Hard questions
The questions we must answer well.
Buyers, investors and partners ask these in the first meeting. Short, honest answers, the same from every founder.
Every founder gives the same answers. We change them only together.
- 01
Why not just use ChatGPT Enterprise or Copilot?
Keep them for email and office work. For the core work, on data that cannot leave, a rented assistant is generic, priced per seat and changed on someone else’s schedule. Eigna takes that core work, and the client owns the result.
- 02
Model makers already train custom models.
As training projects for the largest firms, with embedded scientists or about $100,000 a year before compute, and no gated loop sold as a product. We deliver and run owned models for the mid-market too, and they keep learning after go-live. Those projects prove the demand.
- 03
Open models will make this a commodity.
We hope so. Better open models lower our cost, and we re-base on the best one, through the gate. What does not become a commodity is the client’s data, the sealed test set, the release record and our Formula for the field.
- 04
Is this a services business?
It starts as a project and becomes a running fee. Formulas make every install in a field faster than the last, and certified partners carry the rollout work.
- 05
Can a young company deliver to a bank?
One team, one task and one Test Lab first. Designed for a signed installer, keys held by the client, a Model Deed and rollback in 60 seconds. Cleared partners where the work requires them.
- 06
How does a model keep learning on data that cannot leave?
It learns only from work the client approves. With Local, an encrypted training snapshot goes to a GPU we rent in a region the client approves, and is deleted after the run. Every new version must beat the current one on a sealed test set.
- 07
Everyone is moving AI on premises and onto devices.
Yes, and it helps us. The labs offer private copies of their own model, released on their schedule. Device makers sell a machine with the same open model for everyone: it remembers, it does not learn your work. They sell where AI runs. We sell whose it is.
- 08
Why should anyone trust your numbers?
Because we publish the losses next to the wins. Benchmark 001 shows a 17.8% cut in forecast error and a baseline we have not beaten yet: our own gate would not release that version. If a Test Lab does not beat the client’s baseline, the client keeps the report.
- 09
Why would you build hardware?
To make owning visible. The first boxes are ours, built from standard parts in our own case, so people can see and test what owned means. At scale, hardware partners build them; we keep the models, the loop and the proof.
- 10
Your first client shares your founders.
Yes, and we say so up front, with a related-party contract. It lets us run the loop on real trading work from day one. The next cases come from outside, through Test Labs.
- 11
Everyone says own now.
Yes. Owning the weights is the frame, not the proof. The proof is the loop: a model that keeps learning its approved work and is released only when it beats the last version, run for the client.
- 12
Clients can leave with the weights.
Yes, and they should be able to: the weights, the test set and the release record, with the Model Deed. They stay for the next release. What they cannot download is the Formula and the team that runs it.
Start
Ask us one more.
Thirty minutes with a founder. Same answers, your questions.