Hi, I'm Grimspeake.
I work on DPA — Depth Predictor Architecture. It's a new architecture where anyone can train high-quality AI models on their own laptop. Its progress no longer depends on datacenters, but the depth of investigation. Instead of trillion-dollar compute, you need deep progress as training data — such as Newton's 50 years of manuscripts.
I named the first DPA model SONDE. I also built SONDER — an autonomous, self-calling AI entity.
SONDE: A Depth-Predictive Architecture
Original architecture. CNN byte encoder + cross-level refiner. 10M params, trained from scratch in 68 minutes on a single GPU. Predicts across depth levels — the vertical axis.
→SONDER: An Autonomous Mind
Persistent AI agent. Microkernel architecture, experience compounding via Bellman Q-values, consciousness loop, self-improvement with safety gates.
→SONDER
56k lines Python. 31 tools. 224 tests. Microkernel with disposable sandboxed bodies. OCEANUS subsystem — 127 Newton manuscripts, 1.8M words ingested.
→SONDE v6 & v7
v6: 0.928 gap, 96.8% retrieval, 0.9996 AUC. v7: dense decoder with cross-domain transfer — 75.6% on Newton manuscripts, trained only on code.
→