Loeber.eth
@johnloeber
lowbackgroundsteel.ai Soon, the majority of online content will be AI-generated. But generated from what training data? Well, online content, of course. But if that’s going to be mostly AI-generated already then that’s not real “training data”, that’s just reinforcement… 1/5
1 reply
0 recast
0 reaction
Loeber.eth
@johnloeber
It’s ironic. The better LLMs become, the more LLM-generated content there is, and the less distinguishable it is from human-generated content. This proportionally diminishes human-gen training data (or pollutes it with AI-gen data) thereby slowing progress of new models! 2/5
1 reply
0 recast
0 reaction
Loeber.eth
@johnloeber
An attempt at a solution is to identify and secure as much genuine human-generated data as possible. That’s easy for past data. Stuff from the 1800s is the “low background steel” analogue of AI. But unlike steel, such old data has limited usefulness. Recent data is better… 3/5
1 reply
0 recast
0 reaction
Loeber.eth
@johnloeber
…since when we’re generating content, we care about the current state. We want to summarize text from today and answer today’s questions. Training on text from 1850 doesn’t suffice; that’s a big gap to the modern day. An LLM trained on data from 2000 won’t know what Uber is. 4/5
1 reply
0 recast
0 reaction
Loeber.eth
@johnloeber
The big question: as we go into a future of needing *present* training data, how can we guarantee the authenticity of it as human-generated? Avoiding feedback loops of AI training on AI-generated data seems important. Recursive reinforcement is dangerous in other areas. 5/5
0 reply
0 recast
0 reaction
