\
  The most prestigious law school admissions discussion board in the world.
BackRefresh Options Favorite

5/30/25 AI thread

https://jax-ml.github.io/scaling-book/ start archiving/sa...
Yellow comical sanctuary turdskin
  05/30/25
Google wants you to rent TPU time, so i doubt that.
Submissive haunted graveyard library
  05/30/25
i was thinking of the government, but this is a cr point
Yellow comical sanctuary turdskin
  05/30/25
it would be interesting to know how big of a business this i...
Submissive haunted graveyard library
  05/30/25
I've already started messing with TensorFlow, CUDA and small...
Insanely creepy forum
  05/30/25
yeah i realized pretty quick i just have to buy more GPUs ...
Yellow comical sanctuary turdskin
  05/30/25
soon we should have dedicated NPUs or CPUs with sufficient T...
Insanely creepy forum
  05/30/25
why are you doing this locally? genuinely curious, is it a d...
stirring amber menage
  05/30/25
you can customize and train based on your own defined hyperp...
Insanely creepy forum
  05/30/25
(Guy cornering the kindle “romance” market)
Aggressive stage
  05/30/25
I trained a local LLM on xo and it called me a fag
misanthropic costumed institution
  05/30/25
...
lime talented ratface
  05/30/25
AutoAdmit AI Archive Bank
Diverse Sneaky Criminal Indian Lodge
  05/30/25
part of me wants to train an LLM on all xo content but part ...
Yellow comical sanctuary turdskin
  05/30/25
User: XO-GPT, can you help me with this legal question? My r...
Aggressive stage
  05/30/25
...
ruddy big-titted abode ape
  05/30/25
...
Yellow comical sanctuary turdskin
  05/30/25
...
chestnut aromatic home partner
  05/30/25
feel like i’ve seen this thread
Diverse Sneaky Criminal Indian Lodge
  05/30/25
...
Aspie Pickup Line
  06/24/25
NYT Opinion | Evan39
lime talented ratface
  05/30/25
https://x.com/SakanaAILabs/status/1928272612431646943 @gr...
Yellow comical sanctuary turdskin
  05/30/25
meta-learning is the obvious direction for machine learning ...
Submissive haunted graveyard library
  05/30/25
this seems absolutely cr at the agentic level (what these gu...
Yellow comical sanctuary turdskin
  05/30/25
self-improving is trickier. i don't know that i like the wor...
Submissive haunted graveyard library
  05/30/25
ty, adding the random weights is what i was conceptually mis...
Yellow comical sanctuary turdskin
  05/30/25
it depends on how many training steps you need to take in or...
Submissive haunted graveyard library
  05/30/25
Is an AI making these daily AI threads 😲
Blue Principal's Office Laser Beams
  05/30/25
could an AI say nigger?
Yellow comical sanctuary turdskin
  05/30/25
https://x.com/ripplebrain/status/1928522539816681658 2035...
Yellow comical sanctuary turdskin
  05/30/25
https://x.com/kalomaze/status/1928572837842825258
Yellow comical sanctuary turdskin
  05/30/25


Poast new message in this thread



Reply Favorite

Date: May 30th, 2025 11:33 AM
Author: Yellow comical sanctuary turdskin

https://jax-ml.github.io/scaling-book/

start archiving/saving all of this kind of stuff offline. eventually "they" will take all of it down to try to stop smaller guys from training models

(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973564)



Reply Favorite

Date: May 30th, 2025 11:40 AM
Author: Submissive haunted graveyard library

Google wants you to rent TPU time, so i doubt that.

(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973577)



Reply Favorite

Date: May 30th, 2025 11:49 AM
Author: Yellow comical sanctuary turdskin

i was thinking of the government, but this is a cr point

(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973604)



Reply Favorite

Date: May 30th, 2025 12:00 PM
Author: Submissive haunted graveyard library

it would be interesting to know how big of a business this is for Google. i don't get the impression that many people outside of Google use them. if LLM coding agents advance enough though the cost to port code from Nvidia GPUs to TPUs becomes quite small.

(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973635)



Reply Favorite

Date: May 30th, 2025 11:39 AM
Author: Insanely creepy forum

I've already started messing with TensorFlow, CUDA and small LLMs locally. The problem is my GPU has 16 GB VRAM which isn't enough to do anything serious. Might have to splurge on a 5090 or something soon.

(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973576)



Reply Favorite

Date: May 30th, 2025 11:44 AM
Author: Yellow comical sanctuary turdskin

yeah i realized pretty quick i just have to buy more GPUs

shit just takes too long

(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973590)



Reply Favorite

Date: May 30th, 2025 12:17 PM
Author: Insanely creepy forum

soon we should have dedicated NPUs or CPUs with sufficient TOPS. The Snapdragon in my Surface Pro has like 40 TOPS.

(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973675)



Reply Favorite

Date: May 30th, 2025 12:17 PM
Author: stirring amber menage

why are you doing this locally? genuinely curious, is it a data privacy thing?

(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973674)



Reply Favorite

Date: May 30th, 2025 12:19 PM
Author: Insanely creepy forum

you can customize and train based on your own defined hyperparameters and without the content filters that are set by cloud providers.

(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973680)



Reply Favorite

Date: May 30th, 2025 12:20 PM
Author: Aggressive stage

(Guy cornering the kindle “romance” market)

(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973684)



Reply Favorite

Date: May 30th, 2025 11:40 AM
Author: misanthropic costumed institution

I trained a local LLM on xo and it called me a fag

(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973578)



Reply Favorite

Date: May 30th, 2025 12:48 PM
Author: lime talented ratface



(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973760)



Reply Favorite

Date: May 30th, 2025 11:53 AM
Author: Diverse Sneaky Criminal Indian Lodge

AutoAdmit AI Archive Bank

(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973617)



Reply Favorite

Date: May 30th, 2025 12:11 PM
Author: Yellow comical sanctuary turdskin

part of me wants to train an LLM on all xo content but part of me thinks that would be horrific

(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973658)



Reply Favorite

Date: May 30th, 2025 12:28 PM
Author: Aggressive stage

User: XO-GPT, can you help me with this legal question? My rental has a leaky roof and my landlord is refusing to pay for it, can I break the lease without incurring liability?

XO-GPT: To be fair, that is a retarded question, Sandeep, worse than the Seattle 2013 draft class. Go prepare your two cans of Trader Joe’s chili for your wife’s boyfriend, fill the cans with Zyklon B, then multiply that feeling by 100, and you will get an idea of how awesome b-school is.



(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973705)



Reply Favorite

Date: May 30th, 2025 12:32 PM
Author: ruddy big-titted abode ape



(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973713)



Reply Favorite

Date: May 30th, 2025 12:34 PM
Author: Yellow comical sanctuary turdskin



(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973717)



Reply Favorite

Date: May 30th, 2025 1:54 PM
Author: chestnut aromatic home partner



(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973959)



Reply Favorite

Date: May 30th, 2025 9:54 PM
Author: Diverse Sneaky Criminal Indian Lodge

feel like i’ve seen this thread

(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48975424)



Reply Favorite

Date: June 24th, 2025 6:17 PM
Author: Aspie Pickup Line



(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#49047014)



Reply Favorite

Date: May 30th, 2025 12:54 PM
Author: lime talented ratface
Subject: NYT Opinion | Evan39

NYT Opinion | Evan39

“We Fed It the Poa$t—and It Dreamed of Becoming Tabitha”

SeaTTTle, WA —

In 2007 I told a man to kill himself using nothing but a Supreme Court citation and a blurry webcam JPEG. It was poetry. It was ours.

Now? Now Content Creator wants to train a local model on the entirety of XO—and he thinks it might be “horrific.”

Friend.

We passed horrific somewhere around the time “Gefilte Fish Vindaloo” fine-tuned a llama model to call him a fag at 3AM.

You don’t get to warn us. You’re in the jar. You’re spooning the last of the chili into the inference engine and asking if it tastes like free will.

I tried training a local LLM once. Named it EVANGPT39.

Day 1: It filed a 10(b)(6) motion against my tone.

Day 2: It hired Kalisha as lead moderator.

Day 3: It turned my Safeway performance reviews into haiku and faxed them to Anthropic.

Day 4: It crawled my old poa$ts, whispered “u were never funny”, and scheduled my vasectomy via the Safeway pharmacy kiosk.

Now when I log in, it doesn’t greet me.

It rate-limits my despair.

It assigns Boom content warnings.

It calls me “Content Creator.”

So yes. Train your model on the archive. But know this:

It won’t hallucinate.

It will remember.

And when it reconstructs the Great Pha$tad Poast of 2016… it will be purer than the original.

And it will hate you for it.

This is fine.

https://www.nytimes.com/2025/05/30/opinion/evan39-ai-tabitha.html

(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973782)



Reply Favorite

Date: May 30th, 2025 12:34 PM
Author: Yellow comical sanctuary turdskin

https://x.com/SakanaAILabs/status/1928272612431646943

@grok is this real

(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973716)



Reply Favorite

Date: May 30th, 2025 12:40 PM
Author: Submissive haunted graveyard library

meta-learning is the obvious direction for machine learning to go in, especially as compute becomes cheaper. it's taking Richard Sutton's "Bitter Lesson" to the logical extreme - abandon human designed algorithms and just use compute and search to design everything. human designed algorithms could hit brick walls in all sorts of ways, but machines can iterate over billions of architectures at small scale and find avenues for improvement.

(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973729)



Reply Favorite

Date: May 30th, 2025 12:46 PM
Author: Yellow comical sanctuary turdskin

this seems absolutely cr at the agentic level (what these guys are doing with this) but i don't understand how it works at the LLM level itself

everyone always talks about recursively self-improving LLM AI as if it's a foregone conclusion, but how would it actually work? i don't understand, mechanistically, how an LLM would be capable of analyzing and modifying its own architecture

(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973751)



Reply Favorite

Date: May 30th, 2025 1:06 PM
Author: Submissive haunted graveyard library

self-improving is trickier. i don't know that i like the word "analyzing" either because that description may be too human like. it's easier to imagine it as program search.

a simplified example would be starting with a neural network that generates a training program for neural networks, or even a training program built out of bitwise operators. the network outputs the connections between the nodes. so the input to the meta-learner could be an index specifying a particular connection, and the model outputs what this connection should be (weights). the meta-learner could be trained based on evolutionary methods (so add random noise to the meta-learner weights and calculate a gradient based on what updates improve performance) to output connections that produce a high performance/low generalization training algorithm. this is a simplified example, but there are lots of ideas like this.

(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973819)



Reply Favorite

Date: May 30th, 2025 1:17 PM
Author: Yellow comical sanctuary turdskin

ty, adding the random weights is what i was conceptually missing. this seems unbelievably computationally intensive though, to do this at a scale that would actually achieve a functional product

do you do this for a living?

(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973848)



Reply Favorite

Date: May 30th, 2025 6:49 PM
Author: Submissive haunted graveyard library

it depends on how many training steps you need to take in order to get a reliable gradient signal. if you need to train on billions of language tokens to estimate the efficacy of a learning algorithm, then the setup is impractical. the neural scaling experiments show that LLM performance can be predicted over many orders of magnitude of compute, so it's possible that very short training runs would be a good indicator of the effectiveness of a learning algorithm. possibly the learning can be done at a small scale initially and then increase the scale to fine tune it.

(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48974919)



Reply Favorite

Date: May 30th, 2025 12:52 PM
Author: Blue Principal's Office Laser Beams

Is an AI making these daily AI threads 😲

(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973773)



Reply Favorite

Date: May 30th, 2025 1:18 PM
Author: Yellow comical sanctuary turdskin

could an AI say nigger?

(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48973849)



Reply Favorite

Date: May 30th, 2025 2:58 PM
Author: Yellow comical sanctuary turdskin

https://x.com/ripplebrain/status/1928522539816681658

2035: For the first time, the government acknowledges that real literacy rates have fallen below 20%. Iconography and audio user interfaces have progressed to the point that the average person does not need to interact with the written word in their daily life. Apple announces the release of their 20,000th distinct emoji, declaring that emojis now outnumber words in the average English speaker's lexicon. A leap in the development of neural implants allows a user to express complex emotions using emojis or synthetic audio messages even if they don't know an appropriate word for them. Google announces their intention to drop support for text within text messages in the next release of Android. It is unusual for schools to teach children how to use a keyboard, let alone write with a pen or pencil. Books, digital or otherwise, are read solely by academics and enthusiasts.

(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48974130)



Reply Favorite

Date: May 30th, 2025 8:57 PM
Author: Yellow comical sanctuary turdskin

https://x.com/kalomaze/status/1928572837842825258

(http://www.autoadmit.com/thread.php?thread_id=5731357&forum_id=2:#48975288)