Not scalabale long term, especially once they start selling to clients. Clients will have privacy and security issues with OpenAI having full access (and logging history) of all their previous queries.

That is what adversarial data points are -- people have discovered that it is possible to "confuse" image models using attacks such as these. Take a normal picture of cat, feed it into an image model and it'll label it correctly and say - hey, I'm 97% sure there's a cat in this. Change a small number of pixels using some algorithm (say <1% of the entire image) - to a human, it will still look like a cat, but an image model now thinks it's a stop sign (or something equally unlikely) with 90%+ probability.

Yeah, I couldn't understand what this sentence means either.

Research in fairness is something you can do without too much compute. Small BERTs and stuff are things you can easily experiment with on even a single GPU

As reported by whom? The owner of the platform who has a vested interest in making sure his investment is successful?

Otoh, look at how fast Mastodon is growing, or how many advertisers are no longer on Twitter, or any other external metric and you'll have your answer.

Not to mention the extra $10B+ of debt that Twitter took on after the purchase - servicing that alone is a $10B affair.

Son in law? Sister in law?