You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
CKSM is a sequence model architecture derived from the kinetic dynamics of the NMDA receptor. It is deterministic, runs in constant memory regardless of context length, and learns continuously without retraining. The goal is to found a company around it and sell specialized inference to enterprises that need a model that genuinely improves over time on their domain.
Finish the research paper and run the CKSM-tiny experiment on WikiText-103, comparing against Mamba-3 at identical training FLOPs on a single RTX 4090. This is the first falsifiable empirical result. If it passes, the architecture scales.
Living expenses to work on this full time.
Solo. 22 years old, Paris, no degree. In the past year I built Tachyon, a lock-free shared memory IPC library with 56ns p50 round-trip latency across 7 languages, and Kinapse, a recurrent spiking neural network running a full forward pass in under 5 microseconds on a single CPU core. Both are public on GitHub: github.com/riyaneel/Tachyon and github.com/riyaneel/Kinapse.
The kinetic gate provides no measurable perplexity gain at tiny scale. This is a defined hard stop before any large-scale compute is spent. The architecture gets revised.
Zero.