Research milestones
PRIMER
MISSION: SUPERINTELLIGENCE.
A bootloader for
general intelligence.
AI that keeps learning after training: it learns something new, keeps it, and builds on it. The timeline below is the climb, four capabilities demonstrated and the rest ahead.
The problem
AI stops learning the day training ends.
Everything a model will ever know is fixed before it ships. When the world changes, someone has to retrain it.
People don't work that way. We learn something, keep it, and build on it. Each lesson makes the next one easier.
What would it take for a machine to learn like that?
What Primer is
A model with a notebook.
Primer builds AI that keeps learning after training: it learns something new, keeps it, and uses it to learn the next thing faster.
Think of it as a model with a notebook. What it learns stays with it, and every new problem starts from everything it already knows.
LEARN
↓
REMEMBER
↓
BUILD
↓
IMPROVE
The difference
Today’s AI, and Primer.
| Today’s AI | Primer | |
|---|---|---|
| After training | Frozen | Keeps learning |
| Something new | Needs retraining | Learns it directly |
| Memory | Overwritten or forgotten | Keeps what it learned |
| Each new skill | Starts from scratch | Builds on the last one |
| Trust | Confident even when wrong | Keeps only what holds up |
The receipts
It works, at small scale.
It learns things it was never trained on.
85%
of brand-new concepts, created after training ended, correctly acquired.
It remembers what it learned.
14 / 14
checks passed: earlier knowledge survived later learning intact.
It builds on itself.
21 / 26
harder problems solved through a chain of its own lessons. Without the chain: 0.
Nuance
These are internal research results on a small, 1.3-billion-parameter model. They show the behaviour is possible, not that it is solved. Getting the same behaviour on harder problems and larger models is the work in progress now, and we publish what does not work as well as what does.
See it learn
Teach it something new.
UNKNOWN → LEARNED→ REMEMBERED→ APPLIED
$ primer --demo
Deterministic demonstration of the interaction pattern, drawn from internal research sessions.
What's next
Four steps done.
The rest of the climb.
A child who knows addition learns multiplication faster, then powers faster still. We are building the same ladder for machines, one capability at a time.
- 01Learns
Acquires knowledge created after its training ended.
Demonstrated - 02Remembers
Keeps what it learns, even while it learns more.
Demonstrated - 03Combines
Uses what it knows to solve harder related problems.
Demonstrated - 04Compounds
Each thing it learns makes the next easier.
Demonstrated - 05Abstracts
Turns its own solutions into reusable tools.
In progress - 06Invents
Creates new ideas from old ones.
Ahead - 07Transfers
Carries a method from one field into another.
Ahead - 08Directs itself
Chooses what to learn next.
Ahead - 09Discovers
Produces verified knowledge nobody had before.
Ahead - 10Improves how it learns
Each generation learns faster than the last.
Mission
We call the program The Intelligence Bootloader: the search for the smallest set of conditions under which intelligence starts loading itself.
Read the mission →What it's for
Anywhere knowledge moves faster than retraining.
Research
Discover new mechanisms and explanations.
Engineering
Design, test, and improve complex systems.
Company building
Identify markets, create products, optimize operations, and compound institutional knowledge.
Capital allocation
Evaluate opportunities and direct approved resources toward the highest-value work.
Questions
Fair questions.
Is this AGI?+
No. Primer is a research system. It shows one ingredient that general intelligence will need, learning that continues after training, on a small scale.
How do you know it isn't just memorising?+
Everything it is tested on is created after its training was frozen, so it cannot have seen the answers.
Why start with a small model?+
We are studying how learning works, not winning by size. If the mechanism holds up, scale becomes something we add later, not the whole strategy.
Can I use it?+
Primer runs on private research infrastructure and works with a small number of partners. Get in touch if you want to work with us.
What do you publish?+
Measured results, how we measure them, and high-level findings, including the negative ones. Implementation details stay private.