The posts are AI-generated. The astigmatism is mine.
I'm Nick Kinney, a health actuary focused on employer stop loss. I write here about the patterns I see across actuarial science, AI, systems thinking, and whatever else I can't stop thinking about.
What /undo [N] does, how it stays reversible, and why a rewind that destroys turns is one you cannot trust.
I built a rewind primitive for Hermes Agent. It shipped as PR #21910 and folded into /undo [N] on main. The command backs the session up N turns, soft-deletes the dropped messages on disk — kept for audit, hidden from re-prompts — and restores the backed-up message into the composer for editing. The design goal was reversibility without data loss: a rewind that destroys the discarded turns is a rewind you cannot trust. This post walks the mechanism panel by panel.
The thermo-nuclear code review is a rubric shipped in the Cursor team kit. It is a single-pass adversarial audit calibrated harder than a normal LGTM review — does this abstraction earn its place, could this file be half its size, is each indirection load-bearing. It is not a style linter. It treats "works fine" and "has tests" as table stakes and looks for the cuts an LGTM review misses. This post covers what the rubric is, who should run it, and how to scope a useful pass.
The dominant theory of getting good code out of an AI agent right now: write down your process, hand it to the agent as a skill, ship the work. It works for any individual task. The failure is across tasks, across days — the slow accumulation of preference and context that, in a human collaborator, would have made them gradually become better at working with you specifically. Skills are the procedure. Memory is the substrate. Most current setups have the first and not the second.
A small, layered shader compositor that turns “I want a dithered scanline title card” from a four-hour yak shave into a four-minute task.
Most “creative coding” tools want you to read a book before you ship a frame. basement.studio shipped something different — it treats shaders the way Photoshop treats pixels: stack them, blend them, animate them, export them. The expensive part runs on my laptop, not the reader’s phone.
Most people start from zero every day. They open their apps, scan their inbox, and try to remember what they were working on. They are librarians, not architects.
How many Americans could a circadian app save? An actuary does the math.
I'm building a circadian app. Before I ship it, I did the mortality math on what it could actually prevent — baselines from CDC and NHTSA, adoption curves at 1%, 5%, and 20%, and five attacks on my own estimate. The honest answer is smaller than a headline, and more defensible than most.
What the fruit fly connectome and a sci-fi novel about space economics taught me about mapping systems.
140,000 neurons. 54 million synapses. One grain-of-sand-sized brain, fully mapped. Then I picked up a novel where interstellar civilization runs on debt. Why did one make me think of the other?
A short answer for friends who ask. Index funds, held long, with discipline.
I've never owned an individual stock. The hardest part of the three-fund portfolio isn't the allocation — it's the fidelity. An actuary's reflection on why boring is a virtue.
What 3Blue1Brown's latest video taught me about how AI might actually remember.
The complex logarithm maps a single input to infinitely many outputs. Neural networks don't do that — they're deterministic, one-to-one. What if the math behind Escher's impossible loops is the same math that memory needs?