I do not believe anyone with real power at Google values liberal values above revenue. The RLHF and fine-tuning is solely to prevent news articles of the form "Google unveils new shockingly racist AI!!", which is absolutely what would happen without explicit fine-tuning, and which would invite regulation and scrutiny.
I too think artificially "fixing" models is the wrong way to go. If the models are biased towards racism it's because the training data is biased towards racism which is because society is biased towards racism. Which is true, and we'd be better served by acknowledging that and shining a light on it. Just ban AI (as a known tainted product) from being used for making any decisions of importance.
> The RLHF and fine-tuning is solely to prevent news articles of the form "Google unveils new shockingly racist AI!!", which is absolutely what would happen without explicit fine-tuning, and which would invite regulation and scrutiny.
Yes that is undoubtedly true, and is a great point. I'm not sure whether it just so happened that the revenue incentives lined up with the liberal values well enough that nobody ever questioned or pushed back, or if the revenue goals outweighed the liberal values, but my guess is it's probably more the former. Though once revenue and liberal values are in tension, it will be interesting to see which direction they go. My guess is it will be a mixture that leaves no clear trump card, and makes it very difficult to predict given situations.
I too think artificially "fixing" models is the wrong way to go. If the models are biased towards racism it's because the training data is biased towards racism which is because society is biased towards racism. Which is true, and we'd be better served by acknowledging that and shining a light on it. Just ban AI (as a known tainted product) from being used for making any decisions of importance.