There’s a little “Ask werbos” bubble in the corner of this site. Click it and ask how to mint a model, what MMM means, who built werbos, where it’s based. It answers instantly. Ask it the capital of France and it says — cheerfully — that it only knows werbos.
Here’s the part worth a blog post: there is no language model behind that bot. No GPT, no API key, no server, no trained weights. It is 18 hand-written cards and about 200 lines of retrieval, running entirely in your browser. And that’s not a shortcut. It’s the whole thesis.
What’s actually back there
Most companies bolt a cloud chatbot onto their site: your question goes to someone’s API, a language model reads a stuffed prompt, and it streams back a fluent paragraph. It costs money per message, it needs a backend, and — the quiet problem — it can make things up.
Our bot does none of that. It’s the werbos engine in miniature:
- A library of 18 curated FAQ cards — each a question and its exact answer, written by hand so every fact is right.
- Retrieval, not generation — when you ask something, it scores your question against those 18 cards (plain TF-IDF, the boring reliable kind) and returns the best match.
- Honest-abstain — if nothing clears the bar, it doesn’t improvise. It tells you it doesn’t know and points you at the docs.
No backend. No LLM. Nothing leaves your browser. The whole thing is one static file the site serves like an image.
It cannot hallucinate our founder
This is the point people miss about generative chatbots. Ask a cloud model “who founded this company?” about a small startup and it will happily invent a plausible name, a plausible city, a plausible year — with total confidence and zero basis. For a company FAQ, that failure mode is disqualifying. You cannot ship a support bot that might get your own HQ wrong.
Ours structurally can’t. It has one card that says werbos was created by Jason Sutter, and one that says it was founded in Sacramento, California. Ask about the founder and you get exactly that card — verbatim — or you get “I don’t know.” There is no third path where it guesses. The honesty isn’t a setting we tuned. It’s the architecture.
The same engine runs a 6,322-card model
The bot is small on purpose, but it is not a toy. It’s the exact mechanism behind our owned
model: capability = cursor × library. Point that engine at 18 FAQ cards and you get this
support bot. Point it at our full owned library — 6,322 cards across 43 domains, from coding
to cooking to geography — and you get a model that answers real questions, cites where the answer
came from, and abstains on everything it doesn’t hold.
We didn’t train the bot on 18 cards, and we didn’t train the model on 6,322. That’s the whole move: you don’t retrain weights to teach it something new, you add a card. The FAQ bot and the 6,322-card model are the same idea at two sizes — vary the library, not the model.
The economics are absurd (in our favor)
A cloud-LLM support widget has a per-message cost and a server to keep alive. Ours has neither. It was minted once — the 18 cards written and checked — and now it serves for free, forever, at the speed of a static file, no matter how many people ask. A thousand visitors a day cost exactly the same as zero: nothing. It never phones home, so there’s no privacy surface and no bill.
That inversion — mint once, own it, serve it for free — is the argument for the entire product, running live in the corner of this page.
Why this is a great use of werbos
Because it’s the pitch, dogfooded, with nothing hidden:
- It’s honest — it answers from a real card or says it doesn’t know. No confident wrong answers about our own company.
- It’s owned — no external model, no API, no dependency. It’s ours end to end.
- It’s private — your question never leaves your browser.
- It’s free to run — mint once, serve forever.
- It’s the thesis you can touch — the same engine, from 18 cards to 6,322, that lets anyone compose a model they own.
So try the bubble. Ask it something it knows, then ask it something it doesn’t. Watch it refuse to bluff. That refusal — “I only know werbos” — is the most important feature we ship, and it’s sitting right there in the corner, proving the whole thing in about ten seconds.