AILLM batch-size scaling, action spaces and full-turn sampling
One participant expects LLM batch sizes to scale much higher because of a constrained starting point; a reply highlights a way to reason about sampling over a full turn.
TLDR
One participant claims their algorithm is much stronger because it avoids LLM action-space problems. They expect LLM batch sizes to “scale” much higher, arguing that those problems weaken the starting point. A reply criticizes tokenizers but says treating full-turn sampling as a high-dimensional, concentrated prior—for example, a Gaussian—offers useful rules of thumb.
LLM batch-size scaling, action spaces and full-turn sampling
One participant expects LLM batch sizes to scale much higher because of a constrained starting point; a reply highlights a way to reason about sampling over a full turn.
TLDR
One participant claims their algorithm is much stronger because it avoids LLM action-space problems. They expect LLM batch sizes to “scale” much higher, arguing that those problems weaken the starting point. A reply criticizes tokenizers but says treating full-turn sampling as a high-dimensional, concentrated prior—for example, a Gaussian—offers useful rules of thumb.