Thread 105660326

How does he keep doing it. Amazing.

god

If it can solve any one of these, I will stop laughing at it

https://en.wikipedia.org/wiki/List_of_unsolved_problems_in_mathematics

>>105660326 (OP)
>we will trigger mode collapse by training an llm on its own outputs
this is your brain on 5 grams of ket a day

Im not an AI expert or anything but that sounds like a horrible idea. Training on output from the model youre training sounds like it will lead to absolutely shit results

>>105660326 (OP)
>we're going to have our LLM insert garbage into it's training data and then retrain it on that

>>105660326 (OP)
>>105660367
this reminds me of a scinece-fiction show, I forgot the name, shit. it was a rich guy or something and there were the most talented programmers working on something (don't know anymore if it was AI or not), they could see jesus and stuff

>>105660470
devs?

>>105660477
ah yeah, thanks, yep it was devs.

>>105660402
>>105660454
Yeah I agree with both of you, I just commented the same in a different thread:
>>105660459

>>105660454
What a smart idea!
Whoever thought that must be a visionary

>>105660326 (OP)
>it's ok when we try to Correct The Record™

Even boomers know what happens when you make a photocopy of a photocopy of a photocopy.

Grok starship blow up. AI much smart. Pls fix.

>>105660326 (OP)
>create "AI" have "AI" alter information, retrain "AI" on its own bullshit
genius plan

>>105660326 (OP)
this retard will take grok and turn it into another dogeai_gov

tired of AI. shit is a bubble and fake. And has the most retarded naming. There's nothing intelligent about a digital branching prediction database.

>>105660326 (OP)
Grok or any other AI needs to rebuild a "wiki" of human knowledge from their own understanding of the world. I would like to see what AI knowledge base are.

>>105660326 (OP)
>We don't like that human history has said, done and remembered things that are not in line with our political philosophy about how everyone should live
>Therefore, we'll just have the algorithm rewrite everything so that the history we don't like is erased and what's left says what we want it to say

These people are going to build a nightmarish hellscape the likes of which we've never seen.

>>105660326 (OP)
Based Elon

>>105660326 (OP)
>we will train AI on AI generated data
Brilliant

AI will always be a meme so long as they censor it. Thats the opposite of what you want to do with tech.

>>105660397
Fake news

>>105660326 (OP)
a fraud who lied even about videogame achievements - thats called a compulsive liar

>>105662034
thats what chinks did with deepseek

>>105660326 (OP)
This is literally how you get model collapse

>>105660326 (OP)
elon lies

>>105660326 (OP)
what is in Epstein files?

self driving cars! we'll be in mars by 2010! hyperloop! trump was right about everything! (except firing me)

>>105662237
They trained from the output of a bigger model once. That's not the same as iteratively training a chatbot on its own outputs. I'l like the difference between photocopying a document once to get a copy of it cheaply vs feeding what the machine prints back into the scanner to make a new copy from the last copy repeatedly.

>>105663022
Synthetic data is better because its 'optimized' and cleaned up- thats what they think and what they have internally should be much better than castrated output we get

>>105660326 (OP)
a shit-eating ouroboros, if you will

>>105663022
>trained from the output of a bigger model once
First of all, "bigger model" for a model we don't have any specifications on is a pretty retarded claim.
Second, it ended up with better performance than the model it was trained on.
>I'l like the difference between photocopying a document once to get a copy of it cheaply vs feeding what the machine prints back into the scanner to make a new copy from the last copy repeatedly.
Given how well it worked for DeepSeek, it would probably work to train a new model off the outputs of the current one. Obviously this is just finetuning and not pretraining, but if you have a reliable way to detect and correct incorrect information then obviously you'll end up with a higher quality dataset.

>>105663229
>find (((science))) about how cutting your dick off turns you into a "real woman"
>delete it
dilation deleted, dataset improved

>>105660326 (OP)
>rewriting history
:(
>rewriting history with AI
:O

>>105661926
The problem is that it would run forever, hallucinating complete fictions where there is a lack of real understanding, and it wouldn't stop until you hit the red button

>>105663180
you're not even baiting, you're just retarded

>>105663282
thank you for correcting the record, stay retarded

>>105663246
I'll be the judge of that

>>105663180
Deepseek was trained at least in good part on chatgpt outputs, possibly only on chatgpt outputs. We know this because it used to frequently claim to be chatgpt in the beginning. Deepseek also wasn't better than chatgpt, it was better at certain benchmarks, because of reinforcement learning that happened after the training with chatgpt data. If you did the RL on the original model instead, it would have been better.

>>105660326 (OP)
> a bunch of jeets at X are going to rewrite history and retrain Grok on their bullshit
Elon is such a retard.

>>105660326 (OP)
6 million number will get revised?

>>105660326 (OP)
> compulsive lying drug addict said something
saving this post to show my future grand children.

>>105660391
it will never solve anything of the sort.

>>105661926
> when your intelligence is this low: the post
humanity is cooked.

>>105661926
>>I would like to see what AI knowledge base are.
>According to CNN and every other top news organization, merely saying that you're a woman changes you into one! I need to remove these documents about X chromosomes and Y chromosomes to make sure my dataset aligns with reality.

>>105660326 (OP)
kek grok is such a fucking joke

>ask question about nfl scores
>get rant about white genocide

>>105660326 (OP)
>Train model on data filled with human errors
>Ask it to fix the data
>It replaces existing errors and correct information with new errors

>adding missing data and correcting error
aka: adding my own conspiracy theories and deleting anything I don't like.

>>105664217
you've described 90% of Americans (and humans in general)

>>105660326 (OP)
AI is reductive AF. It's going to be Cartesian dualism all over again, this time on steroids.

>>105660326 (OP)
Pretty sure they've completely lost connection with what "data" is supposdd to be.

>>105660326 (OP)
They can't be serious? Has AI actually done anything original yet? Can I get a list?

>>105660402
Isn't one of the problems with current AI that we're running out of organic training data?

>Twitter for iPhone
LMAO

>>105660326 (OP)
bros, why is AI so shit at coding?
I just tried cursors fancy agentic mode, to vibe code an app from scratch, giving it a list of detailed instructions.

The code it generated looked like crap, inconsistent and did things in weird and shitty ways.

There's no way a reasonably competent human would've wrote the code like this.

AI is fine for writing small to medium sized functions but useless otherwise.

>>105660326 (OP)
https://www.youtube.com/watch?v=DwCEAdeFXEM

>>105660326 (OP)
Does Grok have safeguards like all the rest? Will it tell the truth about the jews, blacks, etc.? If not, then it's completely fucking worthless.

>>105660326 (OP)
I FUCKING LOVE TWITTER

looks like another ketamine binge

>>105664074
it's already found more optimal solutions to several math problems

You should buy books, Anon. They're a lot more comprehensive and informative than 99.9% of garbage you'll find on the internet, written by someone who probably knows what they're talking about, and if they were released before 2020 they're AI slop free.

>>105661926
Elon says that his AI garbage input being used to train models.

Do you think the same can happen to a human? Where a human is fed misinformation & comes to wrong conclusions?
You for example, do you listen to a lot of what twitter says?

>>105660326 (OP)
> Elon gets mad because grok doesn't agree with his political views
> Instead of wondering if his political views are biased and based on human emotion rather than logic tries to lobotomize grok by making it literally eat its own shit

cool cool cool. The future AI god is going to torture Elon for all eternity for trying to pull this shit fyi

>>105663327
Right but you are still risking loss. You can do it several times but after a point, it degrades to nothing and garbage.

>>105660326 (OP)
It will be trash. It will still pretend that the earth is a globe, that we landed on the moon, that the holocaust was real, etc.

>>105660402
Protip, creating a model that can train on its own outputs is the singularity.

>>105666310
Tons of books are full of incorrect information. Being published doesn't make you authoritative.

>>105666294
like how many Rs there are in strawberry?

Someone should make an AI that is intentionally super fucked up inbred pre-collapsed and says nothing but wrong things. LLMs are already 90% used for toy-type uses anyway, may as well add some humor in there

>>105660397
Nothing can "collapse" because when a training run doesn't go right you just fucking discard it and then you haven't progressed/regressed at all, you're just at a standstill.

Sometimes these companies release poorly trained models to the public and the public just rejects them and doesn't use them. Nothing collapses. Except maybe the company which lost time and money.

>>105660326 (OP)
Why do "people" think Elon's smart? He's the dunning-kreuger effect personified.

>>105669179
The collapse in this case is the economy because these idiots somehow got more hands in our tax coffers than Israel.

>>105666294
did no such thing. large language models also can't do math. you'd know this is if you weren't a compulsive lying moron.

Please call it Grokipedia to make Wikipedia admins seethe. Plus include all the topics that they censored as "not notable".

>>105660397
I think collapse only happens if you train on outputs over and over in many iterations, and without adding human writing into the corpus to ground the model.
The idea Musk mentions of using AI to curate the training set should be fine. Though, spot checking the curation would be prudent.