Geometry of Thought and the Greedy Trap: A Deep Dive into the Nature of Intelligence

Overview

This long-form essay (from a Chinese tech forum) investigates the nature of intelligence through two lenses: the geometry of thought (modeling cognition on Riemannian manifolds) and the greedy trap (the systematic failure of locally-optimal strategies). It argues these are two faces of one phenomenon: intelligent agents navigating curved conceptual spaces with limited information.

Key points

1. Cognitive geometry: theoretical foundations

• The most relevant current framework is Alpay & Senturk (Jan 2025), "The Geometry of Thought", proving a deep isomorphism between Transformer self-attention and tropical geometry. In the high-confidence limit (inverse temperature β → ∞), softmax attention converges to tropical matrix multiplication:
\[\text{TropicalAttention}(Q,K,V)_{ij} = \max_k (Q_i \cdot K_k + V_{kj})\]

This is exactly a Bellman-Ford shortest-path update: each Transformer layer performs one relaxation step, and chain-of-thought reasoning is explicit shortest-path search over a latent token graph. Reasoning failures map to graph problems: disconnection, negative cycles, or insufficient iterations.

• Three traditions converge here: Gärdenfors' conceptual space theory, manifold learning (t-SNE, UMAP), and neuroscience cognitive maps.
• A Riemannian framework interprets: the metric tensor g_μν as context-dependent semantic distance, curvature as cognitive complexity/learning load, and geodesics as optimal reasoning paths. The geodesic equation
\[\frac{d^2\gamma^i}{dt^2} + \Gamma^i_{jk}\frac{d\gamma^j}{dt}\frac{d\gamma^k}{dt} = 0\]

distinguishes "deliberate" curved reasoning from "intuitive" straight-line shortcuts.

• AlphaGeometry (DeepMind) solved 25/30 IMO geometry problems (~gold-medal level) via a dual system: a neural language model as fast System-1 hypothesis generator, and a symbolic deduction engine (DDAR) as System-2 verifier, trained on 1B synthetically generated diagrams without human demonstrations.
• Note: claims about a "Shihao Jiu laboratory" Riemannian framework (including quasi-crystal cognition and Betti-number phase transitions) could not be independently verified; the essay flags this explicitly.

2. Experimental evidence for non-Euclidean concept spaces

• Triad-comparison experiments (128 participants, 100 concepts, 500 optimized triads): triangle-inequality violations of only 3.7%; significant deviation from zero curvature (t = 4.89, p < 0.001); mean Gaussian curvature estimated at 0.152 ± 0.031.
• Curvature–cognitive-load correlation: reaction time r = 0.42 (p < 0.001); cross-validated prediction of learning-difficulty ratings at 72.3% accuracy.

3. The greedy trap

• Classical algorithms: Greedy Best-First Search ignores path cost g(n) and fails under heuristic deception (e.g., a 20-cost path vs. an optimal 12); nearest-neighbor TSP degrades to O(log n) approximation ratio; 0-1 knapsack density-greedy fails due to indivisibility. Greedy methods only guarantee optimality on matroid structures — which most AI problems (theorem proving, planning, creativity) lack.
• LLM decision failures: on multi-armed bandits, models like Gemma2 27B show action coverage of only ~45–65% with rapid saturation, linearly growing regret, and frequency bias. The most striking finding is the knowing-doing gap: the model correctly identified the UCB-optimal arm 87% of the time but selected the globally optimal arm only 21% of the time (greedy choice dominated at 58%). Model scale does not automatically fix this — larger models showed premature lock-in.
• CoT limitations: insufficient coverage of the latent space, local lock-in, length limits, and error bootstrapping — geometrically, a coarse geodesic approximation with too few Bellman-Ford iterations.
• Human parallel: Kahneman/Tversky heuristics, student pattern-matching errors in math problem solving (e.g., greedy chain length 10 vs. optimal 8 in alternating-dot configurations) show the trap is a universal property of resource-bounded agents — but humans uniquely possess metacognition to switch to slow thinking.

4. Unification: greedy paths are truncated geodesics

• Greedy gradient descent drops the Christoffel-symbol (curvature) term; it works in flat regions but systematically departs from the global geodesic where curvature is high.
• Failure modes map to geometry: metric misestimation (context-insensitivity), curvature neglect (intuition failure in abstract domains), topological obstacles (incommensurable paradigms), rigid step size (miscalibrated confidence).
• High-dimensional landscapes are dominated by saddle points rather than minima; momentum, adaptive learning rates, and noise are effectively curvature-exploiting "non-greedy" corrections.
• Kahneman's fast/slow thinking corresponds to high- vs. low-temperature Riemannian gradient flow on a cognitive potential landscape V(x).

5. Implications for AI design

• From symbolic/connectionist to geometric intelligence: symbols as constrained submanifolds, learning as function approximation on manifolds, reasoning as geodesic tracking.
• Geometric chain-of-thought: explicit graph construction, A*/Bellman-Ford search replacing greedy decoding, hierarchical planning, and dynamic depth allocation by local curvature.
• Curvature-aware exploration: treat uncertainty as curvature and information gain as distance; local temperature replaces global ε-greedy.
• Fixing the knowing-doing gap: shared manifold representations for knowledge and action, geometric regularization constraining actions to the geodesic neighborhood of cognition, counterfactual training, symbolic verification loops.
• Architecture frontiers: continual learning via manifold extension rather than overwriting (mitigating catastrophic forgetting), multi-agent consensus via Fréchet means on shared manifolds, and value alignment as embedding ethical constraints into the manifold structure.

Bottom line

The essay's central thesis: greedy traps are geometrically inevitable on any curved conceptual manifold, so the path to more reliable AI reasoning lies in explicitly representing and exploiting global geometric structure — curvature, topology, and geodesics — rather than scaling up purely local, next-token optimization.