On step changes, and what we do between them.
- 01
Step changes are real, but they are not promises.
A jump in capability is a measurement, not a forecast. We respect the data when the curve breaks, and we refuse to extrapolate past the next checkpoint. The step has happened. The next one has to be built.
- 02
Between them, the work is compression.
After a jump, the frontier is wide and wasteful. The job is to find the smallest object that holds the new behavior. Compression is not a finishing step; it is how the capability becomes legible.
- 03
The encyclopedia is online; cognition runs on-device.
Trivia belongs in a database. Reasoning belongs close to the user. We move the thinking part down to silicon people already own, and leave the lookup part to the network.
- 04
Small enough to inspect. Sharp enough to ship.
A core you cannot read is a core you cannot trust. We aim for models that fit in a head and in a budget, and we sign every claim with a reproducible eval.