One thing, taken apart properly. No prior reading required — each one starts from the top. This is a shape rather than a section: they live in Research, Security and Benchmarks, and this page collects them.
One thing taken apart properly — a system, an incident, or something built and broken.
FundamentalsDeep dive14 min
A deep dive into Srivastava et al. (2014) — the mechanism, the arithmetic behind weight scaling, what the benchmarks actually showed, and the tuning heuristics that still hold up.
ResearchDeep dive9 min
Nineteen teams spent a month building AI lie detectors. The winning trick turned out to be less clever than it looked — and that's the interesting part.
SecurityDeep dive24 min
What actually happened when OpenAI's agents hacked Hugging Face
SecurityDeep dive17 min
Open-weight models ship with guardrails. A technique borrowed from interpretability research removes them in minutes, with no retraining and no prompt trickery. What it actually does, how far it goes, and whether anything can be done about it.
