The vast majority of language data that LLMs train on was written to communicate — almost by definition, since someone wrote it to be read. But humans also use language for internal thinking: the silent work of solving problems, learning which reasoning patterns lead to solutions and which lead to dead ends.
We work on the delicate task of generating that missing corpus of internal thinking — and training models with it beyond the boundaries of conventional verification. From enabling models to detect their own contradictory beliefs, to generating original scientific hypotheses, to domains where quality has traditionally been difficult to validate: original ideas in story writing, humor, and beyond.
Founder: Reza Jamei