The seam you can't see
Attach your lecture PDF to a chatbot and ask a question. Some of the answer comes from your PDF. Some comes from what the model already believed about the topic — from a different textbook, a different notation, a different course, or nowhere at all. Both halves arrive in the same confident register, in the same paragraph, and there is no seam showing you where one ends and the other begins.
For most tasks that blending is a feature. For studying it is the whole danger, because you are asking about the one subject where you are — by definition — not yet able to spot the error. That is the asymmetry that makes a general assistant a bad fit for revision, and it is why Thaxas keeps the seam visible.
What Thaxas does that a chatbot structurally can't
- Cites the page. Every claim points at a page in *your* document. You can check it in a click.
- Refuses. A similarity floor stops generation before the model is ever called if nothing in your material is relevant. There is no answer to hallucinate because there is no generation.
- Marks the boundary. When an answer runs outside your material, the interface says so, visibly.
- Remembers your course, not the conversation. Your material persists in a space, so you are not re-uploading the same PDF every session and hoping the context survived.
- Schedules your revision. FSRS tracks your memory state per card across weeks. A chat thread tracks nothing.
Be fair: use ChatGPT for these
Explaining a concept that is *not* in your material and that you just need to understand. Brainstorming. Coding. Being talked through something from first principles when you don't care about your lecturer's particular framing. General reasoning of every kind. It is better than us at all of it, and using both is not a contradiction — it is just knowing which tool has the property you need right now.
_Written in July 2026 from publicly available information. These products ship constantly, so check their site before you decide — and tell us if we've got something wrong._