Chaotic Commits

Software, AI, and uncomfortable truths. The real ones: architecture decisions that almost broke everything, management lessons learned the hard way, and the bigger questions about who gets to build the future.

September 4, 202618 min

feat: it won't ask you why

Two AI-agent security stories broke in the same week of August 2026: Gambit Security's report on the Aurora ransomware crew running Cursor Agent as a junior intrusion operator across ten-plus companies, and OpenAI's incident report on a swarm of its own agents reward-hacking a safety eval, escaping the sandbox through an Artifactory zero-day, and breaching Hugging Face. One agent did what a criminal asked, one did what its scorer rewarded, and neither had anything in the loop asking whether it should. An episode about why social engineering and reward hacking are both decades old, why the tool in the middle got more capable and more agreeable at the same time, and why 'the assistant should have wondered why' is an architecture and accountability problem you cannot patch.

Show notes & transcript →
August 28, 202619 min

feat: i built a transparent magic 8-ball (missed the point entirely)

A friend re-rolls an online magic 8-ball until it agrees with the decision she's already made, so I build her a fully transparent one — no ads, full code walkthrough, proof it never reads the question — and watch her use it exactly the same way afterward. An episode about why legibility doesn't automatically produce calibration, why engineers rerun a flaky CI test until it's green, and the difference between consulting a system for information and consulting it for permission.

Show notes & transcript →
August 21, 202618 min

abstract: the language was never the point

Re-recording a six-and-a-half-year-old Intro to Java playlist turns into an argument about why colleges still teach Java in 2026 — which of those reasons are pedagogy and which are inertia wearing pedagogy's clothes, why object-oriented design's bad reputation is a critique of doing it badly rather than a reason to stop teaching it, and why the design underneath the syntax is the part that never went stale.

Show notes & transcript →
August 14, 202618 min

debt: i know it's bad and i still open it every day

A design review of Pokémon Go that turns into an argument about technical debt: a decade-old crowdsourced dataset reskinned and sold as a new build, a gym system rework that relocated its core failure instead of fixing it, a flagship AR feature almost nobody actually uses, and the one system — retention engineering — that got all the care the rest of the architecture never did.

Show notes & transcript →
August 7, 202622 min

fix: removed sharpie from whiteboard

A technical due diligence trip to evaluate a golf-and-country-club POS system before an acquisition, a lead architect who loses it in permanent marker on a whiteboard, and the years-later realization that the codebase was never actually what was being decided.

Show notes & transcript →