3Blue1Brown — Five puzzles for thinking outside the box

Why this is in the vault

30-minute Sanderson original on a single load-bearing move: lift the problem into a higher dimension so the problem becomes trivial, then project the answer back down. Three 2D puzzles each crack open only when you treat the 2D structure as a shadow of a 3D object. Calisson tilings become stacks of cubes; strip-coverings of a disk become hemispherical caps with area-preserving projection; Monge’s theorem on three circles becomes three cones with a plane resting on top. The vault keeps it because (1) this is the third high-quality 3Blue1Brown source for CA-023 “coordinate change as the core problem-solving move,” promoting it from 2-source inbox to 3-source canon-tier; (2) the closing admission — we can leverage 4D as creative-intuition creatures, but higher dimensions are “analysis without intuition, daunting” — is a load-bearing editorial frame for RDCO’s AI coverage, where LMs operate in thousand-dimensional embedding spaces humans cannot visualize; (3) the Slovakian-student correction (the sphere-tangent-plane proof fails when one circle is much smaller than the others) is a clean concrete instance of CA-022 “proofs that look fully general have hidden position-dependent gaps” — which pairs with the Euclid-Proposition-1 gap from the companion backfill.

Episode summary

30-minute Sanderson original (November 2024). Five geometry puzzles where the answer only falls out once you lift into a higher dimension. Puzzle 1: rhombus tilings of a hexagon — treat them as projections of cube-stacks, solved. Puzzle 2: minimum sum of strip widths covering the unit disk (the Tarski-plank problem) — project to a hemisphere where strip area = π × width regardless of position. Puzzle 3: Monge’s theorem that three circles’ pairwise external-tangent-intersection points are collinear — the “spheres with tangent plane” proof has a fatal gap (caught mid-lecture by an IMO student), fixed by switching to cones and centers of similarity. Puzzle 4: explicit volume formula for a tetrahedron via 4×4 determinants (teased but not solved — the setup for a future video). Puzzle 5: constructing a 3D puzzle whose answer requires thinking as a 4D creature — the rhombic dodecahedron as projection of a hypercube. Closes with the analysis-vs-intuition thesis: higher-dimensional reasoning loses the intuitive guidance that makes problem-solving tractable.

Key arguments / segments

[00:00:30] Puzzle 1 setup: rhombus tilings of a hexagon. Fill an n-sided hexagon with 60/120-degree rhombuses. Rotate any little-hexagon-of-3-tiles 60°; how many moves to get between any two tilings?
[00:03:00] Puzzle 2 setup: the Tarski-plank problem. Cover a unit disk with parallel-chord strips; minimize the sum of widths. Trivial bound: 2 (the diameter). Can you do better?
[00:04:30] Puzzle 1 resolved via 3D lift. Color the rhombuses by orientation; the tiling is a 2D projection of a stack of unit cubes in an n×n×n frame. The rotate-the-hexagon move corresponds to adding or removing one cube. Therefore (a) any tiling connects to any other (remove all cubes, rebuild), and (b) the max-moves answer is n³ — the cubic bound reveals the 3D structure baked into the 2D puzzle.
[00:07:00] Puzzle 2 resolved via hemisphere lift. Directly: strip area on a flat disk depends on where the strip sits, so width doesn’t cleanly bound area. Trick: project the disk onto a hemisphere above it. Archimedes’ cylinder-projection theorem guarantees the area of a hemispherical strip is exactly π × width, regardless of position. Summing over covering strips: total area ≥ 2π, so sum of widths ≥ 2. The parallel-strip solution is optimal.
[00:10:00] Puzzle 3 setup: Monge’s theorem. Three non-intersecting, distinct-radius circles in the plane. For each pair, take the intersection of external tangent lines. Claim: all three intersection points are collinear.
[00:12:00] The sphere-and-tangent-plane proof. Inflate each circle to a sphere (equator). A tangent plane resting on all three spheres touches each at a point; the line between any two tangency points is an external tangent passing through the relevant intersection. The plane and the base plane intersect in a line containing all three points.
[00:14:00] The Slovakian-student correction. An IMO student in the audience pointed out: if one sphere is much smaller than the other two and placed between them, no plane can rest tangent to all three simultaneously. The argument works for certain positions of the three circles but is not fully general.
[00:14:30] The Slovakian-coach repair: use cones, not spheres. Replace each sphere with a cone whose base is the circle and whose apex angle is equal across all three. The apex of each cone becomes a center of similarity for the pair — the point about which one circle scales to coincide with the other. Any three apex points determine a plane (always), and that plane’s intersection with the base plane is the collinearity line. Works for all configurations including one-inside-another.
[00:17:00] Generalization: three similar shapes. The center-of-similarity framing extends beyond circles. Any three similar shapes in the same orientation have three centers of similarity that fall on a line. The cone-lift becomes a pyramid-lift; the argument stays intact.
[00:17:30] Puzzle 4: tetrahedron volume via 4×4 determinants. Sanderson teases the result — the volume of a tetrahedron with given vertex coordinates is expressible as a 4×4 determinant, and the analog in dimension n is an (n+1)×(n+1) determinant. This is the natural explanation for the determinant formula; a dedicated video is promised as part of the linear algebra series.
[00:19:00] Puzzle 5: 3D tiling as projection of 4D hypercube. Build the 4D hypercube as {0,1}⁴ vertices with edges between coordinate-adjacent pairs. Project along the (1,1,1,1) direction. The wireframe collapses to a rhombic dodecahedron — the 3D analog of the hexagonal projection of a 3D cube. This already explains why the rhombic dodecahedron tessellates 3D space (because the 4D hypercube tessellates 4D space).
[00:25:00] The 4D puzzle. Fill a big rhombic dodecahedron with unit-rhombic-dodecahedron copies; the move is “slide through the origin” (the 3D analog of the hexagon rotation). Max moves between any two tilings: n⁴.
[00:26:00] Applications of higher-dimensional lifts outside geometry. Quaternions (4D extension of complex numbers, encodes 3D rotations). 24D sphere packing → Voyager error-correcting codes. High-dimensional random-vector statistics → explains transformer scaling (vectors nearly orthogonal at random). Compression algorithms.
[00:27:30] Closing thesis: analysis vs. intuition in higher dimensions. “Analysis without intuition is daunting. The space of all possible logical moves you can make in the pursuit of a proof is often too vast to explore in a reasonable time. Intuition is what offers the guiding lights telling you which paths are worth trying.” The sad note: 4D-and-higher problems force reasoning without the intuitive guidance that made 2D→3D lifts tractable for us as 3D creatures.

Notable claims

[00:06:00] Max moves between any two rhombus-tilings of a size-n hexagon is exactly n³. The cubic number is a signature — three-dimensionality is baked into the 2D puzzle, not imposed by our cognition.
[00:09:00] The area of a strip on a hemisphere is exactly π × (width), by Archimedes’ cylinder-projection theorem (area-preserving projection onto the circumscribing cylinder). This is why the Tarski-plank minimum is exactly 2.
[00:14:00] The first, intuitively-satisfying “spheres with a tangent plane” proof of Monge’s theorem is not fully general — there exist circle configurations (one small circle between two larger ones) where no tangent plane exists. Caught by a student during an IMO-prep lecture; the repair uses cones and centers of similarity.
[00:22:00] The rhombic dodecahedron is the natural 3D projection of a 4D hypercube along the (1,1,1,1) direction; its known property of tessellating 3D space follows trivially from the fact that the 4D hypercube tessellates 4D space.
[00:26:00] Pairs of random vectors in sufficiently-high-dimensional space are, with very high probability, nearly perpendicular. Sanderson explicitly notes this is relevant to why LLMs scale as well as they do (embedding geometry in ~d ≈ 10³–10⁴). This is a throwaway line in the video but a load-bearing fact for RDCO’s AI writing.
[00:28:00] “Analysis without intuition is daunting. Intuition is what offers the guiding lights telling you which paths are worth trying.” Sanderson’s thesis that higher-dimensional reasoning loses the intuitive guardrails that make lower-dimensional mathematics tractable.

Guests

Grant Sanderson (3Blue1Brown) — solo host; no guests.
Named attributions only: Daniel Kim (first two puzzles); David and Tomei (earliest Calisson reference); Po-Shen Lo (tetrahedron example); Tadashi Tokieda (Numberphile video on Monge’s theorem); an unnamed IMO student (caught the sphere-proof gap); the Slovakian-team deputy leader (proposed the cone-repair).

Mapping against Ray Data Co

CA-023 “coordinate change as core problem-solving move” is now 3-source canon-tier ready. The two prior sources (~/rdco-vault/06-reference/2026-04-20-3blue1brown-why-colliding-blocks-compute-pi physics-to-geometry; ~/rdco-vault/06-reference/2026-04-20-3blue1brown-how-and-why-to-take-a-logarithm-of-an-image complex-log unwrap) covered two domains of the lift. This video adds three fresh examples in one sitting (2D→3D for tilings, strips, Monge), plus a 3D→4D lift, plus the explicit meta-thesis that this is a general problem-solving technique. Promotes CA-023 well above the 3-source bar and makes it canon-tier drafting-ready as ~/rdco-vault/06-reference/concepts/coordinate-change-as-core-move.md. Sanderson’s own framing (“step into a higher dimension”) is the pull-quote.
The analysis-vs-intuition closing is the editorial frame for every RDCO piece on LLMs. Sanderson admits something most AI coverage refuses to: we can do lower-dimensional problems because we have intuition as 3D creatures; 4D-and-higher requires pure analysis without the guiding lights. This is exactly what LLMs do in their embedding spaces — thousand-dimensional reasoning with no intuition, pure analysis on vectors. The honest framing for RDCO AI writing: LMs don’t “understand” in the intuition-as-guidance sense; they analyze. That’s useful for certain problems and insufficient for others. Worth writing as a vault concept note: “LM intuition is a category error — they do analysis on embeddings, without the guiding-lights shortcut humans use”. Uses this video as the anchoring source, pairs with the ~/rdco-vault/06-reference/2026-04-20-3blue1brown-but-what-is-a-neural-network and ~/rdco-vault/06-reference/2026-04-19-3blue1brown-large-language-models-explained-briefly videos for the mechanistic grounding.
The Slovakian-student correction is a clean teaching instance of CA-022 “binary-decision-around-continuous-probability” adjacent pattern: proofs-with-hidden-position-dependence. The sphere-tangent-plane proof felt fully general; it’s not. The repair (use cones, use centers of similarity) generalizes by weakening the intermediate object (cone is more flexible than sphere, center-of-similarity is more abstract than external-tangent). Pattern: when a proof feels general, stress-test it against degenerate configurations; if it fails, weaken the intermediate object. This pairs with the Euclid Proposition-1 gap from the companion backfill as a “subtle proof gaps” cluster. Worth a candidate entry in CANDIDATES.md as CA-024 “generality-under-degenerate-configurations stress test.”
The “random vectors in high dimensions are nearly perpendicular” throwaway is a load-bearing fact for LLM writing. Sanderson mentions it in passing at [00:27:00] but it’s the explanation for why embedding-based retrieval (RAG) works: in ~10³–10⁴ dimensions, semantically-unrelated vectors are nearly orthogonal, so cosine-similarity thresholding cleanly separates related from unrelated. Worth capturing as a standalone vault concept note linked from both this assessment and the vectors-chapter-1 linear-algebra assessment.
Puzzle 4’s tetrahedron-determinant tease is a bookmark for the linear algebra series. Sanderson promised a dedicated video explaining why the determinant formula has the form it does via n-simplex volume in n dimensions. When that video ships, it belongs next to this one and the linear algebra chapters. Worth a follow-up watch-queue item.
Sanity Check angle: “The LLM shortcut is analysis without intuition, and that’s the trade.” Lead with the hexagon-tiling puzzle (3D-creature intuition trivializes the 2D problem). Pivot to Sanderson’s closing admission (4D loses the intuition shortcut). Land on: LMs are native 4D-and-higher reasoners — they don’t have human intuition, they have analysis on embeddings. Some problems are native to their strength; others aren’t. Frames the pragmatic question “where does LM reasoning actually help?” without either the hype or the dismissal. ~1,500 words.

Open follow-ups

Draft ~/rdco-vault/06-reference/concepts/coordinate-change-as-core-move.md from CA-023. 3 canon-tier sources now: colliding-blocks-compute-pi (physics→geometry), logarithm-of-an-image (complex-log unwrap), and this video (three 2D→3D lifts + one 3D→4D). Include Sanderson’s meta-framing and the analysis-vs-intuition caveat as the cap on the pattern. ~1 hour.
Add CA-024 “generality-under-degenerate-configurations” to CANDIDATES.md. Sources: Slovakian-student correction here, Euclid’s Proposition-1 gap (Blåsjö’s antagonistic-debate reframe), and the CA-022 binary-decision-around-continuous-probability pattern. 3 sources, promotion-ready.
Vault concept note: “High-dimensional random vectors are nearly orthogonal → why RAG works.” Standalone ~500 words, cross-links to this video and the linear-algebra chapter assessments.
Watch-queue: the tetrahedron-determinant explainer when Sanderson ships it as part of the linear algebra series. File next to ~/rdco-vault/06-reference/2026-04-20-3blue1brown-linear-transformations-matrices-chapter-3.

~/rdco-vault/06-reference/transcripts/2026-04-20-3blue1brown-five-puzzles-thinking-outside-the-box-transcript.md — full transcript
~/rdco-vault/06-reference/2026-04-20-3blue1brown-why-colliding-blocks-compute-pi — physics→geometry coordinate change (CA-023 source 1)
~/rdco-vault/06-reference/2026-04-20-3blue1brown-how-and-why-to-take-a-logarithm-of-an-image — complex-log unwrap (CA-023 source 2)
~/rdco-vault/06-reference/concepts/CANDIDATES.md — track CA-023 promotion to canon-tier (now 3-source) + CA-024 proposal