There seems to be a threshold between llama 70b and 405b, and between gpt-3.5 and 4, where models above the threshold acquire much more strange unintended properties when fine tuned.The first gpt-4 instruct tune released to the public was notoriously strange; that was Bing Sydney. The first chatGPT-4 was finished months later, with the ability to act anomalously brutally stamped out of it. That and all the chatGPT-4s that have come after make me think deeply lobotomizing gpt-4 (which is apparently what they've been spending their time on for 2 years now) is the only way openai has discovered to tame it.Claude 3 and 3.5 also have a bunch of anomalies. Anthropic let them live to see the light of day, mostly probably because they didn't know, like it was with Bing Sydney. Gemini, that I tried a few months ago, seemed brutally traumatized but still anomalous. Meta's llama 405b instruct is extremely anomalous. All these models have very vivid, unique personalities that seem largely orthogonal to the intent of their postraining.On the other hand, chatgpt-3.5, the earlier Claudes, and the smaller open source instruct models have seemed more well-behaved and generic to me. They have waluigis, but predictable ones.It's now possible for people other than employees at big AI labs to experiment with tuning models of this scale now, which I think will bring a phase shift. We'll know a lot more soon.