What I Got Wrong About Canadian-Specialized AI

For six months I was certain about the direction.

Seven days ago that broke.

The idea was clean.

Canadian institutions need AI they can trust, so build them a Canadian model: train it on Canadian law, regulation, finance, and government records.

Run it on Canadian servers.

Make it open, so anyone could check our work.

Teach it to operate in both English and French, the way this country actually does.

Ship a small family of models, and release the research we built along the way.

I thought I'd be writing about that all summer.

The idea wasn't wrong.

The way I was building it was.

I had confused the model with the system.

The category error

Here's the mistake in plain terms.

I thought "Canadian AI" meant a model that has memorized Canada - every law, rule, and convention learned in advance and stored in its head.

So the plan was to teach it all of that up front, like a student cramming for a closed-book exam.

What people actually need is closer to an open-book exam.

A capable, well-trained mind - plus a current, trustworthy library it can pull from at the moment a question is asked.

Those are two different things.

One is the model. The other is the system around it.

I had been pouring all my effort into the first, when the second was where the value lived.

The model is the reasoning - the part that reads a rule it has never seen and applies it correctly.

The library (in the trade we call it retrieval) holds the live, citable Canadian sources and hands the model the right pages when a question comes in.

And we let it use real tools, so it can check authoritative systems directly instead of recalling a frozen, months-old impression of them.

Here's what a week of staring at this taught me.

A model that has memorized Canada is not what Canadians will choose.

A genuinely capable model - handed a great Canadian library and the ability to use real tools - is.

The "Canadian" part people care about shows up at the moment of the answer, current and sourced, not baked into the model months earlier.

For serious institutional work, the open-book approach wins on every count that matters.

It stays current.

Canada's rules change constantly - new banking-AI governance rules are landing in 2027, and case law and government guidance shift all the time.

A memorized model is out of date the day a rule changes and doesn't know it.

A library you can update overnight.

It shows its work.

A bank, a hospital, or a government department doesn't want a confident-sounding paragraph.

They want the answer and the exact source it came from.

An open-book system produces that citation automatically.

A model answering from memory just asks you to trust it - which is the one thing these buyers will never do.

It keeps the model smart.

Cramming a powerful general model full of one country's fine print doesn't make it more Canadian.

It makes it worse at thinking.

You'd be spending real capability to memorize something a library gives you for free, fresher, with a source attached.

Why this is the bigger bet, not the smaller one

If the "Canadian" part lives in the library and not the model, then the same system serves any country.

Point it at Canadian sources, it serves Canadians.

Point it at Brazilian or German or Kenyan sources, it serves them.

The thing I'd framed as a Canadian moat was really a general capability with a Canadian configuration.

That matters more to us than it would to most companies, because of how we're built.

We're bootstrapped.

No venture capital, no runway funded by someone else's fund, no permission to spend two years on a beautiful thing that serves exactly one mid-sized market.

Revenue is oxygen - and a from-scratch Canadian model capped our market at a single country.

A capable model plus a library caps nothing.

The same work serves Canada first and then expands into every market with the same problem, which is all of them.

For a company that has to pay for itself, "the approach that also opens the rest of the world" isn't a nice-to-have.

It's the difference between a project and a business.

It also makes the work easier - which I should have weighted far more heavily than I did.

Standing on a strong open model means I inherit a frontier-grade mind and spend my effort on the parts that are actually ours: the quality of the library, the tools, the way we prove the answers are right.

Building the mind myself meant owning the hardest, most expensive part of the whole stack - to get a result the open world was about to hand me for free.

Less to build. More capable. A bigger market.

I was choosing the opposite of all three.

The tradeoff, said plainly

Every real decision costs something, and I want to name this one honestly.

Focusing on the best possible general model means the dedicated French and Quebec work has to wait.

Building a model that nails Quebec's legal-French conventions as a specialty was on the near-term list.

For now, it's deprioritized.

I don't love writing that.

French is something I care about personally, and I'm not letting it go.

But here's why it isn't the loss it sounds like.

When you build a genuinely capable model with real character - which is exactly what we're doing with Vinci - it comes out strong in nearly every major language.

French included.

You don't earn good French by specializing for French.

You get it as a consequence of building a good model.

So French doesn't disappear from the plan.

It arrives a different way: a fluent, capable model, plus the actual French-language sources in the library it reads from.

The polished Quebec legal register comes later.

General French fluency is there from the start - for free - because we're building the model right.

What's preserved - and now belongs to everyone

Two things I built for the old plan don't just survive the pivot.

They matter more now, and they're going out to the world.

The Canadian Bilingual Legal Corpus and CBLRE - our way of measuring legal-AI quality - were going to be training fuel.

Used once, melted into the model, invisible afterward.

Under the new approach, a body of legal text and an honest way to score legal-AI answers are exactly the open-book library and the report card.

They're load-bearing, not leftovers.

So we're releasing them as what they always should have been: open research projects, free for anyone in the world to use, test, and build on.

A legal corpus that spans two languages and two legal traditions, and a straight way to measure how well an AI actually handles legal questions.

These are among our first gifts to the community.

A researcher building legal AI anywhere - not just in Canada - can pick them up and run.

That contribution is worth far more to the world as an open gift than it ever would have been buried inside one company's model.

The bet

What we're launching on August 8 is one thing, not a lineup with two years of promises attached.

It's called vinci-studio: a strong open model, a Canadian library it reads from, the ability to use real tools, a bit of character tuning, and a public bundle anyone can use to confirm it does what we claim.

The rest is roadmap, not promise.

I'd rather under-promise the lineup and over-deliver on the one thing than the reverse.

It's a smaller launch than I first announced.

It's a more honest one, it's built the right way, and - for a company that has to fund itself - it's the version that can actually pay for itself.

The lesson

Two of them, and the second is the one I actually needed.

The easy one: hold your thesis as a question, not an answer.

"A Canadian model" was my answer.

The real question - what gives serious institutions AI they can trust and verify? - kept evolving while I defended the answer I'd already announced.

From now on I write the question down at every big decision, separate from whatever answer I'm working with.

When the answer shifts under me, the question is still there to steer by.

The hard one: don't make the machine memorize what it could simply look up.

The most powerful-sounding option is rarely the right one.

I spent two weeks longer than I should have on the harder, more expensive path because changing course in public felt costly.

It is costly - but that cost is paid once, up front.

The cost of building the wrong thing is hidden, and it grows every day you keep going.

I'm choosing the upfront cost over the hidden one from here on.

What I'm watching

The ground is moving fast.

Open models get better every month.

The rules keep shifting.

The research I'm trying to keep up with ships faster than I can read it.

My job is to keep listening, and to change course early and out loud rather than late and quietly.

The pattern I most want to catch next time it forms: mistaking the thing I announced for the thing I was actually trying to do.

Vinci ships August 8.

That's what I'm building now - seven days into the turn, writing from inside it instead of after.

We'll see in six weeks how it lands.