Anonymous at Thu, 21 Mar 2024 15:25:44 UTC No. 16089525

>>16089468
You understanding is all wrong. LLMs do not generate images, chatgpt creates a prompt and sends it to Dall-E which is a so called adversarial trained diffusion model. Dall-E is not an LLM so your whole point of view is moot.
It's bad at text because of the way it works. It creates an image of full noise, then randomly changes some pixels and then asks itself "does this look more like a cow now?" If you asked it for a cow. If it does look slightly more like a cow the random changes are committed, this continues iteratively till no significant signal gain is achieved.
This is a simplification of course, there is much more to it, especially regarding training where it is pitted against another neural network.

Anonymous at Thu, 21 Mar 2024 15:32:06 UTC No. 16089531

>>16089525
I don't want you to summarize it for me,
please comprehend my words.
>Can anyone point me towards any scientific papers that explain why LLMs have this problem?
save me the teacher act, and just give me papers please.

Anonymous at Thu, 21 Mar 2024 15:38:03 UTC No. 16089536

>>16089468
bumping for answers,
Please read my request,
- Do not summarize anything for me
- Do not tell me shit about my understanding
Just prove me papers so I can read up, save that other bullshit for your personal blog.

Anonymous at Thu, 21 Mar 2024 15:38:21 UTC No. 16089537

>>16089531
LLMs don't create images, baka.

Anonymous at Thu, 21 Mar 2024 15:56:53 UTC No. 16089556

>>16089537
No problem,
I came across what I was looking for no thanks to you or the other anon.

Anonymous at Thu, 21 Mar 2024 15:58:22 UTC No. 16089557

Mods or Admins, can you delete this thread? I've come across what I was looking for.

Dr. Soystein at Thu, 21 Mar 2024 16:51:05 UTC No. 16089619

>>16089537
>>16089468
What if you asked it to create an image in ASCII?