Research
We trained a 7M parameter LLM exclusively on primary sources from 1924–1946. No Wikipedia. No modern text. No knowledge of what came after. Here's what we learned.
Four models. One GPU. No coordination. We fixed it with a text file.
We had an intuition that past tokens could declare their future relevance, replacing similarity-based attention with direct lookup. After dozens of hours of experiments, the result was clear — and not the one we expected.
We’ve been experimenting with a simple idea: what if memory in language models isn’t just a buffer, but a computational space ?
We explore how event-driven SNNs can replace transformer attention for low-power edge inference, achieving competitive performance at a fraction of the energy cost.
Custom silicon designed natively for SNN workloads. Here's what we're building, why conventional GPUs are inadequate for brain-inspired AI, and the technical roadmap ahead.
Bigger isn't always better. We make the case for small, efficient, private AI — and why sub-billion parameter models are the future of personal computing.