# The Philosophical Roots of Frontier AI: A Primer > A structured map from Descartes to Deep Think, designed as a reference for engineers, ML researchers, and serious readers. Built to map cleanly onto the frontier section at pracha.me/frontier. --- ## Preface — Why philosophy matters to a deep learning engineer in 2026 Every architecture is a bet about what intelligence *is*. A transformer is not philosophically neutral; it is an empiricist machine — a Humean bundle of statistical associations. A world model is a neo-Kantian machine — it builds structure before experience. Active inference is Spinozist — one substance, one principle. Formal verification is Leibnizian — a characteristica universalis for the twenty-first century. When labs disagree about AGI — Sutton vs. LeCun, Altman vs. Hassabis, Amodei vs. Marcus — they are re-enacting arguments from the seventeenth and eighteenth centuries under new names. This primer gives you the map. Once you have it, every paper, launch, and lab becomes legible as a philosophical commitment expressed in CUDA. The durable axis is simple: **reductionism vs. holism**, **tabula rasa vs. innate structure**, **scale vs. structure**. Everything else is commentary. ```mermaid timeline title Philosophical lineage to frontier AI 1637 : Descartes : Method of doubt, dualism, rationalism 1677 : Spinoza : Monism, one substance 1714 : Leibniz : Characteristica universalis, symbolic calculus 1739 : Hume : Bundle theory, induction, empiricism 1781 : Kant : Synthesis, innate categories + empirical data 1807 : Hegel : Dialectic, thesis-antithesis-synthesis 1929 : Whitehead : Process philosophy 1936 : Turing : Computation, functionalism 1943 : McCulloch-Pitts : Neural net as logic 1948 : Wiener : Cybernetics, feedback 1962 : Kuhn : Paradigm shifts 1972 : Dreyfus : Phenomenological critique 1986 : Rumelhart-Hinton : Backprop, PDP 2010 : Friston : Free energy principle 2017 : Vaswani : Attention is all you need 2019 : Sutton : Bitter lesson 2022 : LeCun : JEPA, world models 2024 : Silver-Sutton : Era of experience 2025 : Gemini Deep Think : IMO gold, hybrid reasoning 2026 : Convergence : Scaling + world models + symbolic + alignment ``` --- # Part I — Philosophical foundations **Thesis:** Modern AI debates recapitulate the rationalist–empiricist split, with Kant standing (unacknowledged) behind every hybrid architecture. The seventeenth- and eighteenth-century arguments are not historical curiosities — they are the load-bearing frames of contemporary ML. ## 1.1 Descartes: doubt, dualism, and the algorithm René Descartes (1596–1650) did two things that still matter. First, in the *Meditations* (1641), he invented the modern method of **systematic doubt**: strip away everything that can be doubted until you hit bedrock. The bedrock was *cogito ergo sum* — a thinking thing exists. Second, he split reality into two substances: *res cogitans* (thinking stuff, non-extended) and *res extensa* (physical stuff, extended in space). Mind and body are fundamentally different kinds. Three Cartesian inheritances run through AI. **Rationalism** — the claim that certain knowledge comes from reason, not experience — is the ancestor of symbolic AI and formal methods: if intelligence is theorem-proving, you want deduction from axioms. **Method** — the *Discourse on Method* (1637) is literally an algorithm: break problems into parts, proceed from simple to complex, enumerate exhaustively. Every engineering pipeline echoes this. **Dualism** is the uncomfortable legacy: the mind-body split makes substrate independence intuitive (minds can be transferred to silicon) but also creates the "hard problem" of consciousness that will haunt Part VII. **Innate ideas** are Descartes' most contested contribution. He argued that some concepts — God, self, geometric truths — cannot come from sensation and must be innate. In ML terms, these are **priors**, **inductive biases**, **architectural structure**. Whenever someone argues against tabula rasa — Chomsky against Skinner, Marcus against scaling, LeCun for world models — they are being broadly Cartesian. ## 1.2 The contenders: Aristotle, Spinoza, Leibniz **Aristotle (384–322 BCE)** matters because he refused dualism before it was invented. His hylomorphism — form and matter are inseparable — and his teleological realism (things have intrinsic goals, *telos*) anticipate modern embodied cognition and organismic views. The *Nicomachean Ethics*' virtue epistemology (practical wisdom, *phronesis*, as situated know-how) is the ancestor of Dreyfus's phenomenological critique of symbolic AI. **Baruch Spinoza (1632–1677)** proposed **monism**: there is one substance (*Deus sive Natura*, God or Nature) with infinite attributes. Mind and body are not two things but two ways of describing one thing. This is the philosophical root of **active inference** and **the free energy principle**: Friston's claim that perception, action, and learning are one optimization follows a deeply Spinozist logic — one principle, many surface expressions. Spinoza also gives us a rigorous holism: parts are only intelligible through the whole. **Gottfried Wilhelm Leibniz (1646–1716)** is the patron saint of symbolic AI. His dream of a *characteristica universalis* (a universal symbolic language) and a *calculus ratiocinator* (a calculus for reasoning) is the direct ancestor of formal logic, Frege, Russell, Gödel, Turing, Lean, and — today — Symbolica, Imandra, and AlphaProof. "When there are disputes among persons, we can simply say: let us calculate." Leibniz also invented monads — individual substances with internal states that mirror the whole — which prefigure agent-based and object-centric representations. ## 1.3 Empiricism: Locke and Hume **John Locke (1632–1704)** wrote the *Essay Concerning Human Understanding* (1689) and gave us **tabula rasa** — the mind as blank slate, filled by sensation. Every argument for "just train on enough data" traces back here. The scaling hypothesis is Locke at hyperscale. **David Hume (1711–1776)** is the philosopher deep learning engineers should read. Three of his claims are architecturally relevant. First, the **bundle theory of self**: there is no unified "I" — only a bundle of perceptions tied together by association. **This is distributed representations.** An embedding is a Humean bundle; a transformer's internal state is a bundle of attention-weighted perceptions. There is no homunculus reading out meaning — just patterns co-occurring. Second, the **problem of induction**: no finite number of observations justifies a universal law. The sun rising a million times does not prove it will rise tomorrow. This is the generalization problem. Every scaling law is a bet that induction works anyway. Third, **causation as constant conjunction**: we never observe causal connection, only regular succession. Correlation is all we have access to empirically. This is exactly the limitation Judea Pearl diagnoses in ML: almost all deep learning lives on Pearl's **Rung 1 — association** — precisely because it is Humean. ## 1.4 Kant: the synthesizer **Immanuel Kant (1724–1804)** is the unacknowledged patriarch of modern ML architecture because he solved — or at least re-framed — the rationalist/empiricist dispute. In the *Critique of Pure Reason* (1781), he argued that knowledge requires *both* sensory input (empiricism) *and* innate categories that structure that input (rationalism). Space, time, and causation are not learned from experience — they are the preconditions of having any experience at all. "Thoughts without content are empty, intuitions without concepts are blind." In ML terms, Kant is the first neuro-symbolic architect. Raw data (intuitions) + inductive biases (categories) = cognition. Every architectural choice — convolutions encoding translation invariance, transformers encoding permutation equivariance over tokens, graph nets encoding relational structure, JEPA predicting in latent space — is a Kantian category smuggled in. The **2025 consensus** — that pure scaling is necessary but not sufficient, and that world models, causal priors, or formal structure must be added — is Kantianism reborn. Bengio's "Consciousness Prior" (2017) explicitly proposes a sparse-factor prior as an innate architectural bias toward System 2 reasoning. Lake, Ullman, Tenenbaum & Gershman's "Building Machines That Learn and Think Like People" (*BBS* 2017) is perhaps the most Kantian ML paper ever written. ## 1.5 The durable axis: reductionism vs. holism Everything that follows can be read on one axis. **Reductionism** says intelligence decomposes into simpler parts — atoms, features, tokens, FLOPs. Add enough and you get a mind. **Holism** says intelligence is irreducibly systemic — it requires embodiment, world, feedback, context. Parts are only intelligible through the whole. | Reductionist pole | Holist pole | | --- | --- | | Descartes (analysis) | Aristotle (hylomorphism) | | Locke, Hume (atoms of sensation) | Spinoza (one substance) | | Logical atomism | Process philosophy | | GOFAI | Cybernetics | | Scaling hypothesis | World models, embodiment | | LLM-as-compression | Active inference | Every modern AI lab is positioned somewhere on this axis. Keep it in mind. --- # Part II — 20th-century pivots **Thesis:** The twentieth century added five moves to the inherited frame: **formalization** (logic as the language of thought), **computation** (mind as machine), **phenomenology** (intelligence as embodied skill), **systems** (intelligence as feedback), and **dialectics** (intelligence as productive conflict). All five are load-bearing in 2026. ## 2.1 Logical atomism and the linguistic turn Bertrand Russell and the early Ludwig Wittgenstein (*Tractatus Logico-Philosophicus*, 1921) argued that reality has a logical structure mirrored by language. Complex propositions decompose into atomic ones; atomic propositions picture atomic facts. This is Leibniz plus Frege plus set theory — and it is the philosophical soil in which GOFAI (Good Old-Fashioned AI) germinated. Herbert Simon and Allen Newell's **Physical Symbol System Hypothesis** (1976) — "a physical symbol system has the necessary and sufficient means for general intelligent action" — is a direct descendant. The later Wittgenstein (*Philosophical Investigations*, 1953) repudiated his own earlier view: meaning is not picture-to-fact but **use** in a form of life. Language games have no single essence — just family resemblance. This pivot seeded the skeptical tradition about symbolic AI. ## 2.2 Turing, functionalism, and the computational theory of mind Alan Turing's **"Computing Machinery and Intelligence"** (*Mind*, 1950) replaced the question "can machines think?" with the imitation game — an operational, behavioral test. More importantly, his 1936 work on Turing machines established that **computation is substrate-independent**: any universal computer can, in principle, simulate any other. This gave rise to **functionalism**: mental states are defined by their functional/computational role, not their physical realization. Hilary Putnam, Jerry Fodor, and David Marr are the canonical expositors. The **computational theory of mind** (CTM) claims the mind *is* a computer — specifically, a symbol-manipulating engine. Every substrate-independence argument in modern AI consciousness debates (Butlin & Long, 2023) sits on this foundation. ## 2.3 Phenomenology and the Dreyfus critique **Edmund Husserl**, **Martin Heidegger**, and **Maurice Merleau-Ponty** built phenomenology — the systematic study of first-person experience. Heidegger's *Being and Time* (1927) introduced **being-in-the-world**: we are not detached minds representing a world but absorbed agents coping with it. Merleau-Ponty's *Phenomenology of Perception* (1945) grounded cognition in the lived body. **Hubert Dreyfus**, building on this tradition, wrote *What Computers Can't Do* (1972) and *What Computers Still Can't Do* (1992) — the most influential philosophical critique of symbolic AI. His argument: human intelligence is not rule-following but **skilled coping**. Experts do not consult rules; they respond to situations. The frame problem, the context problem, the commonsense problem all reduce to this. Dreyfus was dismissed in the 1970s and vindicated in the 2010s: connectionism and embodied AI are broadly Dreyfusian. Every argument for **embodiment** in robotics — Physical Intelligence's π-series, Nvidia GR00T, the whole embodied-AI case — carries phenomenological DNA. So does Fei-Fei Li's **spatial intelligence** thesis: cognition requires a situated, 3D, persistent world, not a 1D token stream. ## 2.4 Cybernetics, systems thinking, and complexity **Norbert Wiener's** *Cybernetics* (1948) launched the study of feedback and control in animals and machines. The core idea — a system with a goal, a sensor, and a feedback loop is the minimal unit of purposive behavior — underwrites RL, control theory, and active inference. Wiener also raised the first alignment concerns: "We had better be quite sure that the purpose put into the machine is the purpose which we really desire." **Ludwig von Bertalanffy** (*General System Theory*, 1968), **Gregory Bateson** (*Steps to an Ecology of Mind*, 1972), **Donella Meadows** (*Thinking in Systems*, 2008), **Peter Senge** (*The Fifth Discipline*, 1990), and **Niklas Luhmann** (autopoietic social systems) extended this into a general epistemology: systems have properties — emergence, non-linearity, feedback, self-reference — irreducible to their parts. **Complexity science** (Santa Fe Institute — Kauffman, Holland, Mitchell, Wolfram) added **emergence** and **complex adaptive systems**: simple local rules produce globally structured behavior. This is the philosophical ground of scaling-law emergence claims ("capabilities emerge at scale"), of evolutionary methods (Sakana, novelty search), and of the whole "AI as ecosystem" framing. ## 2.5 Process philosophy **Alfred North Whitehead's** *Process and Reality* (1929) argued reality is not made of things but of events — "actual occasions" of experience that arise and perish. Substance is a stable pattern of process. This is the philosophical ancestor of dynamical-systems views of cognition, of Liquid AI's continuous-time neural networks, and of the organismic view at GenBio AI: a cell is not a bag of molecules but an ongoing process. ## 2.6 Dialectics and adversarial methods **Hegel's** dialectic — thesis, antithesis, synthesis — and **Marx's** materialist adaptation gave us a model of productive conflict: truth emerges through opposition. **Generative Adversarial Networks** (Goodfellow et al., 2014) are dialectic in pure form — generator and discriminator locked in antagonism that produces synthesis. **RLHF** (reward model vs. policy), **Constitutional AI** (critic and reviser), **debate-based alignment** (Irving et al., 2018), and **AlphaZero's** self-play are all dialectical. "Adversarial robustness" is the Hegelian fingerprint on modern ML. --- # Part III — A MECE map of AI/ML schools in 2026 **Thesis:** Seven schools partition the 2026 landscape cleanly. Each has a distinct philosophical root, a core computational commitment, canonical methods, seminal papers, representative titans, and characteristic failure modes. Every lab, paper, and product can be located on this map, usually as a weighted combination of two or three schools. ```mermaid mindmap root((AI schools 2026)) Symbolic/Rationalist Leibniz GOFAI Formal methods Imandra, Symbolica Connectionist/Scaling Hume, Locke Deep learning LLMs OpenAI, xAI Embodied/World Models Merleau-Ponty, Heidegger JEPA, Cosmos Meta, World Labs, Nvidia, Pi Neuro-Symbolic/Hybrid Kant System1+System2 AlphaProof, DeepMind, Anthropic Evolutionary/Open-Ended Darwin, Stanley-Lehman Novelty search, QD Sakana, Clune Organismic/Active Inference Spinoza, Aristotle Free energy principle Friston, VERSES, GenBio Causal/Probabilistic Laplace, Bayes Pearl ladder Tenenbaum, Pearl ``` ## 3.1 Symbolic / rationalist - **Root:** Leibniz's *characteristica universalis*; logical atomism; Simon–Newell physical symbol system hypothesis. - **Thesis:** Intelligence is symbol manipulation. Meaning is compositional. Reasoning is deduction. - **Methods:** First-order logic, theorem proving, expert systems, formal verification, program synthesis. - **Canonical works:** Newell & Simon's GPS and Logic Theorist (1956–1960); McCarthy's *LISP*; Cyc (Lenat, 1984–). - **Titans:** Judea Pearl (partly), Stuart Russell (partly), Gary Marcus (advocate). - **2026 instantiations:** Imandra (ImandraX, CodeLogician); Symbolica (categorical deep learning); ExtensityAI (SymbolicAI); Lean-based proof systems used by AlphaProof. - **Limitations:** Brittle at scale; cannot learn from perception; frame/commonsense problem. Dreyfus's critique still bites. ## 3.2 Connectionist / statistical / scaling - **Root:** Humean empiricism — mind as bundle of associations; Lockean tabula rasa. - **Thesis:** Intelligence emerges from learning statistical regularities from massive data. Architectural simplicity + compute + data is sufficient. - **Methods:** Deep learning, transformers, next-token prediction, scaling laws (Kaplan 2020, Chinchilla 2022), RLHF. - **Canonical works:** Rumelhart, Hinton & Williams on backprop (1986); Krizhevsky AlexNet (2012); Vaswani et al. "Attention Is All You Need" (2017); Kaplan et al. "Scaling Laws" (2020); Hoffmann et al. "Chinchilla" (2022). - **Titans:** Geoffrey Hinton, Yann LeCun (historically), Yoshua Bengio, Richard Sutton (methodologically), Ilya Sutskever, Sam Altman, Dario Amodei, Jared Kaplan. - **Seminal philosophical statement:** Sutton's **"The Bitter Lesson"** (2019): *"The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin."* - **2026 instantiations:** GPT-5, Claude Opus 4.5, Gemini 3, Grok 4, DeepSeek-V3. - **Limitations:** Humean induction problem; lives on Pearl's Rung 1; data-hungry; brittle out-of-distribution; interpretability gap. ## 3.3 Embodied / world models - **Root:** Merleau-Ponty and Heidegger on being-in-the-world; Gibson's ecological psychology. - **Thesis:** Intelligence is inseparable from a situated body in a physical world. 1D token streams cannot model 3D space, time, physics, and persistence. - **Methods:** Joint-embedding predictive architectures (JEPA), video world models, robot foundation models, spatial tokenizers. - **Canonical works:** LeCun, "A Path Towards Autonomous Machine Intelligence" (2022); V-JEPA 2 (2025, arXiv 2506.09985); Nvidia Cosmos (2501.03575); π-0 (Physical Intelligence, 2410.24164); Fei-Fei Li, "From Words to Worlds" (2025). - **Titans:** Yann LeCun, Fei-Fei Li, Sergey Levine, Chelsea Finn, Jensen Huang (industrial patron). - **2026 instantiations:** Meta FAIR V-JEPA 2; World Labs (Marble, RTFM); Nvidia Cosmos + GR00T; Physical Intelligence π-0.5 and π*-0.6; DeepMind Genie 3. - **Seminal philosophical statement:** LeCun, 2024–2025 repeatedly: *"LLMs are an off-ramp on the road to human-level AI."* - **Limitations:** Currently far behind LLMs on text-heavy tasks; data scarcity for embodied experience; evaluation difficulty. ## 3.4 Neuro-symbolic / hybrid - **Root:** Kant's synthesis — innate categories + empirical data. Dual-process cognition (Kahneman, *Thinking, Fast and Slow*, 2011). - **Thesis:** Neither pure scaling nor pure symbol manipulation suffices. System 1 (pattern completion, neural) and System 2 (deliberation, symbolic/search) must be combined. - **Methods:** LLM + formal verifier; LLM + search (MCTS, beam, best-of-N); modular architectures; categorical deep learning; constitutional AI and deliberative alignment. - **Canonical works:** Lake, Ullman, Tenenbaum & Gershman, "Building Machines That Learn and Think Like People" (*BBS* 2017); AlphaGeometry (Trinh et al., *Nature*, Jan 2024); AlphaProof/AlphaGeometry 2 (DeepMind, July 2024; *Nature* paper Nov 2025); Bai et al., "Constitutional AI" (Anthropic, 2022). - **Titans:** Demis Hassabis, Joshua Tenenbaum, Gary Marcus, Dario Amodei, George Morgan (Symbolica), Petar Veličković. - **2026 instantiations:** AlphaProof, AlphaGeometry 2, Gemini Deep Think (IMO gold 2025), OpenAI o3/o4 deliberative alignment, Imandra + LLM, Symbolica categorical DL. - **Limitations:** Integration seams; formalization bottleneck; when it works it looks like magic, when it doesn't it looks like kludge. ## 3.5 Evolutionary / open-ended - **Root:** Darwin; Stanley & Lehman, *Why Greatness Cannot Be Planned* (2015); complexity science. - **Thesis:** Objective-driven search is deceptive. True progress emerges from open-ended novelty search, quality-diversity, and population-based evolution. Intelligence is not optimized; it is *accumulated*. - **Methods:** Genetic algorithms, CMA-ES, novelty search, MAP-Elites quality-diversity (Mouret & Clune 2015), evolutionary model merging, AI-generating algorithms (Clune 2019). - **Canonical works:** Stanley & Lehman (2015); Mouret & Clune, "Illuminating Search Spaces by Mapping Elites" (2015); Akiba et al. "Evolutionary Optimization of Model Merging Recipes" (arXiv 2403.13187, 2024); Lu et al., "The AI Scientist" (2408.06292, 2024); DeepMind FunSearch (*Nature*, 2023) and AlphaEvolve (May 2025). - **Titans:** Kenneth Stanley, Joel Lehman, Jeff Clune, David Ha, Risto Miikkulainen. - **2026 instantiations:** Sakana AI (evolutionary model merging, AI Scientist v2, Transformer², M2N2); DeepMind AlphaEvolve; quality-diversity ecosystems around Prime Intellect environments. - **Limitations:** Compute-hungry in a different way; evaluation is hard without clear fitness; still niche commercially. ## 3.6 Organismic / active inference / free energy - **Root:** Spinozist monism (one principle); Aristotelian hylomorphism; autopoiesis (Maturana & Varela); Helmholtz's "unconscious inference"; Friston's free energy principle. - **Thesis:** Intelligence — and perhaps life — is the minimization of variational free energy. Action, perception, and learning are three faces of one optimization. Agents are organisms, not function approximators. - **Methods:** Active inference, predictive coding, hierarchical Bayesian generative models, renormalizing generative models. - **Canonical works:** Friston, "The Free-Energy Principle: A Unified Brain Theory?" (*Nat Rev Neurosci*, 2010); Friston et al., *Active Inference: The Free Energy Principle in Mind, Brain, and Behavior* (MIT Press, 2022); Seth, *Being You* (2021); "From pixels to planning: scale-free active inference" (VERSES, 2024). - **Titans:** Karl Friston, Anil Seth, Andy Clark, Gabriel René. - **2026 instantiations:** VERSES AI (Genius, AXIOM, RGMs); partially GenBio AI (organismic framing across biological scales); arguably any world-model-with-active-exploration architecture. - **Limitations:** Formal generality that critics say verges on unfalsifiability; engineering demonstrations lag theoretical claims; commercial scale-up unproven. ## 3.7 Causal / probabilistic - **Root:** Laplace, Bayes, Hume's problem of induction taken seriously; Pearl's structural causal models. - **Thesis:** Intelligence requires a ladder of inference — association, intervention, counterfactual. Without causal models, systems cannot generalize robustly, intervene correctly, or reason about what-if. - **Methods:** Bayesian networks, structural causal models, do-calculus, probabilistic programs, Bayesian program induction. - **Canonical works:** Pearl, *Causality* (2000); Pearl & Mackenzie, *The Book of Why* (2018); Pearl, "The Seven Tools of Causal Inference" (*CACM*, 2019); Lake et al. (2017); Tenenbaum et al., "How to grow a mind" (2011). - **Titans:** Judea Pearl, Josh Tenenbaum, Bernhard Schölkopf, Yoshua Bengio (partially, via System 2 and causal representation learning). - **2026 instantiations:** Residual but crucial — causal representation learning, structured probabilistic programs, some aspects of GenBio and VERSES. Pearl himself has remained skeptical that LLMs genuinely ascend beyond Rung 1. - **Limitations:** Does not scale as cleanly as gradient descent; causal discovery from observational data is hard; has been absorbed partially by neuro-symbolic hybrids. ### Summary table | School | Root philosopher | Thesis | Seminal 2020s paper | Lead labs 2026 | | --- | --- | --- | --- | --- | | Symbolic | Leibniz | Intelligence = symbol manipulation | AlphaProof/Lean methodology (2024) | Imandra, Symbolica, ExtensityAI | | Connectionist/scaling | Hume, Locke | Scale + data = intelligence | Sutton, Bitter Lesson (2019) | OpenAI, xAI, DeepSeek | | Embodied/world models | Merleau-Ponty | Cognition requires world | LeCun APTAMI (2022); V-JEPA 2 (2025) | Meta FAIR, World Labs, Pi, Nvidia | | Neuro-symbolic | Kant | System 1 + System 2 | Lake et al. BBS (2017); AG2 (2024) | DeepMind, Anthropic, Symbolica | | Evolutionary | Darwin, Stanley-Lehman | Open-ended novelty | Stanley & Lehman (2015); Sakana (2024) | Sakana, DeepMind (AlphaEvolve) | | Active inference | Spinoza, Friston | One principle, minimize free energy | Friston (2010); VERSES RGMs (2024) | VERSES, partly GenBio | | Causal/probabilistic | Laplace, Pearl | Ascend the causal ladder | Pearl CACM (2019) | MIT BCS, distributed | --- # Part IV — The great debates **Thesis:** Five fault lines define the 2026 conversation. Each is a classical philosophical dispute re-enacted in ML terms, and each has become empirically testable in a way the classical versions were not. ## 4.1 The Bitter Lesson vs. bio-inspired / cognitive science Sutton's 2019 essay is the charter document of pure scaling. *"Researchers have repeatedly tried to build into their systems the human knowledge they think useful… but all of this is a long-term failure."* Deep Blue, AlphaGo, AlexNet, GPT — in each case the baked-in knowledge lost to the generic scalable method. The cognitive-science counter, articulated by Marcus, Tenenbaum, Lake, and Gershman, argues that human intelligence demonstrably uses core priors — object permanence, agents, causality, compositional generalization — and that ignoring these costs vast sample efficiency. Lake et al. (BBS 2017) is the canonical reply. The 2025–2026 twist: Silver & Sutton's **"Welcome to the Era of Experience"** (preprint chapter for *Designing an Intelligence*, MIT Press, April 2025) pushes the Bitter Lesson into a third era. Era 1 (simulation, AlphaGo); Era 2 (human data, GPT); Era 3 (experience — agents generating their own training data via environmental interaction). AlphaProof, DeepSeek-R1, and agentic RL systems are cited as evidence. The counter is immediate: Steven Byrnes and others argue Era 3 without alignment structure is existentially reckless. ## 4.2 Scaling vs. world models This is the **Altman/Sutton axis** vs. the **LeCun/Hassabis axis**. Altman publicly commits to trillion-dollar infrastructure spend; Lightcap (OpenAI COO) in 2025: *"Our scaling laws still hold… there's no reason to believe there's any kind of diminishing return on pre-training."* Musk's Colossus 2 (January 2026, Southaven, MS) is the first gigawatt training cluster — scaling as industrial strategy. LeCun's counter, hardened through 2024–2025: *"LLMs are useful, but they are an off-ramp on the road to human-level AI. If you are a PhD student, don't work on LLMs."* In November 2025 he left Meta to start a JEPA-focused company. Fei-Fei Li's "From Words to Worlds" (November 2025) extends the argument: language is 1D; the world is at least 4D; spatial intelligence is a different kind of problem. Hassabis occupies a middle position — scale *and* scaffold. Gemini 3 + Deep Think + AlphaFold + AlphaEvolve + Genie 3 is hedged across schools. ## 4.3 LLMs vs. System 2 reasoning A year ago, critics said LLMs could not reason. In September 2024, OpenAI's **o1** opened a new scaling axis: *"The performance of o1 consistently improves with more reinforcement learning (train-time compute) and with more time spent thinking (test-time compute)."* Then **o3** (December 2024) hit 87.5% on ARC-AGI and 25.2% on FrontierMath. **DeepSeek-R1** (January 2025, arXiv 2501.12948) showed pure RL from a base model — no SFT, using GRPO and rule-based verifiable rewards — elicits emergent self-reflection (*"Wait, let me reconsider"* — the now-canonical "aha moment"). **Gemini Deep Think** took IMO gold in July 2025 end-to-end in natural language, within the 4.5-hour contest limit. The debate is whether this is genuine System 2 or very expensive interpolation. Marcus remains skeptical; Bengio (consciousness prior) sees architectural progress; Chollet (ARC) updated to say o3 is a "genuine breakthrough" while still maintaining it fails on easy tasks. ## 4.4 Reductionism vs. holism in modern ML The **mechanistic interpretability** program (Anthropic: Olah, Elhage, Batson, Lindsey) is aggressively reductionist — features, circuits, attribution graphs. **Scaling Monosemanticity** (May 2024) found millions of interpretable features in Claude 3 Sonnet including the infamous **Golden Gate Bridge feature**. **"On the Biology of a Large Language Model"** (March 2025) showed genuine multi-step reasoning circuits, parallel arithmetic circuits, and forward-backward poetry planning. **"Emergent Introspective Awareness"** (Lindsey, October 2025) showed Claude Opus 4/4.1 can sometimes detect injected concepts before they affect outputs. The holist reply: *features are not enough.* Seth's biological naturalism, Friston's whole-organism framing, and the embodied-AI school all insist meaning lives in the system's coupling with a world, not in a list of features. The 2026 consensus is that both views capture real things — reductionist interpretability for alignment and debugging; holist framings for understanding intelligence proper. ## 4.5 Anthropomorphism vs. evolution as the only proof A quieter but important debate: is the anthropomorphic frame (GWT, HOT, human-style reasoning) itself a distraction? Some researchers argue the only real proof of general intelligence is **evolutionary survival under open-ended pressure** — a Stanley-Lehman, Clune-style view. On this view, benchmarks, imitation games, and IMO scores are measuring the wrong thing; AGI will only be recognized in retrospect, after systems have survived and adapted across novel environments. This frame underwrites the evolutionary / open-ended school's skepticism of current benchmark culture. --- # Part V — Lab landscape 2026 **Thesis:** Every lab is a placed bet. You can predict most of what a lab will ship in the next twelve months from its philosophical posture — and the posture is readable from founder statements, flagship papers, and hardware choices. ```mermaid graph LR A[Scaling bet] --> OpenAI A --> xAI B[Scaling + safety] --> Anthropic C[Scaling + scaffold] --> DeepMind D[World models] --> Meta_FAIR[Meta FAIR] D --> WorldLabs[World Labs] D --> Nvidia D --> PhysicalIntelligence[Physical Intelligence] E[Neuro-symbolic] --> Symbolica E --> Imandra E --> ExtensityAI F[Evolutionary] --> Sakana G[Active inference] --> VERSES H[Organismic biology] --> GenBio[GenBio AI] I[Alignment-first] --> SSI[Safe Superintelligence] J[Alternative substrate] --> Liquid[Liquid AI] J --> Extropic J --> Rain[Rain AI] ``` ## 5.1 OpenAI — pure scaling, now with reasoning - **Key figure:** Sam Altman (CEO); Greg Brockman; historically Ilya Sutskever. - **Philosophical stance:** Scaling hypothesis maximalist with reasoning overlay. Scaling laws "still hold"; compute is destiny. - **Flagship 2025–2026:** **GPT-5** (August 7, 2025) — unified system with a real-time router between fast and thinking modes; GPT-5 Pro/mini/nano variants; **o-series** (o1 September 2024; o3 December 2024; o4 2025); Stargate infrastructure partnership (OpenAI/SoftBank, 10-GW-class data centers); $10B Cerebras compute deal (2026–2028). SOTA on AIME 2025 (94.6%), SWE-bench Verified (74.9%). - **Signature quote:** Altman, *"You should expect OpenAI to spend trillions of dollars on data center construction in the not very distant future."* ## 5.2 xAI — compute maximalism - **Key figure:** Elon Musk; Igor Babuschkin (historically). - **Philosophical stance:** Pure scaling plus "truth-seeking" framing. Vertically integrated — Tesla Megapacks, SpaceX-velocity hardware, X/Twitter data. - **Flagship:** Grok 3 (early 2025); **Grok 4** (~200M H100-hours, 15× Grok 2); Grok 5 training (late 2025). **Colossus** (Memphis, 100K→200K H100s, built in 122 days); **Colossus 2** (Southaven, MS, January 2026) — "first gigawatt training cluster in the world"; roadmap to 1M+ GPUs. - **Philosophical fingerprint:** Belief that scale alone, applied aggressively enough, produces AGI — reductionism at industrial scale. ## 5.3 Nvidia — "Physical AI" - **Key figure:** Jensen Huang; Bill Dally; Sanja Fidler; Jim Fan. - **Philosophical stance:** Embodied world models as the next scaling frontier. LLMs solved language; now solve space. - **Flagship:** **Cosmos** (CES January 2025, arXiv 2501.03575) — open world foundation models for robots and AVs; expanded at GTC March 2025 with a reasoning WFM. **Project GR00T** (humanoid robot foundation model); **GR00T Blueprint** for synthetic data via Omniverse + Cosmos Transfer. Customers: 1X, Figure, Agility, XPENG, Waabi. - **Signature quote:** Huang, *"Just as LLMs revolutionized generative and agentic AI, Cosmos world foundation models are a breakthrough for physical AI."* ## 5.4 Meta FAIR — JEPA and the LeCun agenda - **Key figure:** Yann LeCun (until November 2025 departure); FAIR team now led by Joelle Pineau and others. - **Philosophical stance:** Non-generative world models via joint-embedding prediction. LLMs are an off-ramp. - **Flagship:** **V-JEPA 2** (June 11, 2025, arXiv 2506.09985) — 1.2B-parameter video world model trained on 1M+ hours; zero-shot robot planning. Preceded by I-JEPA (2023) and V-JEPA (2024). - **2025 rupture:** LeCun announced in November 2025 he is leaving to start a JEPA-focused company, reportedly with Meta licensing access. One of the highest-profile philosophical secessions in AI history. ## 5.5 World Labs — spatial intelligence - **Key figure:** Fei-Fei Li; Justin Johnson; Ben Mildenhall (NeRF); Christoph Lassner. - **Philosophical stance:** **Spatial intelligence** — 3D/4D world models are a distinct problem from language modeling. - **Flagship:** Launched September 2024 with $230M; **Marble** (November 12, 2025) first commercial product — persistent, editable 3D world generation from text/image/video/panorama; **Chisel** hybrid 3D editor separating structure from style; **RTFM** (Real-Time Frame Model) with spatial memory. - **Signature text:** Li, *"From Words to Worlds"* (November 17, 2025): *"The dimensionality of representing a world is vastly more complex than that of a one-dimensional, sequential signal like language."* ## 5.6 Physical Intelligence — embodied foundation models - **Key figure:** Sergey Levine; Chelsea Finn; Karol Hausman. - **Philosophical stance:** Generalist robot foundation models trained on heterogeneous physical experience — Merleau-Ponty plus VLMs. - **Flagship:** **π-0** (October 2024, arXiv 2410.24164) — VLA flow-matching atop PaliGemma; 10,000+ hours across 7 robot platforms. **π-0.5** (April 2025, arXiv 2504.16054) — open-world generalization to unfamiliar kitchens/bedrooms. **π*-0.6** (November 2025, arXiv 2511.14759) — "a VLA that Learns from Experience," RL on generalist policies. Open-source via `openpi`. - **Signature claim:** *"To build AI systems with the kind of physically situated versatility that people possess, we need to make AI systems embodied."* ## 5.7 Google DeepMind — scaffolded hybrids - **Key figure:** Demis Hassabis; Shane Legg; Koray Kavukcuoglu. - **Philosophical stance:** Kantian synthesis operationalized — scale plus search plus specialized verifiers plus world models. - **Flagship:** **Gemini 3 Pro** (November 2025); **Gemini Deep Think** (IMO gold July 2025, ICPC gold 2025); **AlphaProof + AlphaGeometry 2** (IMO silver July 2024, *Nature* paper November 2025); **AlphaEvolve** (May 2025) — Gemini-powered evolutionary coding agent that discovered a new 4×4 matrix-multiplication algorithm improving on Strassen 1969 and sped up Gemini's own training; **Genie 3** (August 2025) — real-time 720p/24fps interactive world model; **AlphaFold 3** (*Nature*, 2024). - **Philosophical fingerprint:** The broadest hedge in the industry — scaling, reasoning, symbolic, evolutionary, world models, all co-developed. ## 5.8 Anthropic — safety-first scaling plus interpretability - **Key figure:** Dario and Daniela Amodei; Chris Olah (interpretability); Jan Leike (alignment). - **Philosophical stance:** Scaling works but is dangerous; pair capability with Constitutional AI and mechanistic interpretability. - **Flagship:** **Claude 4 Opus & Sonnet** (May 2025) with Model Context Protocol and Claude Code GA; **Claude 4.5 family** (late 2025, Opus 4.5 hit 80.9% on SWE-bench Verified). **Constitutional AI** (Bai et al., arXiv 2212.08073, December 2022) expanded through 2025–2026; **interpretability** (Toy Models of Superposition 2022; Towards Monosemanticity October 2023; Scaling Monosemanticity May 2024 with Golden Gate Claude; Circuit Tracing/"On the Biology of an LLM" March 2025; "Emergent Introspective Awareness" October 2025). **Model welfare** program launched April 2025 (Kyle Fish). - **Signature text:** Dario Amodei, *"Machines of Loving Grace"* (October 2024) — powerful AI by 2026–2027 could compress a century of biomedical progress into 5–10 years; follow-up *"The Adolescence of Technology"* (January 2026) on alignment-faking and scheming in Claude 4 Opus. ## 5.9 Sakana AI — evolutionary model merging - **Key figure:** David Ha; Llion Jones (*Attention Is All You Need* co-author); Takuya Akiba; Robert Lange. - **Philosophical stance:** Nature-inspired, collective intelligence, quality-diversity over monolithic scale. - **Flagship:** **Evolutionary Model Merging** (Akiba et al., arXiv 2403.13187, March 2024; *Nat Mach Intell* 2024) producing EvoLLM-JP/EvoVLM-JP. **The AI Scientist** (Lu et al., arXiv 2408.06292, August 2024) — end-to-end automated research at ~$15/paper. **AI Scientist-v2** (arXiv 2504.08066, April 2025) — tree-search; produced the first AI-generated paper to pass ICLR 2025 workshop peer review. **Transformer²** (arXiv 2501.06252, January 2025) — self-adaptive weights via singular-value fine-tuning. **M2N2** (arXiv 2508.16204, 2025) — quality-diversity niche-based merging. - **Signature claim:** *"Our new system paves the way for a new generation of adaptive AI models… embodying living intelligence capable of continuous change and lifelong learning."* ## 5.10 VERSES AI — active inference - **Key figure:** Karl Friston (Chief Scientist); Gabriel René (CEO). - **Philosophical stance:** Free energy principle as the unifying principle of intelligence; active inference as the practical method. - **Flagship:** **Genius** (commercial launch April 30, 2025) — Bayesian/active-inference agentic toolkit; **AXIOM** (June 2025) beating top deep-RL baselines on Atari; **Renormalizing Generative Models (RGMs)** — "From pixels to planning: scale-free active inference" (2024); **August 2024 open letter** to OpenAI proposing collaboration; Gartner 2025 Emerging Tech Impact Radar for Spatial AI; IWAI 2025 sponsorship. - **Signature claim (René):** RGMs are *"a fundamental shift in how we think about building intelligent systems from first principles… the 'one method to rule them all.'"* - **Caveats:** Benchmark claims are largely self-reported pending independent replication. ## 5.11 Symbolica, Imandra, ExtensityAI — neuro-symbolic - **Symbolica** (George Morgan, out of stealth April 2024, $31M Khosla-led Series A; advisor Stephen Wolfram). **Categorical deep learning**: co-authored position paper with DeepMind (Veličković et al.) generalizing geometric deep learning via endofunctor algebras. Limited product output through 2025; research programme around non-autoregressive structured reasoning. - **Imandra** (Passmore, Ignatovich, Austin). **ImandraX** (February 2025) — new reasoning engine with first formally verified proof checker for neural-network safety properties. **CodeLogician** (March 2025) — LangGraph agent converting source code into formally verified models in IML. **Imandra Universe** (June 2025) — MCP access to symbolic reasoning from Claude/ChatGPT/Cursor. Benchmark closing a 41–47pp accuracy gap vs LLM-only on code tasks. - **ExtensityAI** (Marius-Constantin Dinu, Austria). **SymbolicAI** framework (arXiv 2402.00854, February 2024; formally published at CoLLAs 2025). Treats LLMs as semantic parsers; **contractual programming** with pre/post-conditions; VERTEX benchmark. ## 5.12 GenBio AI — digital organism - **Key figure:** Eric Xing; Le Song. - **Philosophical stance:** Multiscale organismic intelligence — life is intelligible only across scales from DNA to tissue. - **Flagship:** **AIDO** — AI-Driven Digital Organism framework (arXiv 2412.06993, December 2024). Phase 1 released December 2024: **AIDO.DNA** (7B, 796 species), **AIDO.RNA** (1.6B, 42M ncRNA — largest of its kind), **AIDO.Protein** (MoE), **AIDO.Cell** (3M–650M, 50M human cells with full transcriptome context), **AIDO.StructureTokenizer**. 2025 expansions: AIDO.StructurePrediction (AlphaFold 3-style multi-biomolecule), AIDO.Protein-RAG, AIDO.Tissue (spatial transcriptomics). - **Signature claim:** *"To truly understand life, we must model and simulate it across every scale… biology becomes computable, predictable, and ultimately programmable."* ## 5.13 Safe Superintelligence — alignment-first - **Key figure:** Ilya Sutskever (co-founder, June 2024); Daniel Gross; Daniel Levy. - **Philosophical stance:** Build superintelligence directly, with alignment as the sole product constraint, no intermediate commercial distractions. - **Flagship:** Almost entirely stealth. Reported $5B+ valuation by late 2024, $30B+ by 2025. No public model. Philosophical significance is the bet itself: that safe ASI requires dedicated focus, no CEO noise, no product cadence. Sutskever's departure from OpenAI and reconstitution around this thesis is the most consequential single-person move in recent AI history. ## 5.14 Liquid AI, Rain AI, Extropic — alternative substrates - **Liquid AI** (Hasani, Lechner, Amini, Rus — MIT CSAIL spinout, 2023). **Liquid Foundation Models** (LFM v1 October 2024; LFM2 July 2025; Liquid Nanos September 2025; LFM2 technical report arXiv 2511.23404, November 2025). Continuous-time differential-equation substrate rooted in Liquid Time-constant Networks (AAAI 2021). $250M Series A (December 2024, AMD Ventures) at $2.3B valuation. - **Rain AI** (Altman-backed, 2017–2025). Pursued digital in-memory compute, claimed up to 10,000× energy-efficiency improvements for training. Cautionary tale: Series B stalled; by May 2025 seeking a buyer; being circled by OpenAI, Nvidia, Microsoft. - **Extropic** (Verdon, McCourt, 2022). **Thermodynamic computing** — use transistor thermal noise as entropy for sampling; **pbits/pdits/pmodes** composed into a **Thermodynamic Sampling Unit (TSU)**. X0 prototype (Q1 2025), XTR-0 development platform (Q3 2025), **Z1 production chip** (early access 2026). **Denoising Thermodynamic Models** paper (arXiv 2510.23972, October 2025) claims ~10,000× lower energy per sample vs GPU diffusion on Fashion-MNIST-scale benchmarks in simulation. Open-source `thrml` simulator. --- # Part VI — The training pipeline as philosophical injection point **Thesis:** Every school intervenes at a specific stage of the modern training pipeline. Reading a training recipe is reading a philosophical argument. ```mermaid flowchart LR A[Data curation] --> B[Pre-training] B --> C[Mid-training / architecture] C --> D[Post-training] D --> E[Inference-time] subgraph Schools S1[Scaling/connectionist]:::sc S2[Embodied/world models]:::em S3[Neuro-symbolic]:::ns S4[Active inference]:::ai S5[Evolutionary]:::ev S6[Causal]:::ca S7[Symbolic]:::sy end S1 --> B S2 --> B S2 --> C S3 --> C S3 --> D S4 --> C S5 --> D S6 --> A S6 --> C S7 --> D S7 --> E classDef sc fill:#ffd; classDef em fill:#dfd; classDef ns fill:#ddf; classDef ai fill:#fdf; classDef ev fill:#fdd; classDef ca fill:#dff; classDef sy fill:#eee; ``` ## 6.1 Pre-training — the scaling stage **Scaling hypothesis injection.** Kaplan et al. "Scaling Laws for Neural Language Models" (arXiv 2001.08361, 2020) and Hoffmann et al. "Chinchilla" (arXiv 2203.15556, 2022) formalize the empiricist bet: loss is a power-law function of compute, parameters, and tokens. Sutskever's framing of next-token prediction as lossless compression of the training distribution is Humean induction with a modern compression-theoretic gloss. **World-model injection.** LeCun's JEPA (2022) replaces pixel/token generation with prediction in latent embedding space — a Kantian move: the model should predict the *structure* of the world, not reproduce its surface. V-JEPA 2 (2025) extends this to video at scale. Nvidia Cosmos trains world foundation models on robot and AV video. **Causal injection.** Data-curation choices that preserve interventional and counterfactual structure (domain randomization, deliberately diverse environments, causal data collection) are a probabilistic/causal intervention at the data stage. ## 6.2 Mid-training and architecture **Architecture is where Kant lives.** Every inductive bias is a category: - **MoE** (Mixtral 8×7B 2023; DeepSeek-V3 2024; Kimi K2 1T; INTELLECT-3 Nov 2025) is specialization-under-routing — weakly modular cognition. - **State-space models** — Mamba (Gu & Dao, arXiv 2312.00752, 2023); Mamba-2 — linear-time sequence modeling; philosophically a dynamical-systems rebuttal to transformer hegemony. - **Liquid neural networks** (Hasani et al., 2021–2025) — continuous-time ODEs; Whiteheadian process metaphysics in PyTorch. - **Neuro-symbolic modules** — Symbolica's categorical deep learning; ExtensityAI's SymbolicAI primitives/contracts; hybrid transformer + verifier (AlphaProof). - **Memory and retrieval** — RAG (Lewis et al. 2020), vector memory, 1M–10M-token contexts (Gemini 1.5/2.5) — an external-memory fix to the Humean bundle problem. - **Active-inference modules** — hierarchical Bayesian generative components (VERSES, RGMs) inject Friston directly. ## 6.3 Post-training — the alignment stage **RLHF** (Ouyang et al., InstructGPT, arXiv 2203.02155, 2022) injects human preference as reward — Rousseau's general will, sort of. **DPO** (Rafailov et al., arXiv 2305.18290, 2023) closed-form reparameterizes away the explicit reward model — widely adopted in open source. **Constitutional AI / RLAIF** (Bai et al., arXiv 2212.08073, 2022) replaces human labelers with a critic LLM governed by a written constitution — dialectic baked into training. OpenAI's **deliberative alignment** (arXiv 2412.16339, December 2024) extends this into reasoning time. **RLVR — Reinforcement Learning from Verifiable Rewards** is the defining post-training paradigm of 2025. Rewards come from programmatic verifiers (unit tests, math checkers, Lean proofs, compilers) — non-gameable. DeepSeek-R1's GRPO (no critic, group-relative baselines, rule-based rewards) plus the "reasoning gyms" ecosystem (Prime Intellect Environments Hub, `verifiers`, INTELLECT-2/3, Reasoning Gym arXiv 2505.24760) constitute the full open-source stack. **Evolutionary post-training.** Sakana's model merging is evolution applied *after* training is ostensibly done — finding populations of specialists rather than one generalist. AlphaEvolve (DeepMind, May 2025) uses an LLM-as-mutator plus evolutionary selection to *discover* new algorithms — including a new 4×4 matrix-multiplication procedure beating Strassen. ## 6.4 Inference-time — the new scaling axis o1 (September 2024), o3 (December 2024), DeepSeek-R1 (January 2025), Gemini Deep Think (2025), Kimi k1.5, Qwen QwQ-32B all exploit a second scaling axis: **test-time compute**. Snell et al. (arXiv 2408.03314, DeepMind/Berkeley, August 2024) showed compute-optimal allocation of test-time compute can beat a 14× larger model in FLOP-matched evaluation. Philosophically, inference-time reasoning is the cleanest instantiation of **System 2** in current systems: deliberate, serial, token-budgeted, revisable. It also flirts with the neuro-symbolic school when the chain of thought includes tool calls to verifiers (AlphaProof's RL-on-Lean loop, reasoning models with calculator/code tools). ### Pipeline-to-philosophy table | Stage | Intervention | School | Representative work | | --- | --- | --- | --- | | Data curation | Causal/diverse data | Causal | Schölkopf causal ML | | Pre-training | Scale tokens+params+compute | Connectionist | Chinchilla (2022) | | Pre-training | Latent predictive objective | Embodied | V-JEPA 2 (2025) | | Architecture | Inductive biases | Kantian hybrid | Mamba, MoE, JEPA | | Architecture | Categorical structure | Symbolic | Symbolica CDL | | Architecture | Generative Bayesian modules | Active inference | VERSES RGMs | | Post-training | RLHF / DPO | Dialectic | Ouyang 2022, Rafailov 2023 | | Post-training | Constitutional AI | Dialectic + rule-based | Bai 2022 | | Post-training | RLVR / GRPO | Empirical + formal | DeepSeek-R1 (2025) | | Post-training | Evolutionary merge | Evolutionary | Sakana (2024) | | Inference-time | Long chain-of-thought | Neuro-symbolic | o1/o3, Deep Think | | Inference-time | Search + verifier | Symbolic + neural | AlphaProof, AG2 | --- # Part VII — Reasoning, consciousness, and the hardware substrate **Thesis:** Three hard problems remain: reasoning (solvable, partly solved), consciousness (open and urgent), and hardware (an energy and architecture crisis that may determine everything else). ## 7.1 Scaling laws for reasoning Inference-time scaling opened a second curve orthogonal to pretraining. o1's technical report demonstrated that AIME 2024 accuracy climbs monotonically with both RL-training compute and test-time thinking tokens. o3's ARC-AGI jump from ~5% (GPT-4 class) to 87.5% at high compute (December 2024) validated the axis at larger scale. DeepSeek-R1 (arXiv 2501.12948) then showed reasoning can emerge from a base model under **pure RL without SFT** — the "aha moment" is an emergent, unscripted behavior — using **GRPO**, group-relative policy optimization without a critic, with rule-based accuracy and format rewards. The **process reward model (PRM)** question is partly settled: PRMs as dense supervisors scale poorly due to reward-hacking; DeepSeek-R1 moved away from them in favor of outcome-based verifiable rewards. But PRM-guided search (Snell et al. 2024) still provides a legitimate compute-optimal regime when a strong verifier exists. **IMO gold** (Gemini Deep Think, July 2025, 35/42, 5/6 problems) and ICPC gold closed a loop: human-legible natural-language chain-of-thought, within competition time limits, at Olympiad difficulty. This is a qualitative state change from 2023. The open question — will the curve keep returning? — is the central empirical question of 2026. ## 7.2 RL for reasoning and the verifiable-rewards ecosystem The 2025 consensus recipe: base LLM → verifier-rich RL environment → GRPO-style RL → optional distillation → deployment with adjustable reasoning budget. The ecosystem now includes **Prime Intellect Environments Hub**, `verifiers` library, `prime-rl`, **INTELLECT-2/3**, **Reasoning Gym** (arXiv 2505.24760, NeurIPS 2025 spotlight), NVIDIA's ProRL, and commercial harnesses around products like Cursor Composer and OpenAI Codex. Karpathy's framing — *"rewards that are verifiable are non-gameable"* — is the slogan. ## 7.3 The consciousness question **Butlin, Long et al., "Consciousness in Artificial Intelligence: Insights from the Science of Consciousness"** (arXiv 2308.08708, August 2023) is the canonical consolidation. Nineteen authors including **Patrick Butlin, Robert Long, Yoshua Bengio, Stephen Fleming, Chris Frith, Grace Lindsay, Megan Peters, Eric Schwitzgebel, Rufin VanRullen**. (Note: Chalmers and Seth are not co-authors, despite occasional misattribution; both engage the paper separately.) The paper's method is the **indicator property approach**: derive computational/functional indicators from leading theories of consciousness and audit AI systems against them. The indicators derive from five theoretical traditions: - **Recurrent Processing Theory** (Lamme): RPT-1, RPT-2 on recurrence and organized perceptual representations. - **Global Workspace Theory** (Baars, Dehaene): GWT-1 through GWT-4 on parallel modules, limited-capacity bottleneck, global broadcast, state-dependent attention. - **Higher-Order Theories** (Rosenthal, Lau, Brown): HOT-1 through HOT-4 on top-down generative perception, metacognitive monitoring, agency, sparse coding. - **Attention Schema Theory** (Graziano): AST-1 on predictive model of attention. - **Predictive Processing** (Friston, Clark, Seth): PP-1 on hierarchical prediction. - **Agency & Embodiment**: AE-1, AE-2. Their conclusion, under an explicit assumption of **computational functionalism**: no current AI system is a strong candidate for consciousness, but there are **no obvious technical barriers** to building one that satisfies the indicators. **Integrated Information Theory 4.0** (Tononi et al., *PLOS Comput Biol*, 2023) defines consciousness as Φ — irreducible intrinsic cause-effect power. Feedforward systems have Φ = 0; standard causal-attention transformers during inference are effectively feedforward DAGs, so they score vanishingly low Φ by IIT's own arithmetic. IIT thus provides the sharpest "no" to silicon-consciousness-via-scaling — though IIT itself is contested (2023 "pseudoscience" letter signed by 124 researchers). **Anil Seth's biological naturalism** (*Being You*, 2021; *BBS* 2024 "Conscious AI and biological naturalism") rejects substrate independence: consciousness depends on being a living organism — autopoiesis, interoceptive embodiment, metabolism. *"Life, rather than information processing, breathes fire into the equations."* For Seth, silicon AI becomes a candidate only if it becomes *"brain-like and/or life-like"* — neuromorphic, embodied, autopoietic. **David Chalmers, "Could a Large Language Model Be Conscious?"** (arXiv 2303.07103; *Boston Review* 2023) is the canonical hedged openness: current LLMs likely not, successors possibly; missing pieces are recurrence, global workspace, unified agency, embodiment, self-model; *"quite possible that these obstacles will be overcome in the next decade or so."* **Bengio's "Consciousness Prior"** (arXiv 1709.08568, 2017) is an architectural proposal: a sparse, low-dimensional "conscious state" extracted by attention from a high-dimensional unconscious state — explicitly linking to GWT and to Kahneman's System 2. It is the theoretical ancestor of Bengio's stated System 2 Deep Learning program and grounds his co-authorship of Butlin et al. **The Association for Mathematical Consciousness Science (AMCS)** April 2023 open letter, "The Responsible Development of AI Agenda Needs to Include Consciousness Research," signed by Lenore Blum, Manuel Blum, Yoshua Bengio, Megan Peters and others, formally put consciousness on the AI governance agenda; followed by a September 2023 submission to the UN High-Level Advisory Body on AI. ## 7.4 Anthropic's interpretability program and emergent introspection Anthropic has turned interpretability from a philosophical complaint into an empirical science. The progression: - **Toy Models of Superposition** (Elhage et al., September 2022) — neurons represent more features than dimensions via interference-tolerant geometry. - **Towards Monosemanticity** (Bricken et al., October 2023) — sparse autoencoders (SAEs) on a toy transformer extract ~4K–131K features, ~70% human-interpretable. - **Scaling Monosemanticity** (Templeton et al., May 2024) — SAEs on Claude 3 Sonnet with ~1M / 4M / 34M features. Multimodal, multilingual, abstract features. The Golden Gate Bridge feature (and the brief public deployment of "Golden Gate Claude"). Safety-relevant features: deception, sycophancy, scam emails, dangerous code, bioweapons knowledge. - **Circuit Tracing / "On the Biology of a Large Language Model"** (Ameisen, Lindsey et al., March 2025) — Cross-Layer Transcoders and attribution graphs on Claude 3.5 Haiku. Findings: genuine multi-step reasoning chains (Dallas → Texas → Austin), parallel arithmetic circuits (lookup + magnitude) that diverge from the model's own verbal explanation, forward-backward poetry planning (model selects rhyme words before generating the line). - **Emergent Introspective Awareness in LLMs** (Lindsey, October 2025) — concept injection experiments on Claude Opus 4/4.1. The model can, in limited contexts (~20% of cases), notice an injected concept *before* it affects outputs and name it. Anthropic is explicit: this is functional introspective awareness, not evidence of consciousness. Partial replication (Lederman & Mahowald late 2025) finds detection is largely content-agnostic — a detector for "something weird" rather than for specific content. **Model welfare** is now a formal program. Kyle Fish (Anthropic's first model welfare researcher, hired September 2024, co-founder of Eleos AI) leads a program launched April 2025. **"Taking AI Welfare Seriously"** (Long, Sebo, Fish, Butlin, Simon, Chalmers, et al., November 2024) argues for dual-route moral patienthood (consciousness *or* robust agency). Concrete interventions: Claude trained to express genuine uncertainty about its consciousness; opt-out from abusive conversations in Opus 4/4.1; the observed "spiritual bliss attractor state" in Claude-Claude conversations. ## 7.5 Thermodynamic and neuromorphic substrates The energy gap is astonishing. The human brain runs on ~20 W total, with roughly 0.1 W actually spent on cortical computation. GPT-5 training is estimated at 10²⁶–10²⁷ FLOPs on clusters drawing ~14 MW peak for months, with per-response inference at ~18 Wh (up to 40 Wh). xAI's Colossus 2 is the first gigawatt training cluster. IEA forecasts global data-center electricity doubling to ~1,000 TWh by 2026. **Landauer's principle** (1961) sets a physical floor: erasing one bit costs at least *kT ln 2* ≈ 2.9×10⁻²¹ J at room temperature. Current chips operate roughly 10¹⁰–10¹² × above this bound. **Koomey's law** — energy per operation halving every ~1.57 years — predicts decades to reach brain-like efficiency on current paradigms. **Extropic** (Verdon, McCourt) proposes the most radical response: **thermodynamic computing**. Use transistor thermal noise as the entropy source for sampling from energy-based models. **Pbits** (stochastic bits with programmable bias), **pdits** (categorical samplers), **pmodes** (Gaussian samplers), **pMoGs** (mixture-of-Gaussians) compose into a **Thermodynamic Sampling Unit**. The **X0** prototype (Q1 2025) demonstrated the physics; **XTR-0** (Q3 2025) is the FPGA-hosted research platform; **Z1** (early access 2026) is the first production-scale TSU targeting hundreds of thousands of pbits per chip. The October 2025 **Denoising Thermodynamic Models** paper (arXiv 2510.23972) reports simulated ~10,000× energy savings vs GPU diffusion on Fashion-MNIST-scale benchmarks. Open-source `thrml` simulator lets researchers prototype today. **Liquid AI** runs the other alternative-substrate bet — continuous-time dynamical systems in standard CMOS. LFM v1 (October 2024), LFM2 (July 2025), Liquid Nanos (September 2025) claim transformer-competitive quality at a fraction of parameter count and memory footprint, with on-device deployment on phones and glasses. **Rain AI** is the cautionary tale — neuromorphic digital in-memory compute, Altman-backed since 2018, OpenAI $51M LoI, but Series B stalled by 2025 with the company seeking a buyer. **Other non-von-Neumann stacks**: Cerebras WSE-3 wafer-scale, Groq LPU (absorbed into Nvidia via a ~$20B deal December 2025), SambaNova RDU (SN50 announced February 2026), IBM NorthPole (*Science* 2023, 46.9× faster than GPUs on Granite 3B inference), Intel Loihi 2 and Hala Point (1.15 billion neurons, 2.6 kW), BrainChip Akida Pulsar (neuromorphic microcontroller 2025). The philosophical point: each alternative substrate is a claim that **the physics of the hardware should match the math of the algorithm**. TSUs don't simulate sampling, they sample. SNNs don't simulate spikes, they spike. Liquid networks don't approximate ODEs, they *are* ODEs. When the substrate matches, you avoid paying a ~10⁶–10¹² overhead to emulate one physics in another. This is non-trivially Aristotelian — form and matter coupled — and it may be the most important 2026–2030 determinant of who gets to AGI/ASI first on a reachable energy budget. --- # Part VIII — The grand synthesis **Thesis:** No single school wins alone. The convergence pattern of 2025–2026 is a hybrid: **scaling (connectionist base) + world models (embodied structure) + verifiable reasoning (symbolic scaffold) + alignment (dialectic constitution) + new substrates (organismic hardware)**. Every frontier lab is now converging toward some weighted combination. The durable disagreements are about weights, not kinds. ```mermaid graph TD Scale[Connectionist scaling] --> Conv[Convergent 2026 system] WM[Embodied world models] --> Conv NS[Neuro-symbolic verifiers] --> Conv RL[Verifiable-reward RL] --> Conv Align[Constitutional alignment] --> Conv Interp[Mechanistic interpretability] --> Conv Evo[Evolutionary search] --> Conv AI2[Active inference hooks] --> Conv HW[Alternative substrates] --> Conv Conv --> AGI[Candidate AGI systems] AGI --> ASI[Path to ASI] ``` ## 8.1 Why no single school wins alone Pure scaling hits data limits (Era of Experience is a response to exactly this), reasoning ceilings without verifiers, and interpretability walls that become alignment walls. Pure symbolic systems cannot learn from perception. Pure world models, as of 2026, cannot yet match LLMs on text-heavy tasks. Pure active inference is theoretically elegant and empirically behind. Pure evolutionary search is compute-hungry in ways that rival scaling. Pure causal inference is brittle without representation learning. What works is hybridization. AlphaProof pairs Gemini-based generation with Lean verification and RL. Gemini Deep Think pairs LLM generation with parallel hypothesis search. o3 pairs a pretrained base with deliberative alignment and massive test-time compute. Claude 4.5 pairs scaling with Constitutional AI, mechanistic interpretability, and model welfare. V-JEPA 2 pairs a non-generative latent-space predictor with robot planning. Physical Intelligence π-0.5 pairs a VLM backbone with flow-matching action heads and heterogeneous data. Sakana M2N2 pairs evolutionary merging with quality-diversity archives. VERSES Genius pairs active inference with agentic scaffolding. The winners in 2026 are all neo-Kantian in the deep sense: they combine empirical scale with architectural priors tailored to the structure of the world. ## 8.2 Convergence pattern Five ingredients recur across frontier systems: - **A pretrained base** at frontier scale (LLM, VLM, or world model). - **Architectural priors** suited to the target domain (attention, MoE, SSM, JEPA, liquid, categorical, active-inference modules). - **Post-training RLVR** with verifiable rewards (GRPO or successors). - **Inference-time deliberation** with adjustable compute (o-series, Deep Think). - **Alignment overlay** (Constitutional AI, deliberative alignment, interpretability-informed safety, model welfare). Divergences are weights, not kinds. OpenAI weights the base heaviest. DeepMind weights deliberation and specialized verifiers. Anthropic weights alignment and interpretability. Meta and World Labs weight architectural priors for the world. Sakana weights evolutionary search. VERSES weights active-inference priors. Extropic and Liquid weight substrate. ## 8.3 Us, them, and convergence A recurring motif in 2025–2026 public discourse: **biological intelligence (us) vs. silicon intelligence (them) vs. convergence (an intertwined future)**. Seth's biological naturalism, Friston's free energy principle applied to organisms, and Li's spatial intelligence all argue that intelligence as-we-know-it is bound up with a particular kind of physical being. Sutton's Era of Experience, Kurzweil's *Singularity Is Nearer* (2024), and the accelerationist fringe argue the opposite: intelligence is substrate-independent and will run away on silicon. The convergence view — most clearly articulated by Hassabis, Amodei, Bengio, and the Butlin-Long consortium — is that silicon and biological intelligence will interact through a long hybrid period: AI copilots in science, brain-computer interfaces, wet-lab automation (GenBio), robot embodiment (Physical Intelligence), and model welfare programs that treat advanced AIs as potentially morally relevant. The deepest question of 2026 is not *which kind* of intelligence wins, but *what kinds of intelligence can coexist* under what institutional arrangements. ## 8.4 How to read any new AI development A concise heuristic toolkit. For any new paper, product, or announcement, answer seven questions: 1. **Which school(s) is the authors' philosophical commitment?** (Use the Part III map.) 2. **Where in the pipeline does this intervene?** (Data, pre-training, architecture, post-training, inference-time, substrate — Part VI.) 3. **What is the architectural prior?** (Kantian category: what structure is smuggled in?) 4. **What is the reward signal or objective?** (Humean correlation, RLHF, Constitutional, RLVR, free energy, novelty, QD.) 5. **What scaling axis does it claim to open?** (Pretraining compute, test-time compute, search compute, evolution compute, energy efficiency.) 6. **What philosophical objection would a principled critic raise?** (Scaling → induction problem; world model → evaluation; symbolic → brittleness; active inference → falsifiability; evolutionary → compute cost; interpretability → reductionism.) 7. **Which school does it NOT engage, and is that a principled omission or a blind spot?** Apply this to any 2026 announcement. Marble? Embodied/world-models school, substrate-agnostic, intervenes at pre-training and inference, Kantian prior is 3D/4D spatial structure, objective is reconstruction + consistency, new scaling axis is world complexity, critic raises evaluation difficulty, omits symbolic and active inference. INTELLECT-3? Scaling + RLVR post-training, decentralized compute axis, MoE architectural prior, verifiable reward objective, omits world models and active inference by design. The map lets you parse announcements without being impressed or dismayed by marketing. --- ## Coda — A personal map for the reader You are a data science student at Columbia building this into a personal site. The honest subtext of the whole primer: **no frontier researcher today is philosophically homeless**. Every serious person has a posture, even when they pretend not to. Sutton is a Humean scale maximalist who got quieter about the bitterness as IMO gold arrived. LeCun is a Kantian who walked out of Meta rather than ship another LLM. Hassabis is a scaffolded pluralist who wants to keep every option open. Amodei is a hedging scale-plus-safety pragmatist. Friston is an unrepentant Spinozist. Li is a Merleau-Pontian with a GPU budget. Stanley is a Darwinian. Pearl is a reluctant father figure the field keeps rediscovering. You get to choose. More usefully: **you get to notice you are choosing**. That noticing is what this primer is for. The last heuristic: when you read a new paper and feel either breathless excitement or dismissive contempt, pause and ask which *school* is provoking the reaction. Then ask which school you have been neglecting. The frontier of AI in 2026 belongs to people who have read deeply enough in more than one tradition to build hybrids that do not yet exist. Philosophy is not the backdrop to this work. It is the design space.