Practical payoff: branch-free code is what GPUs want. Regex + JSON parsing written as single-threaded JS runs as WebGPU kernels over 100k documents. chenglou.me/handwired/

Caveats: compiling programs into weights has precedent (DeepMind's Tracr, 2023). And gradients through step functions are ~zero almost everywhere, so training tunes constants more than it learns logic.

Other choices that transfer: - Loops with runtime-dependent counts live in the host, not the net - State writes back only at events and settle checks, so behavior is identical at any frame rate - A plain-JS reference is the test oracle

They're enforced, too. Every evaluation checks every cell and throws with its name if a value leaves its declared domain. Most "neural nets can compute X" claims leave that domain implicit.

Every value declares a range and a grid (pointer coords are multiples of 1/64 px). With grids, a comparison costs 2 ops. Without, the same comparison over 0 to 16,384 costs 15. Same answer, bigger net. Precision contracts become architecture.

The trick: `a > 4 ? 10 : 12` becomes relu(a-4) - relu(a-5), an exact 0/1 bit. Both branches compute; the bit picks one. A small desktop OS with drag, scroll and spring animation runs on 9,909 hand-set weights across 130 layers.

Cheng Lou's Handwired compiles plain JavaScript into a neural net: no training, exact results, runs in a browser. The part worth stealing isn't the trick. It's that type declarations decide how big the network gets. #MachineLearning #JavaScript #WebGPU

"The best founders understand that taste is a competitive advantage that compounds. It runs deeper than pixels—it’s in your codebase, your culture, your cap table." 🧵3/3 #founders #startups #ProductDesign

"The relentless, almost painful ability to know what should exist, what shouldn’t, and where quality matters. It’s the difference between shipping a product and shipping a point of view." 🧵2/3 #founders #startups #ProductDesign

This essay "Taste" by @saranormous moved me when I first read it but it lands harder now, where our moat as founders is really just our taste. You ARE the moat! So read this because it might be exactly what you need to hear: saranormous.substack.com/p/t… A few quotes below 🧵1/3 :

I think we need a different way to describe building software with agents - it's not "coding" in the traditional sense of the term and I'd rather keep that term to imply that a human built something by hand...we need people good at terms to coin something! @vercel @karpathy

Neat daily lesson from MindStaq today (these are short lessons that get shared in my Today screen based on my role - as a cofounder/CTO) The 90-day promise audit Once a quarter, list every commitment you made out loud — to customers, investors, candidates, your team — with the date you said it. Anything without an owner and a date is a wish wearing a promise's clothes. The audit takes an hour and usually reveals two or three things you've quietly stopped believing in but never withdrew. Suggestion: Block an hour this week; the list doubles as your real roadmap and your next investor update.

I’ve started playing AI subway surfer. Prompt the next instruction into a remote session, lose connection between stops, arrive at the next stop and read the response. Then quickly type in a response if needed to keep it going. Save the big notes for the bridge🤘🏾

MindStaq's daily lesson for me today really hit a point I wanted to share immediately: Reallocation, not replacement — redeploying people when tasks get automated When a task gets automated, the leader's job is not to defend the role but to name the adjacent work the person is now freed for. Practical version: for each automated task, write down (1) the task, (2) the judgement it was hiding, (3) who now owns that judgement, and (4) what the freed hours are for. Leaders who answer (4) before announcing the automation keep trust; leaders who leave it blank get attrition and quiet resistance.