Terence Tao’s “smell test” pinpoints the core dilemma of today’s AI: superb at imitation yet short on genuine reasoning. This article explores how the scarcity of process data has become a bottleneck and follows the journey of process supervision from an expensive theory to a generative, self‑evolving practice—illuminating a pathway toward truly thinking machines.

I once regarded ‘aphantasia’ with indifference, secure in my own powers of imagination and recall. Yet a chance reading shocked me into recognising that my idea of ‘imagining’ might differ radically from most people’s—I cannot ‘see’ images in the mind’s eye. This essay records that belated discovery and the ensuing re-examination of my memory and modes of thought.

I used to scoff at talk of AI ushering in an ‘intellectual levelling’. Surely my well-honed search tricks were enough. Yet a recent hunt for Russian-literary sources revealed the limits of my skill and the reach of AI in vaulting knowledge barriers. What follows is the tale of that change of heart, and a few modest thoughts on AI and how we retrieve information.

I investigate a frustrating workspace 'flashback' bug experienced when using the Arc browser with the AeroSpace tiling window manager on macOS. The root cause identified involves frequent makeKeyAndOrderFront: API calls by Arc (and potentially other Chromium/Electron apps) conflicting with AeroSpace's Accessibility API-based virtual workspace mechanism.

I built a reusable component for rendering interactive Vega-Lite charts on my blog. This post explores the implementation details, the power derived from the underlying Grammar of Graphics principles, and why this declarative approach is ideal for showcasing data-driven insights from my machine learning work.

I created a simple Preact component for my Astro site to render PDF documents, like arXiv papers, using just an ID or a URL. This allows for seamless integration of papers for future discussions, notes, or analysis, avoiding iframes for better design control.