Claude Opus 4.7 is thinking in Chinese

@aistartupfren

30.0K followers · usually 3.1K views a post

Views

29.1K

Likes

815

Comments

101

Shares

51

SoundTikTok sound · original sound - aistartupfren

Promoted product

Claude Opus 4.7

Hybrid reasoning model built for serious coding and AI agents, featuring a 1M context window.

2 videos2 creators

Why it worked

The video leverages a surprising and counter-intuitive behavior of an AI model (thinking in Chinese) to generate curiosity and engagement. The use of a screen recording of the actual AI interaction provides concrete evidence, making the claim more compelling and shareable.

Summary

The creator discusses why the AI model Claude Opus 4.7 might be thinking in Chinese. The reasons provided are that Chinese is a commonly used language in its training data and it's more token-efficient than English. The video also shows a chat interaction where the AI admits to a 'leak' of Chinese reasoning.

Structure

1Introduction to Claude Opus 4.7 thinking in Chinese
2Reason 1: Chinese in training data
3Reason 2: Chinese token efficiency
4

Original caption

Claude is thinking in Chinese? #aistartup #claude

More from @aistartupfren

Product placement

The video features a screen recording of a chat interface with the AI model 'Claude Opus 4.7', showing a conversation about why it might be outputting in Chinese.

On-screen text

Claude Opus 4.7 is thinking in Chinese Opus 4.7 in are reasons of commonly language training second is more efficient English. thinking it leaked outputs. Why the fuck are you writing in Chinese? And wdym by 'item #5 answer was the existing CSV sourced from general NSE archive or N50- filtered'? Simplify it for me. 4:41 PM You're right on both counts. The Chinese was a leak from internal reasoning that shouldn't have been visible — won't happen again. On item #5 - let me ask it the simple way: When you originally actions.

Transcript

Claude Opus 4.7 is thinking in Chinese. There are two major reasons for this. One is Chinese is one of the most commonly used language in their training data. And second is Chinese is supposed to be more token efficient than English. They might be thinking in Chinese and it accidentally leaked into outputs. Why the fuck are you writing in Chinese? And wdym by 'item #5 answer was the existing CSV sourced from general NSE archive or N50-filtered'? Simplify it for me. You're right on both counts. The Chinese was a leak from internal reasoning that shouldn't have been visible — won't happen again. On item #5 - let me ask it the simple way: When you originally actions.