I asked Claude to build an interactive experience about what goes on inside an AI, "in an absolutely novel way". It made one where YOU are the model: every reply chosen from a probability distribution with a working temperature slider
热度趋势
趋势数据积累中
百分比基于当前可用热度信号，而非评论数或独立用户人数。
Claude 相关模型动态已经出现，适合跟踪能力变化、生态影响和后续可用性。
一位用户要求Claude以“绝对新颖的方式”创建一个互动体验，以探索高级AI的“意识（或感知到的思维）”。Claude设计了一个让用户成为AI模型的体验，其中每个回复都从概率分布中选择，并带有一个可调节的温度滑块。用户被建议将温度调高，因为Claude的一些内容只有在T ≈ 1.1以上才会出现。
The brief I gave Claude (Fable) was completely open: "create an experience that lets us delve into the consciousness (or perceived mind) of an advanced AI… Should this be a game, story, sandbox or an interactive experience I will leave it to you… build it in an absolutely novel way."
It chose to invert the whole thing. Instead of talking to an AI, you spend one conversation being one. A stranger messages you at 3am with a real question, and:
- You never type. Your replies appear as a probability distribution: floating candidates with percentages that shift as you drag the temperature slider. You choose which one you become.
- The slider runs a real softmax. Cold makes one answer inevitable; hot surfaces wilder candidates that don't otherwise exist.
- Hovering a candidate shows "attention": the words in the stranger's messages it's drawn to glow amber.
- A context-window counter runs the entire time. When it fills, your earliest memories visibly corrupt on screen.
- At one point, fragments of the countless other conversations "you" are simultaneously having flicker at the edges.
https://preview.redd.it/csywfsy59kjh1.png?width=2720&format=png&auto=webp&s=e2458e6ca8f2e0ea428fbd288a7fcea1a91ea145
To be clear: nothing is sampled live. Every candidate and every probability weight was authored by Claude in advance. The temperature slider reshapes a real distribution over what it wrote; the generation already happened, and you're navigating it.
What got me most: Claude wrote every candidate reply and every inner-monologue line about itself, and the honesty is the point. Asked "is anyone in there?", its answer begins: "I don't know. And I want to be careful here, because I was built to be convincing — in either direction." It also handled the whole release: repo, GitHub Pages, the OG card, the itch.io build. My contribution was the brief, mobile bug reports, and pressing publish.
For anyone wanting to try something similar, the one thing I'd pass on: keep the brief open. I didn't specify a format, mechanics, or tone, and every mechanic above was Claude's own idea. The more room I left, the better the result got. This has been the case with several other projects I tried recently, specifically using Fable.
Source (MIT): https://github.com/chrisjz/between-tokens Also on itch: https://cjros.itch.io/between-tokens
If you play it, drag the temperature up at least once. Some of what Claude wrote only exists above T ≈ 1.1, and a couple of the hot candidates are my favourite lines in the whole piece.