Approach
More progress per GPU dollar.
Data efficiency is the lever the industry hasn’t pulled. We built Ancor to pull it.
The thesis
AI is compute × model × data. Two of the three have an anchor.
Compute has NVIDIA. Models have the frontier labs. Data — the term that decides what all that compute actually learns — has no equivalent. It is still gathered opportunistically, scraped, or synthesised from models that were trained on the same gaps.
The same GPU budget produces more progress when it is spent on data that contains more useful information and better matches the model’s current gaps. That is the whole company.
Data efficiency
Bulk data vs. targeted observation
Bulk
- More of what the model already knows
- Diminishing returns per token
- Synthetic errors compound into the next generation
- Measured by volume
Targeted
- Exactly what the model is missing
- Information-dense; moves the evals it was requested for
- Captured in the world, with provenance attached
- Measured by progress per GPU dollar
The mark
⊨ Entails.
Our mark is the double turnstile — the symbol for entailment. If the observations are true, what the model learns from them must be too. That is the standard every observation is held to.
Principles
How we work.
- Real over syntheticSynthetic data teaches a model what models already know. Observation carries information the model does not have yet.
- Targeted over bulkA gap is a specification. We capture against it, rather than hoping a larger pile happens to contain the answer.
- Continuous over staticA model’s gaps move as it trains. The observation stream should move with them.
- Provenance over volumeEvery observation can be traced to a place, a time, a method and a person who agreed to it.
- Consent by defaultPeople sense the world for AI knowingly. Consent is recorded at capture, auditable and revocable.
Who we are
A small team building the data anchor.
Ancor is built by researchers, engineers and field operators who have spent years watching good models fail on ordinary reality. We are working with a small number of AI labs and research teams. If you are training models and hitting the data wall, we would like to talk.
Next step
Spend the next GPU dollar better.
Thirty minutes to walk through your gaps and how we'd observe our way to closing them.