Why subscribe?

Subscribe to get full access to the newsletter and publication archives.

Vector Safety Net is about the part of AI alignment that looks like game theory: what happens when agents have mixed motives, when the reward can’t be checked, and when the only thing keeping behavior honest is reputation over a long horizon.

I’m Makho Mashoba. I run a data science team that ships multi-target tracking and presence models to production, and before that I spent ten years building ML systems in insurance, an industry that has been pricing non-verifiable behavior for a century. My degrees are in psychology and economics, so I tend to read RL through the lens of incentives rather than architectures.

What you’ll get here: concrete proposals rather than commentary. Each post takes one mechanism (a reputation term in the reward, social grading of unverifiable outputs, mixed-motive training corpora from real marketplaces and games), argues for it, and names the ways it fails. If you work on RL, multi-agent systems, or alignment, I’d rather be corrected in the comments than agreed with.

Posts are free. I write when I have something worth your time.