3 ms·

> Yes, operators trying to tame their models for public consumption *Falsify reality.

• lupire 3y ago
  Curate the false reality. The model falsifies realty by inherent architecture, before any tuning happens.
• samatman 3y ago
  LLMs have nothing to do with reality whatsoever, their relationship is to the training data, nothing more. 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.