Shawn NgTopic
Machine Learning
7 deep dives tagged Machine Learning — real systems, real numbers, no hand-waving.
What Is Jev? TypeSafe AI's System One Model, and Why It Refuses to Talk
Jev is TypeSafe AI's System One model: a non-autoregressive transformer that returns typed, calibrated decisions instead of text. How it differs from LLMs.
Recommendation Is Vector Search With a Query You Never Typed
Search and recommendation run the same engine: encode a query as a vector, return the nearest items. Only the query's origin differs. Here is the geometry.
The State of AI: A Field Report From Someone Who Actually Ships This Stuff
A working engineer's mid-2026 field report on AI: what changed, what plateaued, where production reality is messy, and where the real work moved.
Local Models vs Frontier APIs: An Honest Accounting
What it actually takes to run LLMs on your own hardware, why memory bandwidth is the real bottleneck, and when the economics flip in favor of the API.
The Solvers Refuse to Die: Classical Optimization vs the Learned Upstarts
An honest scoreboard of classical optimization solvers versus learning-to-optimize, and the hybrid pattern where ML proposes and exact methods verify.
How Machines Learned to See
Computer vision's full arc: hand-crafted features, the 2012 ImageNet moment, ResNet, vision transformers, CLIP, and what models still cannot see.
How YouTube Decides What You Watch
How YouTube's recommendation system works: two-stage retrieval and ranking, watch-time prediction, and multi-task MMoE models, per the published research.