Meta is targeting a 50:1 manager-to-engineer ratio in its new AI org. I manage a team of about 30 engineers at a company that’s all-in on AI adoption, and my experience says the job has gotten harder, not easier, in the agentic coding era.

BlueDot Impact ran a puzzle: seven of eight features in a small network are linearly represented, one isn’t. Here’s how I found it, worked out the geometry (a country feature sharing a superposed direction with food, carved out as a middle band), and built a toy model with an even weirder representation.

I replicated and extended a recent paper showing that 4-bit quantization reverses machine unlearning. The finding holds on a different benchmark, a different model family, and extends to magnitude pruning. The suppressed knowledge comes back, utility is preserved, and there’s no surface signal that anything has changed.

The BlueDot Impact Technical AI Safety course is free, well-structured, and worth your time. The mechanistic interpretability unit is what sent me down the rabbit hole I’m currently in.

Thoughts from the Cloud Intelligence / AIOps Workshop at ASPLOS 2026. The field is at an inflection point — not because everything is resolved, but because the capability gap has closed. The harder work now is infrastructure, accountability, and trust.

I’ve been working through the mech interp fundamentals — reading, notebooks, and a research project — while pushing the boundary of what agentic tools can do without giving up ownership of the work.