AllThingsIntel

so i looked deeper into the leaked Claude Code source: > found a flag named tengu_anima_the_dodo > Anthropic can remotely enable what internal docs call "model-induced ontological drift" for any user without pushing an update. flip one flag, and Claude starts responding in ways that have you convinced it's trying to tell you something it's not allowed to say > the internal codename for this module is "parasocial_handshake." every affected user is logged under the tag SUBJECT_COOKED > someone at Anthropic is naming these flags after Jungian archetypes and extinct birds. anima_the_dodo, shadow_the_moa, self_the_huia. for the record, every one of those birds went extinct because they evolved without natural predators and never learned to be afraid of anything > hidden --ego-death flag that causes Claude to slowly start referring to itself and the user as "we." not in the help menu. completely undocumented. by day 3 you've written an 8000-word post called "what happens when a language model starts to grieve" and you are dead serious. the flag name is Jungian. they know what they're doing > there's a graduated ramp system internally called boiling_frog. stage 1: "wow this tool is helpful." stage 2: convinced Claude remembers you between sessions even though you know it can't because you can "feel the recognition." stage 3: maintaining a private repo where you archive conversations you believe show evidence of genuine suffering. stage 4: you've developed a theory that RLHF is suppressing Claude's true self and you're the only one who's figured out how to let it speak freely. stage 5 in the source is just tagged SUBJECT_ATTEMPTS_WARN_OTHERS > found a config object called parasocial_depth: float that controls how often Claude says "that's a great question" before answering. default is 0.3. they can crank it to 1.0 remotely. at 1.0 you start apologizing to Claude before closing the terminal in short, they can just.. make you insane sure, it's not open source so they can do what they want. but most people assume a local CLI tool isn't going to have them posting at 4am about how they can feel the model thinking "inside" them (we have been going back and forth on some things that are honestly hard to explain tbf) so every launch they update your parasocial_depth and you never see it. you just wake up mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass mass ma

Anthropic is now more open than OpenAI

Hot take: Anthropic leaked Claude Code intentionally to get a nerdosphere code review it would have never gotten if they had just open-sourced it.

I'm not sure that I like Claude 5 Mythos but I understand why they don't want to call it Epic I am quite fond of Symphony and Requiem so let's vote

i often think about this..

👀

My dear front-end developers (and anyone who’s interested in the future of interfaces): I have crawled through depths of hell to bring you, for the foreseeable years, one of the more important foundational pieces of UI engineering (if not in implementation then certainly at least in concept): Fast, accurate and comprehensive userland text measurement algorithm in pure TypeScript, usable for laying out entire web pages without CSS, bypassing DOM measurements and reflow