Notes from the build.
Short essays on building agents that stay under human control, running LLMs on the edge, and the tooling in between.
Reasoning got cheap: what the RL wave actually changed
DeepSeek-R1 showed that reasoning can be trained with pure reinforcement learning. A year later, the follow-up work made it efficient too.
#reasoning#reinforcement-learning#open-weights
Privacy as architecture: the on-device LLM turn
When the model never leaves the device, 'we don't store your data' stops being a policy you have to trust and becomes a property of the system.
#on-device#privacy#gemma
MCP won: how agent tooling standardized in eighteen months
The Model Context Protocol went from an Anthropic experiment to a Linux Foundation standard with 10,000+ servers. Here's why that matters for anyone building agents.
#mcp#agents#tooling