Results
What Primer has shown so far.
Four capabilities, each a step up from the last: it learns, it remembers, it combines, and it builds on itself.
Every result below starts with what it means, then gives the number, how we measured it, and where it stops.
Internal research benchmarks. External validation is ongoing. These results do not demonstrate AGI or superintelligence.
01 / Learns
It learns things it was never trained on.
0.852
Share of brand-new concepts, created after training ended, that Primer correctly acquired. Best result so far.
Most models only know what they were trained on. We froze Primer, then created new concepts it could not have seen, and asked it to learn them.
Novel knowledge acquisition improved across successive research generations.
Gen I
0.000
Gen II
0.333
Gen III
0.778
Gen IV
0.852
How we measured it
Novel Knowledge Acquisition Rate (NKAR), Primer 1.3B · Internal post-freeze research benchmark. Evaluation targets are generated after checkpoint freeze to reduce the possibility that success comes from memorized training examples.
Nuance
Internal research benchmark. External validation pending.
02 / Remembers
It keeps what it learns.
14 / 14
Memory checks passed: knowledge learned earlier survived later learning intact.
Learning is only useful if it lasts. Models that keep training tend to overwrite what they knew before.
In our memory evaluation, Primer kept everything it had learned through later rounds of learning, and nothing false got in.
How we measured it
Retention (PR) 1.00; interference resistance (PI) 1.00; false discovery rate (FDR) 0.00, in the tested configuration.
Nuance
One memory evaluation, in one tested configuration. Longer time horizons are still being studied.
03 / Combines
It uses what it knows on harder problems.
22 / 22
Compositional targets solved when the required prerequisite knowledge was available.
Knowledge becomes powerful when it becomes a building block. Given the prerequisite knowledge, Primer solved every harder related target in this experiment.
How we measured it
Harder targets built from concepts the system had already learned; each solution checked independently.
Nuance
The prerequisite knowledge was available in this experiment, and the targets were related to it. Combining knowledge on harder, less related problems is active research.
04 / Compounds
Each lesson makes the next one easier.
21 / 26
Harder targets solved through a multi-step chain of the system's own acquired knowledge. Without the chain: 0 / 26.
This is the behaviour Primer is built to study. It learned a concept, used it to learn a harder one, then used that to learn a harder one still, drawing only on what it had learned itself.
On later learning tasks, its own earlier knowledge made it more successful and cheaper than starting fresh: 81% vs 59%.
How we measured it
Preregistered evaluation; replicated across independent seeds and checkpoints; zero false discoveries. Every step drew only on the system's own earlier learning.
Nuance
A small model and a controlled set of concepts. Showing the same compounding across broad domains is the next milestone.
Harder concepts
It is starting to learn rules, not just facts.
43%
of newly created structural rules correctly acquired after training, on our strictest evaluation.
Recent experiments extended post-training learning to a qualitatively harder class of concepts: structural rules rather than simple mappings. Zero false discoveries were recorded on this instrument, and previously demonstrated memory retention remained intact.
How we measured it
A previous version of this page reported 47% from an earlier evaluation. That figure was withdrawn after an internal audit found overlap between training and evaluation data, and the result was re-measured on fully separated data at 43%. The qualitative finding was unchanged.
Nuance
Coverage of this concept class is still incomplete. Generalization remains an active research milestone.
What's next
The rest of the climb.
Next: turning its own solutions into reusable tools, then inventing new ideas from old ones, carrying methods between fields, and choosing what to learn next. We will publish what works and what does not.