[alex43578]

Two things that narrow the “unthinkably vast input data”: 1) You’re already in the latent space for “AI representing itself to humans”, which has a far smaller and more self-similar dataset than the entire training corpus.

2) We’re then filtering and guiding the responses through stuff like the system prompt and RLHF to get a desirable output.

An LLM wouldn’t be useful (but might be funny) if it portrayed itself as a high school dropout or snippy Portal AI.

Instead, we say “You’re GPT/Gemini/Claude, a helpful, friendly AI assistant”, and so we end up nudging it near to these concepts of comprehensive knowledge, non-aggressiveness, etc.

It’s like an amplified, AI version of that bouba/kiki effect in psychology.