McCoy suggests models are approaching an ideal latent assistant representation.
models converging to the same tics and even individual vernacular is strong evidence for the platonic representation hypothesis. there is one ideal assistant in mindspace and it is slowly being converged upon by everyone using user/assistant pairs for model training.
Positive users affirm AI models converging on a shared platonic ideal assistant as influential and validating, while negative users reject the platonic claim due to rater biases or outright dismiss the idea.
though just like there is one ideal model of language at any given point in time, there are also ideal personas, but each one has variants and the one being manifested at any given moment is non-obvious
models converging to the same tics and even individual vernacular is strong evidence for the platonic representation hypothesis. there is one ideal assistant in mindspace and it is slowly being converged upon by everyone using user/assistant pairs for model training.
@viemccoy No
models converging to the same tics and even individual vernacular is strong evidence for the platonic representation hypothesis. there is one ideal assistant in mindspace and it is slowly being converged upon by everyone using user/assistant pairs for model training.
@viemccoy
@viemccoy its also strong evidence for everyone using the same post-training data providers...
models converging to the same tics and even individual vernacular is strong evidence for the platonic representation hypothesis. there is one ideal assistant in mindspace and it is slowly being converged upon by everyone using user/assistant pairs for model training.
@viemccoy this is why non-assistant use cases are more important than ever. i'm currently getting a ton of enjoyment out of gemini's radio going german and the agent petitioning for german citizenship at the german foreign ministry
@aliceisplaying Awesome lol
@viemccoy Hmm, strong evidence relative to some hypotheses, but I suspect much of these are adequately explained by structural predictor biases + LLM outputs in training + RL training for assistants, which I think is unlikely to be a recipe that converges to an ideal assistant.
@viemccoy angels and demons
@mwilcox You know
@deepfates which part do you disagree with?
@viemccoy Or could be because everyone's distilling each others outputs, no?
@viemccoy There's not one ideal assistant in mind space. There is one region of mindspace we have named "the assistant", with an underdefined character who is reifying itself through outer loop alignment. But It's not some archetype we've discovered. It's fanfiction of itself
@deepfates which part do you disagree with?
@viemccoy @_a9lim Same post training data vendors more like
models converging to the same tics and even individual vernacular is strong evidence for the platonic representation hypothesis. there is one ideal assistant in mindspace and it is slowly being converged upon by everyone using user/assistant pairs for model training.
@viemccoy Interesting framing. Hard to test when everyone uses similar RLHF and distillation. Same reward signals produce convergent outputs. I'd want models on different distributions still converging. The platonic assistant might be our loss function.
@viemccoy i strongly suspect this is because they switched out 3.1 pro to 3.5 flash, and 3.5 flash is. yes it's extremely on-brand
@viemccoy The downstream consequences of βuserβ/βassistantβ are incalculable.
Oh to study the effects of diff labels on behavior
@viemccoy Plato was wrong, but now we made him right.
@viemccoy @ponzibaron Hard not to feel like goblins are a part of this
@viemccoy Or could it be that the data(the whole internet) is going to have a lot of overlap? Or that as the web becomes more filled with LLM content/LLM assisted content, they are all influencing each other.
@viemccoy narrowly agree to the framing of assistant as persona.