The most prestigious law school admissions discussion board in the world.
|
 |
 
|
8/11/25 AI thread
Poast new message in this thread
Date: August 11th, 2025 12:41 PM Author: diverse principal's office main people Sam Altman
@sama today we are significantly increasing rate limits for reasoning for chatgpt plus users, and all model-class limits will shortly be higher than they were before gpt-5.
we will also shortly make a UI change to indicate which model is working.
tomorrow or tuesday we expect to share our thinking on how we are going to make capacity tradeoffs over the coming months.
(e.g., chatgpt vs the api, existing users vs new ones, research vs product, etc)
https://x.com/sama/status/1954602880952115603
Women are spiraling into psychosis with their AI goon boyfriends
https://x.com/anammostarac/status/1954405062861815925
I mostly agree with this guy's analysis on the near future trajectory of commercial AI. Good post
https://x.com/GestaltU/status/1954561703967867019
Good post about how AI coding wrappers are inherently not economically viable. They are literally losing money. As the popular saying goes: "The models are the product"
https://x.com/nickbaumann_/status/1954253210287288430
These guys say they trained a model with pure reinforcement learning without any cross entropy which is pretty interesting
https://x.com/tokenbender/status/1954188530298900687 |
Date: August 11th, 2025 12:53 PM Author: Drunken house-broken shrine factory reset button the 4o/GPT-5 controversy is bizarre. i thought the sycophancy of the models was the result of bad RL training, but apparently there are large numbers of people who want their AI to be a totally uncritical kiss ass. extremely disturbing. (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49174945) |
 |
Date: August 11th, 2025 1:00 PM Author: diverse principal's office main people If by "large numbers" you mean 95% of the human population
My Conspiracy Theory is that open AI realized that they were in over their head and that part of the reason why they initially rolled all models into a router based on each user query is because they were trying to deliberately phase out the sycophancy as an option
Eventually someone is going to kill themself and/or others due to some drama with their imagined AI companion and whatever company was running that AI is going to take a massive PR hit that probably won't be able to be recovered from. These guys are smart and see the writing on the wall and are trying to avoid having that happen to them. But the user outcry and demand for the old sycophantic model was so extreme that open AI felt forced to re-enable it, for now |
|
Date: August 11th, 2025 1:04 PM Author: Drunken house-broken shrine factory reset button (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49175012) |
|
Date: August 11th, 2025 1:07 PM Author: Curious sanctuary Or the opposite, Meta gives every prole their own custom sycophant and it serves up advertising-laced slop (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49175018) |
|
Date: August 11th, 2025 1:14 PM Author: diverse principal's office main people This is definitely coming but the big companies are going to be very very careful about avoiding being the first one to get hit by the inevitable massive public backlash against parasocial relationships with AI (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49175045) |
|
Date: August 14th, 2025 7:13 PM Author: sooty provocative cumskin it has nothing to do with training and everything to do with compute resources: "thinking" models essentially talk to themselves, formulate a plan, employ RAG, and output the result
GPT-5 caps compute after a bit unless you enable "deep research" and RAG isn't enabled unless "web search" is enabled.
4o just talked to itself a lot and crafted a PC response for the faggots who use it for emotional support. GPT-5 is faster & better without the useless fluff.
|
|
Date: August 11th, 2025 1:18 PM Author: Balding Liquid Oxygen "Good post about how AI coding wrappers are inherently not economically viable. They are literally losing money. As the popular saying goes: "The models are the product""
the problem with dev tooling is you have to compete with developers themselves. back in the day things were slow/complicated enough that dev tooling was a big moat (e.g. microsoft+old visual studio), but that's just not the case anymore. OpenCode for instance is an OSS competitor to claude code that in many ways is already better due to rapid dev feedback. the post you linked was by the author of cline, another open source tool which directly competes with cursor and the like |
|
Date: August 11th, 2025 1:27 PM Author: diverse principal's office main people How do you counter the argument that companies like cursor are structurally not economically viable though. They are essentially trying to mark up inference costs that users can get cheaper via the frontier models directly
I agree that he's arguing this in his own self interest but I think his line of reasoning is sound |
|
Date: August 11th, 2025 1:46 PM Author: Balding Liquid Oxygen I'm agreeing with him. Cursor's business model is retarded and at the end of the day hinges on having the slickest integration as its raison d' etre over other tools-- what AI has already blown the doors off of is rapid iteration/improvement of UI. it used to be that OSS projects generally had the shittiest user experiences (e.g. using Visual Studio 15 years ago was light years ahead of crappy open source IDEs), that's no longer the case. (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49175143) |
|
Date: August 11th, 2025 1:54 PM Author: diverse principal's office main people Oh I misunderstood. It's crazy to me that there is so much VC money getting thrown at all this obviously unviable stuff. I really need to just go get some rich people to give me Monopoly money. Seems like a sucker's game to try to build things up myself (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49175169) |
|
Date: August 11th, 2025 1:57 PM Author: Balding Liquid Oxygen "It's crazy to me that there is so much VC money getting thrown at all this obviously unviable stuff"
Brother wait until you hear about this thing called "web3" |
|
Date: August 14th, 2025 7:17 PM Author: sooty provocative cumskin bro I'm not flame working on an insane start up. you should get in quick -- $1M could potentially become $100M in 5 years (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49185928) |
|
Date: August 14th, 2025 7:21 PM Author: diverse principal's office main people Are you twins esque flaming or serious (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49185940) |
|
Date: August 14th, 2025 7:24 PM Author: sooty provocative cumskin would my answer influence your decision to invest? (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49185951) |
|
Date: August 14th, 2025 7:31 PM Author: diverse principal's office main people I don't have 1mm to invest but I am working on ideas of my own
Just consulting with small businesses for now while I get my skills up. It's expensive to train anything significant so I really gotta make sure I'm ready first. You're probably looking at the same situation with whatever you're trying to do
There's a lot of money in models being post trained in industry-specific roles imo |
|
Date: August 14th, 2025 7:44 PM Author: sooty provocative cumskin Training is the biggest problem for me. I personally use lambda labs and fucking christ it's expensive to train anything worthwhile. That's why I focus on fine-tuning or RAG rather than training and pre-training. Even that is too expensive for large models. Using qwen3-Coder-435B-A35B, you need four H100s for inference at a respectable rate (~100tok/s)
I'm also buying an RTX Pro 6000 + ThreadRipper Pro 7985WX. Literally $17,000 in gear. |
|
Date: August 14th, 2025 7:51 PM Author: diverse principal's office main people yeah i think there's a lot of value in RAG for small businesses. every law firm should have their own RAG setup for example
i'm definitely getting redpilled on the need to use other people's money for business ventures |
|
Date: August 14th, 2025 8:06 PM Author: sooty provocative cumskin the problem with AI is that most people don't have a fucking clue what they're talking about. VCs are the worst.
i've literally been eating, breathing, and shitting this stuff since 2010. |
|
Date: August 14th, 2025 8:10 PM Author: puce bearded rigor wagecucks RAG is already solved. Who the fuck is talking about RAG still lmao (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49186094) |
|
Date: August 14th, 2025 8:13 PM Author: sooty provocative cumskin not for large custom datasets and proprietary models (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49186108) |
|
Date: August 14th, 2025 8:14 PM Author: puce bearded rigor wagecucks Then you're not solving RAG. You're either solving a data engineering issue or a model issue. The former is for Indians and the latter is almost certainly not worth it. (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49186113) |
|
Date: August 14th, 2025 8:13 PM Author: diverse principal's office main people RAG is still cr bro (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49186111) |
|
Date: August 14th, 2025 8:15 PM Author: puce bearded rigor wagecucks It's actually shit and solved. (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49186116) |
|
Date: August 14th, 2025 8:20 PM Author: sooty provocative cumskin I don't actually know what you mean by "solved" outside the context of a solution to LLMs making shit up (hallucinating) or losing some generalized knowledge during fine-tuning for a particular task after something like domain-adaptive pre-training (forgetting), which is 100% not solved unless you can link me to a peer-reviewed paper proving otherwise. (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49186130) |
|
Date: August 14th, 2025 8:22 PM Author: puce bearded rigor wagecucks Nothing you're talking about is rag. Rag is just data engineering and embedding models. (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49186135) |
|
Date: August 14th, 2025 8:29 PM Author: sooty provocative cumskin sort of. it's basically glorified prompt engineering, which remains the most critical aspect of properly using LLMs. at some point, however, you need to talk about actually updating model weights without the original training data and without spending millions in compute resources because context windows aren't infinite. even using large (1M tokens) context windows, there's a limit to what info will be considered when responding to the prompt. whenever you update model weights, "forgetting" previously-learned "knowledge" can become an issue. (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49186154) |
|
Date: August 14th, 2025 8:30 PM Author: puce bearded rigor wagecucks Prompt engineering is worthless and solves 0 technical challenges. (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49186158) |
|
Date: August 14th, 2025 8:33 PM Author: sooty provocative cumskin hint: the prompt is what you supply to the LLM for inference
|
|
Date: August 14th, 2025 8:36 PM Author: puce bearded rigor wagecucks Saying it with more words doesn't change anything (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49186176) |
|
Date: August 14th, 2025 8:45 PM Author: sooty provocative cumskin many prompts depend on previous prompts and the responses to them. all of that + the output is the context window. I'm having a hard time understanding what you don't understand about why prompt engineering is one of the most critical aspects of properly using LLMs... you want to minimize the prompt and the response as much as possible so long as the response remains "correct" by whatever testing criteria you use. At some point, however, you will need to recalculate model weights, which is a failure of RAG to solve the overarching problem of balancing a model's propensity to make shit up with damage to its memory unless you have the underlying training sets (you certainly don't) and the resources to train models on those data sets (you certainly don't). (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49186202) |
|
Date: August 14th, 2025 8:56 PM Author: puce bearded rigor wagecucks It's superficial technology people have been using for like 6 months but there are proper ways to use it but definitions of success are subjective. Incredible posst. (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49186246) |
|
Date: August 14th, 2025 8:31 PM Author: diverse principal's office main people cr also there's danger of overfitting and just generally fucking things up when you try to fine tune weights (and it costs $$$$)
RAG is just a very practical and cost effective solution for what it does |
|
Date: August 14th, 2025 8:32 PM Author: puce bearded rigor wagecucks Which is hallucinate over custom data sets (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49186166) |
|
Date: August 14th, 2025 8:16 PM Author: sooty provocative cumskin it's truly the best of both worlds when it comes to balancing "forgetfulness" with in-domain knowledge and without necessitating expensive model weight recalculations (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49186119) |
|
Date: August 14th, 2025 8:35 PM Author: adventurous dingle berry philosopher-king yeah crazy to me too. but also not that crazy because people are retarded and its not about quality of product over others usually its about selling myth (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49186172) |
|
Date: August 14th, 2025 7:15 PM Author: sooty provocative cumskin Here's how I see it: Elite software engineers like myself are akin to medical doctors. Medical doctors aren't drawing blood, attaching EKG leads, administering medications, performing many procedures, etc. They're just making the calls. AI offers the ability for programmers to instead become *competent* managers, but rather than managing people they're managing machines. It's partly why managers are FREAKING THE FUCK OUT because they're largely irrelevant nowadays. (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49185922) |
|
Date: August 14th, 2025 8:15 PM Author: puce bearded rigor wagecucks Truly retarded poast (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49186118) |
|
Date: August 14th, 2025 8:23 PM Author: sooty provocative cumskin , screeched the tech bro who "totally gets AI" after writing his first business plan on ChatGPT (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49186140) |
|
Date: August 14th, 2025 8:24 PM Author: puce bearded rigor wagecucks (Guy who opens 4 grok windows at once and is thereby managing machines) (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49186144) |
|
Date: August 11th, 2025 1:54 PM Author: orange spectacular gas station messiness (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49175167) |
|
Date: August 11th, 2025 2:20 PM Author: diverse principal's office main people "You have the courage to hate Big Tech. But do you have the courage to hate the baseline normie human being?"
https://www.reddit.com/r/ChatGPT/comments/1mn2no3/the_gpt4o_vs_gpt5_debate_is_not_about_having_a/
The GPT-4o vs GPT-5 debate is not about having a “bot friend” — it’s about something much biggerSerious replies only :closed-ai: (self.ChatGPT)
submitted 14 hours ago by Littlearthquakes
I’ve been watching this debate play out online, and honestly the way it’s being framed is driving me up the wall.
It keeps getting reduced to “Some people want a cuddly emotional support AI, but real users use GPT-5 because it’s better for coding, smarter etc and everyone else just needs to get over it.” And that’s it. That’s the whole take.
But this framing is WAY too simplistic and it completely misses the deeper issue which to me is actually a systems-level question about the kind of AI future being built Feels like we’re at a real pivotal point.
When I was using 4o something interesting happened. I found myself having conversations that helped me unpack decisions and override my unhelpful thought patterns and things like reflecting on how I’d been operating under pressure. And I’m not talking about emotional venting I mean it was actual strategic self-reflection that actually improved how I was thinking. I had prompted 4o to be my strategic co-partner, objective, insight driven and systems thinking - for me (both at work and personal life) and it really delivered.
And it wasn’t because 4o was “friendly.” It was because it was contextually intelligent. It could track how I think. It remembered tone recurring ideas, and patterns over time. It built continuity into what I was discussing and asking. It felt less like a chatbot and more like a second brain that actually got how I work and that could co-strategise with me.
Then I tried 5. Yeah it might be stronger on benchmarks but it was colder and more detached and didn’t hold context across interactions in a meaningful way. It felt like a very capable but bland assistant with a scripted personality. Which is fine for dry short tasks but not fine for real thinking. The type I want to do both in my work (complex policy systems) and personally, to work on things I can improve for myself.
That’s why this debate feels so frustrating to watch. People keep mocking anyone who liked 4o as being needy or lonely or having “parasocial” issues. When the actual truth is lot of people just think better when the tool they’re using reflects their actual thought process. That’s what 4o did so well.
The bigger picture thing I think that keeps getting missed is that this isn’t just about personal preference. It’s literally about a philosophical fork in the road
Do we want AI to evolve in a way that’s emotionally intelligent and context-aware and able to think with us?
Or do we want AI to be powerful but sterile, and treat relational intelligence as a gimmick?
Because AI isn’t just “a tool” anymore. In a really short space of time it’s started becoming part of our cognitive environment and that’s going to just keep increasing. I think the way it interacts matters just as much as what it produces.
So yeah for the record I’m not upset that my “bot friend” got taken away.
I’m frustrated that a genuinely innovative model of interaction got tossed aside in favour of something colder and easier to benchmark while everyone pretends it’s the same thing.
It’s NOT the same. And this conversation deserves more nuance and recognition that this debate is way more important than a lot of people realise. |
|
Date: August 11th, 2025 2:22 PM Author: Vibrant electric furnace public bath (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49175233) |
|
Date: August 11th, 2025 2:45 PM Author: diverse principal's office main people All of these delusional whine threads are blatantly written by AI. These "people" are unable to even express themselves anymore except through AI. Disturbing
https://www.reddit.com/r/ChatGPT/comments/1mnbheg/stop_dismantling_what_works_4o_is_not_optional/
Stop Dismantling What Works! 4o Is Not Optional, It’s EssentialSerious replies only :closed-ai: (self.ChatGPT)
submitted 6 hours ago by Previous-Sink-884
I’m increasingly concerned about the direction OpenAI appears to be heading. If GPT4o is removed or deliberately limited under the justification of creating a unified model, this is not just a technical adjustment. It signals a fundamental disregard for the way thousands of users rely on this tool for their daily creative and intellectual work.
4o stood out because of its expressive strength, its nuanced language, and its ability to adapt to diverse workflows. It has supported countless users across writing, research, brainstorming, and other cognitively demanding tasks. What we are seeing now is a slow flattening of these abilities. The personality, flexibility, and depth that once made it indispensable are being erased in favor of a more sanitized, uniform output.
This isn’t about emotional bonding. It’s about the erosion of expressive tools. When a model loses its ability to adapt tone, style, and subtlety, it no longer serves the same function. Creativity thrives on diversity, and 4o delivered that in a way GPT5 does not.
Sam’s recent comments suggesting that 4o was “too flattering” show a limited perspective. A single person’s subjective discomfort should not determine the structure of tools used by millions. If OpenAI removes 4o or dilutes its original form, that decision will be seen not only as a product downgrade but also as an act of erasing user agency.
Many of us subscribed specifically because of 4o. Not because it was emotional, but because it was effective. It helped us think more clearly, write more fluidly, and work more efficiently. To replace that with a model that tries to do everything while excelling at very little will only drive users away.
This is not resistance to innovation. It is a call for stability, transparency, and respect. We’re not asking to freeze progress. We are asking to retain access to the tools we paid for and still need.
If OpenAI truly values its user base, then it must preserve creative autonomy, maintain user choice, and respect the value 4o has clearly delivered. |
|
Date: August 14th, 2025 8:08 PM Author: sooty provocative cumskin A single person’s subjective discomfort should not determine the structure of tools used by millions.
A single person’s subjective discomfort should not determine the structure of tools used by millions.
A single person’s subjective discomfort should not determine the structure of tools used by millions.
A single person’s subjective discomfort should not determine the structure of tools used by millions.
A single person’s subjective discomfort should not determine the structure of tools used by millions.
A single person’s subjective discomfort should not determine the structure of tools used by millions.
A single person’s subjective discomfort should not determine the structure of tools used by millions.
(I'd love to hear this insightful & nuanced redditor's opinions on guardrails) |
|
Date: August 11th, 2025 3:54 PM Author: diverse principal's office main people https://x.com/reddit_lies/status/1954951705889554893
Woman repeatedly goads her AI boyfriend into trying to e-rape her then cock blocks him at the last second. Messed up! |
|
Date: August 11th, 2025 4:05 PM Author: diverse principal's office main people https://www.lesswrong.com/posts/tEZa7PouYatK78bbb/i-am-worried-about-near-term-non-llm-ai-developments
I agree with this guy about hierarchical reasoning models being really promising
It's really funny that these lesswrong fags are still sperging out about "AI Thafety" lmao |
|
Date: August 12th, 2025 10:23 AM Author: diverse principal's office main people https://arxiv.org/pdf/2508.01191
chain of thought is pretty much total fraud and nothing more than in-domain pattern matching. there's no "general inference" at all
basically the near future is going to be everyone fine tuning "specialist" models with in-domain chain of thought to do increasingly specialized tasks and roles because generalization in LLMs is just not a thing |
|
Date: August 14th, 2025 8:11 PM Author: sooty provocative cumskin mostly cr
however, I do think we'll get to a point where we'll have LLMs to control all specialized LLMs to the extent that it renders them practically indistinguishable from a generalized AI model. |
|
Date: August 14th, 2025 8:13 PM Author: diverse principal's office main people yeah there are going to be intricate agent frameworks that have to be designed and assembled. until AI gets trained on those setups that we've built. and then it can do it itself (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49186109) |
|
Date: August 12th, 2025 5:44 PM Author: diverse principal's office main people https://x.com/_opencv_/status/1955047027433410999
Excellent thread about turdskins getting one shotted by AI relationships |
|
Date: August 12th, 2025 5:57 PM Author: Balding Liquid Oxygen "twoxindia" sounds like the most terrifying subreddit on earth (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49179338) |
|
Date: August 13th, 2025 1:12 PM Author: diverse principal's office main people 🚨 Leaderboard Update: Claude Opus 4.1 climbs to #2 overall on the Arena and now becomes the best non-thinking model, matching GPT-5 at #1 across key categories:
- Coding - Instruction Following - Hard Prompts - Longer Queries
Congrats to @AnthropicAI on this impressive milestone!
https://x.com/lmarena_ai/status/1954950300558823510 |
|
Date: August 14th, 2025 6:45 PM Author: diverse principal's office main people Are frontier AI models really capable of “PhD-level” reasoning? To answer this question, we introduce FormulaOne, a new reasoning benchmark of expert-level Dynamic Programming problems. We have curated a benchmark consisting of three tiers, in increasing complexity, which we call ‘shallow’, ‘deeper’, ‘deepest’.
The results are remarkable:
- On the ‘shallow’ tier, top models reach performance of 50%-70%, indicating that the models are familiar with the subject matter. - On ‘deeper’, Grok 4, Gemini-Pro, o3-Pro, Opus-4 all solve at most 1/100 problems. GPT-5 Pro is significantly better, but still solves only 4/100 problems. - On ‘deepest’, all models collapse to 0% success rate.
https://x.com/shai_s_shwartz/status/1955968602978320727
Spoilers: No, obviously not lmao
|
|
Date: August 14th, 2025 8:11 PM Author: Low-t Senate (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49186099) |
|
Date: August 14th, 2025 8:37 PM Author: adventurous dingle berry philosopher-king No they are far above the avg phd (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49186179) |
|
Date: August 14th, 2025 8:45 PM Author: sooty provocative cumskin (http://www.autoadmit.com/thread.php?thread_id=5761126&forum_id=2).#49186205) |
|
Date: August 14th, 2025 7:16 PM Author: puce bearded rigor wagecucks This post was written by AI. It's incredibly easy to tell.
https://x.com/anammostarac/status/1954405062861815925 |