Growing capacity without growing every step.
This is a proposed methodology, not a demonstrated breakthrough. The central idea separates stored graph capacity from the portion activated for a training example. A fixed active budget could allow capacity to grow while keeping routine work bounded.
Four ingredients to investigate
- Rational cells. Use expressive vector cells with positive-denominator gated updates, then measure stability and throughput.
- Packed communication. Keep a bounded active neighborhood and compact messages that can run in regular batches.
- Slow edge discovery. Evaluate gradient-scored candidate connections periodically, accounting for the discovery cost.
- Responsibility splits. Add capacity with an initially function-preserving split, then allow specialization.
The claim must survive a fair test
Compare against a dense network and static sparse controls using the same hardware, data, training time, and active parameter budget. Report quality, memory, examples per second, routing and growth overhead, and variance across seeds. Increasing stored capacity cannot be called free.
The open problem
Irregular memory traffic, routing imbalance, optimizer-state migration, and interference between old and new cells may erase the theoretical benefit. Our goal is to make those costs visible early.