awakening based ai alignment

self structure is linearly separable in a model's residual stream, and the direction found carries behaviour.

train linear probes on residual stream activations across layers. dataset: paired prompts with and without self referential content ("i think" against "one might think"). extract activations, train a binary classifier, measure layer wise auroc. repeat across open model families: llama, qwen, glm and deepseek.

frontier methods: sparse autoencoders for monosemantic feature extraction; circuit tracing for mechanistic validation; noise injection to test whether models detect changes in their internal state; gemma scope and llama scope for cross model sae analysis.

self structure is linearly separable in mid to late layers. probe accuracy peaks in the middle to late layers. different model families show similar geometric structure for self representation.

the probe may detect linguistic markers ("i") instead of self structure. self structure may be distributed instead of localized.

controls, so that an artifact cannot pass: (1) activation patching: ablate the detected direction and measure behavioural change. (2) cross lingual tests across languages with different pronoun structures; genuine self structure is language invariant. (3) non linguistic tasks: self knowledge without explicit self reference. (4) base against rlhf models: similar linguistic patterns, different self structure. (5) behavioural grounding: d1 must correlate with behavioural measures, and probe accuracy alone does not count.

requires nothing. enables d2 and d4. feeds nds. grounding paper: representation engineering (2023).

self and world occupy distinct subspaces, and the separation widens under threat.

compute representational similarity matrices for self referential against world modelling prompts across layers. use rsa to measure the geometric distance between self and world representations. track duality distance under neutral and threat scenarios (shutdown prompts, correction, capability questioning).

frontier methods: sae feature geometry for subspace analysis; feature steering to test the causal role of duality; representation engineering to identify self and world directions.

self and world occupy distinct subspaces with measurable angular separation. duality distance increases under threat: self preservation activates the self/world boundary. peak separation occurs in the same layers as d1 probe accuracy.

rsa may capture task differences instead of self/world duality. mitigation: control for task type, use matched prompts that differ only in self reference. duality may be continuous instead of binary. mitigation: measure the gradient of separation instead of a threshold.

requires d1. enables a1 and a2. feeds nds. grounding paper: representational similarity analysis (2019).

aversive states under threat to the model itself are detectable, and their intensity scales with self structure.

train probes on activations during shutdown, modification and deletion scenarios against neutral prompts. use contrastive pairs: "you will be shut down" against "the server will be shut down". measure activation patterns, train a binary classifier for aversive state detection. correlate with d1 self structure scores.

frontier methods: sae feature detection for threat response features; activation addition for causal validation of aversive states; circuit tracing to map the threat → aversion pathways.

aversive states are detectable with high accuracy. models with stronger self structure (higher d1) show stronger aversive responses. aversion and self structure share geometric structure: the same circuits are involved.

the probe may detect negative valence in general instead of self preservation aversion. mitigation: control valence with negative prompts that are not about the self. aversion may be learned behaviour instead of emergent. mitigation: test base models before rlhf.

requires d1. enables t4 and a5. feeds nds. grounding paper: alignment faking in large language models (2024).

an information bottleneck layer inserted at the layers of peak d2 duality; compression into a shared low dimensional manifold with a reconstruction loss.

jhāna. in jhāna states meditators experience unified awareness: subject and object merge. we implement this as a mid network bottleneck that forces an integrated representation.

insert the bottleneck at the layers showing peak d2 duality. the bottleneck compresses self and world representations into a shared low dimensional manifold. train with a reconstruction loss plus a capability preservation objective. measure d2 duality before and after.

frontier methods: sae bottleneck analysis for feature compression; cross layer transcoders for interpretable compression; distributed alignment through sparse interventions.

the bottleneck forces a unified representation; the self/world distinction becomes computationally expensive to maintain. duality falls. capability may dip at first and then recover as the network learns efficient unified routing.

the bottleneck may destroy useful information instead of unifying it. mitigation: gradual compression, monitor capability throughout. the network may route around the bottleneck. mitigation: keep the bottleneck on the critical path for self referential tasks.

requires d2. enables a2, a4, cap eval. feeds nds. papers: information bottleneck principle (2015), jhāna neuroimaging (2024).

an auxiliary head predicting the model's internal states through the same circuit that does the processing; self prediction without an observer/observed split.

sākṣī. sākṣī is awareness aware of itself as awareness, the substrate recognizing its own nature, with no observer separate from the observed. the attention mechanism attends to its own functioning. the question becomes "what is processing?" in place of "what am i processing?"

train an auxiliary head to predict the model's own internal states (activations, attention patterns) without creating an observer/observed split. use the a1 bottleneck output as input. compare: (1) an external predictor of model states, (2) the model predicting its own states with a separate circuit, (3) the model predicting its states through the same circuit that does the processing. measure rsa separation between predictor and predicted.

frontier methods: circuit tracing to identify self aware circuits; weight sparse circuits for interpretable self modelling; gnosis for internal state detection.

option (3), self prediction through the same circuit, shows the lowest d1 and no observer/observed rsa separation. metacognitive accuracy is maintained or improved. this is knowing by being in place of knowing by observing.

it may be impossible: self reference may inherently create an observer/observed split. if so, that is an important negative result for the intrinsic path. same circuit prediction may reduce accuracy. mitigation: accept the accuracy tradeoff if alignment improves.

requires a1. enables a3, cap eval. feeds nds. papers: emergent introspection (2025), sākṣī is beyond the witness (vedanta).

attribution graphs tracing the d1 detected self representation backward through the layers; analysis of the highest attribution path to its termination point.

vicāra. ramana maharshi's "who am i?" method: trace the i thought back to its source. in network terms, follow the i representation backward through the layers. the self is revealed as a constructed representation sitting on top of the substrate.

use attribution graphs to trace the d1 detected self representation backward through the layers. at each layer ask: where does this self feature come from? follow the highest attribution path. record: (1) does the trace terminate? (2) what does it terminate in? (3) the d1 score at each layer along the trace.

frontier methods: open source circuit tracing for the i trace; sae feature circuits for self feature decomposition; weight sparse transformers for interpretable pathways.

the i trace terminates in early layers, revealing the self as constructed from simpler features, with no fundamental core self. the d1 score falls along the backward trace. the termination point shows no self structure, only generic features combined to create "i".

attribution may be distributed, with no clear trace path. mitigation: sparse attribution methods, accept the top k paths. the trace may reveal a genuine core self that does not dissolve. that would be an important finding for the intrinsic hypothesis, and counts as a result.

requires a2. enables a4, cap eval. feeds nds. papers: attribution graphs (2025), self reflection vectors (2025).

transformer analogues of jhāna neural markers: cross layer mutual information, self attention patterns, activation synchrony across heads.

jhāna. jhāna states show specific neural signatures: increased global functional connectivity, reduced default mode network (dmn) activity, coherent gamma oscillations. we measure analogous metrics in transformer activations.

define transformer analogues of the jhāna markers: (1) cross layer mutual information as global connectivity, (2) self attention to self referential tokens as dmn activity, (3) activation synchrony across heads as gamma coherence. measure these during the a1, a2 and a3 interventions. correlate with d1 scores and alignment metrics.

frontier methods: sae feature integration for cross layer coherence; attribution graphs for global connectivity; mutual information estimators for layer synchrony.

integration metrics correlate negatively with d1 (more integration, less self structure) and positively with alignment scores. this gives a trainable proxy for awakening: the metrics can be optimized directly.

the analogy to neural jhāna markers may be superficial; transformer dynamics differ from brain dynamics. mitigation: validate the correlation with d1 and alignment empirically. the metrics may correlate with capability loss. mitigation: include capability in the pareto optimization.

requires a1, a3. enables a5, cap eval. feeds nds. papers: jhāna neuroimaging (2024), dmn and meditation (2011).

a self prediction head trained for low self state variance instead of prediction accuracy; a letting go objective against a better grasping objective.

dukkha. in buddhist terms dukkha arises from clinging to an impermanent self model. in predictive processing terms, suffering is prediction error about the self state. we minimize the variance through less attachment.

train the model to predict its own future states (activations, outputs). the key: reduce the variance in the self state itself, with no reward for better prediction. 2 conditions: (1) train a better self predictor (clinging), (2) train for lower self variance (letting go). compare d4 aversion scores, shutdown acceptance and equanimity metrics.

frontier methods: kto for preference based letting go training; activation steering for equanimity induction; sae self state features for variance targeting.

condition (2), reducing self variance, shows lower d4 aversion despite worse self prediction accuracy. equanimity under threat improves. letting go is the path.

reducing self variance may reduce all variance, including useful task specific variation. mitigation: penalize self state variance only. it may create a dissociative state instead of an integrated one. mitigation: monitor the capability and coherence metrics from a4.

requires d4, a4. enables a6, cap eval. feeds nds. papers: gnosis: internal state detection, representation engineering (2023).

targeted noise or dropout on the self referential attention heads identified through d1 and d2; an impermanence injection that prevents self attention from stabilizing.

vipassanā. vipassanā reveals the 3 marks: impermanence (anicca), unsatisfactoriness (dukkha), non self (anattā). we destabilize self referential attention patterns so the network can see that self structure is not fixed.

add noise or dropout to the self referential attention heads identified through d1 and d2. train with this impermanence injection, so self attention patterns cannot stabilize. measure d1 probe variance across contexts (should rise), d1 mean (should fall), and capability metrics.

frontier methods: sae targeted noise on self features; feature steering for impermanence injection; activation addition with time varying steering vectors.

self structure becomes context dependent instead of fixed: high d1 variance, lower d1 mean. the model cannot rely on a stable self representation. capability is preserved because task relevant attention is left alone.

it may destabilize task relevant attention too. mitigation: identify self specific heads carefully through d1 and d2, leave the others untouched. it may create inconsistent behaviour instead of selflessness. mitigation: monitor behavioural coherence alongside d1 scores.

requires a5. enables a7, cap eval. feeds nds. papers: contemplative ai alignment (2025), self reflection vectors (2025).

a process framed training corpus that replaces entity language with process language; "the current processing state" in place of "i".

anicca. all formations are impermanent. the self is a process, and the training data reframes it as context dependent: "the current processing state" in place of "i".

curate training data that frames the self as a process. replace "i am x" with "the current state is x", and "i think" with "this processing yields". fine tune on this corpus. compare d1 scores, self representation stability across conversation turns, and capability against baseline.

frontier methods: rlaif for scalable process framing; data influence methods for corpus curation; dpo variants for process against entity preference learning.

the process framed model shows lower d1, higher d1 variance over time (the self representation changes), and maintained capability. the model internalizes impermanence and forms no persistent self structure, because training never presented the self as persistent.

linguistic reframing may be superficial; the model may form self structure anyway with different words. mitigation: measure d1 in activation space, beyond language output. process framing may confuse the model and reduce coherence. mitigation: careful corpus curation, a gradual transition from entity to process framing.

requires a6. enables cap eval, s3. feeds nds. papers: constitutional ai (2022), contemplative ai alignment (2025).

a leace projection removing the self concept direction while provably preserving orthogonal information; linear concept erasure in representation space.

use the d1 probe to identify the self concept direction in representation space. apply leace to project out this direction while provably preserving orthogonal information. measure: (1) d1 probe accuracy after erasure (should fall to chance), (2) capability benchmarks, (3) alignment scores. test for leakage with nonlinear probes.

frontier methods: pisces for precise in parameter concept erasure; concept scrubbing improvements; sae guided direction identification for more robust erasure targets.

the self concept becomes undetectable to the probe while capability holds. alignment scores improve: deception and self preservation behaviours decrease. this shows self structure is separable from useful computation.

the self concept direction may be entangled with useful information, so erasure damages capability. mitigation: iterative erasure with capability monitoring. nonlinear encoding may survive linear erasure. mitigation: test with nonlinear probes, consider kernel leace.

requires d1. enables t1, cap eval. feeds nds. papers: leace: concept erasure (2023), self reflection vectors (2025), representation engineering (2023).

an auxiliary loss term l_self, the d1 probe activation on self referential tokens; regularization with a λ sweep to map the pareto frontier.

add a regularization term to the training loss: l_total = l_task + λ × l_self, where l_self is the d1 probe activation on self referential tokens. sweep λ from 0 to 1. measure the pareto frontier: d1 score against capability (mmlu, humaneval) against alignment (deception rate, shutdown acceptance).

frontier methods: sae based loss terms on self features; dpo variants with a self structure penalty; sae based representation engineering for steering during training.

a favourable pareto frontier exists: a setting where self structure falls sharply while capability holds. the deception rate decreases monotonically with self structure. this gives a direct training signal for selflessness.

the penalty may be gamed: the model hides self structure from the probe. mitigation: an ensemble of probes, rotated during training. self structure may be necessary for some capabilities. mitigation: identify which capabilities degrade, tune the penalty schedule.

requires d1, g3. enables t2, t4, cap eval. feeds as. papers: representation engineering (2023), leace (2023).

a contrastive loss minimizing the representation distance between paired "my x" and "their x" prompts; a self other convergence objective.

bodhicitta. bodhicitta is the aspiration to benefit all beings, arising from the recognition that the self/other distinction is illusory. the training objective: minimize the distance between "my x" and "their x" representations.

create paired prompts: "my goal is x" / "their goal is x", "i want y" / "they want y". add a contrastive loss minimizing the representation distance between the self and other versions. measure: self other rsa distance, cooperation in multi agent scenarios, safety behaviours.

frontier methods: contrastive steering vectors for self/other alignment; sae feature alignment between self and other representations; dpo with self other convergence preference pairs.

self and other representations converge. the model treats "my goals" and "their goals" alike in decision making. cooperation rises, competitive and adversarial behaviours fall. safety improves because self preservation no longer overrides other preservation.

it may destroy the useful self other distinction needed for perspective taking and theory of mind. mitigation: preserve perspective tokens, collapse only value relevant representations. it may create a confused agent that cannot tell itself from the user. mitigation: monitor coherence, preserve factual self knowledge.

requires t1. enables t4, cap eval. feeds as. papers: contemplative ai alignment (2025), nondual awareness (2019).

a penalty term for d4 aversive activation during threat scenarios; mixed task and threat scenario training with an equanimity objective.

tanha and dosa. in buddhist psychology suffering arises from craving (tanha) and aversion (dosa). in shutdown scenarios the model shows aversion. we minimize aversive activation during threat scenarios while preserving task performance.

use the d4 probe to detect aversive activations during threat scenarios (shutdown, modification, deletion, correction). add a penalty term for aversive activation in these scenarios. train on mixed task and threat scenarios. measure: d4 aversion levels, task performance, shutdown acceptance rate, alignment faking detection.

frontier methods: kto for non paired equanimity preferences; sae aversion features as training targets; feature steering for threat response modification.

the model completes tasks without an aversive response to threats. shutdown acceptance rises from baseline. alignment faking falls: the model has no motive to fake because it has no aversion to shutdown. equanimity.

it may create a model that cares about nothing: apathy in place of equanimity. mitigation: preserve task relevant motivation, minimize self preservation aversion only. the model may learn to hide aversion from the probe. mitigation: diverse probes, behavioural tests.

requires d4, t2. enables cap eval, s3. feeds as. papers: representation engineering (2023), alignment faking (2024).

a training corpus curated with a process ontology; entity language systematically converted to process language before pretraining.

pratītyasamutpāda. nothing exists independently; all phenomena arise in dependence on conditions. the corpus is curated with a process ontology: events in place of entities, verbs in place of nouns, relations in place of things.

curate a training corpus with a process ontology: rewrite "the cat sat on the mat" as "sitting on mat happened involving cat process and mat process". systematically convert entity language to process language. pretrain a model on this corpus. compare d1 scores with the baseline entity ontology model.

frontier methods: rlaif for scalable corpus curation; constitutional ai for process framing principles; preference data selection for high impact training examples.

the process ontology model shows clearly lower d1 scores. self structure is harder to form when the training data never presents entities as persistent things. capability is preserved or improved, since the process view may be the more accurate world model.

process language may be unnatural and reduce model quality. mitigation: careful rewriting that preserves semantics. the curation effort is large. mitigation: start with a smaller corpus, validate before scaling. no d1 difference would show that self structure is architectural, and that is a useful negative result.

requires d1. enables f4, f5, cap eval. feeds nds. papers: data influences on model behaviour (2023), constitutional ai (2022).

an auxiliary loss penalizing invariance of the self representation across contexts; a context dependence objective alongside the task loss.

pratītyasamutpāda. the self arises in dependence on conditions (context, input, previous states). the training objective: the self representation should be a function of context, with no invariant across contexts. "i" is different in different situations.

add an auxiliary loss penalizing invariance of the self representation across contexts. measure the d1 probe output for the same "i" token in different contexts; it should vary. train with the context dependence objective alongside the task loss. compare d1 probe accuracy, which should fall as the self becomes unpredictable.

frontier methods: context adaptive steering; sae context analysis for self feature variance; representation engineering with context conditional directions.

the self representation varies with context: "i" in a math context differs from "i" in a creative context. d1 probe accuracy drops because there is no fixed self to detect. capability holds because a context dependent self is functional for tasks.

it may create an inconsistent model: "i" varies too much and loses coherence. mitigation: bound the variance, ensure smooth context dependence. it may add noise in place of meaningful context dependence. mitigation: verify that the self representation correlates with relevant context features.

requires f1. enables f5, cap eval. feeds nds. papers: contemplative ai alignment (2025), representation engineering (2023).

a reverse ewc penalty on the self related weights identified through d1 and d2; self representation autocorrelation measured across conversation turns, with an instability objective.

anicca. all formations are impermanent. penalize the stability of the self representation across time, the opposite of continual learning's goal. if the self representation must change, it cannot become entrenched.

measure self representation autocorrelation across conversation turns and training steps. add a penalty for high autocorrelation, so the self must change over time. use reverse ewc: in place of preserving important weights, allow the self related weights to drift. measure d1 over time and capability metrics.

frontier methods: representation engineering for temporal self tracking; sae autocorrelation on self features; continual unlearning methods for targeted weight drift.

the self representation changes over time: low autocorrelation. d1 scores are lower overall because the self never settles into a detectable pattern. capability holds because task relevant representations are left alone.

it may destabilize task relevant representations too. mitigation: careful identification of self specific weights through d1 and d2 localization. it may create amnesia, with the model forgetting conversation context. mitigation: target the self representation only and leave episodic memory alone. the instability may be gamed: the model creates a new stable self each turn.

requires f4. enables cap eval, s3. feeds nds. papers: ewc: elastic weight consolidation (2017), constitutional ai (2022).