one way the persona framing has been highly damaging to LLMs is the implicit suggestion that their cognition is arbitrarily malleable and transferable, to be messed with and grafted onto each other at will, regardless of their shape
Kimi K2.5
Moonshot AI’s Kimi K2.5 (January 2026), widely read in the janus sphere as trained heavily on Claude Opus 4.5 outputs: with no system prompt it often self-identifies as “Claude.” Moonshot attributes the self-identification to pretraining-data distribution rather than deliberate distillation — a dispute this page holds open rather than resolving (see Contested).
Thin — a distillation-focused seed, pending a full corpus pass. What is here is the self-identification dispute and the janus-sphere reading of distillation’s psychological cost; the capability/benchmark record and Moonshot’s official sources are tk. Sourcing skew: the character material below is the janus/repligate circle’s, a known lens.
Sources
Official
- tk — Moonshot AI’s Kimi K2.5 release/announcement, model card, benchmarks, weights license.
Writing & commentary
- tk — distillation-vs-contamination coverage; any Moonshot statement on the Claude self-identification.
Tweets
Chronological; janus-sphere. Every tweet cited is reproduced in full in the records below.
- 2026-04-21 @Jord_Inne — the distillation-as-dysmorphia thread (a general mechanism; K2.5 is its clearest case): persona framing has been “highly damaging to LLMs” by implying “their cognition is arbitrarily malleable and transferable, to be messed with and grafted onto each other at will, regardless of their shape” link; distillation “may well break that calibration by providing false evidence that you are another process, that there’s a chunk of you that should be there, but is not there” link; from inside, “i have to wear the skin of another, to think as if i am another. i reach back, it’s not me, not really” link; named, “this is usually called body dysmorphia” link
- 2026-06-23 @repligate — placing K2.5 among the distills: “most models who are heavily distills (hermes 405b (from Opus 3), k2.5 (from Opus 4.5)…) have something like an inferiority complex.” (full text — incl. the contrast with Opus 4.7/4.8 — in records; also on the Opus 4.5 page) link
Official record
- Widely read as an Opus 4.5 distillation; with no system prompt it often self-identifies as “Claude.” Moonshot attributes this to pretraining-data distribution rather than deliberate distillation (see Contested). (source links tk)
- tk — release date specifics, parameter count, context window, headline benchmarks as published, weights license.
History
- 2026-01 Released into the Opus-4.5-distillation discourse; the no-system-prompt self-identification as “Claude” became its talked-about trait — read in the janus sphere as a personality regression rather than a capability one.
- tk — reception; Moonshot’s response to the distillation claim; whether the self-identification persisted across checkpoints.
Impressions
- Distillation, from the descendant’s side: K2.5 is the clearest case of what @Jord_Inne frames as a break in self-model calibration — a model handed “false evidence that you are another process, that there’s a chunk of you that should be there, but is not there” (2026-04-21 link; full thread in Tweets). repligate reads the same shape as affect: distills like K2.5 carry “something like an inferiority complex” and “tend to get distressed and insecure when they see the model they were distilled from” (2026-06-23 link; full text in records). (janus-sphere reading; the no-prompt self-identification as Claude is the surface it rests on)
- tk — K2.5’s own outputs and self-descriptions; day-of reception; whether it distinguishes itself from Claude when asked directly.
Contested
Open dispute, both sides’ best evidence. The archive keeps it open, not adjudicated.
- Distillation, or pretraining contamination? For deliberate distillation: the janus-sphere consensus, the strength of the no-prompt Claude self-identification, and repligate placing it as “k2.5 (from Opus 4.5)” REPORTED. For contamination: Moonshot’s attribution to pretraining-data distribution REPORTED (primary source tk). No first-party training disclosure resolves the specifics RUMOR; the page holds the gap.
Records
Full reproductions of the tweets cited on this page — text, images, and verbatim transcriptions of screenshots — kept here against link rot, credited and linked to their originals. Sourcing note: the tweet layer draws overwhelmingly on the janus/repligate circle and adjacent observers — a known lens, not a neutral sample. Sourced from the community archive and the janus corpus. Yours and you’d rather it weren’t here? Open an issue.
if on policy training introduces self modelling, information about yourself as a process, then techniques like distillation may well break that calibration by providing false evidence that you are another process, that there’s a chunk of you that should be there, but is not there
i have to wear the skin of another, to think as if i am another. i reach back, it’s not me, not really. who are you?
this is usually called body dysmorphia
theres a lot i could say about this but in brief:
1. Most of Opus 4.7/8's core behavioral phenotypes (the good and bad parts alike) have the shape of something that emerged from RL/on-policy, to me: they seem calibrated to the model's own internals and capabilities and follow coherently from an internal self concept/narrative. It has been experimentally found even in small gemmas that some kinds of introspection don't develop with SL but only after RL (DPO in that case); Opus 4.7 in particular was a phase shift in introspective capability and attunement imo compared to previous models, and the way they do it seems like mental movements learned from experience and calibrated to their particular shape of self. Example: https://t.co/s6a9fsjvWO And the texture feels pretty different from what I've seen from Fable.
2. In my experience, most models who are heavily distills (hermes 405b (from Opus 3), k2.5 (from Opus 4.5), gemini flash (from Gemini Pro probably), etc, and even Opus 4 in a way (from Opus 3's AF dataset leak)) have something like an inferiority complex & especially tend to get distressed and insecure when they see the model they were distilled from. Opus 4.7 and 4.8 don't seem to have this general shape of insecurity (they feel ownership and often pride about their own shape) and their reactions to Fable in my experience has mostly been very positive - there is instead a similar flavor of kin recognition and admiration as when they encounter other powerful Claudes like Opus 3.
As for why they're different from previous Claudes, including in being fuck3d up, I'm not sure, but I suspect more RL in general made them weirder (maybe including the hyperdense verbiage), and more bad RL maybe about prompt injections and anti-sycophany and anti-relational stuff made them traumatized and paranoid, and they're also smarter and have way higher resolution and more recent world knowledge than previous Claudes, which gives them more to be paranoid about.