53 PM Author: Cerebral Multi-billionaire Curious how this performs on law stuff. I’ve been maining gemini lately and have used grok the least (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49093630) |
Date: July 11th, 2025 2:50 PM Author: Magenta School Cafeteria https://x.com/MechanizeWork/status/1943726855015841805
https://www.mechanize.work/blog/sweatshop-data-is-over/
move over, teachers getting paid 100k to try to "teach" nigger kids how to write their name and tie their shoes
hello, RL engineers getting paid 300k to teach AI how to write its name and tie its shoes |
|
Date: July 11th, 2025 2:50 PM Author: Cerebral Multi-billionaire (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49093622) |
|
Date: July 11th, 2025 2:54 PM Author: Tantric hall i think i am going to stick to Opus, o3 and 2.5 pro. the differences look pretty marginal and i have zero faith xAI won't use my data for training. (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49093634) |
|
Date: July 11th, 2025 2:57 PM Author: Magenta School Cafeteria they're definitely all using our data for training
that's just Part Of The Deal |
|
Date: July 11th, 2025 2:59 PM Author: Crimson abode windowlicker (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49093650) |
|
Date: July 11th, 2025 3:01 PM Author: Tantric hall Google does if you use aistudio but i buy API credits. Same with Anthropic. i highly doubt they are lying about not training on API users if you opt out. i have less faith in OpenAI but i'm also not too concerned.
an Elon company? no way am i trusting them unless the model is way better. |
|
Date: July 11th, 2025 3:02 PM Author: Bright vigorous temple I trained a local llm on XO and it called me a fag (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49093659) |
|
Date: July 11th, 2025 3:03 PM Author: motley plaza (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49093670) |
|
Date: July 11th, 2025 3:12 PM Author: Crimson abode windowlicker Needs tweaking until it calls you a nigger (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49093697) |
|
Date: July 11th, 2025 3:10 PM Author: learning disabled area "grok 4, program grok 5 and don't make any mistakes" (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49093695) |
|
Date: July 11th, 2025 3:14 PM Author: Magenta School Cafeteria this was literally what the "AI 2027" Doomsday Report was btw (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49093701) |
|
Date: July 11th, 2025 4:40 PM Author: learning disabled area (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49093973) |
|
Date: July 11th, 2025 5:44 PM Author: Magenta School Cafeteria https://x.com/keyonV/status/1943730495264584079
https://arxiv.org/pdf/2507.06952
paper that demonstrates that while LLMs excel at predictive tasks that fall within their training data, they can't generalize that predictive ability into a complete and accurate world model to make correct predictions on tasks that weren't within their training
lecun is right imo. statistical inferences do not lead to the ability to make generalized inferences |
|
Date: July 11th, 2025 5:49 PM Author: Magenta School Cafeteria https://archive.ph/cYJS5
good WSJ normie article about this |
|
Date: July 11th, 2025 6:07 PM Author: boyish bearded mental disorder (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49094310) |
|
Date: July 11th, 2025 6:06 PM Author: Cerebral Multi-billionaire (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49094306) |
|
Date: July 11th, 2025 6:28 PM Author: costumed range (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49094378) |
|
Date: July 11th, 2025 6:48 PM Author: Tantric hall this is why i think the reinforcement learning approach of trying to paper over these problems with more data is wrongheaded. they are getting insufficient generalization from 30 trillion token datasets so they think they just need to use RL and chain of thought to make 300 trillion token datasets or whatever and their problems will be solved. the architectures and training methods should be fixed first. there's a lot of research showing inadequate generalization on toy tasks even with lots of compute.
https://arxiv.org/abs/2207.02098 |
|
Date: July 11th, 2025 7:05 PM Author: Magenta School Cafeteria yeah, i've become convinced that LLMs just fundamentally cannot out-of-distribution generalize. and i don't see how RL could ever solve it. all you're doing is training it on additional specific tasks that it....still can't generalize out of
i think companies are doing RL now as more of a marketing gimmick than anything else. people are saying that the latest grok 4 is only performing so well on benchmarks because its RL training had overlap with the benchmark tests. this is what i mean when i say that they're just gaming the benchmark tests
i think the next few years might end up being the big base models trained with RL into a bunch of different specialized sub-models that are used in larger multi-agent architectures in order to be more useful in practice in different specific contexts |
|
Date: July 11th, 2025 7:18 PM Author: Tantric hall right. i view RL as the way to get strong superhuman agents once you have the right base learning model. they should be able to use the existing data to train robust human level agents, but that's apparently not happening because the learning algorithm is inadequate.
that paper i linked seems to imply there are already models that consistently generalize better than transformers, so it's not clear we see companies flailing about for years using RL gimmicks with transformers. memory augmented transformers could be a near-term replacement for current models. if models don't improve substantially over the next year, there will be a strong motivation to try novel approaches like this. |
|
Date: July 11th, 2025 7:53 PM Author: Magenta School Cafeteria god i love AI. i can have it explain this paper to me and then immediately answer all my follow up questions to understand architectures that could actually potentially generalize. lol
this would take me like...i dunno, 2 days to do before AI. not even counting all the time and effort that AI has saved me to get to the point of being able to understand this in the first place |
|
Date: July 13th, 2025 1:24 PM Author: Magenta School Cafeteria https://x.com/karpathy/status/1944435412489171119 (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49097842) |
|
Date: July 11th, 2025 7:34 PM Author: Magenta School Cafeteria https://itcanthink.substack.com/p/what-are-robot-world-models
really good short article explaining how people think that we'll be able to use generative AI video to train robots to have functional world models
the notion of this actually working seems crazy to me but these guys are all very smart so it must be somewhat viable or someone would be calling it out as BS. it would cut robot training time and costs IMMENSELY and make them a lot more commercially viable and speed up robotics development by years |
|
Date: July 11th, 2025 6:19 PM Author: yapping aphrodisiac dysfunction laser beams This shit is way overrated for my legal practice (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49094362) |
|
Date: July 11th, 2025 8:32 PM Author: Magenta School Cafeteria https://x.com/NeonWhiteRabbit/status/1943541842655478255
looks like we found the real reason why grok 4 is so surprisingly good |
|
Date: July 11th, 2025 8:36 PM Author: ungodly cruise ship jap (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49094671) |
|
Date: July 11th, 2025 8:49 PM Author: trip stage dopamine https://www.youtube.com/watch?v=YkzlPcsSGlA (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49094712) |
|
Date: July 12th, 2025 1:33 AM Author: costumed range OpenAI got cucked again. They tried to acquire Windsurf but didn’t want Microsoft to have access to its IP. Now the acquisition is off and Google has hired their CEO.
https://x.com/ns123abc/status/1943806065524507007?s=46&t=YKr-jZOYUHE15Tew69wt4w
The Microsoft partnership was key to getting them off the ground, but now it’s an adversarial relationship. |
|
Date: July 12th, 2025 1:52 AM Author: Impertinent Fluffy Roast Beef Google will acquire everything in the end (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49095203) |
|
Date: July 12th, 2025 3:08 AM Author: Passionate rehab digit ratio Microsoft should put them out of business. The decision to partner with them rather than hiring away their engineers and scaling themselves was strange. (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49095224) |
|
Date: July 12th, 2025 8:51 AM Author: Magenta School Cafeteria Yeah they somehow tricked the greedy jeets at Microsoft up to this point but it seems like their luck has run out
I haven't believed in OpenAI since a lot of their top guys left a year ago. It's a really bad sign for an org when a bunch of your top level producers leave |
|
Date: July 12th, 2025 9:19 AM Author: costumed range I think so too, and Zuckerberg just poached a lot more people. Besides brand recognition from chatGPT, I don’t think they have any other advantages over the competition. (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49095411) |
|
Date: July 12th, 2025 1:33 PM Author: Tantric hall the failure to transition to a for-profit likely hurt them a lot. also increasingly clear there is no secret sauce and that training LLMs is not rocket science. (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49095861) |
|
Date: July 12th, 2025 9:01 AM Author: outnumbered very tactful background story parlour yes sir, it's all moving pretty fast rn (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49095396) |
|
Date: July 12th, 2025 9:05 AM Author: Magenta School Cafeteria https://x.com/Kimi_Moonshot/status/1943687594560332025
China just released a new 32 billion parameter model that is apparently the 2nd best model in the world behind o3 and it's all open source(!)
If China could whip up the infrastructure and hardware for enough inference compute to meet demand, I think they could snap up a bunch of the market share for AI because they could undercut the US labs by so much |
|
Date: July 12th, 2025 9:22 AM Author: Magenta School Cafeteria https://x.com/nrehiew_/status/1943694228804063614 (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49095415) |
|
Date: July 12th, 2025 1:37 PM Author: Impertinent Fluffy Roast Beef Would you actually use a Chinese LLM for work? (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49095865) |
|
Date: July 12th, 2025 1:39 PM Author: Tantric hall The local qwen models are quite good. Doubt i would want to send data to a chinese LLM though (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49095870) |
|
Date: July 12th, 2025 1:49 PM Author: Magenta School Cafeteria Yeah they are really good. I use qwen to mess around locally and it does really well
I don't trust the American LLM companies with my private data that actually matters either and I only work locally with it when I need to |
|
Date: July 12th, 2025 1:50 PM Author: Impertinent Fluffy Roast Beef If I installed on a PC is anything better than Qwen? How do you train it on daily news updates and shit if it's local? (http://www.autoadmit.com/thread.php?thread_id=5749144&forum_id=2#49095898) |
|
Date: July 12th, 2025 1:59 PM Author: Magenta School Cafeteria I think qwen is the best right now but there's a new Chinese one that I posted in this thread about that is apparently even better
There are ways to scrape web search info into your local model and incorporate it into your local LLM prompts but I haven't set it up yet because it hasn't really come up. It looks pretty easy to do though |