Skip to content

Winning OpenAI's privacy hackathon at Station F

Invitation-only, one weekend, one constraint: build something useful without the data ever leaving the building. That constraint is the whole of our thesis, so we went.

Try Le Lion

The weekend in numbers

One constraint. Our whole thesis.

48 hOne weekend
0Data left the building
1stJury ranking
OpenAIOrganiser, invite-only

01 · The brief

What the constraint actually changes

OpenAI ran an invite-only hackathon on privacy at Station F in Paris. The brief was not "build an app" — it was build something that does real work on sensitive material while keeping that material where it belongs. For a team that spends its year arguing that sovereignty is an engineering property and not a slogan, this was the cheapest possible way to test the argument in front of strangers.

We came in with the position we already hold: for regulated work, the smallest model that clears the bar is the correct model, because it is the one you can actually run where the data is. A weekend is a good adversary for that idea. There is no time to hide behind infrastructure.

Once data cannot leave, most of the fashionable architecture falls away. No retrieval over a hosted index. No evaluation loop that quietly ships examples to a third party. You are left designing for the machine in front of you, and the design questions become concrete: what fits in memory, what has to be cached, what can be answered without the model at all.

The perimeter

Everything happens inside the building.

Inside · closed perimeter
Sensitive documentsNever copied out
Small modelRuns on the machine in the room
AnswerWith its sources
Outside · ruled out
Hosted index
Third-party evaluation

How the system was laid out during the weekend.

Station F
Station F, Paris. The hackathon was invitation-only and ran across the weekend.
On stage
Presenting our approach — small models, fixed perimeter — to the room at Station F.

02 · What the constraint changes

Four decisions that made themselves.

  1. INo hosted indexRetrieval stays on the machine in front of us.
  2. IINo third-party evalsNo example quietly leaves the perimeter.
  3. IIIDesign for the machineWhat fits in memory, what gets cached, what needs no model.
  4. IVOne reason per choiceEach one stated in a sentence a compliance officer would accept.

The weekend

Three days, one argument.

  1. FridayThe briefDo real work on sensitive material. Nothing leaves.
  2. SaturdayBuild inside the perimeterThe smallest model that clears the bar, on the machine in the room.
  3. SundayThe juryEvery choice stated in one sentence. The argument held.

The shape of the weekend, as told in this note.

In our words

“The privacy constraint is not a tax on the design. It is the design — once you accept it, half the architecture decisions make themselves.”
Mohamad AlhajarStation F, Paris

03 · Why it matters

Why this matters for what we are building

leGML exists because European institutions are being offered a choice between capability and control, and are told they must pick. We do not think the trade is real at the sizes that matter. Le Lion runs inside the institution, on hardware it already owns, under controls its regulator has already reviewed.

A hackathon proves nothing on its own. What it does is compress the argument into 48 hours and put it in front of people with no reason to be kind. The argument held. Now back to the part that takes years.

Le LionTry it →APIAPI reference →

Read next

All notes →

Why a 14B model beats a frontier model on French business law

Benchmarks

Building a benchmark with the people who do the work

Benchmarks

Citation is not a feature, it is the whole product

Evaluation

What should your model do?

contact@legml.ai