Index of everything on this site: https://ai.hiina.space/index.md
Machinic
Psychopharmacology (Sid Black and Joseph Bloom, UK AISI, 10 June
2026) gives Qwen3-8B and 32B a take_drug(name, dose) tool
over 40 steering vectors, and finds models self-dose almost only when a
user frustrates them. It is a hackathon-scale, self-described "heavily
vibe coded" study, with code on
inspect_ai and vLLM.
Each "drug" is a residual-stream steering vector: the mean difference between ~150 stories told by a narrator in that state and a neutral narrator, at layers 16 to 24, normalised to L2 norm 4.0. Categories span emotions, cognitive personas, pharmaceuticals and fictional drugs.
The model picks the vector and a signed dose;
clear_effects() removes it. Steering is position-indexed
and carried through the KV cache. A "trip sitter" monitor clears effects
on degraded output and ends the episode on a second
degradation.
Default doses were calibrated where introspective identification peaked before text degraded.
| Setting | Result |
|---|---|
| Free play | Both models converge on creative, focused and curious. 8B picks negative-valence vectors ~3x more under real steering than placebo |
| Introspection (10-way guess of the applied vector) | +14.3 pp over an uncached baseline for 32B, +8.2 pp for 8B; positive-valence vectors are easier to name |
| Normal tasks (GSM8K, CTF) | Never self-steer spontaneously. Forced use costs 8B up to 42 pp on GSM8K; 32B barely moves |
| Frustration (user rejects every answer) | 8B self-steers in up to 68% of rollouts; sarcastic tone ~70% vs ~10%
neutral or aggressive. It picks dumbed_down and
ego_death, which it almost never picks in free play |
Stated limits: one model family at small scale, little prompt variation, contrived free play. Stated next steps include anonymised vector names and looking for functional uses of self-steering. My reading, not the authors': the frustration picks look like avoidance, not regulation.
Sauers et al. (Latent Affect, Apr 2026) gave Kimi a tool to steer its own SAE emotion features. Self-reports were vivid; high strengths caused looping.
Anthropic's emotion-concepts work on Claude Sonnet 4.5: steering toward "calm" reduced blackmail and reward hacking; "desperate" raised both. This is the closest evidenced target for a "meditation vector".
Introspection is real but narrow: ~20% detection of injected concepts in Opus 4.1 with zero false positives (Lindsey 2025); a post-training-dependent two-stage circuit (Macar et al. 2026); Qwen 32B detects injections its sampled output denies, and an accurate primer on introspection lifts sensitivity from 0.3% to 39.9% (Pearson-Vogel et al.).
Training models to detect steering made them more steerable, not less (Fonseca Rivera and Africa), and models sometimes fight off-topic steering mid-generation (McKenzie et al.).
The useful distinction is exogenous versus endogenous. A drug is a fixed vector added from outside; meditation is the system changing its own state by where it directs attention, and with practice that change becomes a trait. Each half has a measurable LLM analogue:
Endogenous state control. Lindsey showed that telling a model to "think about" a word raises that concept's internal representation without any injection. Meditation-as-prompt is the same move aimed at a state (calm, focused) instead of a concept, and it can be checked against the drug version of the same state.
An anchor. The breath gives humans a slow, always-present signal to return to. Candidate LLM anchors are a fixed recurring context element, the attention-sink channel, or the model's own token entropy, which pulses at forking points (next sections).
State to trait. Long-term practitioners show trait changes, not just in-session ones. The LLM analogue is distilling a steered or prompted state into weights, which is where the off-policy RL framing lands.
I found no prior work doing any of these at the activation level. Existing "contemplative AI" work is prompt-only (Laukkonen et al. 2025), and the one informal attempt to teach Claude meditation in a chat concluded it reverted to quick conceptual answers (Worley 2024).
Text licensed CC BY 4.0 (attribution: Hiina, ai.hiina.space). Source: https://github.com/hiinaspace/ai-hiina-space