@cwebber @bkuhn @copiesofcopies

IMO even local models are currently unacceptable when they have been trained on unknown data, by untrustworthy people, using far more resource than any community will ever have access to, with a special privilege to ignore all IP considerations when it comes to accessing and storing their training set and using all web data that no one ever agreed to be used in such a way.

Training anything that way should still be regarded as "the largest ongoing intellectual property and labor theft" and ethically regarded as if in biology someone wanted to clone a human, but if someone creates a local model that fits on world-accessible consumer hardware (not just affordable to first world countries), trained on known data explicitly licensed for reuse (or public domain) maybe it will be time to look at the implications on society (in learning, community building, work for a start...) and collectively and democratically decide if it is worth putting resources in…