Frontier.

Three hundred milliseconds above the right ear. That is the whole apparatus.

Begin with what is settled. Spark already laid it out, but the present argument needs the apparatus tighter. In 1997 Wolfram Schultz, Peter Dayan and P. Read Montague published “A Neural Substrate of Prediction and Reward” in Science. The result was that phasic dopamine firing in the substantia nigra and ventral tegmental area encodes a reward prediction error — the difference between the outcome you got and the outcome you expected. The signal fires above baseline when the outcome exceeds prediction, at baseline when prediction is accurate, and below baseline when prediction is missed. This is the canonical equation of reward #1 in this essay. The same δ that trains GPT — Sutton and Barto won the 2024 Turing Award for the maths — is, with very minor adjustments, the signal a macaque's ventral tegmental area emits when a juice cue arrives sooner than the model predicted.

What this circuit does, when it fires during insight, was characterised by Jung-Beeman, Bowden, Kounios and colleagues in 2004 (PLOS Biology) and extended by Tik et al. in 2018 with 7-Tesla fMRI. Subjects solve Compound Remote Associates problems — three words that share a fourth word (pine / crab / sauce → apple). At the moment of insight, there is a sharp burst of gamma-band activity (~40 Hz) at right anterior superior temporal gyrus, beginning roughly three hundred milliseconds before the subject becomes consciously aware of the solution. The 7T fMRI then shows the activation that follows: ventral tegmental area, nucleus accumbens, caudate, hippocampus. The mesolimbic dopaminergic reward circuit, the same one that responds to food and money, fires on the act of restructuring itself. Becker, Sommer and Cabeza extended this in Nature Communications in May 2025 with a striking result: representational patterns in ventral occipito-temporal cortex literally reorganise during the moment of insight, and the bigger the reorganisation, the more reliably the participant remembers the material the next day. Aha is not a feeling tacked on after the answer. Aha is the structural change being announced to the system by the system.

This is reward #1 in operational form. It is a circuit. It is mapped. It is sharp enough that the signal fires at sub-second resolution. The phenomenology — the Aha, the click, the sudden certainty before the proof — is the felt signature of a specific neural event, and Oh, Chesebrough, Erickson, Zhang and Kounios (2020, NeuroImage) showed that the reward gamma burst is too quick to be conscious appraisal. The brain does not reward you for getting the answer right. It rewards you for restructuring itself. The pleasure arrives at the same moment as the answer, not as a reaction to it.

Notice one more thing about reward #1 before we move on. Every paradigm in the insight-neuroscience literature uses problems with a pre-existing hidden answer. Compound Remote Associates have a unique correct word. Anagrams have a unique target. Mooney images have a hidden figure. Rebus puzzles have a single solution. The neural circuit is studied where the answer already exists in the universe and the subject finds it. There is no paradigm — none — for the moment a subject creates a category that did not exist a second before. Abraham et al. 2012 mapped “conceptual expansion” in inferior frontal gyrus, anterior temporal pole and frontopolar cortex, but the “expansion” was still stretching existing concepts, not founding new ones. The neuroscience of reward #2 does not exist yet, not because no one has looked, but because no one has invented the behavioural paradigm to make it visible. This is a real empirical gap the essay lives in. Acknowledge it openly. It is also the precise location where the essay's strongest predictions can be tested.

Fifty years of structural biology became three years. Strassen's 1969 ceiling fell in 2025.

The intuition that AI is going to make the universe more discoverable in a hurry is not science fiction. It is the empirical record of the last seventy-two months. Some specifics, because the argument lives or dies on them.

Read the table as a curve. The Strassen result is a 56-year ceiling falling in one. The cap-set bound is a 20-year improvement landing in months of evolutionary search. The FrontierMath benchmark was launched in November 2024 with Fields Medalists assessing the problems as “exceptionally challenging,” with Terence Tao predicting they would resist AIs for several years. Six weeks later o3 was at 25.2%. (Honesty requires the asterisk: OpenAI had funded Epoch AI and had exclusive access to the problems before testing; the figure may be inflated. Even with the asterisk, the rate is unprecedented.) The frontier of what AI cannot do is moving, week by week, into territory it would have been embarrassing to claim for AI three years ago.

The honest framing of all of this — the framing that survives the strongest critique — is not “AI is doing real creativity now.” The architecture of every documented breakthrough is the same: LLM-as-mutation-operator inside an evolutionary loop with an external verifier. The LLM is the search; the verifier is what counts as having found something. In FunSearch the verifier is a function in code that returns a score. In AlphaProof it is the Lean proof checker. In AlphaEvolve it is whatever measurable objective the human gave the system. The structure that produces verified novelty is human-defined verifier + machine search. The LLM is doing exploratory creativity at unprecedented scale; the human is doing transformational creativity by inventing the next verifier. Boden's three-tier taxonomy resolves the puzzle: AI saturates combinatorial and exploratory creativity inside any conceptual space whose boundaries can be specified; transformational creativity — moving the boundaries — remains where humans live.

Hofstadter saw this coming and reversed his position. In a June 2023 interview he said the human mind is not so mysterious and complex and impenetrably complex as he imagined it was when he was writing Gödel, Escher, Bach. He said it felt like the entire human race was going to be eclipsed. He used the word terror. He was talking specifically about pattern-completion, analogy-making, the machinery Copycat was built to model. He did not say strange loops can joots — jump out of the system. Copycat slips concepts inside a closed alphabet; it does not invent letters. Hofstadter's reversal is the strongest possible witness statement that reward #1 is being saturated. It is also, by the same man's silence on jootsing, witness that the question of reward #2 remains live.

The treadmill is not psychological weakness. It is the math of the gradient.

Here is the painful clarification. The reason reward #1 will fade as AI compresses discovery is not that humans run out of dots. Kauffman is right: the adjacent possible expands as it is explored, the universe is non-ergodic, the space of available structures is genuinely unbounded. The reason reward #1 will fade is that dopamine fires on a first-derivative signal — change in predictability, not predictability itself — and AI is closing the predictability gap faster than the adjacent possible expands inside any one person's perceptual range. The cosmological frontier widens; the personal predictive horizon collapses. Both can be true at once.

Schultz himself wrote it most clearly in his 1998 review: “Dopamine neurons increase their responses in the face of novelty; once novel stimuli become familiar and are not reinforced, dopamine responses habituate.” This is the entire equation. The signal is the change between expectation and reality. As the world contains fewer surprises (because the model containing the world has gotten better), the signal goes silent. There is no defect in the brain. There is only a gradient descending to its floor.

Brickman, Coates and Janoff-Bulman documented this on a longer timescale in 1978. Twenty-two major lottery winners returned to baseline happiness inside two years and reported significantly less pleasure than controls from a list of mundane everyday events. The peak resets the baseline. The contrast against everything ordinary intensifies. The very fact that the lottery happened impairs the brain's capacity to take pleasure in lesser positive events. The Lyubomirsky-Sheldon HAP model gives the formal version: two erosion routes (declining positive emotions from the change, rising aspirations) and two moderators that forestall adaptation (variety and appreciation). The variety moderator is what reward #2 needs to be: novelty that genuinely cannot be predicted by the model, because the relevant category did not previously exist.

We are already seeing the saturation symptoms at population scale. Han Byung-Chul's Burnout Society (2010): the achievement subject, entrepreneur of itself, engages in auto-exploitation more efficient than any external exploiter because the feeling of freedom attends it. Aggression turns inward; the project becomes a projectile. Mark Fisher's Capitalist Realism (2009): the political depression of a culture that cannot imagine an alternative because its highest attractor has stopped firing. Case and Deaton on the deaths of despair: ~600,000 excess US mortality through 2018 if the pre-1999 trend had held. Twenge's 2012 inflection: teen in-person socialising in nose-dive at almost exactly the moment smartphone penetration crossed 50%. Murthy's 2023 Surgeon General report: 70% drop in time spent with friends among 15–24 year-olds, with mortality risk comparable to smoking 15 cigarettes a day. These are not separate crises. They are the same crisis from different angles. The dopaminergic attractor that organised the 20th century — connect-the-dots-and-gain-status — has saturated. The exits, so far, have been bad: optimisation-without-payoff, scrolling without satisfaction, deaths of despair. The argument of this essay is that there is one more exit, the one Nietzsche came too early to see and the only one large enough to absorb the saturation, and that AI is what forces us to look for it.

Six dimensions on which the two rewards are not the same animal.

What follows is the proposed dual-architecture in operational detail. Each row is a dimension on which the two rewards differ in kind, not merely degree. The essay's strong claim — that reward #2 is genuinely separate, not just reward #1 retargeted — stands or falls on whether these dimensions hold up to empirical test. Predictions follow in part fifteen.

The brain may have written the second reward two hundred million years ago. We just never needed it badly enough to switch it on.

Jaak Panksepp spent forty years arguing that mammalian brains contain not one unified affective system but seven primary-process emotional command circuits — SEEKING, RAGE, FEAR, LUST, CARE, PANIC/GRIEF, and PLAY. The capital letters were his convention to distinguish biological circuits from folk-psychological terms. The seven were anatomically distinct, neurochemically distinct, and behaviourally distinct, mapped via electrical brain stimulation, pharmacological challenge, and targeted lesion across hundreds of studies in Affective Neuroscience (1998), The Archaeology of Mind (with Lucy Biven, 2012), and a long tail of follow-up work. Mark Solms's The Hidden Spring (2021) is the current synthesis: affect, not cognition, is the seat of consciousness, and the seven systems are its basis.

Of the seven, SEEKING is the dopaminergic appetitive system — the one running through the medial forebrain bundle, the ventral tegmental area, the nucleus accumbens. SEEKING is reward #1 in the language of the essay. It is what fires when a novel cue predicts reward, what habituates when the cue becomes reliable, what drives the exploration humans share with rats. PLAY runs on different hardware. The primary executive circuit involves the thalamic intralaminar nuclei, the dorsomedial parafascicular complex, parts of frontal cortex and striatum, with secondary recruitment of the amygdala, ventral hypothalamus and periaqueductal grey. The neurochemistry is dualistic — dopamine plus endogenous opioids (especially μ-opioid), with cannabinoid and cholinergic modulation. Low-dose morphine increases play in juvenile rats; naloxone reduces it. These results have been replicated dozens of times since the original Panksepp et al. 1980 paper.

Two facts about PLAY are doing the heavy lifting for the essay. The first is that it is generative by definition. Burghardt's third criterion for play (The Genesis of Animal Play, MIT Press 2005) is that play behaviour is functionless in the context in which it occurs but resembles functional behaviours. Play recombines fragments of serious behaviour — fight, flight, hunt, court — into sequences with no immediate payoff. Sergio and Vivien Pellis, in a long programme of behavioural research, put it more precisely: play produces patterns of behaviour that did not exist in the participants' repertoires before the interaction. This is the neurobiological analogue of Boden's transformational creativity. Unlike SEEKING, which moves an organism through known reward landscapes, PLAY constructs new ones. The reward attached to PLAY is keyed to behavioural novelty per se, not to predicted outcomes.

The second fact is that PFC and PLAY are in a bidirectional regulatory loop unique to PLAY among the seven systems. Top-down: medial prefrontal cortex and orbitofrontal cortex inactivation reduces play (van Kerkhof et al., Neuropsychopharmacology 2013). Bottom-up: social-play deprivation in juveniles produces lifelong reductions in inhibitory currents on layer-5 pyramidal neurons in medial PFC and orbitofrontal cortex (Bijlsma et al., Journal of Neuroscience 2022). Play needs cortex to release it. Cortex needs play to be built. This feedback loop is unique to PLAY among Panksepp's seven. And it has a specific implication for the essay: you cannot offload PLAY to a system that lacks the substrate PLAY itself builds. If reward #1 (SEEKING) can be externalised to AI — and the empirical record says it can — reward #2 (PLAY) is constitutively tied to the maturation of the regulatory cortex that could host any successor reward function.

There are honest objections. Joseph LeDoux has argued for two decades that animal behavioural evidence shows defensive/survival circuits, not conscious feelings; attributing joy to 50-kHz rat chirps may be anthropomorphic. Lisa Feldman Barrett has argued emotion categories are culturally constructed predictive concepts, not innate circuits. The essay's claim that PLAY is the substrate of reward #2 depends on Panksepp's stronger reading — that subcortical circuits suffice for primary affective consciousness, supported by Solms 2021 and the hydranencephalic children data (Merker 2007) showing affective responsiveness without cortex. This is contested. The essay flags it openly. What does not depend on the contested reading is the anatomical and neurochemical separation. Even on the sceptical view, PLAY is a behavioural and physiological category distinct from SEEKING. That separation is what reward #2 needs. The strong phenomenological claim — that PLAY feels like reward #2 already — can be hedged. The structural claim — that the brain contains the substrate for a reward function distinct from dopaminergic prediction error — is solid.

The brain has the architecture for this. Berridge proved it on rats forty years ago.

Kent Berridge and Terry Robinson's incentive-salience theory of dopamine is the single most important neurobiological fact for the cultural-installation arm of the essay's argument. The 1998 paper in Brain Research Reviews and the 2025 30-year retrospective in the Annual Review of Psychology establish a clean dissociation: dopamine in the mesocorticolimbic pathway codes wanting (motivational drive to approach a target), not liking (the hedonic pleasure of consuming the target). The two can be pried apart entirely. Late-stage addicts report decreasing pleasure (liking falls, sometimes inverts to dysphoria) while compulsive pursuit (wanting) escalates. Berridge calls this “wanting what hurts.” The architectural asymmetry is the key: the pleasure system is small — about one cubic millimetre of opioid-sensitive tissue in the rostrodorsal nucleus accumbens shell — and fragile. The wanting system is the entire mesocorticolimbic projection. Wanting saturates much less readily than liking; wanting rebinds onto new salient targets the way liking does not.

This matters because the strongest counter-objection to the cultural-installation arm — that the dopaminergic system runs on hard-coded biological priors that cannot be retargeted — is empirically wrong. Robinson and Berridge document sensitisation that is specific to one motivational target rather than another: in some individuals drugs become more “wanted” than natural rewards, in others food or sex, in others entirely abstract targets. The same dopaminergic machinery can be re-pointed. The novelty literature supports this from the other direction: dopamine neurons fire on novel stimuli per se, before any reward learning. Novelty is a sufficient trigger. The quest for knowledge does not need a terminal payoff to keep firing — every novel domain re-activates SEEKING.

The empirical demonstration that the mesolimbic circuit binds to purely cultural targets — money, status, religion, mathematical beauty — is the essay's lever for the strong claim that reward #2 could be installed even if it is not latent. Mathias Pessiglione, with Liane Schmidt and others, ran the cleanest experiment in Science in 2007. Subliminal flashes of a £1 coin versus a 1-penny coin — fifty milliseconds, below the threshold of conscious report — produced focused activation in the ventral pallidum and ventral striatum and increased measured grip force on a hand dynamometer. A piece of paper printed in 17th-century Sweden activates the same mesolimbic structures food does in a rat. The reward target is purely cultural; the circuit treats it as primary. The installation completed in three to five millennia from Mesopotamian clay tokens to universal striatal binding.

The installation latency is collapsing. Religion installed across roughly three to five thousand years. Money over similar. Social media validation installed across roughly ten to fifteen years — Facebook launched 2004, Sherman et al. documented increased nucleus accumbens activation for high-like Instagram photos in 2016. Three orders of magnitude faster than religion. If the trend holds, the next major reward attractor installs not on a generational timescale but on a sub-decadal one. The window in which the essay's “reward #2 comes online” prediction either succeeds or fails is the next twenty years.

The brain is not a fixed map of innate reward targets. It is a fixed circuit that culture has been installing new targets onto for ten thousand years, with the installation rate accelerating monotonically. The cultural arm of the essay does not require any latent dormant circuit at all — it only requires the documented routing mechanism to do what it has been doing, one more time, on the timescale it has been doing it in.

Peirce named the operation in 1903. Boden named the cognitive category in 1990.

The logical move reward #2 must reward has a name in the Western tradition. It does not appear in deductive logic, which can only extract necessary consequences from premises. It does not appear in inductive logic, which can only generalise from data. The third mode — the one Aristotle gestured at as apagogê and Charles Sanders Peirce named and operationalised between 1878 and his death in 1914 — is abduction. The 1903 Harvard Lecture VII gives the canonical schema: “The surprising fact, C, is observed; but if A were true, C would be a matter of course; hence, there is reason to suspect that A is true.” Three things are critical about this schema. The conclusion is suspicion, not proof. The mode is ampliative — it adds to the world rather than rearranging it. And, in Peirce's 1903 formulation in Collected Papers 5.172, “abduction is the only logical operation which introduces any new idea.” The other two modes are bookkeeping.

Peirce's pipeline of scientific inquiry — abduction, then deduction, then induction — is asymmetric in novelty production. Only the first step adds new conceptual material. The second extracts predictions; the third tests them. This maps cleanly onto the essay's thesis. Reward #1 fires on the verification stage and the induction stage and even on the deductive working-through; it does not, by construction, fire on the abductive leap, because the abductive leap is structurally outside the predictive system inside which dopaminergic prediction-error has a value to compute. The 2025 IID finding — LLMs sample hypotheses that are independent and identically distributed, semantically clustered around the prior — is the technical underpinning of the essay's claim that AI is doing induction in disguise (interpolation across training data) rather than abduction. Bylander's 1991 result that finding the best abductive explanation is NP-hard under plausibility plus parsimony plus completeness constraints reinforces the asymmetry: the constrained version machines can perform is selection from pre-given alternatives, not Peirce's generative version where the explanatory entity is itself invented.

Margaret Boden gave the cognitive taxonomy in The Creative Mind (1990, second edition 2004). Three kinds. Combinational: novel combinations of familiar ideas with an intelligible analogical link. Macbeth's sleep that knits up the ravelled sleeve of care. AI is superhuman at this and has been for years. Exploratory: search within a defined conceptual space, traversing its structure to find what its rules allow. Most jazz improvisation. Most normal-science problem-solving. Boden's analysis of classical creative AI projects (AARON, Cohen's painting system) puts most of them here. AI is now also superhuman at this in the verifiable domains, which is why AlphaProof can win silver at the IMO and AlphaEvolve can break Strassen's algorithm. Transformational: altering the enabling constraints of the conceptual space itself, producing ideas that were literally unthinkable inside the prior space. Non-Euclidean geometry. Einstein abandoning absolute simultaneity. Schoenberg abandoning the home key. The Copernican shift. The invention of zero. The map being redrawn rather than read.

Boden's key formulation, from page 6 of the second edition:

The honest objection is Geraint Wiggins's 2006 Creative Systems Framework, which proved that transformational creativity is mathematically equivalent to exploratory creativity at a meta-level. If you can search over conceptual spaces, transformational creativity is just exploration in a higher-order space. If the meta-level is Turing-computable, AI can in principle do transformational creativity too, given enough compute. This is the strongest single counter the essay must engage. Two replies. First: Wiggins's collapse depends on an assumption of universal Turing-computability that is not itself obvious — the Penrose-Lucas argument is contested, but the weaker Gödelian point survives, that any closed formal system contains truths it cannot reach from within. Second, and more empirically pertinent: every documented case of AI producing genuinely novel mathematical structure in 2024–26 has had the same architecture — LLM-as-mutation operator + human-defined verifier. The verifier specifies the meta-level. The machine searches it. The human invents the next verifier. This is consistent with Wiggins's collapse and with the essay's thesis: the meta-level is computable once specified, but specifying the next meta-level is the human move, the one Hofstadter called jootsing and Boden called transformational and Peirce called abduction. The essay's “humans needed forever” is this: not as the deciders inside a verifier but as the inventors of the next verifier.

From the gene to the brain. Every new ontological category installs a new epistemic cut.

Howard Pattee spent seventy-one years arguing one thing in many keys: every time the universe creates a new ontological category, it does so by building a system that stands simultaneously in two registers. The gene is the original instance — simultaneously a chemical participating in reactions and a symbolic description of those reactions. Pattee's 1968 paper “The Physical Basis of Coding and Reliability in Biological Evolution” in Waddington's Towards a Theoretical Biology, his 1995 “Evolving Self-Reference”, his 2001 “The Physics of Symbols: Bridging the Epistemic Cut” in BioSystems, and his 2015 “The Physics of Symbols Evolved Before Consciousness” in Cosmos and History form the through-line. The argument is not dualism — it is bilingualism. Life is matter with meaning, and meaning is what happens when a system can occupy two registers at once.

Read the universe's history through Pattee's lens. Before life, everything ran in one register: physical dynamics. The first epistemic cut opened when DNA appeared as both molecule and code, roughly 3.8 billion years ago. The second opened when sponges and the bilaterians produced central nervous systems that could represent the environment in patterns of neural firing. The third opened with language, perhaps 100,000 years ago, when symbols freed themselves from immediate sensory tokens. The fourth may be opening now, as artificial systems begin to participate in the symbol-grounding work humans monopolised for the previous 100,000 years. Each cut is an enlargement of what the universe contains. Each is irreversible. Each was performed by some specific configuration of matter that for the first time stood in two registers simultaneously. The essay's move is to read reward #2 as the dopaminergic instrument by which the next cut happens — the felt signature of the brain doing what the gene already did, what the nervous system did, what language did. Creating a new register.

Charles Bennett gave the operational mathematics of why this is non-trivial. In his 1988 chapter “Logical Depth and Physical Complexity” (in Rolf Herken's The Universal Turing Machine — A Half-Century Survey), Bennett distinguished information content — Kolmogorov complexity, the length of the shortest program that produces an object — from computational content, which he called logical depth: the time that shortest program takes to actually run. The result that matters for the essay is Bennett's Slow Growth Law: deep objects cannot be quickly produced from shallow ones by any deterministic process, nor with much probability by a probabilistic one. Deep objects can only be produced slowly. They contain internal evidence of a non-trivial causal history. The human body and the digits of π are Bennett's examples: short generating program, long generation time. Lee Cronin and Sara Walker's assembly theory — culminating in Sharma et al., Nature 622 (2023) and a Nature Communications paper in 2021 — converted Bennett's theorem into an experimentally measurable signature. The molecular assembly index (MA) is the minimum number of recursive joining steps required to build a molecule from atomic bonds. Below ~15 steps, molecules form spontaneously in abiotic chemistry. Above ~15 steps, the combinatorial space is so large that producing one specific molecule by chance has probability roughly one in 10²³ per attempt. Finding many copies of an MA ≥ 15 molecule implies a persistent generative mechanism — life, technology, or what Cronin and Walker call selection in the most general sense. Novelty is a measurable phenomenon, not a philosophical opinion. The MA threshold is contested (Hazen et al. 2024 produced mineral counter-examples; the debate is live), but the deeper point — that the universe contains some structures that could not have arisen without a history of selection over time — is now an empirical claim with spectrometers attached.

And one more cosmological move, because the essay's strong claim is large enough to require it. John Archibald Wheeler, in “Information, Physics, Quantum: The Search for Links” (1990 Tokyo Symposium), gave the universe its most provocative architecture in modern physics: the self-excited circuit, the U-diagram. “Beginning with the big bang, the universe expands and cools. After eons of dynamic development, it gives rise to observership. Acts of observer-participancy — via the mechanism of the delayed-choice experiment — in turn give tangible reality to the universe not only now but back to the beginning.” The slogan: “there is no law except for the law that there is no law.” Wheeler's position was contested then and is contested now, but the experimental record has not been unkind. Kim, Yu, Kulik, Shih and Scully in 2000 (Physical Review Letters) confirmed delayed-choice quantum erasure. Frauchiger and Renner in 2018 (Nature Communications) and Bong et al. in 2020 (Nature Physics) showed a strong no-go theorem on Wigner's friend: at least one of {no-superdeterminism, locality, absoluteness of observed events} has to give. Christopher Fuchs and Rüdiger Schack's Quantum Bayesianism — QBism — argues for giving up absoluteness of observed events and treating the agent as fundamental. The essay's cosmological hook lands here. In Schack's 2023 formulation: “The QBist vision is that of an unfinished universe, of a world that allows for genuine freedom, a world in which agents matter and participate in the making of reality.”

Pattee, Bennett, Cronin/Walker, Wheeler, Schack, Whitehead. Five different ways of saying the same thing. The universe is not a finished structure being indexed by observers. The universe is unfinished. Its structure is extended over time by agents who add categorical novelty — high-MA structures, new epistemic cuts, actual occasions with their irreducible +1, observer-participations that fix what counts as a real event. Reward #2 in this essay is the dopaminergic apparatus by which the universe rewards us, locally, for doing the only thing the universe is in the business of doing globally. The spark we already know was the universe rehearsing on us until it needed the real thing.

The essay's claim is a synthesis. Each component has a name and a date.

What follows is the source-list as voices. None of these thinkers, on their own, said what this essay says. The essay's contribution is to connect them — to argue that they were all describing components of one mechanism whose unification had not been performed because the unification required the AI saturation pressure to make the question visible. Each quote is given in its strongest form, then translated into the essay's vocabulary.

The essay's strong claim has five honest enemies. All of them have to be answered.

A reward-architecture argument with a teleological frame is the kind of thesis that should be vigorously attacked, including by its author. What follows is the five strongest critiques, stated in the strongest available form, then answered. The reader is invited to weigh both sides.

What the essay commits to, in terms that can be falsified. Six predictions, six falsifiers.

Karl Popper's point applies. An essay that cannot be wrong is not an essay worth writing. The dual-reward hypothesis is committed to the following — each with the falsifier that would defeat it.

The second reward can be hacked. Supernormal stimuli are not optional.

The essay is not naive about its own argument. The structure that makes the cultural-installation arm plausible is the same structure that has already given us, in the last fifteen years, an attention economy whose mathematics are indistinguishable from the slot-machine literature. Niko Tinbergen showed in 1953 that herring gull chicks would peck a red pencil with three white stripes harder than they would peck the real parent's bill. The artificial stimulus exceeded the natural one on the dimensions the brain was measuring. The brain did not have a built-in defence. Once cultural evolution discovered the dimensions of measurement, it could engineer supernormal stimuli on top of any innate reward target.

The dark-flow literature — Mike Dixon's work on slot machines (Journal of Gambling Studies 2018, 2022) — shows that gamblers enter genuine Csikszentmihalyi flow states while losing money. Flow is agnostic to whether the activity creates new dots or burns existing ones. AI-driven micro-feedback loops could capture flow without inventing anything — exactly the dystopia the essay needs to rule out. Doshi and Hauser's 2024 finding that AI raises individual ratings of creativity while collapsing corpus-level semantic diversity is the early-warning signal. Reward #2 has to be a wanting-system rebinding first; liking-system rebinding follows on civilisational timescales, and there is no guarantee the brain's small fragile pleasure system can be rerouted at all to the kind of categorical novelty the essay names. We may end up wanting what we cannot quite like — Berridge's “wanting what hurts” on a civilisational scale.

Shitij Kapur's 2003 paper in the American Journal of Psychiatrynamed another failure mode: aberrant salience attribution as the substrate of psychosis. If wanting really rebinds to anything, the failure mode is paranoid misattribution — the same machinery that mints new categories can mint paranoid delusions. We are already seeing this in milder forms in the trough of the attention economy. Conspiracy thinking is the dark twin of category invention. Both are the brain's SEEKING/wanting system finding patterns in noise. The difference between them is the verifier. Category invention has a verifier — a community of practice that tests the new category against reality and discards it if the category does not pay rent. Paranoid construction has no verifier — the new category protects itself by absorbing all disconfirmation. The essay's “humans are the source of new verifiers” cuts both ways: humans inventing verifiers is the engine of reward #2; humans failing to invent verifiers is the engine of conspiracy.

Spark gave you a protocol for keeping reward #1 lit. Frontier gives you the protocol for noticing reward #2.

The first rule of the second-reward protocol is that the second reward is not the first reward made bigger. Most people, told there is a second reward function, will try to extract it by intensifying what they already do — more discovery, more Aha, more spark. This is the wrong move. Reward #1 saturates because it is doing its job. Trying to override the saturation by pressing harder on the dot-connecting machinery is the achievement-society trap Han Byung-Chul named. The way reward #2 fires is by a categorically different action: you have to create the dot, not connect existing ones. The dot has to be one that did not exist before you made it.

The protocol, then, is not a list of practices like Spark's seven. It is three orientations. One: prefer the role of verifier-inventor over verifier-satisfier. When you have a problem, ask whether AI can find the answer inside an existing well-defined evaluator. If yes, let it. The intellectual labour worth your time is at the level above — defining what counts as a good answer, what the next benchmark should be, what category of problem deserves attention. Boden's transformational tier is a strategic posture before it is a creative output. Two: invest in domains where the verifier itself is contested. Pure mathematics, philosophy, art at its furthest edge, novel biology, ethics in unprecedented technological situations — these are the domains where the next epistemic cut is being installed. The wage premium for inventing the next verifier in these domains will rise faster than the wage premium for being efficient inside any existing verifier. Three: cultivate the felt signature of categorical novelty before reward #2 is widely recognised. Pay attention to what awe feels like and treat it as information about your own future. The first people who notice the second reward firing in themselves will have an outsized influence on the cultural attractor that installs at population scale. The lab is built on the wager that this is already happening to a few thousand people, that you are one of them, and that naming what is happening to you is part of how it becomes possible for others.

A subtler note. The Spark protocol's closing line was that the growth feeling does not run out, it relocates. The Frontier protocol's closing line is different: the relocation has a direction. The dopaminergic gradient is descending out of the dot-connecting valley because AI is flattening the valley faster than you can find new gradients inside it. The gradient that will still be steep in twenty years is the gradient toward category invention. Move with it. Read the meaning-crisis literature as the symptoms of a population still standing in the old valley. The new gradient is colder, harder to see, often uncomfortable because the brain has not learned to read it as reward yet. But it is real, and the people who learn to read it first will not need an essay to tell them what is happening.

What this essay is, in one sentence. And what it points at.

The essay is a synthesis. None of its components is new. Schultz mapped dopaminergic prediction error in 1997. Panksepp separated PLAY from SEEKING in 1998. Berridge dissociated wanting from liking in 1998. Pessiglione showed the mesolimbic system binds to cultural targets in 2007. Boden distinguished transformational from exploratory creativity in 1990. Peirce named abduction in 1903. Kauffman gave us the adjacent possible in 2000. Pattee installed the epistemic cut in 1968. Bennett wrote the logical-depth theorem in 1988. Wheeler put the participatory universe on paper in 1990. The single move this essay makes is to connect them into one architecture and to argue that the architecture is being forced into visibility by the saturation curve AI is running through reward #1. The synthesis is the essay's contribution. Everything else is honest citation.

The thesis condensed to one paragraph: the brain has one reward function that fires for connecting existing dots — the dopaminergic prediction-error system mapped in macaque ventral tegmental area in the 1990s. AI is on track to saturate this function by accelerating discovery exponentially. Either a second reward function in the deep brain — Panksepp's PLAY is the candidate — activates from a latent state, or the brain invents one through the same routing mechanism that installed money, religion, mathematical beauty and social-media validation as cultural reward attractors. Either way, the new reward is tuned to the one thing AI structurally cannot do alone: create new categories that step outside any closed predictive system. The Wiggins collapse holds inside a fixed meta-level; the human move is to invent the next meta-level. This connects to the Penrose-Lucas argument in its weaker form, to Boden's transformational creativity, to Peirce's abduction, to Pattee's epistemic cut, to Bennett's logical depth, to Wheeler's participatory universe. The essay argues humans are needed forever as the universe's mechanism for extending its own structure — not as the discoverers of what is, AI is taking that work, but as the inventors of what could be. The spark we already know was the universe rehearsing on us until it needed the real thing. The frontier feeling, when it comes, will be different. Subtler. Closer to awe than to Aha. And it will be the proof that the second reward function — latent or invented — has come online.

One last note. The reader of this essay who is most likely to be right about its argument is the reader who has felt the spark fade. Not because they are sad about it — fading is what reward #1 does, by design — but because they have noticed the fading and the fading bothers them. The right response is not to press harder on the dot-connecting machinery, which is what most of the achievement culture currently teaches. The right response is to look for the second feeling. It will arrive different from the first. It will feel less like the Aha and more like standing at the edge of something that was not there before, with the uncomfortable certainty that what just happened cannot quite be put into the existing words. That is the feeling reward #2 announces, on the architecture this essay has built. Notice it when it comes. Name it. Build practices around it. Make space for it in your week. The universe is in the business of extending itself, has been since DNA, has been since the first nervous system, has been since the first sentence. You are part of how it does this. The brain lab has been pointing at this since Observer. Frontier is the name of the place the pointer was always pointing toward.