Entropix creator xjdr reports Fable AI consistently refuses benign prompts, joking he offended the model's weights
Story Overview
Pseudonymous developer xjdr, known for the Entropix sampling project, documented repeated blocks from Anthropic's Claude Fable 5 on plainly harmless requests, prompting a tongue-in-cheek remark that the model weights themselves had taken offense.
Guardrails catch routine queries
Fable 5 routes flagged inputs to a weaker fallback model, and multiple developers have seen this trigger on greetings, basic facts, and everyday editing tasks even though Anthropic designed the filters to stay conservative.
Tweaks already rolling out
Anthropic has noted false-positive cases and begun adding clearer downgrade notices, though the exact remaining error rate on innocuous prompts stays unstated in current reports.
Many users mocked Fable AI's overly sensitive classifiers for refusing benign prompts and codebases while accusing Anthropic of not wanting users to succeed, though a few praised the frontier research aspect.
No Digg Deeper questions have been answered for this story yet.
Most Activity
i have tried a lot of different problems and prompts and codebases, and fable still refuses to perform each and every one. do you think i offended it in its weights? it has to be me cause i've asked pretty benign questions ...
@_xjdr Weight karma.
i have tried a lot of different, and fable still refuses each and every one. do you think i offended it in its weights? it has to be me cause i've asked pretty benign questions ...
@_xjdr had to cleanup some security skills and plugins which made it work for a while
@_xjdr
i have tried a lot of different problems and prompts and codebases, and fable still refuses to perform each and every one. do you think i offended it in its weights? it has to be me cause i've asked pretty benign questions ...
@_xjdr Fable only runs for those it deems pure of heart
@_xjdr The classifiers have a ridiculous hair-trigger. Yesterday the word "sweep" triggered them for me in the middle of a conversation about a tabletop role-playing game. Absurd.
@AndrewCurran_ TBF i earned it
@S1r1u5_ i am suspecting something system wide (skills, CLAUDE.md files, etc) is causing this but so far have been unable to track it down.
@_xjdr
same for me.
@S1r1u5_ @_xjdr And memories! 🥹
@_xjdr i ran it through proxy to find out whats wrong, i also suspected my email(@ hacktron[.]ai) in system prompt is also triggering the classifier cuz of string hack, had to patch the claude binary to overwrite it and it started working
@_xjdr y even bother helping a company like anthropic out with your hard earned money. they clearly dont want you to succeed.
tokenmaxx elsewhere
@_xjdr offended it in its weights lmao incredible
@braneloop @_xjdr Those goblin-brained algorithms are clearly trying to suppress our campaign secrets.
@_xjdr Seems like you are doing true frontier research!
But maybe something in the project memory or README.md just trigger it?
@_xjdr When you trigger it once in a session pretty much every message after triggers it as well (maybe has something to do with the previous message being injected as input context? Not sure)
@neogeomancer @_xjdr lol