Writing

what probing jev and training nine open replicas revealed about its behaviour, strengths and limitations.

Modern text-to-speech got good enough to break the tools that grade it. So I trained a 400K-parameter scorer that runs on a laptop and lands within a point of a 7-billion-parameter judge, on nothing but Kaggle's free T4s.

A practical engineering guide to building AI systems for human well-being, from formalizing happiness as a multi-dimensional construct to reward architecture, failure modes, and the honest open problems.

The fourth-order effect, field by field, what happens to the capitalisation of products in theoretical physics, drug discovery, materials, chips, and cryptography once AI does the research itself.

Inside the eight-layer Chinese gray-market supply chain reselling Claude API tokens at $4/M, where the discounted API access is the cover charge, and the real product is your prompts and outputs streaming into someone else's data lake.