PSYCHOLOGY · SYSTEMS · AI
Your Past Self Has Nothing to Teach You
Life is neither optimization nor noise. What transfers across life's non-repeating cycles is not the policy but the learner.
Life is neither optimization nor noise. Here is what actually transfers across the cycles that do not repeat.
Most people frame life one of two ways.
Either it is an optimization problem with a payoff function to maximize, or it is a story you wander through and decorate with meaning. Both are wrong. Both fail the same test.
The test is this. Take someone who has just become a parent. Take someone who has just survived cancer. Take someone who has just moved continents and started over at thirty-eight. Ask them whether the policies that worked in the previous life cycle are working now. They will laugh.
The reward function changed. The state space changed. The optimizer changed.
Optimization assumes the function holds. It does not.
The lazy alternative is to call life a random walk and say meaning is something we paint onto the noise. This also fails. Because wisdom is real. A person at fifty navigates a novel crisis better than a person at twenty, and not because they have seen this exact crisis before. They have not. Something accumulated. Something carries.
If life were noise, fifty-year-olds would be no better at the new than twenty-year-olds.
They are.
So a third mode exists.
This post is about what that third mode is, and why every serious tradition (existential philosophy, depth psychology, modern machine learning) has independently arrived at the same answer under different vocabularies.
The Thesis, Stated Flat
What transfers between cycles of a human life is not the policy. It is the learner.
The behavior that solved cycle t will not solve cycle t+1, because the distribution has genuinely shifted. What solved cycle t was the act of figuring out a policy under a distribution. That act, repeated, sharpens the capacity to figure out future policies under future distributions. Capacity is the unit of accumulation. Behavior is not.
Wisdom is meta-learned initialization.
Identity is a learned learner.
This is the third mode. The rest of this post is the same insight viewed from five different surfaces.
The General Frame
A life epoch is defined by a relatively stable joint distribution over states, rewards, and relationships. Childhood is one. Undergraduate life is another. First job. Marriage. Parenthood. Mid-career. The transition from one to another involves what statisticians call concept drift, what reinforcement learning calls non-stationarity, and what existentialists called crisis.
Inside an epoch, you can learn policies. You can get good at your job, at parenting infants, at being a graduate student. Inside the epoch the function is stable enough that gradient descent works.
Between epochs the function does not just shift. It is replaced.
The cached policies from the previous epoch are not just suboptimal. They are actively misleading. The reflexes that made you a brilliant solitary student make you a poor team lead. The vigilance that kept you alive in a hostile environment makes you a defensive partner in a safe one. The optimizer that worked is now a saboteur.
If you cannot transfer the policy, what can you transfer?
The thing you learned about how to learn.
Capacity transfers.
This is the only viable answer, and it has been discovered, in some form, by every tradition that has taken the problem seriously.
Surface One: Kierkegaard, Jung, MAML
In 1843, Søren Kierkegaard wrote a small book called Repetition. He was trying to distinguish two ways of relating to your past.
The first he called recollection. Recollection is backward-looking. It reaches for the past as a static artifact and tries to return to it. The Greeks called this anamnesis. It is what you do when you mourn a youth that cannot come back.
The second he called repetition. Repetition is forward-looking. You retrieve the self that acted, not the action. The self is brought into the new situation and renewed. The action is not.
Kierkegaard wrote that repetition is “recollected forward.” A strange phrase. It means the past comes with you, but it does not dictate.
Carl Jung gave this the clinical name individuation. Individuation is not linear progress toward a finished self. It is a spiral. The same archetypal challenges arrive again and again under different costumes, and each integration reshapes the architecture that meets the next one.
The ego that survives a lifespan is not the ego that performed well at twenty-two. It is the ego that learned, repeatedly, how to integrate what twenty-two could not.
In 2017, Chelsea Finn published Model-Agnostic Meta-Learning. The setup is this. An agent is trained on a distribution of tasks. It is not evaluated on those tasks. It is evaluated on a structurally distinct new task using a handful of gradient updates from a fresh initialization.
What gets optimized is not the task-specific policy. What gets optimized is the initialization. The weights you start from.
The math separates two things. The policy parameters are the behavior. The meta-parameters are the prior that allows to be discovered quickly when the new task appears.
The unit of transfer is not what you knew. It is the prior that makes new knowing fast.
Kierkegaard, Jung, Finn. Three vocabularies. One mechanism.
You bring the learner forward.
Surface Two: Phronesis, Internal Working Models, JEPA
Aristotle distinguished two kinds of knowing in the Nicomachean Ethics. The first was techne, which is craft. Techne can be written down. It is the recipe, the algorithm, the explicit instruction. Anyone with the manual can execute techne.
The second was phronesis. Practical wisdom. Phronesis cannot be written down. It is the capacity to recognize what this situation calls for, where this situation has never appeared before, and where the rules from previous situations do not apply.
Phronesis is acquired only through empeiria, which is direct experience. You cannot study your way to it.
John Bowlby, working in attachment theory, found something structurally identical in his concept of Internal Working Models. An IWM is a schema for predicting how a relational situation will unfold. It is not a script. It is a probability distribution over latent dynamics. You walk into a new room and within thirty seconds you have predicted who is dangerous and who is safe. You did not consciously deliberate. The IWM ran.
Michael Polanyi called this tacit knowledge. We know more than we can tell.
The computational name is Yann LeCun’s Joint-Embedding Predictive Architecture. Standard predictive models try to predict the literal future. Pixel by pixel. They fail in any environment with meaningful stochasticity, because the literal future is mostly noise.
JEPA succeeds by predicting in latent space. The model abstracts the input into a compact representation, and the prediction lives in the representation, not in the raw input. The high-frequency surface noise is discarded by design.
The world model is a structural prior. It tells you what kind of thing is happening, not what color the walls are.
Phronesis is human latent-space prediction. The wise elder is not running a faster algorithm. They are running prediction on a coarser, more invariant representation. Surface details are dropped.
The structure is what they see.
This is also why phronesis cannot be taught by lecture. The representation has to be built. Telling someone the right abstraction is not the same as their brain having compiled it.
Surface Three: Wu Wei, Wayfaring, Open-Endedness
The optimization framing has a specific failure mode in complex spaces. Kenneth Stanley demonstrated it formally in his work on novelty search. If the path to a high-value outcome does not resemble the outcome at any intermediate step, optimizing toward the outcome will trap you in deceptive local optima. The objective function lies to you.
The solution is to abandon the objective and search for novelty. Step into what Stuart Kauffman called the adjacent possible. Take the move that opens new moves, not the move that closes the distance to a goal you cannot actually compute the gradient toward.
This is a strict mathematical result. In complex search spaces, objective-driven optimization is provably worse than non-objective novelty search.
Daoism arrived here two and a half thousand years earlier.
Wu wei is usually mistranslated as “non-action.” That is wrong. Wu wei is unforced action. It is action that aligns with the unfolding of the situation rather than fighting the situation toward a predetermined endpoint. The Daoist sage does not lack goals. The Daoist sage does not let goals override the topology of the moment.
Tim Ingold, the anthropologist, calls this wayfaring. Wayfaring is what walking actually is. You move along, the path emerges, you do not transport yourself from point to fixed point by the most efficient line. The way is made by walking.
The contrast Ingold draws is with what he calls transport. Transport is when the destination is fixed and the journey is friction to be minimized. Most modern life-optimization advice is about transport.
Most modern life is not transportable.
When the objective function is unreliable, the only viable strategy is to look for moves that open more moves.
Wu wei, wayfaring, novelty search. The same posture. Drop the fixed objective and trust that capacity acquired along the way is its own form of progress.
Surface Four: Heidegger, Post-Traumatic Growth, Antifragility
The fourth convergence concerns how the system actually grows. Not how it persists. How it expands its capacity.
Martin Heidegger argued that the average person lives in what he called the they-self. You adopt the inherited scripts of your culture. You marry when people marry. You take the job that people take. You repeat the form of a life that has been pre-approved. Nothing in this mode requires you to be anything in particular.
Heidegger thought authentic existence began only when this scaffolding broke. The encounter with mortality, with anxiety, with the realization that no inherited script is going to do the work for you. Only then does what he called Wiederholung, authentic retrieval, become possible. You retrieve your own possibilities. You stop running the borrowed program.
The clinical version is Tedeschi and Calhoun’s framework of Post-Traumatic Growth. A trauma is, by definition, an event that shatters your existing schematic model of the world. Most attempts to handle trauma try to restore the previous model. PTG occurs when the shattering is metabolized into a new model that is not a restoration but an expansion. The new model holds what the old could not.
Nassim Taleb gave this the mathematical name antifragility. A system is antifragile if its response to volatility has positive convexity.
In plain English, the upside from disorder exceeds the downside. Small shocks make the system stronger. The second derivative of the payoff curve, with respect to volatility, is positive.
Fragility is . Robustness is . Antifragility is .
The reason this maps to human development is that the cognitive cost of shattering an old model is enormous, but the new model that emerges is structurally more capable than the old one would have become through smooth optimization. The trauma is convex.
This is also why protected lives produce fragile people. If you optimize for the absence of shocks, you optimize away the only mechanism that builds higher-order capacity. The optimization is self-defeating.
The Daoist accepts the storm. The depth psychologist integrates the shadow. The mathematician shows that convexity to volatility is the only structural mechanism by which expansion outpaces decay.
Same insight. Three vocabularies.
Synthesis: The Same Insight Across Every Surface
Across all four surfaces the structure is the same.
What persists across non-stationary epochs is not policy. It is the abstraction that lets new policies be discovered quickly under new distributions. The transferable unit is the learner. The non-transferable unit is the learned.
The traditions that have taken this seriously have all arrived at four sub-claims.
First, the past does not dictate the present. It conditions the prior. The prior is what you bring forward, not the answer.
Second, surface details should be discarded. Wisdom lives in the latent representation, not in the raw input. The wise elder is not faster. They are coarser, in the structurally correct way.
Third, fixed objectives are unreliable when the function is non-stationary. Move toward what opens moves, not toward the imagined endpoint.
Fourth, growth requires destabilization. There is no smooth gradient up the capacity curve. Convexity to volatility is the only mechanism that scales.
Together, these are the operating manual of what the German Romantics called Bildung and what modern complex systems research calls open-endedness. The research project that produced this post named the synthesis Open-Ended Bildung. The compound captures both halves. Bildung holds the humanist load. Open-ended strips out the buried teleology.
One phenomenon. Four vocabularies.
The Uncomfortable Implication
If this is right, then most career advice from successful people is useless. They are reporting their cached policies. The cached policy was good for the distribution they faced, which is not the distribution you face. They are recommending the action, and what they should be teaching is the learner.
Resumes overstate what transfers. The job you did at twenty-eight is not the thing you actually own at thirty-eight. The thing you own is the abstraction you built while doing it. If the abstraction is good, the job was worth the time. If the abstraction is shallow, you spent ten years optimizing weights you will have to throw away.
Stability inside an epoch is often the enemy of growth across epochs. If your environment is too smooth, your second derivative goes to zero and you become structurally robust but not antifragile. You will survive the next regime, but you will not be expanded by it.
And the worst implication.
The thing you are most proud of accumulating may not be the thing you actually own. Resumes, achievements, identities tied to specific roles. These are policies. They were the answer to a distribution that no longer exists. What you actually own is whatever shape your prior has been bent into by the cycles you went through. That shape does not show up on a CV.
You are not the sum of what you achieved.
You are the prior those achievements left in you.
The One-Line Version
The unit of life is not the policy. It is the learner.
Sources
This post synthesizes a Deep Research report on what the source material calls “the third mode,” triangulating across philosophy, psychology, and computational science. Primary references include Kierkegaard’s Repetition (1843), Jung’s writings on individuation, Aristotle’s Nicomachean Ethics, Heidegger’s Being and Time, Bowlby’s attachment theory, Tedeschi and Calhoun’s framework of post-traumatic growth (1995 onward), Finn et al.’s Model-Agnostic Meta-Learning (2017), LeCun’s work on Joint-Embedding Predictive Architectures (2022 onward), Stanley and Lehman on novelty search and open-endedness, and Taleb’s Antifragile (2012). The compound term “Open-Ended Bildung” is offered by the source synthesis as the closest existing nomenclature for the phenomenon.