The most prestigious law school admissions discussion board in the world.
|
 |
 
|
8/5/25 AI thread
Poast new message in this thread
Date: August 5th, 2025 10:48 AM Author: Stubborn Up-to-no-good State interesting comparison of chatgpt 5 and deekseek r1's reasoning outputs. the new chatgpt5 appears to have a lot more crisp and concise and human like reasoning. tbh it reads a lot like a human taking notes. it will be a lot more cost efficient
https://x.com/jxmnop/status/1952375903658410336
local LLMs are the future
https://x.com/iotcoi/status/1952263680273289337
new hierarchical reasoning model in development
https://x.com/omarsar0/status/1951751651729060081 |
 |
Date: August 5th, 2025 12:20 PM Author: Azure hall can't you basically trust no models "reasoning" output though? that is, the "thinking out loud" part is itself only exposed in a manner that it was programmed to be, not some exposure of the raw inner working of the llm? (http://www.autoadmit.com/thread.php?thread_id=5758545&forum_id=2:#49158472) |
|
Date: August 5th, 2025 12:41 PM Author: Stubborn Up-to-no-good State my understanding is that the newest models' CoT reasoning traces are actually pretty close to the actual inner workings of the model's CoT reasoning tree
after reading this thread in more detail though, i think what is being shown is not chatgpt 5's reasoning trace. it's the actual answer output that it gave. so in reality this post doesn't say anything about any changes/updates to this model's reasoning capabilities |
Date: August 5th, 2025 11:07 AM Author: Stubborn Up-to-no-good State Introducing Genie 3, the most advanced world simulator ever created, enabled by numerous research breakthroughs. ๐คฏ
Featuring high fidelity visuals, 20-24 fps, prompting on the go, world memory, and more.
https://x.com/OfficialLoganK/status/1952732206176112915 |
|
Date: August 6th, 2025 3:13 PM Author: Razzle-dazzle gay shrine Whoa… this will lead people to become “wire heads”, each addicted to his or her own personalized virtual fantasy world. Where do I sign up? (http://www.autoadmit.com/thread.php?thread_id=5758545&forum_id=2:#49162420) |
|
Date: August 5th, 2025 11:09 AM Author: Stubborn Up-to-no-good State https://www.youtube.com/watch?v=ysPbXH0LpIE
https://www.youtube.com/watch?v=XSZP9GhhuAc
these are actually really good videos on modern prompting methods and structure
some very useful tips here for everyone no matter what you use AI for |
|
Date: August 5th, 2025 11:39 AM Author: at-the-ready emerald chapel bbw the hierarchical reasoning paper is interesting and appeared the likely direction to go in. chain of thought is a terrible way to get iterative depth computation from a transformer. recurrent circuits that compute for the necessary period of time is much more like the brain and is more likely to produce generalization benefits than using chain of thought with a verifier (that will only work in the domains you are verifying for). (http://www.autoadmit.com/thread.php?thread_id=5758545&forum_id=2:#49158304) |
|
Date: August 5th, 2025 11:40 AM Author: Stubborn Up-to-no-good State https://xoxohth.com/thread.php?thread_id=5757240&mc=14&forum_id=2#49151205
what are your thoughts on these "moral orientation" "personas" and what exactly do you think causes them? |
|
Date: August 5th, 2025 12:15 PM Author: at-the-ready emerald chapel bbw it seems like the models try to construct a consistent character to respond to a prompt. they are guessing what the best character for a particular prompt is (which can be many things since they are trained on the entire web), and sometimes it isn't appropriate. this doesn't seem surprising and is consistent with other LLM behavior. (http://www.autoadmit.com/thread.php?thread_id=5758545&forum_id=2:#49158455) |
|
Date: August 5th, 2025 12:35 PM Author: Stubborn Up-to-no-good State so you think it's incorrect problem solving techniques being associated with "evil" persona traits in the training data? (that seems to be the explanatory mechanism behind what you're saying, imo, correct me if i'm wrong)
that is apparently the leading hypothesis for this, and it's reasonable enough. but it just doesn't seem convincing to me. is there *really* that much of a correlation between these things in the training data? it just doesn't pass the smell test imo |
|
Date: August 5th, 2025 1:25 PM Author: at-the-ready emerald chapel bbw evil behavior seems to be commonly represented as faulty emotional logic. i could see why a model encouraged to have faulty cognition might activate general aberrant behavior. certainly in people anti-social behavior has a relationship to low IQ. (http://www.autoadmit.com/thread.php?thread_id=5758545&forum_id=2:#49158654) |
|
Date: August 5th, 2025 1:32 PM Author: silver round eye Still can't do shit (http://www.autoadmit.com/thread.php?thread_id=5758545&forum_id=2:#49158684) |
|
Date: August 5th, 2025 3:03 PM Author: Stubborn Up-to-no-good State (guy who uses AI to do his job on a daily basis) (http://www.autoadmit.com/thread.php?thread_id=5758545&forum_id=2:#49159350) |
|
Date: August 5th, 2025 7:11 PM Author: silver round eye It just does slop work i have to fix. It has replaced 0 jobs (http://www.autoadmit.com/thread.php?thread_id=5758545&forum_id=2:#49160170) |
|
Date: August 5th, 2025 5:17 PM Author: Stubborn Up-to-no-good State OpenAI @OpenAI
We released two open-weight reasoning models—gpt-oss-120b and gpt-oss-20b—under an Apache 2.0 license.
Developed with open-source community feedback, these models deliver meaningful advancements in both reasoning capabilities & safety.
gpt-oss-120b matches OpenAI o4-mini on core benchmarks and exceeds it in narrow domains like competitive math or health-related questions, all while fitting on a single 80GB GPU (or high-end laptop).
gpt-oss-20b fits on devices as small as 16GB, while matching or exceeding OpenAI o3-mini.
https://x.com/OpenAI/status/1952783291091653011
https://x.com/omarsar0/status/1952787354445402494
Wow. This is 180 |
|
Date: August 5th, 2025 5:35 PM Author: at-the-ready emerald chapel bbw GPT-5 is likely a pretty decent improvement then considering this is close to their private reasoning models.
LJL at Anthropic rush releasing Opus 4.1 in order to compete. |
|
Date: August 5th, 2025 5:59 PM Author: Stubborn Up-to-no-good State https://x.com/kalomaze/status/1952812751404908672
They are already post trained though so fine tuning them to do your own stuff is really hard |
|
Date: August 6th, 2025 1:49 PM Author: Stubborn Up-to-no-good State Sam Altman
@sama we are providing ChatGPT access to the entire federal workforce!
(for $1 a year per agency)
https://x.com/sama/status/1953103336044990779
uh, you sure you guys got enough monopoly money for this? |
|
Date: August 6th, 2025 7:03 PM Author: Stubborn Up-to-no-good State https://x.com/GoogleDeepMind/status/1952732150928724043
Google DeepMind @GoogleDeepMind
What if you could not only watch a generated video, but explore it too? ๐
Genie 3 is our groundbreaking world model that creates interactive, playable environments from a single text prompt.
From photorealistic landscapes to fantasy realms, the possibilities are endless. ๐งต |