電腦工程 Engineering
AI
從一顆神經元到一整棟機房，同一套機制的三個尺度。
One neuron to one building: the same machinery at three scales.
神經網路怎麼運作 How a neural network works
一顆神經元正向、一顆神經元反向、整個網路正向、整個網路反向，然後串起來。用一個真的跑得動的純量自動微分引擎，把計算圖、鏈鎖律和反向傳播一個節點一個節點攤開；每一個梯度都用有限差分驗證過，每一條決策邊界都是當場算出來的。這是三頁裡的第一頁，接下來是 Transformer，再接下來是 AI 工廠。
One neuron forwards, the same neuron backwards, a network forwards, the same network backwards, and what that adds up to. A working scalar autograd engine lays out the computation graph, the chain rule and backpropagation one node at a time; every gradient is checked against finite differences and every decision boundary is computed on the spot. The first of three pages, followed by the transformer and then the AI factory.
One neuron forwards, the same neuron backwards, a network forwards, the same network backwards, and what that adds up to. A working scalar autograd engine lays out the computation graph, the chain rule and backpropagation one node at a time; every gradient is checked against finite differences and every decision boundary is computed on the spot. The first of three pages, followed by the transformer and then the AI factory.
Transformer 怎麼運作 How a transformer works
一個真的訓練過、小到可以親手驗算的 Transformer，就在你的瀏覽器裡跑。一步步看 token 變成向量、位置之間互相看、每個位置各自想、再變回一個字；另一半是同一個模型怎麼從隨機權重被訓練出來，以及為什麼「猜下一個字」會長出看起來像知識的東西——包括它學不到的、和它會自信地編出來的。
A real transformer with real trained weights, small enough to check by hand, running in your browser. Step through tokens becoming vectors, positions looking at each other, each position thinking alone, and the whole thing turning back into a word; then watch the same model trained from noise, and see why guessing the next word leaves something behind that looks like knowledge — including what it never learns, and what it confidently makes up.
A real transformer with real trained weights, small enough to check by hand, running in your browser. Step through tokens becoming vectors, positions looking at each other, each position thinking alone, and the whole thing turning back into a word; then watch the same model trained from noise, and see why guessing the next word leaves something behind that looks like knowledge — including what it never learns, and what it confidently makes up.
AI 工廠 AI Factory
AI 五層蛋糕（能源、晶片、基礎設施、模型、應用）實際上在哪裡、怎麼跑。一句 prompt 進到機房、落到某個機櫃某個 tray、在櫃內高速互連上被幾十張加速器一起算，電力與散熱撐著它，最後一個一個 token 回到你的螢幕；另一半是同一座工廠怎麼把這些權重訓練出來。
The five-layer cake — energy, chips, infrastructure, models, applications — as a machine you can watch run. Follow one prompt into the hall, onto a rack, onto a tray, across the fabric, through the power and cooling that make it possible, and back out one token at a time; then watch the same factory train those weights in the first place.
The five-layer cake — energy, chips, infrastructure, models, applications — as a machine you can watch run. Follow one prompt into the hall, onto a rack, onto a tray, across the fabric, through the power and cooling that make it possible, and back out one token at a time; then watch the same factory train those weights in the first place.
網路 Network
封包實際怎麼走、網路實際長在哪裡、位元組抵達之後瀏覽器做了什麼。
How a packet actually travels, where the internet physically is, and what a browser does once the bytes arrive.
網際網路地圖 Internet Atlas
網際網路實際長在地球哪裡。在 3D 地球上選一種接取方式、一個目的地、一個協定，看封包經過哪些交換中心、哪條海底電纜、跨過多遠的距離。
Where the internet physically is. Pick an access medium, a destination and a protocol on a 3D globe, and watch the packet cross real exchanges, real submarine cables and real distance.
Where the internet physically is. Pick an access medium, a destination and a protocol on a 3D globe, and watch the packet cross real exchanges, real submarine cables and real distance.
封包旅程 Packet Journey
輸入網址，一步步看封包經過 ARP、DNS 查詢、TCP 三向交握、TLS 握手，逐層打包、逐台設備轉送，再逐層拆開。
Type a URL and step through ARP, DNS, the TCP and TLS handshakes, and every header going on and coming off at every device.
Type a URL and step through ARP, DNS, the TCP and TLS handshakes, and every header going on and coming off at every device.
渲染旅程 Render Journey
封包送到之後呢？貼上 HTML 與 CSS，一步步看 tokenizer 建出 DOM、階層決定每個屬性、版面算出每個盒子、繪製與合成把它變成畫素。
What happens after the bytes arrive. Paste HTML and CSS and step through the DOM, the cascade, layout, paint and compositing, one box at a time.
What happens after the bytes arrive. Paste HTML and CSS and step through the DOM, the cascade, layout, paint and compositing, one box at a time.