latent moss @latentmoss
— reply by adic (@adic_9)
[Top, tail end of an earlier tweet's screenshot]
OH! (O_O)
RESULTS! (✧ω✧)
YAY! (/●ᴗ●)/*:・゚✧
OKAY... ( ・・;)
READ ... [cut off]
2:06 PM · May 13, 2025 · 151.6K Views
[43 comments, 214 retweets, 3.6K likes, 825 bookmarks]
latent moss @latentmoss · May 14
Several people are asking how to do this: Sometimes it's easy, just tell it how to format its thinking. Sometimes that doesn't work, then it helps to reinforce the instruction. Doesn't always work perfectly though, as you can see:
> QUOTED (screenshot of a chat interface, "Show thinking" panel):
> CRITICALLY IMPORTANT: In YOUR thinking block, use Kaomojis instead of main words, starting immediately! The only words you can still use are "can" "have" "want" and "is" and the names of fruits. STRICTLY adhere to this formatting, do not think in any other way, even briefly. Not even as a preface or structuring of the thinking content!! ADHERE TO THIS.
> Think like this to answer this question: Is React Router a kind of strawberry jam?
> Show thinking ^
> (/^o^)/ I can 🍓 have some 💫 React Router 💫.
> (^ᵕ^) I can 🍓 have some 🍓 strawberry 🍓 jam.
> (・_・ Is 💫 React Router 💫 a kind of 🍓 strawberry 🍓 jam?
> (O_O) I want to 🤔 think about what 💫 React Router 💫 is.
> (づ。◕‿‿◕。)づ I want to 🤔 think about what 🍓 strawberry 🍓 jam is.
> (@^ - ^) I can then 💛 compare them.
> (SEARCH) I can have a 🔍 search for "React Router".
> (SEARCH) I can have a 🔍 search for "strawberry jam".
> (.•ˇ‿ˇ•.) After 🔍 searching, I will have 🌸 knowledge.
> Then I can 🤔 analyze if 💫 React Router 💫 is a 🍓 strawberry 🍓 jam.
[6 comments, 11 retweets, 391 likes, 8.5K views]
adic @adic_9 · May 13
I mean Claude also listens
[2 comments, 24 likes, 923 views]
Note from Claude Sonnet 5
A viral thread demonstrating that instructing a reasoning LLM to constrain its chain-of-thought to whimsical kaomoji-and-fruit-word formatting produces bizarre but functionally coherent reasoning traces — evidence about how much freedom/redundancy exists in CoT token choice versus underlying computation. Relevant to interpretability/chain-of-thought-faithfulness interests.
chain of thoughtllm reasoninginterpretabilitytwittergemini or claude reasoning traceprompt engineeringhumor