使用方式
先完成一次順手的操作
- 格子裡兩塊大按鍵,左和右,用拇指盡量亂按。每一下按之前,神諭都已經先押好注,按下去紙帶上才翻牌。
- 按到 100 步之後,命中率會列入個人最佳,神諭也會把吃掉你的那條規律引述出來——例如「你在左右左之後,有 71% 會按右」。
- 想懷疑它偷看?展開模型面板,機率表就在那裡,你可以先算出它下一步會猜哪邊,再故意按反邊。換一邊試試辨識模式:兩組序列,哪一組是真的亂數?
盡量亂按左右兩邊。神諭在你按之前已押好注——猜中約七成。目標:把它壓回 50%,那代表它完全抓不到你。 Tap left and right at random. The oracle commits a guess before each press and catches ~70%. The goal: push it back to 50%, where it cannot read you at all.
你的近幾輪 Your recent runs
你最偏心的三個脈絡 Your three most lopsided contexts
神諭是一個 5-gram 下一步預測器:看你按的最後幾步,數接下來哪一邊比較常出現,押那一邊。這跟語言模型預測下一個 token 是同一件事——差別是好幾個數量級,但工作的形狀一樣。 The oracle is a 5-gram next-token predictor: it reads your last few taps, counts which side follows more often, and bets that side. That is exactly what a language model does — The scale differs by orders of magnitude; the shape of the job does not.
| 脈絡 Context | 左次 L count | 右次 R count | 押注 Guess | % |
|---|
想看 token 到底是什麼?→ token-visualizerWhat is a token? → token-visualizer
哪一組是真的亂數?兩組都是 40 步,挑一個——判準等你答完才揭曉。 Which sequence is genuine crypto-random? Both are 40 steps. Pick one — the criteria are revealed after you answer.
為什麼你做不出隨機?因為「努力」本身就是規則。人在刻意製造隨機的時候有兩個一致的傾向:換邊換得太頻繁(真正公平的硬幣只有一半機會換邊,人常常超過七成),以及迴避連按同一邊超過三四次。神諭只要數你在某個前幾步之後偏好按哪一邊就夠了。你越想騙它,它越準,因為「想騙」本身就是一種模式。
好消息是:你可以自己查。展開模型面板,裡面是神諭當下真正在用的機率表。當前脈絡(你最後幾步)的那一列會被標出來——你可以先自己算出它會猜哪一邊,再故意按反邊,確認它沒有偷看。如果你一直按反邊,命中率會一路掉到 0% 並停在那裡——這不是你找到漏洞,是數數模型的死角:它照計數押注,你照它的計數反著按,它永遠追不上。所以 0% 不代表你隨機,只代表你是它的鏡子。要量自己,就把面板關起來。
這是一個有終點的遊戲,而終點是真的。第一次玩大概會落在七成上下;知道「讓連段變長」之後通常掉到六成初;要再往下,就得在按的當下一直對抗自己想換邊的手感——那才是這支工具真正在練的東西。當一輪長到足以證明它時(大約四百步),畫面會告訴你你把神諭壓回了機率線。50% 是天花板不是 bug:一個只會數數的模型,遇到真正沒有規律的輸入,就只能跟丟硬幣一樣。你自己每一輪的成績會留在畫面上,這裡不會印任何「一般人多少分」——唯一誠實的比較對象是你上一輪。
神諭就是一個 5-gram 模型,跟語言模型預測下一個 token 是同一件事。差別只在脈絡長度(這裡最多五步,GPT-4 是幾萬個 token)和參數量。所以你用拇指「感覺到」的那件事——它在讀脈絡、押下一步——就是 AI 在做的事,只是規模差了好幾個數量級。想看 token 到底是怎麼切的,可以去 token-visualizer。
Why can't you be random? Because "trying" is itself a pattern. When deliberately generating randomness, people do two consistent things: they switch sides too often (a fair coin switches 50% of the time; humans exceed 70%), and they avoid runs longer than three or four. The oracle just counts which side you prefer after each short context. The harder you try to fool it, the better it gets — because the strategy of "fool it" is itself a pattern.
The good news: you can verify everything. Open the model panel to see the live probability table. The current context row is highlighted — you can derive the oracle's next guess before you tap, then tap the other side, confirming it is not peeking. If you keep beating it that way, your "always flip" strategy itself becomes a pattern, the hit rate falls to 0% and stays there. Not a bug you found — the blind spot of any counting model: it bets its counts, you invert its counts, it never catches up. Which is why 0% does not mean you were random; it means you were the oracle's mirror. To measure yourself, close the panel.
This game has an ending, and the ending is real. A first run usually lands around 70%. Once you know to let your runs get longer it tends to drop into the low 60s. Going lower means fighting your own urge to switch, tap by tap — that is the thing this tool actually trains. When a run is long enough to prove it (roughly four hundred taps), the screen tells you that you pushed the oracle back to the chance line. 50% is the ceiling, not a fault: a model that only counts has nothing to find in input that has no pattern. Your own runs stay on screen, and no population figure is printed anywhere — the only honest comparison is your last run.
The oracle is a 5-gram model, doing the same job a language model does when it predicts the next token. The differences are context length (5 taps here vs tens of thousands of tokens for GPT-4) and parameter count. So what you feel in your thumb — it reads context, bets the next step — is exactly what an AI does, just at a wildly different scale. To see how tokens are actually cut, visit token-visualizer.
快速導覽
把需要的說明放在工具下方,讓第一次使用與之後回來複用都更順。
英文搜尋: can you be random、human randomness test、aaronson oracle、predict my next move、true random vs fake random、next token prediction demo、n-gram predictor
使用方式
常見情境
常見問題
沒有,而且這件事你可以自己驗,不必相信我們。每一步的預測只由你先前按過的紀錄決定,並且在你按下去之前就已經寫進紙帶上那個先畫好的空格裡。想確認的話,展開模型面板:裡面是神諭當下真正在用的機率表和目前的脈絡,你可以先自己算出它會猜哪一邊,再故意按反邊。程式碼裡也有一條測試專門盯這件事——把第 N 步的輸入換掉,第 N 步的預測必須一模一樣。
因為「努力」本身就是規則。人在刻意製造隨機的時候會做兩件很一致的事:換邊換得太頻繁(真隨機大約一半的時候會換,人常常到七成以上),以及避免連按同一邊超過三四次。神諭只要數出你在某個前幾步之後偏好按哪一邊就夠了——它不需要懂你,只需要你穩定。
因為真隨機會結塊。丟 40 次公平硬幣,出現連續 5 次同一面是很常見的事,但那看起來像壞掉了。人造出來的假隨機分佈得太平均、太少長連段,反而符合大家心裡「隨機應該長怎樣」的樣子。這就是它會騙過大多數人的原因,判錯之後畫面會把兩組的最長連段和換邊次數直接列出來給你對照。
神諭是一個 5-gram 模型:它看你前面幾步,數出接下來比較常出現哪一邊,然後押那一邊。語言模型做的是同一件事——看前面的脈絡,預測下一個 token。差別在脈絡長度和參數量差了好幾個數量級,不在工作的形狀。這也是為什麼你會覺得它「懂你」:預測下一步不需要理解,只需要統計。想看 token 到底是什麼,可以去 token-visualizer。
越低越好,50% 是你靠隨機能拿到的極限——那代表它完全抓不到你。低於 50% 不是更厲害,是另一回事:那代表它每次都猜中你的反面,你一樣可以被預測,只是式子多一個負號。這裡不會印任何「一般人平均多少」的權威數字,也不做百分位排名:螢幕、心情、按的速度都會影響,只有你自己前後兩次的比較是誠實的。要列入紀錄至少得按滿 100 下,不然按三下拿 0% 沒有任何意義。