Index of everything on this site: https://ai.hiina.space/index.md
Start with X1 (one day on a single 24 GB consumer GPU, reuses the AISI repo) because it answers the literal question: whether a model can reach a drug-like state by attending instead of by injection. X3 is the most promising for a real result, since it fills the gap on long-horizon agents and tests whether "flow" is more than calm.
| # | Question | Model | Compute | Time |
|---|---|---|---|---|
| X1 | Can prompted meditation reach the same internal state as an injected calm vector, at lower cost? | Qwen3-8B | 24 GB consumer GPU | 1 day |
| X2 | Do successful reasoning traces have a different "fork rhythm" from failed ones? | Qwen3-8B (thinking) | 24 GB consumer GPU | 1 to 2 days |
| X3 | Is there a flow latent on agent rollouts, and does gated steering with it beat prompting? | Qwen3-8B, then 14B quantised | 24 GB consumer GPU; optional rented H100 | 4 to 6 days |
| X4 | Does distilling flow-steered rollouts (VSPO-style) beat GRPO, and does it forget more? | Qwen3-1.7B locally, Qwen3-4B rented | ~2 H100-days, roughly $100 to $250 | 1 to 2 weeks |
| X5 | Is a "meditation break" for a stuck agent more than a context reset? | Claude via API, plus Qwen3-8B | API budget, 24 GB consumer GPU | 3 to 5 days |
Compute figures are my estimates: Qwen3-8B in bf16 is ~16 GB and fits a 24 GB card for batched vLLM generation and HF activation capture; 14B needs 4-bit. A builder who already runs Qwen3-8B under vLLM for other agent experiments can reuse that setup for X3 and X5.
Setup. Clone llm-self-steering. Build "calm" and "focused" vectors with its story recipe. Four arms on its frustration tasks and GSM8K: baseline; a meditation preamble the model runs before each answer (attend to the goal as an anchor, note distraction, return), with no vector; injected calm vector at the calibrated dose; random vector of equal norm.
Metrics. Projection of the residual stream onto the calm vector by layer and token, per arm; GSM8K accuracy; capitulation rate when the user rejects correct answers; self-steer rate when the tool is offered.
Interesting result. The preamble moves activations a meaningful fraction of the way along the calm direction (Lindsey-style endogenous control), keeps accuracy where injection costs it, and lowers self-dosing under frustration. That would be a clean "meditation is endogenous pharma without the side effects" finding. A null (prompt moves nothing internally) is also worth writing up.
Setup. Generate ~2,000 thinking traces on MATH and AIME-style problems at fixed temperature. Per trace, record the token-entropy series and the residual projection onto a success probe direction at 60 to 80% depth.
Metrics. Inter-fork interval distribution, autocorrelation and power spectrum of entropy, burstiness, and recurrence quantification; compare correct versus incorrect traces matched on length and problem.
Interesting result. Correct traces show a more regular cadence of decision points, a literal "groove", or failed traces show runs of forks with no settling (thrashing) before collapse into repetition. Either gives a timing feature for gating X3. A flat null says the rhythm framing is decoration and the action is in where forks happen, not when.
Setup. Run Qwen3-8B as a tool-using agent on short coding tasks with tests (or ALFWorld for comparison with NPM). Collect paired success and failure rollouts on the same task. Extract a direction at action-decision tokens and at forking tokens, mean-difference and first PC, at 60 to 80% depth. Project out length and difficulty directions.
Arms. Gated steering when a failure probe fires (AdaRAS-style); always-on steering; prompt baseline ("you're making good progress, stay with it"); Anthropic-style calm vector; random vector; a strength sweep for each.
Metrics. Probe AUC and lead time; task success; trajectory length; KL to the unsteered policy; cosine between the flow direction and calm, focused and confident vectors.
Interesting result. Gated flow steering beats the prompt baseline at matched KL, and the flow direction is only partly explained by calm (say cosine below 0.5), meaning there is a task-competence component beyond affect. Scale to 14B quantised once it works at 8B, since the published methods stop at 14B and nothing exists for 30B+ coding agents.
Setup. Take the X3 direction and plug it into a VSPO-style loop: sample GRPO groups at several steering intensities, filter by true reward, update the unsteered policy with a KL anchor. Compare against plain GRPO and rejection-sampling fine-tuning at matched rollouts.
Metrics. Task success; forgetting on held-out benchmarks; forward KL to the base model, the predictor from RL's Razor.
Interesting result. Faster early learning than GRPO with no extra forgetting. That would be the first evidence on how a steered behaviour policy affects forgetting, which RL's Razor leaves open. This is the "state becomes trait" half of meditation.
Setup. In an agent loop, detect rumination: repeated actions, rising self-similarity of recent turns, or entropy collapse. When it fires, apply one of: a meditation prompt (re-read the goal, state what you notice about your last few steps, drop current hypotheses, return); a plain context reset; a calm vector (open model only); nothing. Use tasks known to induce loops, such as debugging with misleading errors.
Metrics. Loop-escape rate, task success, tokens to success, and whether the agent re-enters the same loop.
Interesting result. The meditation prompt beats a matched reset, meaning the attentional instruction adds something beyond clearing context. It also tests the one popular anecdote in this space with controls.
Text licensed CC BY 4.0 (attribution: Hiina, ai.hiina.space). Source: https://github.com/hiinaspace/ai-hiina-space