So I get to use the platform for free, but I also get paid to post on the platform? I'm not sure that makes sense. Like I hate to take the side of big tech, but they can't literally be paying users to use their platform. Just use something else, there are a million social media sites
Google indexes your website for free, and it will pay you to put ads in it.
That's also what all social media do , they put ads on your thoughts. They dont even need to index your thoughts because you submit them directly. It has nothing to do with being free, it's about incentives. Users are so foolish , they give everything for free, unlike webmasters.
Most posts are ignored and are an absolute loss to the company. Which is why platforms like Twitter only allow you to make money from posting once you reach a certain threshold.
They're not an "absolute loss" since they cost bytes to store, and raise engagement and data metrics.
It's just that they don't want to share the fractions of pennies with everyone, so the fractions accumulate for them.
Then they pay a bit to the higher tiers, so they create the illusion that X is a parallel income source, and gives the lower tiers something to aspire to.
Carrot and stick, or rather glass beads and the hope thereof.
I wanted to do some quick math on this idea- supposed we trained a vanilla transformer model from scratch, as GPT2/GPT3 was done- the number of seen input tokens is known perfectly, as is the sources of those training tokens (since then, everyone has either kept quiet about the sources post-Books3-fiasco, or have been finetuning on top of previous models making this more difficult of a calculation)
GPT-3 was trained on approximately 300 billion tokens.
An small sized technical textbook might contain something like... 130,000 tokens? (1 token ~= 0.75 words, ~100k words in the book).
Thus, say you wrote a textbook on quantum mechanics that was included in the training corpus. A naive computation of the fraction of your textbook's contribution to the total number of training tokens would be 300B/130K = 0.0000004333333333, or 0.000043%.
If our hypothetical AI company here reported, say $500M in yearly profit, if all of that was distributed 100% based on our naive training token ratio (notice I say naive because it isn't as simple to say that every training token contributes equally to the final weights of a model. That is part of the magic.) then $500M * 0.000043% = $215.
You could imagine a simpler world where it was required by law that any such profitable company redistribute, say, %20 (taking the 'anti-VAT' idea) back to the copyright holders / originators of the training tokens. So, our fictitious QM textbook author would receive a check in the mail for $43 for that year of $500M in revenue. Not great, but not zero.
Since then, training corpuses are much, much larger, and most people's contributions would be much smaller. Someone who writes witty tweets? Maybe 1/100th the length of our above example in am model with now 100x the training corpus.
So fractions of a penny for your tweets. Maybe that is fitting after all...
the payment would probably be based on the usage of that source in generating LLM output for the LLM user. This would probably require training a parallel network that connects LLM network nodes to sources. Then the activation of those nodes could be a surrogate for the contribution of the source
