- Entry date
- 15 April 2025
- Category
- Models
- Lead over the world
- ~1 year ahead
- Access
- 🔓 Public
We fine-tuned a model on one person's own browsing, entirely on their own machine, and asked it questions — a private AI, a year before that became the pitch everyone gives.
- LoRA
- local fine-tune on consumer hardware
- 0 bytes
- personal data leaving the device
- RAG > FT
- the honest result for personal data
Honest evaluation
The end-to-end private, on-device fine-tune + RAG system worked as built, and the comparative finding — retrieval beats fine-tuning for personal, freshness-sensitive data — is a real, measured result, not an aspiration.
What would prove or disprove it further
What would disprove or narrow it further: a personal corpus large enough, or static enough, that fine-tuning's frozen-snapshot cost stops mattering and its speed or personalization benefits outweigh RAG's freshness advantage — the report doesn't test that boundary, so the RAG > FT finding should be read as bounded to "personal, frequently-changing data," not as a universal claim about fine-tuning versus retrieval.
The evidence — full reasoning behind the verdict
Verdict: proven. Two claims sit inside this report, and both hold up. The first is that a fully private, on-device personal AI — capture your own browsing, fine-tune a small model locally with LoRA, query your own history, nothing leaving the machine — was buildable in early 2025, a year ahead of when that pitch became common. It ran, end to end, with every part present: extension, local fine-tuning pipeline, retrieval path, and a local app tying it together.
The second claim is the more valuable one, and it's a negative result the report doesn't soften: for personal, freshness-sensitive data, retrieval beat fine-tuning. The reasoning is stated plainly — a fine-tune is frozen the moment it finishes, while your data changes daily; retrieval can point at its source rather than half-remember it; and LoRA on consumer hardware is slow and prone to overfitting a small personal corpus. That's a genuinely falsifiable comparison, and it was actually run, not assumed. Expecting the fine-tune to be the magic and finding the retrieval was the workhorse is exactly the kind of result that only counts as proven because it was allowed to go against the initial expectation.
Runnable proof — see it work
The pitch that every AI company now gives — an assistant that knows you, and keeps your data private — was, in early 2025, still an experiment nobody had shipped well. We built one to find out what was real about it and what was marketing.
What it was
MyBrain was a privacy-first personal knowledge system. A browser plug-in captured your own browsing; a small language model was fine-tuned on it locally; and you could then ask natural-language questions of your own history. The constraint that made it interesting was absolute: nothing left the device. No cloud, no upload, no third party holding your life.
What we built
A working end-to-end system with every part present:
- A Chrome extension that captured browsing data on-device.
- A local LoRA fine-tuning pipeline (Python, transformers, PEFT) that adapted a small model to that personal corpus.
- A retrieval (RAG) path for asking questions against the same data.
- A Node/Express + React app tying it together, with SQLite for local storage.
It ran. You could fine-tune on your own browsing and query it, with the model and the data both sitting on your own hardware. That was the whole thesis — ahead of the curve on personal, private AI — and it held.
What we learned — including the honest negative
The most valuable finding was a negative one, and it changed how we build.
That lesson — prefer a system that can cite what it knows over one that has memorised it — is a straight line to the calculated-mind work, where the whole design goal is a model that reasons over knowledge and cites it, rather than bluffing from baked-in weights.
Where it went / status
Archived as a research prototype, its lessons absorbed. The private, on-device, "your data never leaves your machine" principle became a founding commitment rather than a feature — it is the reason the north-star product puts a model that runs on your own hardware at the centre. We proved the private personal AI was buildable in 2025; we also proved which half of the popular recipe actually carries the weight.
What is still open — kept visible
The honest edges, next to the wins. This is what turns 🔬 into 🟢 — honestly.
- LoRA fine-tuning on consumer hardware has real limits — it is slow, fiddly, and easy to overfit on a small personal corpus.
- For most personal-data questions, retrieval (RAG) beat fine-tuning on both freshness and honesty — the fine-tune's value was narrower than we hoped.
- Capturing browsing is a serious privacy surface; keeping it local is the whole point, and it constrains how much you can do.
Proofs & sparks
We demonstrate rather than assert. Each ✅ proof is a visible result with a hard figure.
- On-device LoRA taxonomy + hot-swapdesign-stage taxonomya four-kind on-device LoRA taxonomy with lock-free hot-swap and structural user-LoRA privacy — thematically matched to the on-device personal-AI work, design-stage.
Where this connects
Sources
- vocabotics project audit — MyBrain (local fine-tune + RAG), Apr 2025vocabotics internal project history · as of April 2025
- LoRA: Low-Rank Adaptation of Large Language Modelsthe public fine-tuning technique the experiment used · as of 2021