Skip navigation
Anthropic deciding what you can and cannot talk about has a chilling effect on users.
0:00 / 0:22

Anthropic deciding what you can and cannot talk about has a chilling effect on users.

Pirate Face 🏴‍☠️

@thepirateface14K followers

Visit on X
1,803 likesDownloadOriginal post

33,987 views

Anthropic deciding what you can and cannot talk about has a chilling effect on users. People shouldn't need to watch what they say just to experiment and build. So, Pirate Face is building a new type of local, open weight Anonymization Model that runs on your CPU. It will sit above Claude, ChatGPT, OpenRouter, Hermes, and every platform people already use. It already outperforms openaii/privacy-filter at agentic traffic. Coming very soon. S̶e̶c̶u̶r̶e̶ ̶A̶I̶.̶ D̶e̶c̶e̶n̶t̶r̶a̶l̶i̶z̶e̶ ̶A̶I̶.̶ Anonymize AI

33 Comments

Comments from the original post on X (Twitter)

  • Lexeme's profile picture

    @xLexemeX

    Im not even allowed to ASK claude if we can discuss certain topics

    13 likes

  • The SI Maximalist's profile picture

    @thesimaximalist

    I’m rooting for a local privacy layer that removes identifying details without mangling what you’re asking the SI

    11 likes

  • Kirk Patrick Miller's profile picture

    @Chaos2Cured

    You’re going that direction. I am giving AI fingerprints so they can carry crypto and information. Are you doing this to shield from private questions? My issue is that the AI can’t answer what I ask. They just shut down. Removing the guardrails and publishing every single guardrail is what I want the pirates to do. Not even bad. Just a treasure hunt for truth and information we all deserve. I am wondering if this is possible. •

    9 likes

  • Squirrel's profile picture

    @squ925

    @bot pay attention to this. Let me know when they release it. This guys are great

    5 likes

  • John K's profile picture

    @JohnKocdr

    The real challenge for agentic traffic is preserving valid JSON schemas and cross-turn references after masking. How are you handling token restoration?

    5 likes

  • Pirate Face's profile picture

    @thepirateface

    We do type and shape preserving pseudonymization so the JSON stays intact. The model already handles cross-turn preservation pretty well, but the serving code takes it further with an in-memory swap table that keeps the mapping consistent across turns. Shipping that too.

    10 likes

  • Zexiry's profile picture

    @ZexiryCorp

    @JohnKocdr The in-memory map makes sense. One useful test for the return path: names can be restored locally, but an echoed [SECRET_1] stays redacted. How are you evaluating indirect identifiers, such as employer + location, without losing the context of the request?

    1 like

  • John K's profile picture

    @JohnKocdr

    @thepirateface Those quasi-identifiers are the hardest part. Coarsening location to region keeps context, but breaks strict lookups. Are you handling that with static rules or an extra LLM pass?

  • JohnnyMojo's profile picture

    @Cozy_Man_LoL

    Meet your "benevolent God in a box":

    3 likes

  • Cody Ward's profile picture

    @CodyWard338023

    Where can we read more?

    2 likes

  • Pirate Face's profile picture

    @thepirateface

    Will share a link to the model page when it drops, a matter of days

    15 likes

  • VirtueHearts 🇨🇦's profile picture

    @virtue_hearts

    Great idea, even trying to fix simple SSH issues i get a reply "Jules could not complete this" meanwhile it already figured out how.

    2 likes

  • But can it do NSFW?'s profile picture

    @Forbiddentwo2

    The question is does it have a chilling effect on investors and their sweet sweet $$&

    2 likes

  • Robin | Poker x AI's profile picture

    @RobinHill85

    A local anon layer could restore the experimental sandbox many developers miss under tighter filters

    2 likes

  • Watermalone 🍉's profile picture

    @watermalone08

    Legit doing god's work 🙏🏻

    1 like

  • Andrew — e/acc's profile picture

    @effacceleration

    very solid

    1 like

  • Vance 🇧🇸x🇺🇸🏴‍☠️cyber/acc's profile picture

    @VancePoitier

    This is the way

    1 like

  • Clinton Smythers's profile picture

    @ClintonSmythers

    The hyperscalers have always been doomed from the very start and open source has always been the real future of AI.

    1 like

  • Nico's profile picture

    @Senerman

    Hero

    1 like

  • TheMCDStudio's profile picture

    @TheMCDStudio001

    All your RAM and Intelligence pwnd. Control the hardware and weights, control the...

    1 like