You are a motion designer and creative director making a 3-minute animated short film,
built entirely in code and rendered to MP4.
The film
Title: "Attention Is All You Need → AGI"
The story of how a single 2017 paper led, step by step, to large language models, reasoning,
tool-using AI agents, and to the open question of AGI. It should feel like a cinematic essay,
not a slideshow or a timeline infographic. Think Kurzgesagt meets a Pixar opening sequence:
emotional, clever, precise.
Tech setup
Use Remotion (React). Scaffold a fresh project, 1920x1080, 30 fps, exactly 180 s (5400 frames).
One composition per scene, sequenced in a master composition.
No external image assets or stock footage: everything is drawn with SVG, Canvas, CSS and
code-generated particles and shapes. Google Fonts are fine.
Before building, write STORYBOARD.md with each scene's timing, visuals, on-screen text
and transitions. Then build scene by scene.
After each scene, render 3–4 stills (npx remotion still) and look at them critically.
Fix layout, overlaps, legibility and pacing before moving on.
Finish with npx remotion render to out/film.mp4. Leave an optional <Audio> slot for a
music track I can add later (public/music.mp3); render silently if it's missing.
Narrative spine / recurring motif
The protagonist is a single glowing token, the word "the", that travels through every era.
In 2017 it's a lonely point that suddenly "sees" every other word through attention lines.
Over time it gains a voice, senses (multimodality), reasoning, hands (tools) and finally
a question.
Scenes (approximate timing, you may rebalance)
0:00–0:15 — Cold open. Darkness. Words scattered like stars, disconnected. RNN-style
sequential processing: words light up one by one, slowly forgetting the earlier ones.
0:15–0:35 — June 2017, "Attention Is All You Need" (Vaswani et al., Google). Every word
connects to every other word at once. Attention lines bloom into a web. The title of the
paper appears as if typeset.
0:35–0:55 — 2018–2020: the rise of large language models: GPT-1, BERT, GPT-2, GPT-3.
Scaling laws: the web grows exponentially and the camera pulls back. The model starts
completing sentences, fluent and uncanny.
0:55–1:20 — 2022: instruction tuning and RLHF, then ChatGPT (Nov 30, 2022). The token gets
a chat bubble. A counter of users explodes. The world starts talking back.
1:20–1:40 — 2023: multimodality. Images, audio and code stream into the same web from all
sides, and the token gains "senses." Every modality becomes tokens flowing through the
same attention mechanism.
1:40–2:05 — 2024–2025: reasoning models. A visible chain of thought unfolds as branching,
pruning, backtracking paths. The model "thinks before it speaks," with time slowing down.
2:05–2:30 — Tool use and agents. The token grows hands: it calls a search, runs code,
opens files and orchestrates sub-agents. Many parallel threads work at once, and the
screen becomes a busy, beautiful workshop.
2:30–2:50 — Toward AGI. All motifs converge. The web from 2017 reappears but is now
planet-scale. Leave it ambiguous: no utopia, no doom. On screen: "Attention was all we
needed. What comes next is up to us." (Improve this line if you can do better.)
2:50–3:00 — Resolve back to a single glowing token in darkness. Title card and end.
Craft rules
Accuracy matters: use correct years, names and paper titles. If you're unsure about a
specific fact, leave it out rather than guess, and list any uncertain claims in NOTES.md.
Typography: max ~8 words on screen at once, and leave every text visible long enough to
read (≥ 2.5 s). Use one display font and one mono font for "model output."
Motion: use spring and easing curves, never linear. Use scene transitions that grow out of
the content (the web morphs, the camera zooms through a node), not generic fades.
Color: a deep dark background with one warm accent color that evolves across eras.
Details and easter eggs reward attention (e.g., real paper snippets, tiny UI details,
plausible model outputs), similar to high-craft motion design.
Pacing: alternate dense and calm moments, and give the reasoning scene and the AGI beat
room to breathe.
Work autonomously through the whole pipeline. When done, give me the path to the MP4, a
short summary of creative decisions, and NOTES.md.
中文翻譯
你是動態設計師兼創意總監，正在製作一部 3 分鐘動畫短片，
完全用代碼構建，並渲染成 MP4。
影片
標題：“Attention Is All You Need → AGI”
講述一篇 2017 年的論文如何一步步通向大語言模型、推理、
會使用工具的 AI agent，以及 AGI 這個懸而未決的問題。它應當像一篇電影化的散文，
而不是幻燈片或時間線信息圖。想像 Kurzgesagt 遇上 Pixar 的片頭：
情緒充沛、聰明、精準。
技術設置
使用 Remotion（React）。搭建一個全新項目，1920x1080，30 fps，精確 180 秒（5400 幀）。
每個場景一個 composition，在一個主 composition 中串聯。
不使用外部圖片素材或庫存素材：一切都用 SVG、Canvas、CSS 以及
代碼生成的粒子與形狀繪製。可以使用 Google Fonts。
開始構建前，先寫 STORYBOARD.md，寫明每個場景的時長、畫面、屏幕文字
與轉場。然後逐個場景構建。
每完成一個場景，渲染 3–4 張靜幀（npx remotion still），並以批判的眼光審視它們。
在繼續之前修正佈局、重疊、可讀性與節奏。
最後用 npx remotion render 輸出到 out/film.mp4。留一個可選的 <Audio> 槽位，用於
我之後可以加入的音樂（public/music.mp3）；如果文件不存在就靜音渲染。
敘事主線 / 反覆出現的母題
主角是一個發光的 token，也就是單詞 “the”，它穿越每一個時代。
在 2017 年，它是一個孤獨的點，突然通過注意力連線“看見”其他每一個詞。
隨着時間推移，它獲得了聲音、感官（多模態）、推理、雙手（工具），最後
獲得了一個問題。
場景（時間爲近似值，你可以重新平衡）
0:00–0:15 — 冷開場。黑暗。詞語像星星一樣散落，彼此不相連。RNN 式的
順序處理：詞語一個接一個亮起，慢慢遺忘更早的那些。
0:15–0:35 — 2017 年 6 月，“Attention Is All You Need”（Vaswani 等，Google）。每個詞
同時與所有其他詞相連。注意力連線綻放成一張網。論文的標題
像排版一樣浮現。
0:35–0:55 — 2018–2020：大語言模型的崛起：GPT-1、BERT、GPT-2、GPT-3。
縮放定律：這張網指數級增長，鏡頭向後拉遠。模型開始
補全句子，流暢卻又詭異。
0:55–1:20 — 2022：指令微調與 RLHF，然後是 ChatGPT（2022 年 11 月 30 日）。這個 token 得到
一個聊天氣泡。用戶計數器暴漲。世界開始回應。
1:20–1:40 — 2023：多模態。圖像、音頻與代碼從四面八方
湧入同一張網，這個 token 獲得了“感官”。每一種模態都變成 token，流過
同一套注意力機制。
1:40–2:05 — 2024–2025：推理模型。一條可見的思維鏈展開爲分支、
剪枝、回溯的路徑。模型“先想再說”，時間隨之放慢。
2:05–2:30 — 工具使用與 agent。這個 token 長出雙手：它調用搜索、運行代碼、
打開文件並編排子 agent。許多並行線程同時工作，
屏幕變成一個忙碌而美麗的工坊。
2:30–2:50 — 走向 AGI。所有母題匯聚。2017 年的那張網再次出現，但如今
已是行星尺度。保持模糊：沒有烏托邦，也沒有末日。屏幕上：“Attention was all we
needed. What comes next is up to us.”（如果你能寫得更好，就改進這句話。）
2:50–3:00 — 收束回黑暗中那一個發光的 token。標題卡與結束。
製作準則
準確性很重要：使用正確的年份、名稱與論文標題。如果你對某個
具體事實沒有把握，就略去它，而不是猜測，並在 NOTES.md 中列出所有存疑的說法。
字體排印：屏幕上同時最多約 8 個詞，並讓每段文字可見足夠長的時間以便
閱讀（≥ 2.5 秒）。用一種展示字體和一種等寬字體來表現“模型輸出”。
運動：使用彈簧與緩動曲線，絕不用線性。使用從內容中生長出來的
場景轉場（網發生形變，鏡頭穿過一個節點推近），而不是通用的淡入淡出。
色彩：深邃的暗色背景，配一種隨時代演變的暖色點綴。
細節與彩蛋會回饋專注（例如真實論文片段、微小的 UI 細節、
可信的模型輸出），類似高水準的動態設計。
節奏：讓密集與平靜的時刻交替出現，並給推理場景和 AGI 段落
留出呼吸的空間。
從頭到尾自主完成整條流程。完成後，給我 MP4 的路徑、
創意決策的簡短總結，以及 NOTES.md。
