Jake Cuth. — Notes
Field notes from an AI engineer.
https://www.jakecuth.com/feed.xml
2026-05-17T00:00:00Z
Jacob Cuthbertson
https://www.jakecuth.com/
© Jacob Cuthbertson
jakecuth.com
LeWorldModel (LeWM) Explained · Yann LeCun's JEPA with SIGReg
https://www.jakecuth.com/notes/leworldmodel-jepa-explained/
2026-05-17T12:00:00Z
2026-05-17T12:00:00Z
LeWorldModel (LeWM) is the JEPA world model that replaces seven loss terms with one Gaussian-matching regularizer (SIGReg). Architecture, training, and GitHub code walkthrough.
DeepSeek-V4 Architecture Explained · Muon, CSA, FP4 KV Cache
https://www.jakecuth.com/notes/deepseek-v4-architecture/
2026-05-17T12:00:00Z
2026-05-17T12:00:00Z
DeepSeek-V4 architecture explained in plain language. The Muon optimizer, Compressed Sparse Attention (CSA), FP4 KV cache, hyper-connections, and what changed from V3.2.
What is Subquadratic? The SubQ Company, Paper, and Architecture
https://www.jakecuth.com/notes/subquadratic-explained/
2026-05-17T12:00:00Z
2026-05-17T12:00:00Z
Subquadratic AI is a Miami startup with a $29M seed and a one-million-token LLM called SubQ. Architecture, benchmarks, paper, and the case for and against.
33 OG images in 80 lines of Python
https://www.jakecuth.com/notes/og-images-with-pil/
2026-05-05T12:00:00Z
2026-05-05T12:00:00Z
Per-page OpenGraph cards without Figma. A one-shot Python script over your existing og:title tags.
The k=2 problem in cluster selection
https://www.jakecuth.com/notes/segmentation-k2-honesty/
2026-05-04T12:00:00Z
2026-05-04T12:00:00Z
Silhouette at k=2 is almost always highest and almost always meaningless. Why the segmentation lab grays it out instead of pretending it's a real choice.
Why my A/B simulator says false positives are 26%, not 5%
https://www.jakecuth.com/notes/ab-test-fpr-26-percent/
2026-05-03T12:00:00Z
2026-05-03T12:00:00Z
Peeking every 50 users turns a 5% nominal false-positive rate into 26% empirical. The Monte Carlo lives on the A/B Test Simulator lab page.