Stanford Faculty Club · 2026
Humanity & AGI Summit 2026
Sanscritic founder Akash Saraf presented an emerging way to design AI for expert work — where the reasoning is the product, not a means to an end.
Today's LLMs are built for problems with one verifiable answer, while real expert work — across law, finance, robotics safety and medicine — is nuanced and context-heavy, and demands reasoning an expert can inspect, challenge and rework. Sanscritic treats reasoning as a structured debate: a metareasoner probes a tuned model from multiple task-specific perspectives, producing traces that are legible, addressable and forkable — like formulas in a spreadsheet.
What we covered
- The gap — how LLMs are trained versus what expert work actually demands.
- Amplification, not automation — why expert work runs on forkable reasoning.
- Owning your reasoning — if you don't own your AI's reasoning, you rent it and pay in expertise.
- The metareasoner — probes, not prompts; generated by AI; no answer datasets or reward models.
- Expert Foundational Models — small enough to deploy on a single GPU, trained to debate internally.
- The Reasoning Workbench — a Cursor-like engine for expert work.
