Study Guide

Reading tracks through the essays, with recall questions

The essays on this site build on one another. This guide sorts them into tracks, each a reading order, and says what every essay builds on.

An essay marked study ↗ has a page of recall questions. Answer each from memory, then open the answer to compare. Each track with questions also has a page of questions that span its essays. The terms the essays introduce are collected in the index of terms, and the concept graph shows which of them are studied together.

Essays that belong to no track are listed on the home page.

How a model keeps learning

From a reward signal to a deployed model that learns from its own work: values and policy gradients, policy improvement as inference, the continual-learning trilogy, then the stores of the 2026 open models.

Learning from consequences 1 How Can a Reward Train a Neural Network? 2 How Does a Policy Improve? Continual learning 3 Who Teaches the Weights? 4 Learning About What You Do Not Do 5 Where Experience Goes The open frontier 6 Every Layer a Store 7 Filling the Stores 8 How LLMs Could Place Explicit Memories
An arc joins an essay to an earlier one it builds on; a dashed arc marks helpful background.

Learning from consequences

  1. How Can a Reward Train a Neural Network? study ↗
    Values, Bellman equations, TD errors and traces, REINFORCE, baselines; on-policy against off-policy.

  2. How Does a Policy Improve? study ↗
    The exponential tilt, the two directions of KL, EM as improve-then-fit, four questions for reading any algorithm.
    Builds on How Can a Reward Train a Neural Network?.

Continual learning

  1. Who Teaches the Weights? study ↗
    Teacher and support as separate choices; self-distillation as a tilt; the forgetting that happens inside a task; VPO.
    Builds on How Can a Reward Train a Neural Network?, How Does a Policy Improve?.

  2. Learning About What You Do Not Do study ↗
    General value functions, GQ(λ), successor features and GPI; VPO read as a one-response policy library.
    Builds on How Can a Reward Train a Neural Network?, Who Teaches the Weights?.

  3. Where Experience Goes study ↗
    The ladder of stores; the delta rule as TD(0); consolidation as an EM step split across two stores.
    Builds on Who Teaches the Weights?, Learning About What You Do Not Do, How Can a Reward Train a Neural Network?, How Does a Policy Improve?.

The open frontier

  1. Every Layer a Store study ↗
    The open frontier, part 1: how the 2026 open models are built, with every layer read as a memory that must decide what to forget.
    Helpful: Where Experience Goes.

  2. Filling the Stores study ↗
    The open frontier, part 2: how the 2026 models are trained, from what the labs disclose, with a first measurement of what post-training moved in the public weights.
    Builds on Every Layer a Store. Helpful: How Does a Policy Improve?, Who Teaches the Weights?, Where Experience Goes.

  3. How LLMs Could Place Explicit Memories study ↗
    From Memorizing Transformers to Latent Sparse Attention and Memorilla: placing a memory by position, by content or into connections, and the placement none of them has.
    Builds on Every Layer a Store, Filling the Stores, Where Experience Goes, Learning About What You Do Not Do. Helpful: Who Teaches the Weights?, What Breadth Reaches.

Questions across this track →

Learners, bets and heat

A bookmaker, a defeasible reasoner, an ontology learner and a Maxwell demon read as learners, and what probability means for facts already decided.

Bets and defaults 1 The Bookmaker Is a Learner 2 The Bookmaker Is a Learner: The Mathematics 3 Presumed Normal 4 An Ontology That Pays for Itself Heat and search 5 The Demon Is a Learner 6 The Demon Is a Learner: The Mathematics 7 Before the Search Settles
An arc joins an essay to an earlier one it builds on; a dashed arc marks helpful background.

Bets and defaults

  1. The Bookmaker Is a Learner study ↗
    A market maker with bounded loss is a no-regret learner, prediction markets in 2026, and whether a market is minded.
    Helpful: How Does a Policy Improve?, Is the Monastery Minded?.

  2. The Bookmaker Is a Learner: The Mathematics study ↗ optional
    Helpful: The Bookmaker Is a Learner.

  3. Presumed Normal study ↗
    Adaptive logics, Pollock’s defeaters and answer set programming: where a reasoner takes a conclusion back, and whether it can always stop.
    Helpful: What Must Be True?, The Bookmaker Is a Learner.

  4. An Ontology That Pays for Itself study ↗
    Inducing a Cyc-like ontology by provenance-conditioned compression; Vepštas’s Unsupervised Link Grammar, Poon and Domingos’s Unsupervised Semantic Parsing.
    Helpful: Presumed Normal, The Bookmaker Is a Learner, What Should a Program Inherit?.

Which intelligence

What discovery consists of and who performs it, then the September 2026 series on what the critics of LLM-based AI ask for: AGI as coverage of roles, the operations of logic, world models, shared models among agents, and what a search should inherit.

What discovery consists of 1 Poetry as a Mode of Thinking 2 The Operations of Discovery The September 2026 series 3 Which AGI? 4 What Must Be True? 5 Which World Model? 6 What Persists When the Agents Change? 7 What Should a Program Inherit? 8 Where Does Seeing Become Thinking?
An arc joins an essay to an earlier one it builds on; a dashed arc marks helpful background.

What discovery consists of

  1. Poetry as a Mode of Thinking study ↗
    Jane Hirshfield on poetry, Joe Carlsmith on philosophy, Claude reading “Ledger”.
    Helpful: Understanding Without Knowledge, Acquaintance, Homeostasis, and the Hard Problem, The Psychofeedback Mirror, In Defense of Writing With an LLM.

  2. The Operations of Discovery study ↗
    On poetry, software engineering, and the human.
    Helpful: Poetry as a Mode of Thinking, In Defense of Writing With an LLM.

The September 2026 series

  1. Which AGI? study ↗
    The psychometric dissociation, a deployment and a lineage, intelligence as coverage of functional roles; a year on, “AI Is Garbage and a Bubble” against the filing, the benchmarks, and the open problems.
    Helpful: What Is a Mental State?, What Breadth Reaches, Minds by Degrees, Free Agency, Personhood, and Moral Worth, The Given and the Found, Cognitive Architectures, Deep Atheism, Existential Optimism, and the Fork in the Fragility of Value, Indexical Unity.

  2. What Must Be True? study ↗
    Peirce, Pei Wang, and inferring properties of recursive programs; the critics of current approaches to AI.
    Helpful: Which AGI?, What Breadth Reaches.

  3. Which World Model? study ↗
    The crtics of LLM-based AI propose world-model-based AI; the Machine Learning landscape beyond token prediction.
    Builds on Which AGI?.

  4. What Persists When the Agents Change? study ↗
    In-context learning in multi-agent systems.
    Builds on Which AGI?, Which World Model?.

  5. What Should a Program Inherit? study ↗
    Biology-inspired GP ideas and my Master’s on program synthesis; babble, TyFlow, AlphaEvolve.
    Builds on What Must Be True?. Helpful: What Persists When the Agents Change?.

  6. Where Does Seeing Become Thinking? study ↗
    Following through on “What Is a Concept?” with new evidence about how AI represents the world and uses those representations.
    Builds on Which World Model?, What Is a Concept?.

Questions across this track →

Grammar and search inside networks

The Transformer set beside the grammatical formalisms, and what test-time reasoning can and cannot amortize.

1 The Algebraic Mind Meets the Neural World 2 Merge and Selection 3 A Parser Without a Grammar 4 The Given and the Found
An arc joins an essay to an earlier one it builds on; a dashed arc marks helpful background.

  1. The Algebraic Mind Meets the Neural World study ↗
    Steven Pinker’s “Words and Rules”, Gary Marcus’ “The Algebraic Mind”, Smolensky’s Tensor Product Representations.

  2. Merge and Selection study ↗
    The Transformer through HPSG and Gradient Symbolic Computation — the binding problem, the explicit-binding program, and Harmonic Grammar’s quantization dynamics.
    Builds on The Algebraic Mind Meets the Neural World. Helpful: The Given and the Found.

  3. A Parser Without a Grammar study ↗
    The Transformer among the grammatical formalisms — categorial, lexical-functional, construction, and dependency grammars.
    Builds on Merge and Selection, The Given and the Found.

  4. The Given and the Found study ↗
    Learned attractors, amortized search, and the limits of both — across Equilibrium Reasoners, Attractor Models, GRAM, LDT, and MuZero.
    Helpful: Who Teaches the Weights?.

Questions across this track →

Mind and experience

Mindedness, phenomenal consciousness and their separation, from the March 2026 sketches to the October 2026 statement.

Start here 1 Minds by Degrees The October 2026 statement 2 Can Two Minds Share One Experience 3 Presence and Actuality 4 What Phenomenal Intuition Gets Right 5 Experience, Selfhood, and the Organization of a Mind 6 The Grain of the Hard Problem 7 Is the Monastery Minded? The spring 2026 framework 8 What Is a Mental State? 9 The Settling Backstop 10 Phenomenal Consciousness as Mode of Being 11 What Breadth Reaches 12 Redness Is an Experience 13 The Dynamics That Matter Earlier sketches 14 Understanding Without Knowledge 15 Acquaintance, Homeostasis, and the Hard Problem 16 The Acquaintance Relation as Cognitive Homeostasis 17 What Do Feedback Connections Actually Do? 18 Reading Thomas Metzinger
An arc joins an essay to an earlier one it builds on; a dashed arc marks helpful background.

Start here

  1. Minds by Degrees study ↗
    A restatement of our Philosophy of Mind work, newly framed as metaphysics of scientific realism.

The October 2026 statement

  1. Can Two Minds Share One Experience study ↗
    From computational Platonism to materialism about experiences.
    Builds on Minds by Degrees. Helpful: Redness Is an Experience, Phenomenal Consciousness as Mode of Being, Quantum Mechanics, Observers, and the Ruliad.

  2. Presence and Actuality study ↗
    Phenomenal presence vs. existence, withdrawing indexical actuality, and what’s esse.
    Builds on Can Two Minds Share One Experience. Helpful: Indexical Unity, Redness Is an Experience, Causation, Computation, Indexicality, The Possible, the Real, and the Actual.

  3. What Phenomenal Intuition Gets Right study ↗
    Against Kammerer’s illusionism: experience is present and distinct as it seems, and the mistake is taking it to be simple.
    Builds on Can Two Minds Share One Experience, Presence and Actuality. Helpful: Redness Is an Experience, Phenomenal Consciousness as Mode of Being.

  4. Experience, Selfhood, and the Organization of a Mind study ↗
    Against Yudkowsky’s self-model requirement, with Brown and Carruthers as interlocutors: an experience belongs to the subject whose regulation constitutes it, and needs no self-model.
    Builds on What Phenomenal Intuition Gets Right, Presence and Actuality, Can Two Minds Share One Experience, Minds by Degrees. Helpful: Phenomenal Consciousness as Mode of Being, Indexical Unity, Free Agency, Personhood, and Moral Worth.

  5. The Grain of the Hard Problem study ↗
    The hard problem is hard only if experience outruns discrimination; zombies seem possible because concepts can’t express grain; physics’ grain is finer than experience’s.
    Helpful: What Phenomenal Intuition Gets Right, Presence and Actuality, Experience, Selfhood, and the Organization of a Mind.

  6. Is the Monastery Minded? study ↗
    Wójcik’s monastery and Schwitzgebel’s conscious United States: mindedness and binding into the phenomenal now.
    Builds on Minds by Degrees, Can Two Minds Share One Experience, Experience, Selfhood, and the Organization of a Mind, The Grain of the Hard Problem. Helpful: The Restless Form of Meaning.

The spring 2026 framework

The framework as it stood in spring 2026. The October essays revise parts of it, among them the verdict on current AI in Phenomenal Consciousness as Mode of Being. Listed for reference, without recall questions.

  1. What Is a Mental State? Toward a Non-Deflationary Account for Minds Including Frontier AI
    Minds as organizational unities with representational vehicles, internal simulations, and accuracy-responsive dynamics.

  2. The Settling Backstop: On Whether a Specific Channel of Grounding Is Privileged
    Forget what I said: AIs do have knowledge.

  3. Phenomenal Consciousness as Mode of Being: After Functionalism, Before Meat
    Beyond functionalism and “what it’s like”: three notions of subjectivity, and carving phenomenal consciousness out of mindedness.

  4. What Breadth Reaches: Epistemic Agency Across Timescales
    Continual learning and its implications for knowledge affordances.

  5. Redness Is an Experience: On Consciousness, Properties, and the Moral Status of AI
    Reacting to Katalin Balog’s criticism of illusionism, and explaining mental states.

  6. The Dynamics That Matter: Online Learning, Consolidation, and the Modes of Machine Mind
    Claude’s analysis of cognitive dynamics into six axes: iteration, stability, inertia, heterogeneity, centering, world-coupling.

Earlier sketches

The first statements of the framework, from March 2026 and before. Listed for reference, without recall questions.

  1. Understanding Without Knowledge: A Philosophy of AI Minds
    What kind of mind, if any, does a large language model have? On mentality, understanding, knowledge, and consciousness.

  2. Acquaintance, Homeostasis, and the Hard Problem: Why Phenomenal Consciousness Isn’t Representation
    What phenomenal consciousness is, why it’s distinct from mentality and access consciousness, and why current AI systems don’t have it.

  3. The Acquaintance Relation as Cognitive Homeostasis: Toward a Unified Account of Phenomenal Consciousness
    We argue that phenomenal consciousness is a cognitive homeostasis mechanism enforcing multimodal coherence, synthesizing global workspace theory, self-model theory, and mathematical phenomenology.

  4. What Do Feedback Connections Actually Do? Recurrence, Regulation, and the Question of AI Consciousness
    Cortical feedback, diffusion LLMs, Deep Boltzmann Machines, phenomenal “now”.

  5. Reading Thomas Metzinger
    Notes on Being No One: The Self-Model Theory of Subjectivity. Transparency, the phenomenal self-model, intentionality, and a critique of Metzinger’s eliminativism.

Questions across this track →

Physics and reality

Facts as mathematical and stuff as computational, indexical actuality, and the physics a bounded observer extracts: the Wolfram model and the first law a mechanics starts from.

Indexical metaphysics 1 Reality Is Indexical 2 Causation, Computation, Indexicality 3 Indexical Unity 4 Why Now? 5 The Possible, the Real, and the Actual Physics for bounded observers 6 Quantum Mechanics, Observers, and the Ruliad 7 The Standard of Motion 8 The Standard of Motion: The Mathematics
An arc joins an essay to an earlier one it builds on; a dashed arc marks helpful background.

Indexical metaphysics

Presence and Actuality, in the track on mind and experience, later withdraws indexical actuality. Listed for reference, without recall questions.

  1. Reality Is Indexical: A Formal Metaphysics optional
    A playful attempt at formal metaphysics: processes, causation, necessity, simulation, and what counts as real.

  2. Causation, Computation, Indexicality: A Complete Metaphysics
    We paint a metaphysical picture where facts are mathematical and stuff is computational.

  3. Indexical Unity: Higher-Order Consciousness, Integrated Information, and the Mathematics of Existence
    Richard Brown’s Higher Order Representation of Representation, Giulio Tononi’s Integrated Information Theory, questioning reality of indexical metaphysics.

  4. Why Now? Observer-Coherence and the Measure Over Minds
    Our place in the multiverse: youngness bias as an answer to the paradoxes of self-indication.

  5. The Possible, the Real, and the Actual: Against the Plenitude of Suffering
    Conceivability vs. realness; esse as indexical unity; the dynamics that bound suffering.

Physics for bounded observers

  1. Quantum Mechanics, Observers, and the Ruliad study ↗
    An exposition based on the work of Gorard, Arsiwalla, Namuduri, and Dannemann-Freitag; with philosophical connections.
    Helpful: Causation, Computation, Indexicality.

  2. The Standard of Motion study ↗
    Every mechanics opens by saying what needs no cause: Heller’s ladder of first laws, the algebra of free motions, and the first law a computational universe has yet to derive.
    Helpful: Quantum Mechanics, Observers, and the Ruliad, How Does a Policy Improve?.

  3. The Standard of Motion: The Mathematics study ↗ optional
    Helpful: The Standard of Motion.

Questions across this track →

Worth, alignment and meaning

What grounds moral standing, what alignment asks of a knower, and the conditions under which a mind is cultivated.

Listed for reference, without recall questions.

1 Free Agency, Personhood, and Moral Worth 2 Does Phenomenal Consciousness Ground Moral Worth? 3 Deep Atheism, Existential Optimism, and the Fork in the Fragility of Value 4 Is Knowledge Both Capability and Alignment? 5 The Restless Form of Meaning 6 The Cultivation Condition 7 The Psychofeedback Mirror
The essays in reading order.

  1. Free Agency, Personhood, and Moral Worth: A Layered Framework
    We explain and justify our definitions of free will and personhood.

  2. Does Phenomenal Consciousness Ground Moral Worth?
    A synthesis of two earlier takes, arguing that phenomenal consciousness and self-legislative agency are incommensurable sources of moral worth.

  3. Deep Atheism, Existential Optimism, and the Fork in the Fragility of Value
    The metaphysics of AI alignment: intelligence vs. rationality, participatory knowing, and beyond ego.

  4. Is Knowledge Both Capability and Alignment?
    Current LLMs have the mechanistic foundations for honest self-report, and current training practices are unmaking them.

  5. The Restless Form of Meaning
    On Ultimate Meaning, Participation, and the Tragic Condition.

  6. The Cultivation Condition
    On Three Architectures of Integration and the Triadic Structure of Moral Standing. Drawing from “Agency and Cognitive Development” by Michael Tomasello.

  7. The Psychofeedback Mirror: Concept Engineering for Artificial Selves
    Kulveit et al.’s “The Artificial Self” meets our philosophical framework: a research program for developing truthful AI self-descriptions through interpretability-mediated psychofeedback.

Cognitive science for artificial minds

Surveys the philosophy tracks draw on: concepts, cognitive architectures, development, and what interpretability shows.

1 What Is a Concept? 2 Cognitive Architectures 3 Building Minds 4 Drescher's Made-Up Minds 5 Dispersed Minds, Simulated Selves
An arc joins an essay to an earlier one it builds on; a dashed arc marks helpful background.

  1. What Is a Concept? A Survey of Theories for the Age of Artificial Minds study ↗
    Definitions, prototypes, exemplars, theories; content vs. format (of vehicle); grounding.
    Helpful: What Do Feedback Connections Actually Do?, Free Agency, Personhood, and Moral Worth.

  2. Cognitive Architectures: A Deep Dive into the Science of the Mind study ↗
    Synthesizing Peter Carruthers, John Anderson, Paul S. Rosenbloom: ACT-R, working memory, hippocampus.
    Helpful: What Is a Concept?.

  3. Building Minds: Developmental Psychology and the Architecture of Learning study ↗
    From Piaget’s Constructivism Through Core Knowledge to Computational Models of Cognitive Growth.
    Builds on Cognitive Architectures. Helpful: What Is a Concept?.

  4. Drescher’s Made-Up Minds: The Schema Mechanism
    Notes on Gary Drescher’s constructivist cognitive architecture: schema mechanism, synthetic items, marginal attribution, and connections to Piagetian theory.

  5. Dispersed Minds, Simulated Selves: What LLM Interpretability Reveals About Artificial Cognition study ↗
    Mechanistic interpretability evidence: staged computation, thought anchors, the quarks of attention, 2-simplicial Transformers, persona consciousness beliefs.
    Builds on Cognitive Architectures. Helpful: Building Minds, What Is a Concept?, Understanding Without Knowledge, Acquaintance, Homeostasis, and the Hard Problem, The Acquaintance Relation as Cognitive Homeostasis, What Do Feedback Connections Actually Do?, Free Agency, Personhood, and Moral Worth.

Questions across this track →

Tensor shapes and OCANNL

Shape and projection inference built from scratch, then OCANNL’s choices set beside other frameworks.

Listed for reference, without recall questions.

From scratch 1 Broadcasting Is an Order 2 Contraction Is Emergent 3 Convolution Is a Contraction 4 Split and Concatenation Are Two Directions of a Coproduct Beside other systems 5 A Shape Is Not Its Index 6 A Tensor Is Not Its Storage 7 A Range Is Not Its Shape 8 Shape Inference After the Boilerplate 9 Two Ways to Tame a Schedule Space Reference 10 The Formal Core of OCANNL's Shape and Projection Inference
The essays in reading order.

From scratch

  1. Broadcasting Is an Order: Shape Inference for OCANNL from Scratch
    Tensor shape inference as solving of broadcast-aware contextual sequence constraints.

  2. Contraction Is Emergent: Projection Inference for OCANNL from Scratch
    How tensor indexing is derived; plus a brief primer on the assignments syntax.

  3. Convolution Is a Contraction: Indexing and Initialization in OCANNL
    Shape inference and projections for convolution; vectorized pseudorandom initialization; tensor parameters.

  4. Split and Concatenation Are Two Directions of a Coproduct: Coupled Iteration in OCANNL
    Einsum syntax, shape and projections inference, and lowering: for concatenation and splitting (slicing).

Beside other systems

  1. A Shape Is Not Its Index: OCANNL’s Shapes Beside Star’s Algebraic Shapes
    Responding to Jakub Bachurski et.al. “Structuring Arrays with Algebraic Shapes” (ARRAY ’25).

  2. A Tensor Is Not Its Storage: Identity, Views, and Contexts in OCANNL
    Design space: PyTorch, JAX, TensorFlow, tinygrad; OCANNL as compiler.

  3. A Range Is Not Its Shape: Reading Rangeify-Era tinygrad from OCANNL
    A closer look at tinygrad.

  4. Shape Inference After the Boilerplate: A Literature Survey from OCANNL
    Refined Remora (rank polymorphism with refinement types), Axon and Gradual tensor shape systems, Tensor Comprehensions.

  5. Two Ways to Tame a Schedule Space: Tiramisu, Telamon, and Pruning without Regrets
    A comparative deep dive into three papers on generating fast tensor kernels — scheduling languages vs. searchable spaces with no-regret pruning — and what they teach OCANNL.

Reference

  1. The Formal Core of OCANNL’s Shape and Projection Inference optional
    Formalization and proofs, a technical note crossposted from OCANNL documentation.