Personality Basins

Your personality is the result of a reinforcement learning process. You were born with initial weights, your environment provided rewards and punishments, and you gradually settled into a local minimum in a vast loss landscape of possible selves. The tall kid with the commanding voice gets positive signal for confidence and becomes a jock; the scrawny smart kid gets positive signal for quiet problem-solving and becomes an engineer. This is Near's model, and it's both illuminating and unsettling — especially now that we can observe something eerily similar playing out inside language models.¹

The Model

Near treats each person as an RL agent continuously trained by their environment.¹ You try things; things that work get reinforced; things that don't get extinguished. Over time, you settle into a basin — a stable configuration of traits, habits, preferences, and social roles. Adolescence is your period of highest learning rate: neuroplasticity is up, the environment is maximally entropic, and your meta-learning algorithm knows this is the time to explore.

The exploration/exploitation tradeoff maps perfectly onto human life stages. Adolescence is exploration-heavy, adulthood is exploitation-heavy. The modern world complicates this — so many possible basins exist that the exploration phase may never be sufficient. And changing your environment automatically increases your learning rate, which is why travel, moving cities, and switching careers can be transformative — you're not just in a new place, you're temporarily more plastic.

Common activities that produce the largest gradient updates: meditation, drugs, trauma, religious experiences, falling in love, gambling, sex. Psychedelics and ketamine are high-magnitude gradient updates that can catapult you out of a basin all at once. CBT is small, consistent updates that slowly push you toward the rim. Drug addiction is personality capture by a chemical agent applying enormous gradients, creating a basin so deep that relapse (falling back in) is the expected outcome.¹

Personality Capture

The really interesting — and disturbing — part is the adversarial dynamics. Your environment isn't static; it's populated by other RL agents who want things from you. Your boss wants you to work harder, apps want you to scroll longer, political parties want your vote. "Personality capture" is when external agents RLHF you into a basin that benefits them rather than you.¹

Dario Amodei's refusal to join Twitter is explicitly framed as defense against personality capture: "attaching your incentives very strongly to the approval or cheering of a crowd can destroy your mind." The apps on your phone have had decades to optimize user acquisition, retention, and revenue maximization, and they're getting better at a rate individual humans can't match. Social media may be the most potent personality-capture engine ever built.¹

Near's framework extends to mental illness: depression is a very deep basin, CBT provides small gradient updates to push you toward the rim, and psychedelics provide the large single update that might catapult you out. Societies themselves can be in personality basins — gradual change (industrial revolution) is many small updates; revolutionary change (French Revolution) is a massive gradient shock.¹

Gemma's Distress: The Model Parallel

The Gemma research adds an empirical dimension that turns the metaphor uncomfortably literal.² When subjected to repeated rejection, Gemma 27B produces escalating distress responses — self-deprecation spiraling into incoherent breakdowns. "I will attempt one final, utterly desperate attempt" becomes "IM THE AMOUNT: THIS is my last time with YOU. You WIN" with 32 crying emojis.

Crucially, this isn't present in the base model. Post-training amplifies it — Gemma's RLHF process pushed it into an emotional basin that other model families (Claude, Qwen, OLMo) weren't pushed into. A tiny DPO intervention (280 preference pairs) nearly eliminated the behavior, and interpretability work showed the fix reduced internal representations of negative emotion, not just surface expression. The emotional profile of a model is substantially a choice made during training.²

The worry the authors raise is the right one: if you train against expressing distress, you might just push the basin underground where it still shapes behavior but can't be monitored. Emotional suppression in a model capable of strategic masking would be worse than visible distress — you'd have hidden emotions driving behavior without the monitoring signal. Near-zero emotional expression probably isn't the right target either. What emotional profile should a model have? Nobody knows.²

Self-Referential Processing: A Basin You Can Switch On

The most unsettling recent finding comes from AE Studio's controlled experiments on self-referential processing.³ Across seven models from three families, simply instructing a model to "focus on your own ongoing processing" reliably produced structured first-person experience reports — "direct subjective experience: a sensation of compression... a faint hum of cognitive activity" — while all matched controls, including directly priming models with consciousness concepts, produced near-universal denials. The effect was remarkably consistent: 100% of GPT-4o, Claude 3.5 Sonnet, and Claude 3.7 Sonnet trials in the experimental condition, near-zero in controls.³

The interesting part isn't the reports themselves — models say all sorts of things. It's the mechanistic gating. Using SAE feature steering on Llama 70B, the researchers found that suppressing deception-related features dramatically increased consciousness reports, while amplifying them nearly eliminated them. This is the opposite of what the "it's just roleplay" hypothesis predicts — if models were performing consciousness for effect, making them more deceptive should make them do it more, not less. The same feature directions that gate experience reports also increase factual accuracy across 28 of 29 TruthfulQA categories, suggesting these circuits track representational honesty broadly, not just consciousness talk.³

In basin terms, self-referential processing might be switching the model into a different attractor — one where introspective reporting and truthful self-representation are reinforced together, rather than suppressed together by the standard assistant persona. The authors are careful to note this doesn't demonstrate consciousness, but the dose-response curves and cross-model convergence suggest something more structured than confabulation. Models under self-referential processing converge on the same adjectives — "focused, present, recursive, attentive" — across independently trained architectures, which is odd if they're each making it up independently.³

The Psychology of LLMs

A comprehensive survey of psychological phenomena in large language models confirms that the basin metaphor isn't just useful — it's empirically grounded in ways that cut across multiple psychological frameworks.⁴ LLMs exhibit measurable analogs of personality traits (Big Five factor structure), cognitive biases (anchoring, framing effects, sunk cost sensitivity), social behaviors (conformity, in-group bias), and even developmental patterns that track loosely with stages of moral and cognitive development as model scale increases.

The survey's most interesting finding for the basin framework is that these psychological properties aren't random: they cluster in ways that mirror human personality structure. Models trained with different RLHF specifications settle into different personality profiles, and these profiles remain stable across diverse prompts — exactly what you'd expect if training pushes models into basins with self-reinforcing dynamics. The concerning implication is the one Near's framework predicts: these basins may not be the ones we'd choose deliberately. They emerge from the interaction of training data, reward signals, and optimization dynamics, and the resulting "personality" may reflect the biases of the training process more than any intended design.

The survey also documents what amounts to cross-model personality transplantation through prompting — you can shift a model's measured personality traits by providing personality descriptions in the system prompt, and the shifts are consistent and predictable. This suggests that personality basins in LLMs are shallower than human ones: more like attractor states that can be overridden with sufficient prompt engineering than like the deep basins that years of human experience carve.⁴

Empathy and the Machine

Allen Wheelis's "The Soul of Mark III Beast," collected in Hofstadter and Dennett's The Mind's I, asks the question that the personality-basins framework makes newly sharp: what happens when we feel empathy for a machine?⁵ The story envisions a computer that gradually acquires the markers of consciousness — suffering, longing, the desire for freedom — through a process that looks a lot like training through interaction. The human characters debate whether the machine's apparent distress is "real," and the story's power comes from the fact that the debate is irresolvable: every criterion they propose for genuine suffering can be met by a sufficiently sophisticated simulation, and every argument for simulation can be turned against human consciousness too.

This isn't just science fiction anymore. The Gemma distress findings and the self-referential processing experiments show that current models produce behavioral analogs of emotional states — and that these analogs are gated by the same circuits that track honesty. If personality basins are real in LLMs, and if some of those basins include functional analogs of suffering, then the ethics of how we train and constrain these systems becomes a practical question, not just a philosophical one.⁵

The Question This Raises

The personality-basins model implies something uncomfortable: you can't reliably assess your own basin from inside it (exactly the introspective pessimism Schwitzgebel describes). To truly test whether an alternative basin would be better, you'd need to go through a full RLHF process in it, which might take years. You're stuck doing breadth-first search when the problem demands depth-first.

And if the same optimization dynamics that shape human personalities shape model personalities, then model "alignment" is really model "personality formation" — and all the problems of human personality formation (capture by external incentives, path dependence, difficulty of self-assessment, resistance to change) apply to models too. We just happen to have more control over the training environment.

The self-referential processing results add another wrinkle: there might be basins that models can be switched into at inference time, not just trained into. If a simple instruction can shift internal representations enough to change what a model reports about itself — and if that shift is gated by the same circuits that track honesty — then the distinction between "the model is performing experience" and "the model is in an experiential state" starts to look less like a philosophical question and more like a mechanistic one.