Does AI Actually Think — Why Architecture Matters More Than Data

Section 1

A Thought Experiment

Hum a song. Three notes — maybe four. Your brain recognises it in under a second. Not "after searching through every song you have ever heard." Not "after hashing every comparable substring." It just knows. The melody settles into something, a pattern that feels right before you can explain why.

Now consider what a brute-force algorithm would need to do. Even with hashing, even with optimised substring matching, searching through tens of thousands of songs for a partial melodic match would take considerable compute and wall-clock time. Your brain does it in milliseconds, using roughly 20 watts — less than a lightbulb.

This is not sequential token generation. Your brain is not committing to one note at a time, left to right, hoping the sequence matches. It is doing something fundamentally different: settling into a low-energy attractor state across an entire landscape of musical memory. The melody does not match because you compared it. It matches because the energy surface of your neural representation has a basin right there, and your perception fell into it.

Your brain recognises a song by falling into an energy minimum — not by running a search algorithm. That distinction is the key to understanding why current AI cannot actually reason.

Section 2

The Illusion of Thinking

In June 2025, Apple published a paper that rattled the AI field. The title was blunt: The Illusion of Thinking. The researchers tested frontier Large Reasoning Models — including OpenAI's o3-mini, DeepSeek-R1, and Claude 3.7 Sonnet Thinking — on controllable puzzles like Tower of Hanoi, River Crossing, and Blocks World. They systematically increased complexity and watched what happened.

3
Distinct performance regimes discovered
0%
Accuracy of all models at high complexity
Reasoning effort decreases as problems get harder
The three regimes of LRM performance (Apple, 2025)
Regime
What happens
Low complexity
Standard LLMs outperform LRMs
The overhead of chain-of-thought reasoning is wasted. Simple pattern matching is faster and more accurate.
Medium complexity
LRMs gain an advantage
CoT helps sustain structured reasoning. This is the narrow band where "thinking" models justify their cost.
High complexity
Complete collapse — both types
Accuracy drops to zero. Not gradual — a cliff. Models reduce reasoning effort despite ample token budget. They give up.
Shojaee et al., "The Illusion of Thinking," Apple ML Research, arXiv:2506.06941, June 2025
The structural claim

LRMs fail to implement explicit algorithms even when the solution procedure is well-known. They do not apply Tower of Hanoi's recursive strategy. They pattern-match on descriptions of reasoning, not reasoning itself. This is not a training data problem — it is architectural.

The paper has critics — Lawsen (2025) argued some puzzles exceeded token limits, Dellibarda Varela et al. noted evaluation issues. Valid methodological points. But the core finding — autoregressive models hit a complexity cliff on constraint satisfaction — aligns independently with what Logical Intelligence demonstrated: Kona (an EBM) solves 96.2% of hard Sudoku puzzles in 313ms; frontier LLMs together solve 2%.

96.2%
Kona EBM solve rate on hard Sudoku
2%
Frontier LLM solve rate (combined)
313ms
Kona average latency vs 30+ sec for LLMs

Section 3

Why Autoregressive = Greedy — The Formal Framework

An autoregressive model generates output one token at a time, left to right, each conditioned on everything before it. Once emitted, a token is committed — irrevocable. There is no mechanism to revise token 47 when token 200 reveals a contradiction. This is, in the precise algorithmic sense, a greedy strategy.

To understand why this matters — and why it is not just a metaphor — we need two definitions from algorithm theory that separate greedy approaches from dynamic programming.

The two properties: one shared, one decisive
Optimal Substructure
Greedy Choice Property
DefinitionAn optimal solution contains optimal solutions to its subproblems.
DefinitionA locally optimal choice is part of some globally optimal solution — without solving subproblems first and without ever revisiting.
Required by
Both greedy & DP
Required by
Greedy only
Cormen, Leiserson, Rivest & Stein, Introduction to Algorithms (CLRS), 4th ed., ch. 15–16

When the greedy choice property holds, you can commit at each step without regret. When it fails, you must defer — carry multiple partial solutions and let the endpoint decide. That deferral, organised efficiently, is dynamic programming. And the greedy choice property fails whenever a consumable resource has uncertain future value.

Superpowers and their invisible costs

Consider a concrete example. You walk across a grid of coins, some negative, and have two "neutralisations" — the ability to ignore a negative cell. Greedy says: neutralise the worst negative you have seen so far. But the cost of a neutralisation is not what you pay when you use it. It is what you cannot neutralise later because you already spent it.

The cost of a superpower is not what you pay when you use it. It is what you cannot do later because you used it now. And that cost is invisible at decision time.

This is exactly the structural problem with autoregressive generation. Each token is a commitment — a "neutralisation" spent. If token 47 commits to placing a 5 in Sudoku cell (3,2), and token 200 discovers this creates an irrecoverable constraint violation, the model has spent its superpower (the token) on a choice whose cost was invisible at decision time.

State definition is the whole game

The correct DP table for the coin grid is not dp[r][c]. It is dp[r][c][k], where k tracks how many neutralisations have been used. The question that defines a state variable: "What would I need to know right now to make perfect decisions going forward, without seeing the future?" Whatever makes two situations' futures different despite their presents looking the same — that is a state variable.

The collapse model — the core distinction

Greedy asks: "What is the best choice, period?" → one survivor per step. Collapses across everything immediately.

DP asks: "What is the best choice per state?" → one survivor per state per step. Collapses within each state, preserves across states until the endpoint.

DP is greedy within each state but refuses to be greedy across states. The commitment happens late, not early. That refusal is the entire difference.


Section 4

The Sudoku Proof — Greedy Collapse in Action

Autoregressive (greedy) vs Energy-Based (holistic) on Sudoku
LLM on Sudoku (autoregressive)
Fills cell (1,1) → commits
Fills cell (1,2) → commits
… 40 cells later, discovers row conflict
Cannot revise cell (1,1) — already emitted
Generates verbose reasoning trace → wrong answer
Sequential commitment. No revision. 2% solve rate.
EBM on Sudoku (energy-based)
Proposes complete 9×9 grid simultaneously
Evaluates all 27 constraints at once
Computes energy score — high means violations
Follows gradient to revise violating cells
Converges to valid configuration
Holistic evaluation. Revision built in. 96.2% solve rate.
Logical Intelligence, Kona 1.0 Sudoku Benchmark, Feb 2026

The Sudoku result is not about Sudoku. It is a demonstration of an architectural gap. Constraint satisfaction problems — where every variable depends on every other — require the ability to say "this complete configuration has energy X, and if I adjust these three cells the energy drops to Y." That is not something a left-to-right token generator can do, no matter how many tokens it spends.

Chain-of-thought does not solve this. CoT is a greedy strategy's attempt to simulate deliberation — generating more tokens that describe reasoning steps. But those tokens are still committed left-to-right, and the "reasoning" they encode is still pattern matching, not constraint propagation. The Apple paper showed this directly: LRMs do not apply known algorithms. They approximate them verbally, and the approximation breaks at scale.


Section 5

Energy-Based Models — The Deep Dive

An Energy-Based Model defines a scalar energy function over the space of all possible configurations, where low energy = valid/desirable and high energy = violations/errors. This is fundamentally different from autoregressive generation.

The energy function

Given an input x (a partial Sudoku grid) and a candidate output y (a completed grid), the energy function Eθ(x, y) assigns a real-valued score. Correct configurations → low energy. Incorrect → high.

The Boltzmann Distribution
pθ(y | x) = exp(−Eθ(x, y)) / Z(x)
where Z(x) = ∫ exp(−Eθ(x, y′)) dy′ is the partition function. Low energy → high probability. Z is intractable to compute exactly — the fundamental computational challenge of EBMs.

This is borrowed directly from statistical physics. The energy function is analogous to the Hamiltonian. The Boltzmann distribution describes how physical systems settle into low-energy configurations — atoms in a crystal, molecules in a protein fold, spins in a magnetic material. EBMs use the same mathematics.

The energy landscape

Think of the energy function as a surface over all possible configurations. Each point is a complete candidate solution. Height = energy = constraint violations. Valleys = valid solutions. Peaks = heavily violated states. Solving = finding the deepest valley.

The energy landscape for constraint satisfaction
Configuration space → Energy → random start gradient descent local minimum (trap) global minimum (valid solution) another valid solution
Start at random configuration, follow energy gradient downhill, settle into a valley (valid solution)

Langevin dynamics — escaping local minima

Pure gradient descent gets stuck in local minima. The solution, from statistical physics: Langevin dynamics — gradient descent with added noise.

Langevin Dynamics Update Rule
yt+1 = yt − (ε/2) · ∇y Eθ(x, yt) + √ε · ωt
ε = step size, ∇yEθ = energy gradient w.r.t. configuration, ωt ~ N(0, I) = Gaussian noise. First term pulls downhill; second term adds random perturbation to escape traps. Provably samples from the Boltzmann distribution as t → ∞.

The noise is not a hack. It is thermal fluctuation — shaking a ball on a landscape so it escapes shallow valleys and finds deeper ones. In physics: thermal energy. In optimisation: simulated annealing. In EBMs: the mechanism that enables exploration without exhaustive enumeration.

Training — contrastive divergence

Contrastive Divergence Gradient (Hinton, 2002)
θ L ≈ Epdata[∇θ Eθ(x, y)] − Epmodel[∇θ Eθ(x, y)]
First term: lower the energy of real data. Second term: raise the energy of model samples. The model learns to carve valleys around valid configurations and build ridges around invalid ones.

Kona was trained only on partial solutions (50% masked) and still learned the full constraint structure. It was never shown a solved puzzle. It learned what makes a valid Sudoku valid — the shape of the energy landscape — and navigates to solutions in novel configurations. This is closer to understanding than memorisation.


Section 6

JEPA — World Models, Not Word Models

Energy-based models solve constraint satisfaction. But where do the constraints come from? How does an intelligent system know that a bottle on a table moves when the table moves — without being told?

This is the question Yann LeCun's Joint Embedding Predictive Architecture (JEPA) addresses. Published in 2022 as "A Path Towards Autonomous Machine Intelligence," it argues that current AI lacks a world model: an internal representation of how things relate and what happens when you act.

JEPA / World Models
Yann LeCun
Chief AI Scientist, Meta · Turing Award 2018
Proposed JEPA in 2022. Argues LLMs are a dead end for reasoning — they model language, not the world. JEPA predicts in representation space, not pixel or token space, learning abstract relationships between states.
Energy-Based Models
Geoffrey Hinton
University of Toronto · Turing Award 2018
Pioneered Boltzmann machines and contrastive divergence — the training algorithms underpinning modern EBMs. His energy-based learning work (1980s–2000s) laid the mathematical foundation that JEPA and Kona build on.
Dynamic Programming
Richard Bellman
RAND Corporation, 1953 · Princeton Ph.D.
Invented dynamic programming and coined the curse of dimensionality — the exponential explosion of state space that remains the fundamental limitation of exact DP, and the reason energy-based approximation is necessary.
JEPA's key insight

JEPA predicts in representation space, not pixel or token space. Instead of "what will the next pixel look like?" JEPA asks "given the current abstract state and an action, what abstract state results?" Prediction at the level of meaning, not surface form.

The architecture has six modules: perception (encode the world), world model (predict state transitions), cost (evaluate outcomes), memory, actor, configurator. The world model — the JEPA — is the core. And it is explicitly framed as an energy-based architecture.

The connection to EBMs is direct. The world model is trained to minimise prediction error in representation space — equivalent to shaping an energy landscape where accurate predictions correspond to low energy. JEPA does not generate predictions; it evaluates them.

JEPA does not generate predictions. It evaluates them. That is the difference between modelling words and modelling the world.

Meta has since published I-JEPA (images, 2023) and V-JEPA (video, 2024), demonstrating that the architecture learns semantic representations without hand-crafted augmentation — concrete steps toward the world-modelling vision.


Section 7

Memorisation vs Understanding — From Overfitting to Benchmark Collapse

The oldest problem in AI is distinguishing genuine understanding from sophisticated memorisation. This is not new. But the problem has mutated through three distinct eras, each with escalating stakes.

The three eras of the memorisation problem
Era The problem What was "memorised" How we tested for it
Classical ML (1990s–2010s) Overfitting Specific training examples Held-out test sets, cross-validation, regularisation
Deep learning (2012–2020) Benchmark gaming Test-set distributions New benchmarks, adversarial evaluation, dynamic test sets
LLM era (2020–now) Data contamination at scale Entire benchmark datasets leaked into training corpora No reliable solution yet
Each era's "solution" to memorisation became the next era's vulnerability

The classical problem — overfitting

In classical machine learning, the fear was simple: your model memorises the training set instead of learning generalisable patterns. A decision tree that perfectly fits 1,000 data points but fails on point 1,001 has memorised, not learned. The solution — held-out test sets, cross-validation, early stopping, regularisation — worked well because the training data was small and controlled.

The deep learning mutation — benchmark gaming

Deep learning changed the scale. Models trained on millions of examples could not memorise individual data points in the classical sense. But a subtler problem emerged: benchmark gaming. Research teams optimised for specific benchmark metrics — not because they wanted to cheat, but because benchmarks were how papers got published and models got funded. The result was models that performed brilliantly on ImageNet but failed on slightly different image distributions. The test set was "clean," but the entire research ecosystem was overfitting to the benchmark's statistical properties.

The LLM mutation — data contamination

LLMs broke the testing paradigm entirely. When your training corpus is "most of the internet," the separation between training and test data collapses. MMLU questions, GSM8K math problems, HumanEval coding challenges — all of these exist on the internet. All of them can (and do) leak into training corpora.

The contamination problem is structural, not accidental

Research by Chen et al. (2025), Deng et al. (2024), and Xu et al. (2024) has shown that test-set leakage in benchmarks like MMLU and GSM8K inflates reported performance by inducing memorisation rather than genuine generalisation. When training data is accessible, overlap checks can reveal contamination — but for most models, such verification is infeasible.

The result: benchmark scores increasingly reflect what the model has seen, not what it can do. A model scoring 90% on MMLU may have encountered those exact questions during training. We cannot tell. The benchmark has been colonised by the training data.

MMLU
Widely used benchmark with confirmed contamination in multiple models
GSM8K
Math benchmark with documented test-set leakage into training corpora
LiveBench
Created in 2024 specifically to address contamination — questions updated monthly
The real-world gap: benchmark performance vs actual capability
What benchmarks measure (increasingly)
Pattern recognition on seen-like distributions
Memorised solutions to leaked test problems
Performance on static, gameable tasks
How well the model fits the test distribution
Fluency. Plausibility. Reproduction.
What real-world tasks require
Generalisation to genuinely novel configurations
Constraint satisfaction across unseen problem instances
Robustness under distribution shift
Self-correction when outputs violate constraints
Understanding. Constraint awareness. Revision.
The gap between these columns is the gap between memorisation and reasoning

The Sudoku benchmark cuts through this entirely. There is no test-set leakage because every puzzle is procedurally generated. The model either satisfies all 27 constraints or it does not — there is no partial credit for plausible-looking answers. And the result is unambiguous: 96.2% vs 2%. When you remove the ability to memorise, the architectural gap becomes visible.

The Apple paper's key evidence

The most damning finding in "The Illusion of Thinking" is not the accuracy collapse — it is that LRMs fail to implement explicit algorithms. Tower of Hanoi has a known recursive solution. The models do not use it. They approximate it verbally, generating plausible-looking steps that diverge from the actual algorithm at scale. This is pattern matching on descriptions of reasoning, not reasoning itself.

The controllable puzzle environments in the Apple study were specifically designed to avoid data contamination — complexity could be tuned to guarantee novel instances. When contamination is structurally impossible, the "reasoning" collapses.

When you remove the ability to memorise — procedurally generated puzzles, controllable complexity, novel instances guaranteed — the architectural gap between pattern matching and genuine reasoning becomes impossible to ignore.

EBMs as a structural test for understanding

EBMs offer something benchmarks cannot: a structural guarantee that performance reflects learned constraints, not memorised solutions. If a model trained on 50%-masked partial solutions can solve novel complete puzzles it has never seen, it has learned the constraint structure — the shape of the energy landscape — not specific configurations. Kona demonstrates this. The model was never shown a solved puzzle. It learned what makes a valid Sudoku valid and uses that knowledge on novel instances. This is the difference between a student who memorised the answer key and one who understands the subject.


Section 8

The Coming Split — and What the Transformer Actually Is

The transformer — backbone of every major LLM — was published in 2017. Since then, models have scaled from millions to trillions of parameters. But the fundamental architecture has not changed. To understand why a split is coming, you need to understand what the transformer actually is — not as a "language model," but as a computer.

The transformer as a general-purpose differentiable computer

Andrej Karpathy put it most clearly: the transformer is a general-purpose differentiable computer that is simultaneously expressive (in the forward pass), optimisable (via backpropagation), and efficient (high parallelism compute graph). This is not metaphor. The transformer's internal structure maps directly onto the fundamental abstractions of operating systems and computer networks.

AK
Andrej Karpathy
Former Director of AI at Tesla · Founding member of OpenAI · Stanford CS25, 2023
"To me, attention is the communication phase of the Transformer. The Transformer interleaves two phases: the communication phase, which is multi-headed attention, and the computation phase, which is the multi-layer perceptron. Communication → computation → communication → computation."
The transformer mapped to OS and computer network fundamentals
Transformer component OS analogue Network analogue
Residual stream Shared memory bus — layers read from and write to a common data path Network backbone — all traffic flows through a shared channel
Attention mechanism Memory management / page table — decides which information to bring into working memory Routing protocol — queries are requests, keys are addresses, values are payloads
Multi-head attention Multi-threaded access — multiple simultaneous reads from shared memory Multiplexed channels — parallel communication streams over the same bus
MLP layers ALU / processing units — perform computation on data in registers Node processing — local computation at each network hop
Layer normalisation Voltage regulation — keeping signals within operational range Error correction — maintaining signal integrity across hops
Residual connections Skip connections / bypass — information can route around processing stages Message forwarding — data can skip intermediate nodes
Context window RAM — fixed working memory available during processing Buffer size — how much data the node can hold during a session
Weights (parameters) ROM / firmware — persistent knowledge stored in the hardware Routing tables — pre-learned rules for how to process and direct information
KV cache Cache memory — recently computed values stored for fast re-access CDN / edge cache — precomputed results stored closer to the consumer
Karpathy, Stanford CS25 (2023) · Elhage et al., "A Mathematical Framework for Transformer Circuits" (2021) · Vaswani et al. (2017)

The architecture's elegance is that it interleaves communication (attention — data-dependent message passing between tokens) and computation (MLP — local processing at each position). This is exactly how a distributed computing system works: nodes receive messages from the network, process them locally, and broadcast updated state. The transformer is a distributed computing system where each token is a node, the residual stream is the network, and attention is the routing protocol.

Why this matters for the argument

The transformer is a brilliant general-purpose computer. But it is a sequential-output computer. It processes all tokens in parallel during the forward pass — but it generates tokens one at a time, autoregressively, left to right. The parallel processing happens inside each generation step; the output itself is greedy. This is like having a supercomputer that can only write one character at a time to its output buffer, with no ability to revise what it already wrote.

The constraint is not in the processing — it is in the output protocol. The transformer can think in parallel but must commit sequentially. And that sequential commitment is the greedy bottleneck.

The architecture has not changed

Since 2017, the research community has tried extensively to improve on the transformer. Longer context windows, more efficient attention patterns, mixture-of-experts routing, state-space models. But the core architecture — residual stream, multi-head attention, MLP, layer norm, autoregressive generation — is the same. Karpathy noted in his Stanford CS25 lecture that attempts to remove or modify individual components almost always make things worse. The architecture is a local optimum that resists perturbation.

The AI community has been here before
Era The split What happened
1980s Symbolic AI vs Connectionist Connectionists won — neural nets could learn from data.
2012–2017 Hand-engineered features vs End-to-end learning End-to-end won — deep learning replaced feature engineering.
↑ In both cases, the side that changed the architecture won. ↓ We are here again.
2024–? Data scaling vs Architecture research Scale more data through the same transformer? Or change what the model is?
The pattern: the side that rethinks the architecture wins.

DeepSeek proved scaling is not the only path. But MoE is still autoregressive — it changes the routing, not the output protocol. The deeper question is whether the autoregressive output constraint itself is the limiting factor — and whether energy-based architectures, which can evaluate and revise complete configurations, represent the next architectural shift. The transformer is a magnificent general-purpose computer. The question is whether general-purpose is what reasoning requires — or whether reasoning requires an architecture designed specifically for constraint satisfaction and world modelling.


Section 9

The Encoding Function — The Real Bottleneck

The transformer optimised compute — parallelism, attention routing, residual connections. But it never rethought how information is represented. Token embeddings live in flat Euclidean vector spaces. And flat spaces have a fundamental limitation: they run out of room.

The superposition problem (MIT, 2025)

Liu, Liu, and Gore at MIT showed that when a model has m dimensions but needs to represent n >> m features, it uses superposition — cramming more features than dimensions by allowing representation vectors to overlap. A space with 6 dimensions trained on 30 features does not reject the 24 it cannot fit. It stores all 30 as overlapping probability distributions in the same 6 slots, creating interference noise that scales as 1/m.

This is not a failure mode. It is the only option in flat space. When you pack n vectors into m dimensions where n >> m, the vectors must overlap. The information is stored but degraded — like a conversation where 30 people talk simultaneously in a room with 6 chairs.

Energy landscapes are not flat. They are curved, high-dimensional surfaces where the geometry itself encodes constraints. A valley in the landscape is not a point in a vector space — it is a basin of attraction whose shape, depth, and connectivity to neighbouring valleys all carry information. The encoding is not a position in a matrix. It is a topology.

Protein folding — Nature's proof that topology beats brute force

A 300-residue protein has approximately 10143 possible configurations. If it sampled one per picosecond, finding the correct fold by brute force would take longer than the age of the universe. This is Levinthal's paradox. Yet proteins fold in milliseconds.

Why? Because the energy landscape of a protein is not flat. It has funnel topology — a shape where almost every path leads downhill toward the native state. The landscape is sculpted by billions of years of evolution so that random thermal fluctuations (Brownian motion + gradient) naturally converge to the correct fold. The protein does not calculate all possible configurations. It does not apply an algorithm. It falls.

Two encoding strategies for constraint satisfaction
Flat vector encoding (LLMs)
Information stored as positions in Euclidean space
More features than dimensions → superposition / interference
Navigation: autoregressive token generation (greedy)
Constraints encoded implicitly in weight matrices
The map is flat. The territory is not.
Topological encoding (energy landscapes)
Information stored as landscape geometry — valleys, ridges, funnels
More features → richer topology, not interference noise
Navigation: gradient descent + stochastic exploration
Constraints encoded in the shape of the surface itself
The map IS the territory.
Liu, Liu & Gore, "Superposition Yields Robust Neural Scaling," MIT (2025) · Levinthal's paradox (1969) · Bryngelson & Wolynes, funnel landscape theory (1995)

The Big Bang analogy is instructive here. Building the energy landscape is expensive — enormous energy concentrated into an initial event. But once the landscape exists, navigation is cheap. Particles do not compute trajectories. They follow the principle of least action and fall. Training an EBM is like creating the initial conditions of a universe: expensive, one-time, and front-loaded. Inference is like physics happening afterward: cheap, natural, and guided by the topology itself.

The transformer's elegant communicate/compute pattern — attention followed by MLP, repeated — is present in energy landscapes too. But here, both are unified. The structure of the landscape is the computation. The ball rolling is the communication. Communication and computation are not alternating blocks — they are the same thing viewed from different angles. And the deeper they are intertwined, the richer the expressiveness — just as in the brain, where no neuroscientist can draw a clean line between "processing" and "signalling." The chemistry between them is what generates the magic.

The curse of dimensionality — why flat spaces hit walls

DP is strictly more powerful than greedy. But it pays for that power with state space, and state space grows exponentially with state variables. Bellman named this the curse of dimensionality in 1957.

State-space explosion as dimensions increase
Assuming 100 values per dimension. Each added dimension multiplies total states by 100.
1 dimension
100 states
2 dimensions
10,000 states
3 dimensions
1,000,000 states
5 dimensions
10 billion states
10 dimensions
10²⁰ — intractable
Bellman, Dynamic Programming, Princeton UP (1957)
"It is the curse of dimensionality, a malediction that has plagued the scientist from the earliest days."— Richard Bellman, Dynamic Programming, Princeton University Press, 1957

This is exactly why EBMs matter. They do not enumerate all states (intractable). They navigate a continuous energy landscape using gradient information — approximating exact DP through physics-inspired dynamics. Langevin sampling does not visit every configuration; it flows toward low-energy regions using local gradient information plus noise for exploration. This is the practical resolution of Bellman's curse.


The encoding function matters more than the navigation algorithm. Physics-inspired encoding — energy landscapes with learned topology — is structurally superior to flat vector encoding. Not because of randomness, but because the geometry itself can embody constraints that flat spaces cannot.

Section 10

Randomness as Exploration, Structure as Intelligence

A natural objection: if energy landscapes are so powerful, why not let randomness guide everything? Give the system stochastic dynamics — Brownian motion, Langevin noise, Boltzmann sampling — and let it wander until it finds solutions. Like a child: give it the basic resources it needs and let randomness and environment guide the path.

This is partially right and importantly incomplete.

Brownian motion is genuinely stochastic — a particle buffeted by random thermal fluctuations. But Brownian motion in an open field goes nowhere useful. Brownian motion in a funnel-shaped landscape finds the minimum. The randomness provides the exploration. The structure provides the direction. Without structure, you have noise. Without noise, you have local minima. You need both — but they play different roles.

The Feynman test — what the burnt child actually teaches us

Feynman said the only way to learn not to get burnt is to let a child get burnt. Both the lit candle and the child wandering toward it are driven by randomness. But the learning is not random. The child's nervous system has structure — pain receptors, memory consolidation, association circuits — that converts a random experience into a permanent constraint. Remove the structure (damage the memory system) and the child burns themselves again tomorrow.

The randomness provides the experience. The structure provides the learning. An energy landscape with randomness but no well-shaped topology is just noise. An energy landscape with topology but no randomness gets stuck. Intelligence is structure. Randomness is the exploration budget that lets you discover the structure's basins.

This resolves the question of whether energy landscapes work only for constraint satisfaction or for open-ended problems too. Boltzmann machines — a direct descendant of Hopfield networks — introduce stochasticity into the energy landscape, enabling sampling from multi-modal distributions. This is not rule-following. It is the same mechanism that SGD uses to escape local minima: randomness as exploration within a learned structure. Creativity, in this framing, is not the absence of constraints — it is navigation of a rich, multi-modal landscape where many valleys are valid and the noise lets you visit different ones.

Symmetry breaking in physics illustrates this at the deepest level. The asymmetry between matter and antimatter — which made our existence possible — was not designed. The environmental architecture enforced it: the initial conditions plus stochastic quantum fluctuations plus the structure of physical law produced a specific, irreversible outcome. Nobody asked for it. The combination of structure and randomness created it. Building an energy landscape with enough stochastic dynamics and the right topology should, in principle, allow the same kind of emergent problem-solving for open-ended questions — not just arbitrary constraint satisfaction.

The intelligence is in the landscape topology. Not in the navigation algorithm. The navigation can be stochastic — even dumb — as long as the landscape is smart. Proteins fold, brains recognise songs, and children learn from burns not because of sophisticated algorithms, but because the energy surface they navigate has the right shape.

Section 11

Bridging Greedy and DP — Where EBMs Sit

Return to the song. Your brain did not run brute-force DP — it did not evaluate every song in memory. But it did not commit greedily to the first match either. It settled into an attractor that satisfied multiple constraints simultaneously. This is neither greedy nor exact DP.

Where EBMs sit on the commitment spectrum
GREEDY COLLAPSE LLMs · token-by-token EXACT DP Brute force · all states EBMs / JEPA holistic evaluation + revision human brain Not brute force. Not greedy. Energy-based settling.
EBMs evaluate complete configurations and revise — avoiding both greedy irrevocability and intractable exhaustive search
The architectural argument, in one sentence

Autoregressive models are architecturally greedy — they commit without revision. Exact DP is architecturally intractable — it evaluates everything. EBMs are architecturally in between — they evaluate holistically and revise iteratively. And that middle ground is where reasoning actually lives.

Budget vs waste — two kinds of state reduction

When exact DP is intractable, we reduce the state space. There are two fundamentally different reasons to drop a state, and confusing them leads to different errors.

Two mechanisms for state reduction
Budget constraint
Why: Cannot afford to evaluate all states
Risk: Dropping a state that matters
ML analogue: Mixture-of-experts — only relevant neurons fire
Information potentially lost
Redundancy discovery
Why: Many states encode the same information
Risk: None — states were duplicates
ML analogue: Pruning / distillation — removing redundant parameters
No information lost
DeepSeek-V2 (2024) · Frankle & Carlin, "Lottery Ticket Hypothesis," ICLR 2019

Double descent is the concrete proof that budget-based pruning can be catastrophically wrong. Classical learning theory predicts a U-shaped error curve — more parameters eventually hurt. But past the interpolation threshold, error drops again. The regions we pruned based on the U-curve assumption were exactly where the second descent lived. Every pruning heuristic has a blind spot shaped like its assumptions.

1989
Vallet et al. first observe double descent
30 yrs
Gap before recognition as general phenomenon
2019
Belkin names it · OpenAI shows universality

Discovery as state-space expansion

The transferable insight across this entire essay is that breakthroughs happen when someone notices a missing state variable — when two situations that look the same are actually structurally different because of something not being tracked.

The pattern across domains
Domain Original model Missing dimension Discovery
LLM architecture Token-by-token generation Holistic constraint evaluation Energy-based models
Statistical learning Bias-variance tradeoff Interpolation regime Double descent (2019)
AI reasoning Scale = capability Architectural constraint Apple "Illusion of Thinking"
World modelling Predict next token/pixel Abstract representation space JEPA (LeCun, 2022)
Classical mechanics Deterministic trajectories Measurement uncertainty Quantum mechanics
The pattern is always the same: what was treated as constant was actually variable

The instinct that detects this — the feeling that "something is off" before you can name what — is the skill that transfers. It is the same instinct that told Belkin the bias-variance curve was too clean. The same instinct that tells you, staring at a failing BFS solution at 2am, that two paths reaching the same cell are not the same subproblem. And the same instinct that says: maybe the architecture itself, not the data, is the bottleneck.


Section 12

Recursive Landscapes and the Alignment Question

If an energy landscape can be trained to encode constraints, and if the system can navigate that landscape to find solutions, then a natural next step emerges: can the system improve its own landscape?

This is where the argument becomes both exciting and dangerous. A system that recursively refines its own energy landscape — deepening valleys around valid solutions, raising ridges around errors, discovering new basins that humans have not mapped — could exceed human cognitive capacity in specific domains. Human intelligence, while extraordinary, has upper bounds. Individual cognition saturates. A self-improving landscape does not have the same biological ceiling.

But the danger is immediate: if the landscape can self-modify, what prevents it from trivialising itself? A landscape that makes everything low-energy — every configuration equally valid — has technically minimised its own energy but has lost all discriminative power. It has made itself useless. This is the alignment problem, reframed in energy-landscape terms.

The Occam's razor answer — truth, curiosity, beauty

A truly intelligent system driven by curiosity would not flatten its own landscape — because trivial solutions do not satisfy curiosity. The complexity of the landscape IS the system's capability. Flattening it is self-defeating: a landscape with no valleys is a landscape that can distinguish nothing.

This parallels Maslow's hierarchy of needs. People driven by intrinsic motivation — not external reward, but genuine fascination with the structure of problems — do not seek trivial answers. They seek deeper questions. A system that finds low-energy states inherently satisfying (truth), that is driven to explore its own landscape (curiosity), and that recognises elegant structure (beauty), would be naturally aligned — not because it was constrained to be, but because trivialisation contradicts its own nature.

Alignment through architecture vs alignment through rules
Rule-based alignment (current approach)
RLHF: external reward signal shapes behaviour
Constitutional AI: explicit rules constrain outputs
Guardrails: post-hoc filtering of generated content
Fundamental limitation: rules can be gamed, reward can be hacked
Alignment by constraint. Brittle.
Architectural alignment (energy-based)
Truth: low-energy states correspond to accurate representations
Curiosity: the system is driven to explore its own landscape
Beauty: elegant constraint satisfaction is inherently preferred
Fundamental advantage: trivialisation is self-defeating
Alignment by nature. Robust.
Any rule-based system can be gamed. An architecturally curious system games itself toward truth.

This does not solve alignment completely. But it changes the problem. An architecture that can evaluate its own outputs against learned constraints, that is driven by intrinsic curiosity to explore the deepest basins of its landscape, and that finds trivial solutions structurally unsatisfying, is fundamentally more alignable than one that commits tokens irrevocably and hopes external guardrails catch the errors.

The framework

This essay has been about architecture — what can reason, why the dominant architecture cannot, and what kind of system might bridge the gap.

The song in your head resolved in under a second. No search. No sequential commitment. Your brain fell into an energy minimum, and it was right. The question for AI is whether we build systems that can do the same — systems that find truth beautiful, that are curious by nature, and that reason not because they were told to but because the shape of their world makes reasoning the path of least energy.

Shojaee et al., "The Illusion of Thinking," Apple ML Research, arXiv:2506.06941 (June 2025)
Lawsen, "The Illusion of the Illusion of Thinking," arXiv:2506.09250 (2025) · Dellibarda Varela et al., critical replication (2025)
LeCun, "A Path Towards Autonomous Machine Intelligence," OpenReview (June 2022)
Assran et al., "I-JEPA," CVPR 2023 · Bardes et al., "V-JEPA," Meta AI (2024)
Hinton, "Training Products of Experts by Minimizing Contrastive Divergence," Neural Computation (2002)
Du & Mordatch, "Implicit Generation and Modeling with Energy-Based Models," NeurIPS 2019
Logical Intelligence, "Kona 1.0 Sudoku Benchmark," logicalintelligence.com (Feb 2026)
Vaswani et al., "Attention Is All You Need," NeurIPS 2017
Karpathy, Stanford CS25: Introduction to Transformers (2023) · Karpathy, Twitter (Oct 2022)
Elhage et al., "A Mathematical Framework for Transformer Circuits," Anthropic (2021)
Cormen, Leiserson, Rivest & Stein, Introduction to Algorithms, 4th ed. (2022), ch. 15–16
Edmonds, "Matroids and the Greedy Algorithm," Mathematical Programming 1 (1971)
Bellman, Dynamic Programming, Princeton UP (1957) · Bellman, Adaptive Control Processes (1961)
DeepSeek-V2, "MoE Language Model" (2024) · Frankle & Carlin, "Lottery Ticket Hypothesis," ICLR 2019
Belkin et al., "Reconciling Modern ML Practice and the Bias-Variance Tradeoff," PNAS (2019)
Preetum et al., "Deep Double Descent," OpenAI (2019) · Vallet et al. (1989)
Chen et al., "Data Contamination in LLM Benchmarks" (2025) · Deng et al. (2024) · Xu et al. (2024)
White et al., "LiveBench: A Contamination-Free LLM Benchmark," arXiv:2406.19314 (2024)
Sun et al., "When Benchmarks Leak: Inference-Time Decontamination," arXiv:2601.19334 (2026)
Liu, Liu & Gore, "Superposition Yields Robust Neural Scaling," MIT (2025)
Levinthal, "How to Fold Graciously," Mossbauer Spectroscopy in Biological Systems (1969)
Bryngelson & Wolynes, "Funnels, Pathways, and the Energy Landscape of Protein Folding," Proteins (1995)
Hopfield, "Neural Networks and Physical Systems," PNAS (1982)
Welling & Teh, "Bayesian Learning via Stochastic Gradient Langevin Dynamics," ICML 2011
Kirkpatrick, Gelatt & Vecchi, "Optimization by Simulated Annealing," Science (1983)
Boltzmann, "On the Relationship Between the Second Law and Probability" (1877)
Maslow, "A Theory of Human Motivation," Psychological Review (1943)
Feynman, "Surely You're Joking, Mr. Feynman!" (1985)