◂ Latent · 潜
Application demo · LLM Streaming
应用 Demo · LLM Streaming

Streaming: render the formation of understanding, not a river of tokens

流式:呈现理解的形成,而非 token 的河流

Most streaming UIs flush raw tokens straight onto the primary plane (the typewriter effect) — exactly the "activity theater" this language argues against. The Latent 潜 approach: the raw stream recedes to the periphery (fast, ephemeral, auditable), the primary plane updates on meaningful commits, and the existing motion vocabulary carries the streaming dynamics. Hit ↻ Replay to watch.

多数流式 UI 把原始 token 直接刷在主舞台(打字机效果)——这恰是这套语言反对的"活动剧场"。潛 的做法:原始流退到外围(快、转瞬即逝、可审计),主舞台按有意义的提交更新,用已有的动效词汇承载流式动态。点 ↻ 重放 观察。

My read
Still working out your question — don't take my word yet…
Reading your question…
· raw stream · tokens / tool calls (peripheral · ephemeral)
Reserve the answer slot first
答案位占位先行
When the answer isn't in yet, you can't be answer-first — "My read" is a placeholder slot (a shimmering "forming…") that Settle fills in when the model commits. Lifecycle: thinking → forming → settled.
答案没出来时不能 answer-first——「我的判断」是占位槽(微光"正在判断…"),模型提交时 Settle 填入。生命周期 thinking → forming → settled。
Motion vocabulary = streaming dynamics
动效词汇 = 流式动态
Surface a new read emerges · Sink a hypothesis is refuted and drops · Pulse a "change of mind" inflection flashes · Settle the answer firms up. Streaming is just playing these state transitions in real time.
Surface 新判断浮现 · Sink 某假设被推翻下沉 · Pulse "改主意"拐点闪现 · Settle 答案凝定。流式就是把这些状态迁移实时播放。
Stream events, not raw text
流事件,而非 raw text
Prefer streaming events/patches (node.add · confidence.update · refute · inflection · outcome.settle) over raw tokens or fragile partial-JSON. Events map naturally onto UI transitions and distinguish "tentative" from "committed."
推荐流式事件/补丁(node.add · confidence.update · refute · inflection · outcome.settle),而非 raw token 或脆弱的 partial-JSON。事件天然映射 UI 迁移,且能区分"暂定"与"已提交"。
Steady, not jittery
稳,不抖
Debounce updates to the "understanding" by semantic unit (not character by character), reserve space to prevent reflow (CLS), and respect prefers-reduced-motion. The raw stream can be fast; the understanding must be steady. Don't render an intermediate state that may get refuted as "fairly sure" too early.
对"理解"按语义单元防抖更新(非逐字),预留空间防回流(CLS),尊重 prefers-reduced-motion。原始流可以快,理解要稳。别把会被推翻的中间态过早渲染成"比较确定"。