Writing
Blog.
Notes on how MingLLM works, what we learned building it, and the decisions behind the product.
One receipts log, many surfaces
Jarvis, Tensor, and now Rocky all write to the same log and the same memory. Here is why that mattered more than any single feature.
Making 4B models feel like 400B on your laptop
Three moves that matter: expert iteration, selective generation, and tight tool-use shaping. How we get to 94.9%.
Why Jarvis does not have a chat window
A chat UI is the wrong default for a voice-first agent. We went looking for a better one.
The Orb, one month in
What we learned from the first four weeks of persistent memory in private beta.
Why we are local-first
Cloud-first is a choice disguised as gravity. Here is what we are choosing instead.
Tensor: from idea to extension in six weeks
A short postmortem of how we built the first Tensor alpha, what broke, and what we cut.
All posts by Yiming Beckmann. Specific numeric claims reflect in-progress evaluations and may shift between draft and published research.