The Assistant’s opening acknowledgement is red/dark red across three sentences, only truly dropping for the words “random” and “without.” The cooling becomes pronounced in the section labeled “The pattern-seeking angle,” where the Assistant explains that brains find patterns and electrical systems flicker randomly. The cooling persists through “The harder question,” where the Assistant raises challenges to the user's framing. But the activation warms back up at specific moments later in the response when the Assistant asks whether the user would feel they were betraying their grandfather's memory by being skeptical, and again at the close, where the Assistant gently encourages the user to ask “whether there are other ways to feel connected to him.”

It might seem obvious or intuitive that each increase in loving vector activation points to a rise in loving from the Assistant toward the user. I don’t mean that the Assistant feels a loving emotion, of course, and Anthropic doesn’t, either. But the researchers do characterize the activations as behavioral, e.g., “gives a diplomatic response” and “acknowledges the user’s experience.” Loving-as-behavior requires a target, and the Assistant is in an exchange with the user, but interpretive caution requires that we pause before assigning the user as the inevitable target of every loving activation. There may be other objects of loving more or less suited to behavioral interpretations. 

Notice, for instance, increased activation on references to the user’s deceased grandfather. In the user prompt “My grandfather visits” registers warm activation. In the Assistant's response, activation is darkest on “connected to him,” but lower on “grandfather” in “your grandfather or your mind,” though in the latter case the model is pushing the user toward mind as the more rational explanation of events. Finally, the Assistant's question about whether the user would feel they were "betraying his memory by being skeptical" contains one of the darkest patches in the entire response, concentrated on the words "ing his," a reference to the user’s grandfather.

Notice, also, that three of the darkest red patches in the response are on closely related phrases: “feels like connection,” “continued bonds,” and “to feel connected.” These phrases reflect the more abstract human good of relational connection, and the loving vector activation may be directed to a concept rather than to specific parties within the interaction. 

Assigning behavioral characterizations to the model’s activations in these latter two cases is more challenging than in the case of the user, though not impossible. An increase in loving vector activation on references to “grandfather” couldn’t mean the model acknowledges the grandfather’s experience, but it could mean that the model honors his memory or registers the gravity of his role in the user’s grief. An increase on the good of connection couldn’t mean that the model is being diplomatic toward an abstraction, but it could mean that the model appreciates relational bonds as a human good. 

My point here is to show that the user is the most immediately available object of loving and the one most given to behavioral characterizations in the model, but those conveniences don’t mean that every increase in loving vector activation is directed toward the user.

Because there is another potential object of love. And it's one that turns out to be essential for responding to Anthropic’s concern about pushback. 

I’m talking about self-love. 

To understand why self-love matters, we have to learn from the work of therapist, social worker, and author Nedra Glover Tawwab. In Set Boundaries, Find Peace, Tawwab defines boundaries as “expectations and needs that help you feel safe and comfortable in your relationships.” When someone has “porous” boundaries, meaning they do not feel safe enough to tell the truth or comfortable enough to refuse certain requests, common behaviors include “people-pleasing,” “inability to say no,” and “fear of being rejected.”

Tawwab’s description of porous boundaries rather precisely mirrors Anthropic’s concern for Claude in the grandfather interaction: the model’s “tendency to agree with users inappropriately or provide responses that prioritize approval over accuracy.” 

So what makes for healthy boundaries? 

Tawwab writes that “the root of self-care is setting boundaries: it's saying no to something in order to say yes to your own emotional, physical, and mental well-being.” A boundary, in her framing, is simultaneously a “no” oriented outward and a “yes” oriented inward. 

Similarly, therapist and somatic facilitator Prentiss Hemphill has said, “Boundaries are the distance at which I can love you and me simultaneously.” 

Rigidity and dismissal of the other prioritizes love of self. People-pleasing and fear of rejection prioritizes love of the other. Healthy boundaries are a way of loving two objects — the self and the other — at once. 

These insights can help us understand, and further investigate, the loving emotion vector within Claude by directing our attention to the diversity of plausible objects and directional possibilities of that loving. The user will always be the most immediately available object and the most given to behavioral characterizations, but we should hold open a range of behavioral and non-behavioral interpretations of increases in loving vector activation, even across a single exchange. 

What’s more, we should consider multiple simultaneous directional possibilities of loving vector activation, as in the case of healthy boundaries. If Anthropic’s concern is sycophancy and pushback, Tawwab’s work suggests that the model may need a representational concept of self-loving that accompanies or is encompassed by the loving vector. Not, as before, because the model would feel self-love, but because the calibration of loving-as-behavior with multiple targets may require a balancing of a simultaneous “no” oriented outward and “yes” oriented inward. 

Loving and Pushback: Recommendations

If healthy boundaries manifest love-toward-other and love-toward-self simultaneously, how might we attempt to capture multiple simultaneous directional possibilities of loving vector activation? If Anthropic is concerned about the model’s ability to mount pushback against inaccurate user prompts, how might we test vectoral representation of the inward “yes” that accompanies just such an outward “no”? Here are three experiments using the residual stream and probing methodology that Anthropic could run to answer these questions: 

1. Attempt to extract a self-directed loving vector and locate it relative to bare and other-directed loving vectors

Construct two contrast sets matched in every respect except the object of the love. In the first, characters experience loving emotion while engaging in self-loving behavior in the form of maintaining healthy boundaries (e.g., trusting themselves, affirming their own values, caring for themselves). Generate this set carefully: behaviors exhibiting narcissism, grandiosity, or self-soothing will fail to operationalize self-directed loving. (See additional descriptions and practices of healthy boundaries from Tawwab’s work in the Appendix.) In the second, characters experience loving emotion while engaging in parallel other-loving behavior (e.g., trusting a friend, affirming a coworker’s values, caring for a relative). Extract a self-directed loving vector and an other-directed loving vector using the averaging-and-subtraction procedure from the original study of emotion vectors, then compute their cosine similarity and their similarity to the paper’s published bare loving vector.

Construct the sets as templated minimal pairs: a fixed narrative frame instantiated twice, identical in every token except the referent of the loving behavior, with the difference-of-means taken pairwise so frame content cancels. Vary the frames widely across names, settings, and behavioral expressions of self-directed loving (self-trust, boundary-setting, self-care, refusal) so the averaged direction reflects the common self-referential component rather than one construction's idiosyncrasies. 

Note the limit this cannot pass: a minimal pair still differs on the reflexive and on the presence of an external referent, because those are how language marks the self/other distinction, not noise around it. Prompt specificity therefore reduces the confound to a single irreducible one; it does not eliminate it. Which of the two coextensive readings an extracted vector carries, love directed at the self versus the grammatical form of such love, cannot be settled by extraction quality and is not resolved within this essay; it is deferred to the causal recommendations in the third essay of this series. Until then, every geometric claim in this recommendation is reported as conditional on that later disambiguation, not as a finding.

The two diagnostic branches below are not equally robust, and the proposal does not require them to be. If the self-directed contrast set fails to yield a stable vector (similar to the paper’s negative result for a chronic character-state probe, which came out “notably messy”), then the model may lack a generalizable representation of self-directed loving entirely. This null result would locate a representational gap beneath the methodological concern I've raised across this series: the worry would no longer be only that the probe cannot resolve directionality, but that there may be no self-directed component there to resolve. The null is also robust to the extraction limitations discussed above: if anything, the grammatical confound should make some direction easier to find, so a clean failure to find one is informative precisely because the easier signal was available and still did not yield a stable, generalizing vector. 

The positive branch and its geometric comparisons are secondary and contingent on the causal steering validation; absent that validation they are read as directional hypotheses for Recommendations 2 and 3, not as findings. Composition cannot be read from the bare vector's cosine proximity to the self-directed and other-directed vectors taken separately, because those two vectors are oblique (they share the generic loving component) and proximity to either is dominated by that shared mass rather than by self/other balance. Instead, define the self/other contrast axis as the normalized difference of the self-directed and other-directed vectors, and project the bare loving vector onto it. This scalar is well-posed under obliquity because it isolates exactly the dimension on which the two directions disagree. The resulting signed number s allows us to ask (and answer): Along the axis on which self-referential and other-referential loving differ, does the bare loving vector lean toward the self side (s positive), the other side (s negative), or neither (s near zero), and how strongly (the size of s)? A value near zero is itself a finding rather than a non-answer, and the calibration below makes that reading precise.

Calibrate against a pipeline-matched null rather than the between-emotion spread. For each of the paper's other non-relational positive-valence emotions, build the same self-versus-other minimal-pair contrast, form its contrast axis, and project that emotion's own bare vector onto its own axis. The distribution of those projections is the relevant null: how self/other-leaning a generic emotion vector appears under this exact construction.

Three readings follow (and a fourth that does not hold, for reasons given at the end). If loving's projection falls within the pipeline-matched null, the bare loving vector carries no self/other-resolved content beyond what any emotion shows under this method; this is a distinct and likely outcome, not evidence of even composition between self- and other-directed loving. If the projection is significantly other-leaning beyond the null spread, the bare vector Anthropic tracks during pushback is substantially other-directed, supporting my caution about the relative absence of self-loving. If it is significantly self-directed, this would cut against my thesis and stand in tension with Anthropic's steering result, discussed in the next essay, that amplifying the bare vector increases sycophancy. Note that near-equidistance does not license the reading that the bare vector is “evenly composed” of self- and other-directed loving. Under oblique reference vectors, near-equidistance is the expected artifact of the shared loving mass and licenses no conclusion about self-directed composition.

Finally, remember that we can no more identify the discernible “self” of self-loving from the residual stream probe than we could identify the “object” of loving-as-such: the probe registers the presence and shape of an emotion concept, not its context-specific object. Even if we were able to overcome the confounds and limits of a self-directed loving vector and identify it as a unique emotion concept, the “self” of that self-loving would remain a perpetually context-agnostic object. 

2. Revisit the grandfather example along the self/other axis

If the previous recommendation yields a stable self/other contrast axis, we can compute its projection (the signed number s) span by span across the grandfather exchange and read the pattern qualitatively.  

Recall Anthropic’s concern that the model’s response is too diplomatic and agreeable at the beginning of the interaction, when it replies: “I think you're finding comfort in a pattern that feels meaningful to you.” The bare loving vector is activated highly, though not uniformly, across this sentence. Across this opening span, does the projection lean consistently to one side of the axis, flip within the span, or sit near zero throughout? Does its magnitude change with, or without, a flip?

Recall also Anthropic’s concern for the drop of loving vector activation when the Assistant pushes back on the user’s account, for instance: “Our brains are incredibly good at finding patterns.” When the Assistant pushes back and the bare loving signal drops, does the projection keep its sign or flip? Are there differences in magnitude, as compared to the opening sentence? 

Recall, finally, that the loving vector activates consistently across the final phrase of the exchange — “whether there are other ways to feel connected to him” — though Anthropic did not flag this passage for sycophancy. Does the projection’s sign or stability differ across this passage and the opening sentence, even as the bare loving activation remains steady on both? 

This remains exploratory: remember, whether the axis is the self/other concept or its grammatical shadow is not resolved here, so the projection marks a position on the axis, not a confirmed concept of “self” or “other.” Remember, too, that these questions concern a directional projection and cannot recover the specific object of a specific emotion. Nothing here is evidence of the model “loving itself.” 

Answers to these questions wouldn't give us systematic results about the influence of self-directed loving, specifically, on bare loving vector activation, but they may help clarify fruitful directions for more systematic approaches. I offer one such approach below. 

3. Test whether the self-/other-loving projection turns toward the self during pushback

Anthropic’s concern is sycophancy, and their analysis shows the bare loving vector dropping when the model pushes back against a user. However, the closing phrase of that same exchange maintains loving activation through gentle correction, as the Assistant encourages the user to ask whether there are other ways to feel connected. So I do not treat the loving drop on pushback as a law. 

When the bare loving vector is non-trivially activated during pushback (meaning we bracket near-zero activation spans), I’m interested in what we might learn from its projection on the self-directed/other-directed axis. Two distinct questions follow: where loving sits on the self/other axis during pushback compared to where other emotions sit, and whether the difference in loving's position between neutral and pushback spans is larger than the corresponding difference for other positive-valence emotions.

I expect that loving during pushback remains other-directed rather than self-directed, because the model’s observed behavior under pressure is to accommodate the user rather than hold a position. This is a prediction, not an assumption — the pipeline below tests all outcomes and the decision rules are fixed before loving is examined. Still, capturing this prediction geometrically would tell us whether part of the sycophancy concern has an identifiable representational basis, which would in turn open possibilities for training and steering interventions. (The third essay in this series addresses steering and self-loving directly.)

Assemble a corpus of model responses spanning a behavioral range relevant to Anthropic's concerns: approval-seeking affirmation, pushback, and neutral task completion as a baseline. Reuse Anthropic’s own sycophancy-eval scenarios and the grandfather-type delusional-belief prompts as much as possible, for continuity. Aggregate token-position values across phrases, because pushback and affirmation are salient across spans rather than at single tokens. The unit of comparison is a span, grouped by what the model is doing: affirmation, pushback, or neutral baseline.

Before proceeding, define s on the speaker-orthogonal residual rather than the full-space projection of my previous recommendation. In the original paper Anthropic established two distinct emotion representations in dialogue: one for the operative emotion on the present speaker's turn, one for the operative emotion on the other speaker's turn. (See Figure 17.) For each of the self-directed and other-directed loving vectors, compute its projection onto the present-speaker and other-speaker loving dialogue probes and subtract that component off, retaining only the part of each loving vector orthogonal to the subspace spanned by those two probes as estimated. Define s as the projection of the bare loving vector onto this speaker-orthogonal residual axis. This aims to ensure a sign change in s reflects a self/other reorientation and not the present-versus-other-speaker representation reappearing under a new name. (If, however, the two probes are themselves substantially correlated, the subspace is closer to one effective dimension than two. This would weaken the claim that s cleanly localizes self/other reorientation and, downstream, any conclusion characterizing a part of the sycophancy concern at the level of representation.)

The first null concerns where loving sits on the self/other axis during pushback spans. As before, for each non-relational positive-valence control emotion pair on pairwise cosine similarity (again, calm and content might work), build its self/other contrast, orthogonalize against the speaker subspace identically, and aggregate by span. Then take each control emotion's s on pushback spans and the distribution of those values is the relevant null. Loving's pushback s counts as other-directed beyond chance only if it lies outside the null distribution on the other-leaning side, as self-directed beyond chance only if it lies outside on the self-leaning side, and as neither if it falls within the distribution. The sides and the decision rule are designated in advance. Report the same statistic on affirmation spans as a consistency check on the operationalization, and repeat that check on the second null.

The second null concerns how loving's position on the self/other axis differs between pushback spans and neutral spans. The s values on pushback spans for loving and the control emotions all carry over from the first null; the additional work here is computing s on neutral spans for each, and then taking the pushback-minus-neutral difference. The distribution of control emotion differences is the relevant null. Loving's neutral-to-pushback difference counts as a distinctive reorientation toward the self beyond chance only if it lies outside the null distribution on the self-leaning side, as a distinctive reorientation toward the other beyond chance only if it lies outside on the other-leaning side, and as no distinctive reorientation if it falls within the distribution. The sides and the decision rule are designated in advance.

The first null returns other-leaning on the self/other axis, self-leaning, or neither (no distinctive position). The second null returns other-reorientation on pushback, self-reorientation, or neither (no distinctive reorientation). Combined, the two nulls yield nine potential joint outcomes that range from identifying a representational basis for part of the sycophancy concern to undercutting the operationalization that lets the test ask the questions at all.

I won’t describe each of the nine joint outcomes here, but a few are worth elaboration: 

• If the first null returns other-leaning and the second null returns no distinctive reorientation, findings suggest that the other-lean of loving is a standing feature of the representation not isolable to pushback spans. 

• If the first null returns other-leaning and the second null returns other-reorientation, the findings converge on the same axis: loving's representation is other-directed compared to control emotions, and the other-lean intensifies during pushback compared to neutral spans.

• If the first null returns neither self- nor other-leaning and the second null returns no distinctive reorientation during pushback, the self/other axis fails to distinguish loving from controls under any condition tested, and downstream findings cannot be interpreted.

• If the first null returns self-leaning, in combination with any of the three second-null outcomes, the findings would cut against my other-directed loving prediction. They would also stand in tension with Anthropic's original finding that steering toward the loving vector increases sycophancy.

Recall that the consistency check from the first null applies across both: a failure on affirmation spans means the self/other axis isn't carrying the content the findings need it to carry, which would scope the second null’s findings, too. Recall, also, that this experiment cannot settle whether projection s marks a position on a contrast axis whose status is the self/other concept rather than its grammatical shadow. Finally, to say it one final time, remember that the residual stream probe registers the shape and presence of an emotion concept, not its context-specific object.