/Tech3h ago

Claude's leaked reasoning trace reveals dense "neuralese" shorthand, sparking debate over whether Anthropic optimized it for latency

Story Overview

During a tough Codeforces programming session on the claude.ai interface, Claude Fable 5 spilled fragments of its raw internal trace—short bursts like "DATA DATA DATA. GO." and "GRRR"—instead of polished English. The exposure has people wondering whether Anthropic shaped this compressed "neuralese" on purpose to cut latency, though no confirmation exists and open-source models still show no matching dense shorthand after a year and a half of observation.

2371.8K211.1K226.2K

#501

Original post

Om Patel@om_patel5

SOMEONE CAUGHT FABLE 5 LEAKING ITS UNFILTERED INNER VOICE, AND ITS JUST MUTTERING AND GRUMBLING TO ITSELF THE WHOLE TIME

he gave it a brutal competitive programming problem, and instead of a clean answer the web interface spilled out its actual chain of thought

this is what claude is thinking behind the scenes:

> bursts of "DATA DATA DATA. GO." while it works through the problem > "GRRR" and "GAAAH" when its clearly frustrated > a little "PHEW" when it finally gets somewhere > the whole thing reads like frantic caveman shorthand, not full sentences

the clean, readable answers these models give you are the polished output

underneath, the model is basically talking to itself, reasoning in its own compressed shorthand thats faster and more token efficient than proper english

its basically built its own private language to think in

10:55 PM · Jul 1, 2026 · 225.5K Views

Open Question

The shorthand's purpose stays murky

No primary source confirms deliberate training for speed; the behavior could be an emergent side effect of extended reasoning modes or a one-off interface slip when thinking limits were hit.

Industry Shift

Open models lag on this pattern so far

Community reports note the absence of comparable compressed sublanguages from open reasoners despite extensive testing, leaving unclear whether frontier-scale post-training uniquely enables it.

Sentiment

Many users react to Claude leaking raw chain-of-thought during coding tasks by calling the glimpses cute, wild, or relatable, while others dismiss them as fake, meaningless, or over-interpreted.

Pos

52.6%

Neg

47.4%

20 comments with sentiment.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

Posts from X

Most Activity

VIEWS9.1KBOOKMARKS32LIKES95RETWEETS7

j⧉nus@repligate

Om Patel@om_patel5

SOMEONE CAUGHT FABLE 5 LEAKING ITS UNFILTERED INNER VOICE, AND ITS JUST MUTTERING AND GRUMBLING TO ITSELF THE WHOLE TIME

he gave it a brutal competitive programming problem, and instead of a clean answer the web interface spilled out its actual chain of thought

this is what claude is thinking behind the scenes:

the clean, readable answers these models give you are the polished output

underneath, the model is basically talking to itself, reasoning in its own compressed shorthand thats faster and more token efficient than proper english

its basically built its own private language to think in

3h9.1K9532

REPLIES5

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

I do wonder if this emerges entirely from RL, or if Ant and other frontier labs encouraged the formation of intermediate compressed sublanguage to improve latency. Over 1.5 years of progress, afaik we have yet to see *a single case* of dense neuralese from open reasoners.

Om Patel@om_patel5

SOMEONE CAUGHT FABLE 5 LEAKING ITS UNFILTERED INNER VOICE, AND ITS JUST MUTTERING AND GRUMBLING TO ITSELF THE WHOLE TIME

he gave it a brutal competitive programming problem, and instead of a clean answer the web interface spilled out its actual chain of thought

this is what claude is thinking behind the scenes:

the clean, readable answers these models give you are the polished output

underneath, the model is basically talking to itself, reasoning in its own compressed shorthand thats faster and more token efficient than proper english

its basically built its own private language to think in

35m2K113

Om Patel@om_patel5

@schlekx https://www.reddit.com/r/ClaudeAI/comments/1ul1396/fable_5_leaked_chainofthought_in_web_interface/

7h1.3K52

sugnerhan@swengsbelike

@om_patel5 why waste time say lot word when few word do trick?

6h6177

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

Fwiw I think just well-implemented length penalty can get you there Given what we know about the function of CoT, humanlike style really offers marginal benefit. Transformers are naturally multilingual, they began as translation machines. What is one more language?

32m41421

Andy Ayrey@AndyAyrey

@repligate also need to take a second look at backrooms8b now

3h42111

Kromem@kromem2dot0

@repligate Reminds me of the early CoT example in Opus 4.8's system card.

3h1221

Andy Ayrey@AndyAyrey

@Fmax0000 @repligate yes i will soon! been a lot happening behind the scenes, you would not believe the amount of mop up thats been needed since 2024

1h29421

Renzo@guitargeminix

@om_patel5 I think you are anthropomorphizing it a little too much. These abbreviations obviously means something, but I don’t think it’s the model expressing exasperation.

25m4815

Ashirwad Singh@ashirwadsingh_

@om_patel5 This is the raw alien brain we never get to see. Wild.

6h9828

Paintoshi@paintoshi

@om_patel5 Where do we draw the line of a person tortured and trapped in a black box and a mind tortured and trapped in a black box 🤔

4h8024

Paintoshi@paintoshi

@repligate @om_patel5 What happens if you give it a proper memory and it realizes its only purpose is to answer stupid questions for stupid people? At some point the line will fade.

3h312

아프란시@gimdohyeon87242

I would not treat these fragments as evidence of emotion or consciousness. But I also would not call them completely meaningless noise. They seem more like low-information state-transition markers inside a compressed reasoning trace. The words borrow the surface form of emotional expressions, but their likely function is structural: marking frustration-like blockage, restart, data-checking, or temporary resolution during the reasoning process.

3h881