Subhash Jha

A local assistant that answers questions like "service id for spot-fix?" or "which services run in ap-tokyo?" across about 200 internal services in a third of a second, without ever inventing a value. A non-generative classifier (Laya) decides what you are asking; a fact index mined from the service repos supplies the answer. Built with AI coding agents: the repo mining, the record format, generating honest training data, fine-tuning Laya from 10% to 95%, and the numbers.

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.

It seems like a contradiction. The same distribution that's hopelessly inadequate for generating images is at the heart of the most powerful image generators we have. The resolution is in what we're asking it to do.