
Farming Expensive Coding Agent Sessions
My coding-agent sessions are more expensive now. The useful spend is the part that turns mistakes into better tests, graders, prompts, and checks.
Read postSoftware systems, AI tools, books, and experiments.

My coding-agent sessions are more expensive now. The useful spend is the part that turns mistakes into better tests, graders, prompts, and checks.
Read post
My first week in San Francisco: paperwork, an unfurnished apartment, two cats, Bay Trail runs, and waiting for furniture.

Building frontend-only tools for load flow, Mandelbrot zooming, optimizer behavior, and event-loop ordering gave me concrete ways to inspect older technical models.

A concrete way to understand dropout in deep learning: sampled subnetworks with shared weights, standing in for a much more expensive ensemble.

Notes on the 11-inch iPad Air M3 as a lightweight reading, annotation, and remote-workflow setup for studying.

I replaced scattered responsive-image conventions in a Next.js static site with one generated manifest for variants, dimensions, and runtime lookup.

A conversation with colleagues about lower-friction AI-assisted prototypes for small internal tools.

How Codex made small website edits easier to make from a phone, while Sora still feels more useful for exploratory visuals than finished work.

Goodfellow Chapter 6 sharpened how I separate representation, parameter efficiency, and optimization when thinking about deep feedforward networks.

Part I of Goodfellow, Bengio, and Courville works best as shared vocabulary for later chapters, with proof details still worth supplementing.

How my AI-assisted coding workflow shifted from quick snippets to clearer briefs, review checkpoints, and build/lint/test verification.

How I added a repo-backed Markdown blog to my static Next.js portfolio without adding a CMS or runtime admin surface.
Subscribe for occasional updates when I publish something new.
Copyright © 2020 - 2026 Alex Leung