From the blog

Deep dives on Verbalized Sampling, LLM behavior, and exploring the full probability landscape of AI models.

The Balance Is Cents

A top-up adds a fixed number of cents. A run subtracts the model price plus seven percent, rounded up.

Sharing stores an id on the prompt. The link, Explore, and the homepage all read that one column.

drift() writes the sampling instructions and parses the string that comes back. The model that runs is the one inside send.

A collection stores run ids. Open the shelf, and those runs are grouped back under the prompt that produced them.

The prompt asks a run to sum to about one. Models often miss, and the tiers still use the numbers as written.

A finished run is a map. Drift deeper puts one answer back in the box, so the next sample follows that branch instead of the original question.

Facet mode breaks one question into distinct angles, then answers each angle once. Use it when more responses would only be rewrites.

Not all models distribute their creativity the same way. Here's how to measure and compare output diversity across different LLMs.

A deep dive into the technique that unlocks the full probability distribution of an AI model, surfacing responses a model would normally suppress.