Writing
Blog
Core Research
Three series, one system.
The work I’d point to first: an open research program on making A.I. safe, personal and fair — each series written alongside the open-source code it describes.
A System for A.I. Safety
How Rao Studios approaches A.I. Safety — media, text, LLM harnesses and human verification combined into one system. A five-part series, starting with media.
Started October 27, 2026
1 of 5 parts published
- 01Media Safety via Veil, Veil8, VeilGarde
- 02Text & Code Safety via AII, Braid (coming soon)
- 03Harnessing a Reduction in Hallucinations (coming soon)
- 04Verification within Human in the Loop Systems (coming soon)
- 051 System, all domains, safely secured by Rao (coming soon)
A Personal & Private Database for A.I.
Ambient Intelligence Index: a database that lives on your own machine and personalizes your A.I. A four-part series, starting with Thread.
Started October 25, 2026
1 of 4 parts published
- 01Thread is a User's Personal Database
- 02Sewn & Sinatra Personalizes Your A.I. (coming soon)
- 03App Records Govern Data (coming soon)
- 04Seam Connects & Decentralizes Retrieval (coming soon)
A 'Fair Use' Distributed Language Model
Personal language models that contribute to a larger one from our own machines — cited at the token level, no data-center required. A four-part series, starting with Braid.
Started October 23, 2026
1 of 4 parts published
- 01The Path Towards 'Fair Use' in Generative A.I.
- 02Ablations: Mamba, Falcon H1, & more (coming soon)
- 03A Framework Around Fair Use (coming soon)
- 04A Protocol to Decentralize Inference (coming soon)
All writing
Perspectives
When Federation Meets Mary
FATE federates the computation between institutions. Mary, the umbrella over AII, Braid, Craft, Ambient and Veil, starts at the device: a private sanctuary for the data, a model trained beside it, and federation on top. What changes if federation goes mainstream, and how Mary can do FATE's job from the device up.
November 8, 2026
Ambient Learning
What building Loom taught me about federation, the four principles I keep circling back to, and why the index Rao is building is ambient.
November 1, 2026
The Case for Personal A.I.
From ad retargeting to chat models that could steer without a user knowing: how the feed's persuasion problem carries into A.I., and a case for personal A.I. that protects identity, psyche, and privacy first, then grows only on your own craft.
October 1, 2026
Royalty Language & Diffusion Models
A year of prototypes — Seer on iPhone, a Chrome extension over all of arXiv, a payout ledger — converging on one layer that attributes every generation, text or image, to the people it draws from before it's served.
September 18, 2026
STTP
What if devices prepared semantic tokens on every creation and edit, shared them peer-to-peer, and a shared aggregation layer — not one model — predicted the completion? Notes on Product of Experts, PPM, and a decentralized alternative to monolithic frontier models.
September 7, 2026
Fair Use
Two words define when and how someone can use another's data. The phrase is invoked as a defense for training on the open web, but the balance inside it, the one that made the internet work in the first place, is the part that tends to get skipped.
September 5, 2026
Seer's Journey to Mary
How a project about the royalty problem in written works became a distributed storage layer, then a pair programmer, then Mary — told close to how it actually happened, in order.
September 4, 2026
AI Safety & Deepfakes
Authentic captures kept in a secure, invisibly watermarked space, a capture codec at the hardware level, and a public upload history: keep your mind local and let your sharing be trackable.
July 31, 2024
From First Push to Now
Loom began as an experiment with a Lemmy instance and grew into a client for aggregating any feed. Notes from 2023 on federation, owning your identity, and who decides the block lists.
September 5, 2023
Accessibility & LLMs
From Cleverbot to Pi, and Clarke, a 2016 mood-diary chat bot still missing a piece. Then the larger question: how language models could open new routes to sight and hearing for people who are blind or deaf.
July 5, 2023
ML Architecture
Agents You Can Read
California is writing the first statutory definition of agentic AI. Rao's answer is an open-source, on-device schema standard for agents: the counterpart to server-based MCP.
November 4, 2026
A System for A.I. Safety
How Rao Studios approaches A.I. Safety — media, text, LLM harnesses and human verification combined into one system. A five-part series, starting with media.
October 27, 2026 · 2 pages
A Personal & Private Database for A.I.
Ambient Intelligence Index: a database that lives on your own machine and personalizes your A.I. A four-part series, starting with Thread.
October 25, 2026 · 2 pages
A 'Fair Use' Distributed Language Model
Personal language models that contribute to a larger one from our own machines — cited at the token level, no data-center required. A four-part series, starting with Braid.
October 23, 2026 · 2 pages
VisionAX
A GPL, OpenCV-based engine that gives Mary a synthetic accessibility tree wherever the real one refuses to appear — vision as a fallback that never leaves the machine, not a replacement for AX as ground truth.
September 3, 2026
Linguistic State Machines
Why SKILL.md was a great demo and a poor harness — and what changes when you treat a language model as a state machine you have to program, not persuade.
August 19, 2026
Ambient Intelligent Systems
How three programs taught each other a paradigm — Mary acts, Seer reasons, Totem remembers connections.
August 16, 2026 · 12 pages
Engineering Culture
The New Silicon Valley
Silicon Valley was never a place, it was a set of defaults. A look at what a year of research on attribution, owned memory, accountable agents and safety at the source adds up to, what a valley built on receipts could look like once it's all shipped, and the ways it could still turn into the old one.
October 28, 2026
Kohinoor
The Koh-i-Noor disappointed London in 1851, so they recut the stone instead of fixing the light. The discourse about AI and engineering is making the same move on the people entering the field.
August 21, 2026
Adversarial ML
Veil & Scorpion
Detection alone is reactive by definition. Scorpion is meant to run as an upload-time memorization check on platforms like Hugging Face — a virus scan against a registry of protected people. Veil catches what that check can't: a model that can synthesize a likeness without ever having memorized it. CLR is the safety net for whatever shipped before the gate existed.
September 12, 2026
Attribution Decay
An MIT study measured that no single training image is necessary to a diffusion output, and the coverage concluded nothing is responsible. That is causation failing under overdetermination.
August 20, 2026
The CLR Journey
From a perspective-warp curiosity to an on-device provenance stack — my research log building a physics-first detector of AI-generated imagery.
August 16, 2026 · 8 pages
HCI
Ambient Abilities
Inside the Ability Studio: how Mary turns intent into Abilities, Skills, and Recipes — and a rehearsal tool that shows the router's work instead of asking you to trust it.
September 2, 2026
Ambient Clients
Meta and OpenAI both shipped Mac assistants that watch your screen. Mary answers the same three questions differently — and the answers are the whole architecture.
August 20, 2026