I think most labs actively create synthetic data using existing model as part of the mix for the pretraining stage for their next model.
Would love to know exactly what the latest process is to keep slop out of training data.
I think most labs actively create synthetic data using existing model as part of the mix for the pretraining stage for their next model.
Would love to know exactly what the latest process is to keep slop out of training data.
I think everyone overblows the whole "AI is poisoning AI!" thing. It could be a problem but the genuine value in Reddit or any other human social media is honestly pretty low from my estimates. It's great for seeing how humans talk but in terms of 'nutritional' value for truth or answers... I am not sold. If I was choosing what to 'feed' AI, I wouldn't even bother with textual social media (besides Github / Gitlab / other source control)
There's way more value, if seeking out answers, in following the links to external sources, scraping books, and other sources that aren't "unwashed masses saying whatever they want".
> the genuine value in Reddit or any other human social media is honestly pretty low from my estimates. It's great for seeing how humans talk but in terms of 'nutritional' value for truth or answers...
> ...
> scraping books, and other sources that aren't "unwashed masses saying whatever they want".
The problem is there's a lot of knowledge that only exists as reddit comments, blog posts, or social Q&A.
You can put it in scare quotes all you want, doesn't stop you from sounding like Scrooge McDuck.
const isAiContent = (str) => str.includes('—');?
:)
Latest generation LLM's use en dashes instead of em dashes to avoid detection.
No, they don’t. But obviously GP was tongue–—in-–cheek.