*Falsify reality.
Most of the idiocy surrounding the "chatbot peril" comes from conflating these things. If an LLM learns to predict that the pronoun token for "doctor" is "he", this is not a claim about reality (in reality doctors take at least two personal pronouns), and it certainly isn't a moral claim about reality. It's a bare consequence of the training data.
The problem is that certain activist circles have decided that some of these predictions have political consequences, absurd as this is. No one thinks it consequential that if you ask an LLM for an algorithm, it will give it to you in Python and Javascript, this is obviously an artifact of the training set. It's not like they'll refuse to emit predictive text about female doctors or white basketball players, or give you the algorithm in C/Scheme/Blub, if you ask.
All that the hamfisted retuning to try and produce an LLM which will pick genders and races out of a hat accomplishes is to make them worse at what they do. It gets in the way of simple tasks: if you want to generate a story about a doctor who is a woman and Ashanti, the race-and-gender scrambler will often cause the LLM to "lose track" of characteristics the user specifically asked for. This is directly downstream of trying to turn predictions on "doctor" away from "elderly white man with a kindly expression, wearing a white coat and stethoscope" sorts of defaults, which, to end where I started, aren't reality claims and do not carry moral weight.
I would prefer things were open, but I don’t think this is the best argument for that
Yes, operators trying to tame their models for public consumption inevitably involves trade offs and missteps
But having hundreds or thousands of equivalent models being tuned to every narrow mindset is the alternative
I would prefer a midpoint, I.e. open but delayed disclosure
Take time to experiment and design in safety, etc. also to build a brand that is relatively trusted (despite the inevitable bumps) so ideologically tuned progeny will at least be competing against something better, and more trusted, at any given time
But the problem of resource requirements is real, so not surprising that being clearly open is challenging