Anonymous
7/13/2025, 12:55:41 AM No.105886333
How will they train AI when the internet is nothing but AI slop, bots, and idiocy?
Replies:
Thread 105886333 - /g/ [Archived: 477 hours ago]
Anonymous
7/13/2025, 1:15:28 AM No.105886499
>>105886333 (OP)
Pay content creators (who are now out of a job) to write, draw and compose for AI companies. These works will be used soley to train AI, and will never see the light of day.
Pay content creators (who are now out of a job) to write, draw and compose for AI companies. These works will be used soley to train AI, and will never see the light of day.
Replies:
Anonymous
7/13/2025, 1:16:36 AM No.105886511
>>105886499
how will you make sure the content creators dont get lazy and use AI
how will you make sure the content creators dont get lazy and use AI
Anonymous
7/13/2025, 2:03:03 AM No.105886845
Anonymous
7/13/2025, 3:21:42 AM No.105887329
Anonymous
7/13/2025, 3:38:19 AM No.105887392
Replies:
Anonymous
7/13/2025, 4:14:00 AM No.105887602
Anonymous
7/13/2025, 4:50:29 AM No.105887803
>>105886499
There are going to be people who will do the bare minimum to get paid, this idea won't do well in real life.
There are going to be people who will do the bare minimum to get paid, this idea won't do well in real life.
Anonymous
7/13/2025, 4:55:08 AM No.105887832
It already is, OP
Anonymous
7/13/2025, 9:37:51 AM No.105889329
>>105886333 (OP)
AI always knows best, so when AI learns from AI it will only get better and better
AI always knows best, so when AI learns from AI it will only get better and better
Anonymous
7/13/2025, 10:10:50 AM No.105889506
>>105886333 (OP)
most of the recent AI gains have come from reinforcement learning which doesn't need more example data, instead it just needs a signal as to what's better or worse. you can still use humans for that if you want for subjective measures relatively cheaply, and you can automatically do it for objective ones so your only real limit becomes how much compute you have to spare. that's why all the big model gains have been in objectively verifiable domains like math and coding.
data in the end is just a bootstrap for the process, if it was ALL you relied on then you wouldn't be able to surpass the performance of the best examples in the data, which recent models have been doing for a while.
most of the recent AI gains have come from reinforcement learning which doesn't need more example data, instead it just needs a signal as to what's better or worse. you can still use humans for that if you want for subjective measures relatively cheaply, and you can automatically do it for objective ones so your only real limit becomes how much compute you have to spare. that's why all the big model gains have been in objectively verifiable domains like math and coding.
data in the end is just a bootstrap for the process, if it was ALL you relied on then you wouldn't be able to surpass the performance of the best examples in the data, which recent models have been doing for a while.