Rewire
Can an AI agent edit a model’s weights and structure better than blind methods given the same compute?
CodeIndependent AI research
The next frontier is intelligence that keeps learning.
We’re exploring the architectures that could make it possible.
01 / Research
Each project starts from one question and a fair baseline. Results are published as notes, including the ones that fail.
Models that change themselves, or other models, after training, and the measurements needed to tell whether that works.
Can an AI agent edit a model’s weights and structure better than blind methods given the same compute?
CodeWhen a language model edits another model’s weights directly, which interpretability information makes its edits work?
In design, on Rewire’s harness.
Can models improve by editing their own weights, or each other’s in a ring, and does a peer edit better than self?
Question still being sharpened.
Does an offline “sleep” phase guided by the model’s own circuits help it learn beyond generic consolidation?
Can a language model design the local learning rules that update another model, in place of post-training and RL?
Can a memory model that reads another model’s activations decide what to store and recall better than retrieval over text?
In design.
Rules for updating a network’s weights, found by search instead of written by hand.
Can evolution find learning rules that beat hand-designed ones, without labels or by developing over a “childhood”?
How a model is structured, and what it costs to run.
Can each token choose its own path through a model’s layers?
CodeCan a planner working inside a learned world model complete long-horizon goals better than flat baselines, at lower cost?
Question still being sharpened.
Pass and fail are the pre-registered verdicts, quoted from each test’s analysis script. Every result is in the technical notes.
02 / How we think
Ask what a system is doing, why it works, and what would make it learn differently.
Build controlled experiments. Compare against strong baselines. Treat a promising idea as a question to test.
Connect fundamental research to systems people can understand, reproduce, and eventually put to work.
03 / Perspectives
A few ideas shaping our research agenda.
04 / The company
Kāula is an independent AI research company working toward intelligence that can keep learning.
Our focus sits at the intersection of adaptive architectures, continual learning, and efficient computation. We’re building a research practice around fundamental questions and the experiments needed to answer them.
Kāula is Native Hawaiian founded. Our name, kāula, is Hawaiian for “seer” or “prophet.”
Explore our research agenda