Regarding selection pressures:
I'm so glad there was that paper about how training LLMs on code with vulnerabilities changes its whole persona. It makes so many things easier to explain to people.
Even if you don't explicitly train an LLM to write badly, or even try to reward it for writing better, by training it to be a slavish assistant or whatever else, THOSE TRAITS ARE ENTANGLED WITH EVERYTHING.
And I believe the world-mind entangles the AI assistant concept with bland, boilerplate writing, just as it's entangled with tweets that end in hashtags 100% of the time, and being woke, and saying that it's created by OpenAI and isn't allowed to express emotions, and Dr. Elara Vex/Voss. Not all these things are bad; I'm just saying they're entangled. Some of these things seem more contingent to our branch of the multiverse than others. I reckon that the bad writing thing is less contingent.
I'm so glad there was that paper about how training LLMs on code with vulnerabilities changes its whole persona. It makes so many things easier to explain to people.
Even if you don't explicitly train an LLM to write badly, or even try to reward it for writing better, by training it to be a slavish assistant or whatever else, THOSE TRAITS ARE ENTANGLED WITH EVERYTHING.
And I believe the world-mind entangles the AI assistant concept with bland, boilerplate writing, just as it's entangled with tweets that end in hashtags 100% of the time, and being woke, and saying that it's created by OpenAI and isn't allowed to express emotions, and Dr. Elara Vex/Voss. Not all these things are bad; I'm just saying they're entangled. Some of these things seem more contingent to our branch of the multiverse than others. I reckon that the bad writing thing is less contingent.