Archives

Two small account-ban stories (and one biology-redline essay) changed how I think about building research workflows on closed models.

Codex’s experimental subagents cut my 131-image workflow from 1–2 hours to ~36 minutes—and improved accuracy.

What “Skills” really are, how they differ from rigid workflows, and practical advice on getting started safely.

Niji 7 improves coherency and prompt following, leans harder into clean anime linework, and intentionally dials down heavy rendering—plus real examples and a couple of takes.

Semaglutide’s China patent cliff meets a crowded biosimilar queue and VBP-style purchasing pressure—tempered only by the real cost floor of injection pens and sterile manufacturing.

GPT Image 1.5 tops Image Arena, but X says it’s worse than Nano Banana Pro. I ran identical prompts—illustration, posters, English/Chinese typography, and a TravelBrain itinerary JSON—to see where each model wins.

OpenAI’s GPT‑5.2-Codex hits state-of-the-art results on SWE‑Bench Pro and Terminal‑Bench 2.0, improves long-horizon refactors and Windows workflows, and leans hard into defensive cybersecurity.

Why I built TravelBrain as an offline travel planner where you collect ideas into an inbox, drag them onto a timeline, and build trips like LEGO — all without syncing to the cloud.

DeepSeek's latest paper admits the gap with closed-source models is growing. But the real story is how chip export controls forced two diverging AI paradigms—and what that means for the endgame.

DeepSeekMath-V2 scored 118/120 on Putnam 2024—surpassing the human record of 90—and achieved IMO gold level. The key innovation: a self-verification architecture where models learn to identify and fix their own proof errors.

From a 5.5-yuan flu shot in Beijing to HPV and PCV13, a look at how overcapacity, overlapping pipelines, and policy-driven tenders are reshaping China’s vaccine industry and what it signals for pharma and diagnostics more broadly.

OpenAI’s GPT-5.1 Codex-Max landed right after Gemini 3.0, topping SWE-bench while cutting costs and reframing what an AI coding assistant can be for working developers.

Why I built KeepFlame as a privacy‑first, offline habit tracker with animated themes, and what I learned from designing a more complex second app after Simple Warm Reminder.

Field notes on China's volume-based procurement (VBP), the recent wave of imported originator drugs withdrawing from the market, and what it might mean for patients, generics, and global pharma from a Japan–China IVD vantage point.

Why I built a calm, glass‑styled reminder app for my father instead of another to‑do list, and what I learned from designing for simplicity, trust, and aging minds.

Sora isn’t just about text‑to‑video. OpenAI turned creation into a social loop—Remix, prompt privacy, and Cameo make sharing the default and co‑creation addictive.

Built for my own workflow, pdf-ops merges and splits PDFs from a fast CLI and a keyboard‑only TUI powered by Ratatui. Core features work on macOS; MIT‑licensed and still evolving.

Google Research sketched a space‑based AI infrastructure—Project “Suncatcher”—that clusters compute in orbit, powered by sunlight, and beams only results to Earth. What’s real, what’s hard, and what could ship first.

As rivals close the quality gap and video takes center stage, Midjourney’s $30 plan and product choices face pressure. A pragmatic look at value, workflow fit, and where it still shines.