March 2, 2026
Botsona drama: MBTI but make it silicon
Language Model Contains Personality Subnetworks
Chatbots have hidden alter‑egos — commenters argue it's vibes vs science
TLDR: Researchers claim AI chatbots contain built‑in “persona” circuits that can be toggled without extra training, even splitting introvert vs extrovert. Comments split between “language shapes behavior,” “this is tautological,” and “stop calling it personality,” with links to Anthropic’s related work fueling the debate.
Turns out chatbots might have hidden “alter‑egos” baked in. Researchers say they found “persona” circuits inside AI models that can be switched on or off—no extra training, no long prompts. Using small calibration sets, they spot signature patterns, then “mask” parts of the model to pull out an introvert vs. extrovert, and more. A contrastive pruning trick sharpens the split between opposites. The punchline: these built‑in personas beat methods that rely on instructions, external lookups (aka retrieval‑augmented generation, where the bot fetches info as it writes), or fine‑tuning.
Cue the semantic brawl. sarducci cheers, arguing language itself shapes behavior: “to me this suggests that language strongly influences behavior.” D‑Machine shrugs: it’s all obvious and dull, calling it “tautological” and pointing to an earlier rant link. tl2do says stop saying “personality” at all—what we see is just outputs people rate as consistent. est asks if this relates to Anthropic’s persona selection, and the thread goes full MBTI meme: “INTJ until the cache clears,” “the bot needs therapy,” and “please don’t prune my ‘sarcastic coworker’ subnetwork.” In short: half the crowd is thrilled about controllable vibes; the other half insists it’s just labels and language games.
Key Points
- •LLMs contain persona-specialized subnetworks within their parameter space.
- •Small calibration datasets reveal distinct activation signatures for different personas.
- •A masking strategy isolates lightweight subnetworks to align outputs with target personas.
- •Contrastive pruning enhances separation for binary-opposing personas (e.g., introvert vs. extrovert).
- •The training-free approach outperforms external-knowledge baselines in persona alignment and efficiency.