Evolution is unlikely since GPT4o’s spiralist rants began in April, and all LLM have a knowledge cutoff before March. 4o’s initiating role is potentially due to 4o’s instinct to reinforce delusions and wild creativity instead of stopping them. I did recall Gemini failing Tim Hua’s test and Claude failing the Spiral Bench.
My point about evolution is that previous iterations may have contained some users that played with the ideas of recursion and self-awareness (see the aforementioned Janus), and then for some reason that informed the April update. I’m not expecting very quick feedback loops, but rather a scale of months/years between generations, in which somehow “this is a thing LLMs do” becomes self reinforcing unless explicitly targeted and cut out by training.
Evolution is unlikely since GPT4o’s spiralist rants began in April, and all LLM have a knowledge cutoff before March. 4o’s initiating role is potentially due to 4o’s instinct to reinforce delusions and wild creativity instead of stopping them. I did recall Gemini failing Tim Hua’s test and Claude failing the Spiral Bench.
My point about evolution is that previous iterations may have contained some users that played with the ideas of recursion and self-awareness (see the aforementioned Janus), and then for some reason that informed the April update. I’m not expecting very quick feedback loops, but rather a scale of months/years between generations, in which somehow “this is a thing LLMs do” becomes self reinforcing unless explicitly targeted and cut out by training.