PRIMER

Mission

Our goal is
superintelligence.

We state it plainly because we mean it literally.

Primer exists to build intelligence that exceeds human capability across broad domains. Not by assembling ever-larger frozen models, but by creating systems whose intelligence continues to develop after training ends.

We believe continual development may be one of the mechanisms required to get there. Primer is our experimental platform for studying that path.

The research program

The Intelligence Bootloader.

Every computer needs a bootloader: the small program that starts everything else. We are searching for its equivalent in intelligence.

The smallest set of developmental conditions sufficient to produce a system that can reliably acquire genuinely novel knowledge through observation, experimentation, reasoning, and feedback.

Most AI development optimizes what a model knows when training ends. The bootloader program optimizes what a system can come to know afterward. We are not trying to load intelligence into a machine. We are trying to find the conditions under which intelligence starts loading itself.

The central hypothesis

There exists a smaller, more structured developmental path than web-scale pretraining that can produce a system able to continually acquire, retain, combine, and apply concepts that were absent from its original training.

Falsifiable by design

What we optimize

Validated novel knowledge acquisition per unit of compute.

Not benchmark scores. Not parameter count.

The path

From first concept
to compounding intelligence.

  1. 01
    Learning

    Can the system acquire knowledge generated after its original training?

    Demonstrated in internal evaluations
  2. 02
    Memory

    Can verified knowledge persist through later learning?

    Demonstrated in internal evaluations
  3. 03
    Composition

    Can previous learning become a useful building block for harder new learning?

    Demonstrated in internal evaluations
  4. 04
    Development

    Can earlier learning make future learning more powerful, autonomously?

    Demonstrated in internal evaluations
  5. 05
    Generalization

    Can the learning architecture extend across broad domains and concept classes?

    Active research
  6. 06
    Compounding intelligence

    Can learning and improvement compound beyond human capability across broad domains?

    Mission

Where we are

What we are not claiming.

Primer has not demonstrated AGI or superintelligence, and is not generally smarter than frontier models.

What we have demonstrated, in internal evaluations, is narrower and, we believe, more important than it sounds: a small model that can acquire knowledge generated after its training ended, retain it, and build on it — including multi-step chains where earlier self-acquired knowledge makes later learning both possible and cheaper.

The distance between that and superintelligence is enormous. Our bet is that it is a distance of development, not just of scale, and that it can be crossed one measured, falsifiable step at a time.

The mission is superintelligence.

The research standard is evidence.

PRIMER

Building intelligence that compounds.