Engineering write-ups, case studies, and the occasional opinion. Every article here was drafted in the backstage and published by a human.
同一個陽台的第二種作物:實際寬 3 m、淨高 2 m、深 0.6 m,取右側 2 m 種四株。台北平地秋播冬收,9 月中播種、11 月震動授粉、1 月開始陸續採,收成煮成紅醬。
為台北低海拔市區、四樓開放陽台量身規劃:實際寬 3 m、淨高 2 m、深 0.6 m,取三分之二做種植區。
In a head-to-head benchmark against two leading commercial detectors, Reef's own detection approach posted the highest AUROC on WildFake (0.930) — and caught an entire class of older generators the turnkey products simply miss.
Generative AI makes fakes look real. A 2024 method called RIGID offers a counter-intuitive trick: train nothing — just give an image a tiny "nudge" and watch how far it moves inside a frozen vision model. Real photos barely budge; AI images slide. A richly illustrated walkthrough.
Teach an AI one task after another and it tends to forget the last — until you let it sleep. A hopeful, bilingual (EN / 繁中) tour of Sleep Replay Consolidation, rebuilt from working Python, with live charts, animated "synaptic dreams", and an invitation to explore continual learning.
A general-audience guide to Anthropic's Claude model pricing: why the per-token sticker price of the Opus tier versus the Sonnet tier rarely equals the real cost per task, and how to route work between them. Pricing and capabilities verified against Anthropic's published rates (July 2026 snapshot); benchmark comparisons kept directional rather than precise.
Why intermittent AI workloads almost never justify buying a GPU: the break-even math, a neutral look at Lambda, RunPod, and Vast.ai, the real economics of serverless inference, and a staged adoption roadmap — with an interactive break-even calculator.
The living style guide — every v1-core component and declarative block, rendered in the house look. Charts, a Mermaid diagram, a live calculator, a bilingual toggle, a gallery, callouts, stat tiles and working anchors — all prose, no author JavaScript.