Independent · New York
TenTwo Labs takes dense, unstructured material — coursework, scientific literature, experimental data — and gives it a structure you can actually work with. Our first product, Sumichi, is in beta now.
Most tools hand back prose about your material. We return the artifacts you work against — outlines, question sets, flashcards, mind maps — because understanding is built on structure, not on more paragraphs to read.
Correctness alone is a weak signal. We capture confidence alongside it, so the system can separate knowing something from guessing it — and so a learner sees precisely where they are wrong about being right.
Review timing is computed per item by a spaced-repetition scheduler (FSRS), not set by a fixed calendar. Different material decays at different rates for different people; the schedule should follow the evidence rather than the syllabus.
One person’s material is the easy case. The same operation at the scale of a whole field is the interesting one.
Many sources, claims that disagree, and evidence that accumulates until it occasionally overturns what came before.
That is where the work is heading: knowledge held as a structure with lineage rather than a pile of documents, so a claim’s origin, the evidence beneath it, the confidence it warrants, and the way it has shifted over time are all first-class. Nothing to announce yet.
Building on a model you cannot see inside is a guess. We would rather know.
Ongoing work in mechanistic interpretability studies how language models represent information internally, currently attribution graphs in open-weight models.
It is not a side interest. Software that declines to overstate what it knows has to be built by people who understand how these models fail.
For partnerships, press, and grant or program inquiries. Reaches the founder directly.
