Comparing adaptive learning rate (dopamine-inspired) against traditional fixed learning rate on a 2D grid pathfinding task with obstacles—exploring whether modulating learning rate by improvement signal helps long-term learning.
Building a minimal automatic differentiation engine from first principles to truly understand how modern ML frameworks compute gradients and train models.
GraphRAG uses entity extraction to build navigable graphs. What if we used embedding geometry instead? A speculative exploration of TDA for document retrieval.
How to recover from AI dependency by believing in your coding abilities and reading code instead of relying on AI assistance.
An experimental approach using JEPA Attach to test whether Joint Embedding Predictive Architectures can help Large Language Models learn from abundant multimodal data when high quality text becomes scarce.
How we instrumented Sammy 3 to feel like a teammate that sees context, speaks naturally, and acts faster than humans.
Lessons learned building AI features that users actually want to use—avoiding the trap of building cool tech instead of solving real problems.