Skip to content
Demo with sample data. Nothing here is a real matter.Back to site →

On device model distillation

No plan yetNot yet filed · Disclosure sent Nov 21, 2026
All matters
We distill our server model into a student small enough to run on the phone. The trick is how we sample the teacher: by task, weighted by how often each task runs in the product, so the student keeps its accuracy where people actually use it. [attached: design_doc_v2.pdf]
Tao AI

I read your design doc. The task weighted sampling is the heart of it, and your benchmarks show the student close to the teacher on the five tasks that matter most.

Two questions:

  1. Are the task weights fixed, or do they change as usage changes?
  2. Has any of this been shown outside Northwind, in a talk, a paper or a demo?
The weights update every week from usage. Nothing is public yet. We may present it at a conference in the spring, date not set.
Tao AI

Your invention disclosure is written: all 12 sections, with a readiness score of 78. Two points are only partly covered, the date of that conference talk and benchmarks beyond your top five tasks; your attorney can go through them with you. When you are ready, send it to your attorney.

Please review the invention disclosure on this matter. I would like your view on filing.
Sent. Your attorney replies here and by email within one business day.

Invention disclosure

Software output, not legal advice
Invention summary
A method for distilling a server model into a device sized student by sampling the teacher’s outputs per task, weighted by how often each task runs in the product, so the student keeps its accuracy where it is used. The weights update weekly from usage. Inventor: Tomás Reyes.
Gaps to fill
The date of the planned conference talk, and benchmarks beyond the top five tasks.
AI generated responses should not be treated as legal advice.

AI generated responses should not be treated as legal advice.