Subhash Jha

Every image model you have heard of is built from the same handful of ideas stacked in cleverer and cleverer ways. A walk up that staircase, one architecture at a time, from a single neuron to the MLP, the CNN, ResNet, the U-Net, attention, Stable Diffusion, SDXL, and finally the diffusion transformer. Nine hand-drawn diagrams, and by the end none of it looks like magic.

One line of the CM-GAN paper reads backwards: when a hole covers most of an object, they carve the object back out. Working out why unlocks a clean way to think about when a generative inpainter fills with background versus when it invents a new object, and why that matters for object removal and blemish retouching.

One India is a village I left after the third grade: mango trees, monsoon rain, and a rooftop full of stars. The other is where my professional life began, where I traded climbing trees for climbing stairs. A quiet essay on freedom, independence, and which of the two is the real me.

Three environments, more internal services than I can count, and endpoints scattered across a dozen repos, wikis, and old Slack threads. Here's the first Claude Code skill I built that actually stuck, a router that figures out where an endpoint lives instead of memorizing it.

You type one letter and ten suggestions appear before your finger leaves the key. Behind that dropdown is one of the tightest latency budgets in consumer software, an in-memory trie sharded across thousands of machines, and a two-speed offline pipeline. A full walk through the system, with the hand-waved parts drawn out.

CEOs are declaring AGI and pricing in a workforce of models. But the thing they're racing to replace runs on twenty watts, adapts to every moment, and never collapses when it thinks about itself. A working engineer's case for biological intelligence.

Clean code, design patterns, readable names. Those are ergonomics for a human brain with seven slots of working memory. The AI doesn't have that brain. So why is it still writing code for it?

A Raycast extension, a folder full of selfies, and a midnight scheduled job that quietly rotates my Slack avatar. Friday gets the most photos. This was not an accident.

A year of aggressive AI adoption has taught me something the productivity dashboards don't show. The machine got faster. The humans didn't. And the gap between them is where most of the new chaos is coming from.

A new Raycast clone ships every week now that AI has collapsed the cost of building one. Most of them are solving the wrong problem. The actual moat was never the launcher, it was what users could build on top of it.