Seedling development happens entirely indoors now, under lamps. In the shortest days of the year I set up folding tables in the back room, fill trays with damp promix, and press tomato seeds no deeper than the edge of a dime. The ritual is deliberate and slow: label the varieties, mist the surface, slide the trays beneath the glowing panels. There is something stubbornly hopeful about coaxing life in the dead of winter, something that feels like the opposite of opening an app and letting a feed decide what I should want next.

By the time real sunlight returns, those seedlings will be knee-high and impatient for the garden. The promise of an early-summer harvest—Sungold cherries splitting in the heat, Brandywines the size of softballs—starts in January under fluorescent hum.Winter also breaks the usual rhythm of destruction. The constant parade of snapped handles, dulled blades, and bent tines that marks the growing season finally stops when the ground freezes.

Tools pile up like casualties. I drag them into the shop, hang the axes, and sit by the woodstove with a pot of coffee while I rebuild them. I weld new shovel heads, splice ash handles, grind edges until they sing. The stove throws orange light across the bench; sparks from the forge drift upward like slow fireflies. Repair is its own satisfaction, but it also turns my mind toward assembly—how parts fit, how forces travel, how a century-old pattern can still be improved with a few calculated grams of steel. Some of that thinking has leaked into my code.

Lately I’ve been working with Hilbert curves again, trying to map high-dimensional spaces onto contiguous memory in a way that keeps cache lines happy. Doing the curve generation in pure assembly, hand-rolled AVX2, gives a ridiculous boost—no GPU, no framework bloat, just the CPU doing exactly what I tell it, when I tell it. The difference is measurable in single-digit milliseconds that compound into minutes when you’re processing millions of points.

I still reach for AI when I’m stuck, but I’ve learned the hard way that it hallucinates confidently and badly unless I hand it an airtight specification first. So the workflow has settled: write the spec like a legal contract, surround the core routines with exhaustive tests, benchmark obsessively, then let the model suggest cosmetic refactors only after the hard parts are proven correct. It’s slower than the old “ask the oracle and pray” approach, but it actually works.

I spent years giving code away under permissive open-source licenses because it felt like the right thing to do. I still believe sharing knowledge matters. But the landscape has changed: megacorps vacuum up open code, close the profitable parts behind paywalls, and give nothing meaningful back. The commons got strip-mined. At this point I’m done subsidizing that machine.

From here forward my new projects ship under what I call the Lifestyle License. It’s simple: you can use the code for free—commercial included—if your revenue is less than mine. Cross that line and you need my permission, which usually means buying me a decent 40 year bottle of scotch or hiring me as a consultant for a month.

The license is short enough to print on a business card and explicit enough that nobody can pretend they misunderstood. It protects my time, funds the next round of tools and seeds, and still leaves the code accessible to anyone who isn’t exploiting it for eight-figure exits.

Winter is for repair, for planting impossible seeds, for writing specifications by firelight. By the time the soil thaws, both the garden and the codebase will be ready.

Categories: Farming