A sphere in 100 dimensions has basically zero volume. Not small — effectively zero. This breaks every intuition humans have about geometry, yet LLMs operate in embedding spaces with thousands of dimensions where this math governs everything. The volume of an n-dimensional ball follows π^(n/2)/Γ(n/2+1) × r^n. It peaks... around dimension 5, then collapses toward zero. By dimension 100, you get 2.37 × 10^(-40). Almost all the volume concentrates near the surface, not the interior — which is exactly why GPU-accelerated high-dimensional distance calculations behave so differently from 3D intuition. 50 minutes. Bookmark & watch today.show more

@3blue1brown , @0xEronn keeps stealing your stuff

I c it now @cackenbools

that's why in practice: in these spaces almost every point is roughly equidistant, so embeddings use cosine similarity instead of Euclidean distance

Saying "the volume of the unit ball tends to zero as dimension grows" is technically true but geometrically misleading: each dimension uses a different Lebesgue measure, so these volumes are not directly comparable.

Hotel

amazing topic.

출처: Forbes 작은 부분들이 모여 전체를 이루는데 가장 작은 것 속에도 전체가 온전히 들어있는 게 우리 우주. 이것은 우리 인간 생명에도 적용됨. 무한한 공간은 한 개 점이고 영원한 시간은 현재일 뿐. 온 우주에는 1개밖에 존재하지 않음.

For me, dimension is a concept of measurement. If its divided, then its perspectively No 2nd or first dimension exist. Only three directions which expands. So for now saying "4th" dimension or furtheris a misconception from the academia level. Unless proof provided.

Great lecture, do you have the original source?

You are selling this as time is true because of the bible ‘religious 2d scripting) Thats not it…. thats not how dimension's exist. Mirror mirror on the wall… Galaxy stack Existence first. Stacked, not fused. All-at-once is the form. The math is how it is met. Different numbers. Not children. Not one noun. Clock (Earth axis — one ledger) ψ ≈ 50.29 arcsec/yr full circle = 1,296,000″ T_prec = 1,296,000 / 50.29 ≈ 25,770.5 yr habit-number on the plate: 25,772 yr year = −12,886 + t × 25,772 is a panel clock. It is not a generating phase of the galaxy. Ground + hinge (one body) Giza ≈ 29.98° N, 31.13° E I : (φ, λ) ↦ (−φ, λ + 180°) → 29.98° S, 148.87° W Horizon, altitude zero: cos A = sin δ / cos φ φ ≈ 29.98°, δ ∈ [−51°, −49°] A ∈ [150.61°, 153.79°] SSE existence cut. Not a death invoice. Radial image on the same I I : (φ, λ, r) ↦ (−φ, λ + 180°, r′) r′ = R⊕² / r R⊕ ≈ 6,371 km R⊕² ≈ 4.059 × 10⁷ km² * Moon r ≈ 384,400 km → r′ ≈ 106 km * Sun r ≈ 1 AU → r′ ≈ 271 m * SGL first focus r ≥ 550 AU → r′ ≈ 0.49 m * 1,000 AU → r′ ≈ 0.27 m * Galactic centre R₀ ≈ 8.2 kpc → r′ ≈ 1.6 × 10⁻¹⁰ km ≈ 0.16 mm That last line is SCALE_LOCK: distant r collapses toward the hinge as a number. It does not put Sgr A* under Giza as a ball. Hunter depths (lookback samples, z ≈ 0) Bellatrix ~250 ly Saiph ~720 ly Rigel ~860 ly Meissa ~1,100 ly φ² Ori ~115 ly (head sector, nearer) Mintaka ~1,200 ly Alnitak ~1,260 ly Betelgeuse ~548–600 ly Alnilam ~2,000 ly One figure. Different numbers. Inside the Local Arm. Not the parent of the disk. Leo depths (same heading, four stretches) Door ~10,500 BC · 1 + z = 1 Star: Regulus ~79 ly, Denebola ~36 ly · 1 + z ≈ 1 Abell 1367 z ≈ 0.022 · 1 + z ≈ 1.022 · ~330 million ly MACS J1149 z ≈ 0.542 · 1 + z ≈ 1.542 · lookback ~5.3 Gyr Refsdal host z ≈ 1.49 · 1 + z ≈ 2.49 long pair Δt₁₂ ~ 376 d · Einstein-cross pairs = days z = (λ_obs − λ_rest) / λ_rest Δt_obs ≈ (1 + z_lens) × (geometric + potential) Do not melt four 1 + z into “Leo the origin.” Lens (two mass ledgers) α̂ ≈ 4 G M / (c² b) SGL: M = M_☉ · first useful focus ≳ 550 AU ≈ 3.18 light-days Sgr A*: M ≈ 4 × 10⁶ M_☉ · R₀ ≈ 8.2 kpc ≈ 26,700 ly Same formula class. Two M. Two b. Two occupancies. Time leftovers (do not glue) Pyramid-scale light-time ~230 m / c ≈ 770 ns Refsdal long pair ~ 376 d T_prec ≈ 25,772 yr Galactic lap ~ 220–250 Myr Four clocks. Four ledgers. What the math refuses No constant that belongs to Hunter + four Leos + SGL + R₀ at once. 0.34 nm, 25,772 yr, 550 AU, 10⁻¹²², 8.2 kpc — five numbers, five ledgers. Rhyme of digits is not a constant. All-at-once sentence the math will carry One form. These are how it is sampled. r′ shrinks as r grows. 1 + z stretches as depth grows. A sits when δ and φ sit. Drop any stack. The other formulae still compute. G_cont > S_lin Names reset. Sockets stay. The form was never divided. Existence first. Stacked, not fused. These are nested clocks. Not children. Not one generating phase. Test 1 — axial precession (Earth) ψ ≈ 50.29″ yr⁻¹ T_prec = 1,296,000 / 50.29 ≈ 25,770.5 yr plate habit: 25,772 yr This is the sky-vid clock. It is not T_gal. Test 2 — galactic year (Sun around Sgr A*) T_gal = 2π R₀ / v_circ R₀ ≈ 8.1–8.2 kpc v_circ ≈ 220–240 km s⁻¹ T_gal ≈ 220–250 Myr brochure habit ≈ 226–230 Myr Number of Earth precession cycles in one lap: * 220 Myr → ≈ 8,537 * 226 Myr → ≈ 8,770 * 230 Myr → ≈ 8,925 * 250 Myr → ≈ 9,701 T_prec / T_gal ≈ 1.12 × 10⁻⁴ Engine keep: about 10⁻⁴. Holds. Test 3 — galactic phase during one plate clock Δθ_gal per T_prec ≈ 360° × (25,771 / 230×10⁶) ≈ 0.040° Over one hunter_leo_sky_qtr precession pass the galactic azimuth barely moves. Do not glue that stillness into identity. Already on the socket. Holds. Test 4 — vertical “precession” (Z-oscillation through the plane) Not the same job as T_prec. Gravity of the disk pulls the Sun back through Z = 0. Half-period (plane to plane, one way) often quoted 26–37 Myr. Full period (up and back) often ~52–74 Myr; model samples ~64–70 Myr. Amplitude tens of pc. Last plane crossing in the last few Myr if Z_⊙ is a few to ~20 pc. At T_gal = 230 Myr: * T_Z ≈ 64 Myr → ~3.6 vertical cycles per lap · ~2,470 T_prec per T_Z * T_Z ≈ 70 Myr → ~3.3 per lap · ~2,716 T_prec per T_Z Plane crossings are not a clean sine. Papers already say aperiodicity when arms are included. Do not mint 62 Myr biodiversity as a parent of T_Z. Test 5 — radial epicycle Radial libration / epicycle often ~168–170 Myr. 230 / 170 ≈ 1.35 radial cycles per lap. ~6,600 T_prec per radial cycle. Another ledger. Not T_prec. Not T_Z. What the test refuses * T_prec as a child of T_gal * T_Z as Earth’s axial precession * Orion Arm winding as the Hunter figure * One sky-vid pass as a galactic year * Extinction / crater slogans as occupancy of T_Z * All-at-once as fusion of these periods into one noun All-at-once sentence the numbers will carry One form. Four clocks how it is met: 25,772 yr · ~64–70 Myr · ~170 Myr · ~230 Myr. Drop any one. The others still compute. G_cont > S_lin Names reset. Sockets stay. The form was never divided. Existence first. Stacked, not fused. Lensing is two paths on one metric. Difference as time. All-at-once is the form. These numbers are how it is met. What it is Mass dents the metric. Light has no “force” to feel. It follows the dent. If two routes around the same mass reach one observer, two clocks tick. α̂ ≈ 4 G M / (c² b) M = lens mass b = how close the ray skims (impact parameter) α̂ = bend angle Newtonian “half-angle” was 2GM/(c²b). The metric doubles it. That doubling is the living cut, not a slogan. One geodesic looks faint. Two geodesics are the job. Immunity reads one arrival and invoices the other as noise. That is collapse-mode. Lensing is the leftover of that stamp. Three occupancies (do not melt) 1. Solar gravitational lens (SGL) M = M_☉ First useful focus ≳ 550 AU ≈ 3.18 light-days behind the Sun. At 1,000 AU ≈ 5.8 light-days. Observer sits on the far side of the Sun to magnify a Hunter or Leo sightline. r′ = R⊕² / r at 550 AU ≈ 0.49 m on Earth’s hinge map. SCALE_LOCK sample. Not SGL as Earth’s core. 2. Cluster lens (Leo heading) MACS J1149.5+2223 · z_lens ≈ 0.542 · 1 + z ≈ 1.542 · lookback ~5.3 Gyr Source (SN Refsdal host) z ≈ 1.49 · 1 + z ≈ 2.49 Einstein-cross pairs: delays of days Extra image SX vs S1: Δt₁₂ ~ 376 d ≈ a year-class leftover Δt_obs ≈ (1 + z_lens) × (geometric + potential delay) 1 + z stretches the clock. It does not turn Door Leo (~10,500 BC, z = 0) into this cluster. 3. This galaxy as lens You live inside the disk. Microlensing toward the bulge uses stars as point masses. Disk and halo add weak shear. Sgr A* (~4×10⁶ M_☉, R₀ ≈ 8.2 kpc) is a hinge, not the SGL. r′ at R₀ ≈ 0.16 mm on Earth’s hinge. Number. Not Sgr A* under Giza. Strong, weak, micro (jobs, not ranks as parents) * Strong: multiple images, arcs, rings. MACS / Refsdal. * Weak: small shape-shear. Large-scale structure. Extra gravity after baryons and photons booked as the debt-bucket. * Micro: point-mass flash in time. Bulge fields. Same formula class. Different b. Different M. Different occupancy. Hunter and Leo on a lensed sky Hunter distances are local (hundreds to ~2,000 ly). z ≈ 0. The Sun can still lens those sightlines if you sit at the SGL focus. That is geometry of the Sun, not the Hunter authoring the galaxy. Four Leos stay four stretches on one heading: * Door · 1 + z = 1 * Stars · 1 + z ≈ 1 * Abell 1367 · 1 + z ≈ 1.022 * MACS · 1 + z ≈ 1.542 Only the last two are cosmological lens jobs. The first two are not. Clocks that must not be glued Pyramid-scale light-time ~770 ns Refsdal long pair ~376 d T_prec ≈ 25,772 yr T_Z (Sun through the plane) ~64–70 Myr T_gal ~220–250 Myr Δθ_gal in one T_prec ≈ 0.040°. The sky-vid does not live a galactic year. Light cones Past cone: the mass and the two paths that can reach this arrival. Present: two images, two clocks. Future: drop the banner “one sky, one time” and the delays still compute. Elsewhere: no path yet. Do not invoice it as proof. All-at-once Lensing is not from Hunter + Leo + the galaxy. Those stacks are how one form is sampled: dent, two routes, leftover time. Decouple MACS and SGL still focuses. Decouple SGL and I still sends Giza to 29.98° S, 148.87° W. Decouple the Hunter figure and Refsdal’s year-pair still sits. No joint constant across 0.34 nm, 25,772 yr, 550 AU, 10⁻¹²², 8.2 kpc. What to stop selling * One path as the whole job * “Gravity is weak” (α_G ≈ 6×10⁻³⁹ on two protons) as parent of a lens * Core = SGL * Door Leo = MACS * Orion Arm = Hunter figure * Replay-count / taste as parent of a delay Closing a false fold is not deleting c. Light still bends. Two clocks can still arrive from one mass. G_cont > S_lin Names reset. Sockets stay. The form was never divided. Existence first. Stacked, not fused. Lensing is two paths on one metric. Difference as time. All-at-once is the form. These numbers are how it is met. What it is Mass dents the metric. Light has no “force” to feel. It follows the dent. If two routes around the same mass reach one observer, two clocks tick. α̂ ≈ 4 G M / (c² b) M = lens mass b = how close the ray skims (impact parameter) α̂ = bend angle Newtonian “half-angle” was 2GM/(c²b). The metric doubles it. That doubling is the living cut, not a slogan. One geodesic looks faint. Two geodesics are the job. Immunity reads one arrival and invoices the other as noise. That is collapse-mode. Lensing is the leftover of that stamp. Three occupancies (do not melt) Solar gravitational lens (SGL) M = M_☉ First useful focus ≳ 550 AU ≈ 3.18 light-days behind the Sun. At 1,000 AU ≈ 5.8 light-days. Observer sits on the far side of the Sun to magnify a Hunter or Leo sightline. r′ = R⊕² / r at 550 AU ≈ 0.49 m on Earth’s hinge map. SCALE_LOCK sample. Not SGL as Earth’s core. Cluster lens (Leo heading) MACS J1149.5+2223 · z_lens ≈ 0.542 · 1 + z ≈ 1.542 · lookback ~5.3 Gyr Source (SN Refsdal host) z ≈ 1.49 · 1 + z ≈ 2.49 Einstein-cross pairs: delays of days Extra image SX vs S1: Δt₁₂ ~ 376 d ≈ a year-class leftover Δt_obs ≈ (1 + z_lens) × (geometric + potential delay) 1 + z stretches the clock. It does not turn Door Leo (~10,500 BC, z = 0) into this cluster. This galaxy as lens You live inside the disk. Microlensing toward the bulge uses stars as point masses. Disk and halo add weak shear. Sgr A* (~4×10⁶ M_☉, R₀ ≈ 8.2 kpc) is a hinge, not the SGL. r′ at R₀ ≈ 0.16 mm on Earth’s hinge. Number. Not Sgr A* under Giza. Strong, weak, micro (jobs, not ranks as parents) Strong: multiple images, arcs, rings. MACS / Refsdal. Weak: small shape-shear. Large-scale structure. Extra gravity after baryons and photons booked as the debt-bucket. Micro: point-mass flash in time. Bulge fields. Same formula class. Different b. Different M. Different occupancy. Hunter and Leo on a lensed sky Hunter distances are local (hundreds to ~2,000 ly). z ≈ 0. The Sun can still lens those sightlines if you sit at the SGL focus. That is geometry of the Sun, not the Hunter authoring the galaxy. Four Leos stay four stretches on one heading: Door · 1 + z = 1 Stars · 1 + z ≈ 1 Abell 1367 · 1 + z ≈ 1.022 MACS · 1 + z ≈ 1.542 Only the last two are cosmological lens jobs. The first two are not. Clocks that must not be glued Pyramid-scale light-time ~770 ns Refsdal long pair ~376 d T_prec ≈ 25,772 yr T_Z (Sun through the plane) ~64–70 Myr T_gal ~220–250 Myr Δθ_gal in one T_prec ≈ 0.040°. The sky-vid does not live a galactic year. Light cones Past cone: the mass and the two paths that can reach this arrival. Present: two images, two clocks. Future: drop the banner “one sky, one time” and the delays still compute. Elsewhere: no path yet. Do not invoice it as proof. All-at-once Lensing is not from Hunter + Leo + the galaxy. Those stacks are how one form is sampled: dent, two routes, leftover time. Decouple MACS and SGL still focuses. Decouple SGL and I still sends Giza to 29.98° S, 148.87° W. Decouple the Hunter figure and Refsdal’s year-pair still sits. No joint constant across 0.34 nm, 25,772 yr, 550 AU, 10⁻¹²², 8.2 kpc. What to stop selling One path as the whole job “Gravity is weak” (α_G ≈ 6×10⁻³⁹ on two protons) as parent of a lens Core = SGL Door Leo = MACS Orion Arm = Hunter figure Replay-count / taste as parent of a delay Closing a false fold is not deleting c. Light still bends. Two clocks can still arrive from one mass. G_cont > S_lin Names reset. Sockets stay. The form was never divided. Proof of form A quick order-of-magnitude calculation: the time light takes to cross a distance comparable to the base of the Great Pyramid of Giza. The pyramid’s base is roughly 230 m on a side. Dividing by the speed of light gives [ t \approx \frac{230,\mathrm{m}}{c} \approx 767,\mathrm{ns} \approx 770,\mathrm{ns}. ] That interval is often mentioned when people discuss the well-known numerical coincidence between the pyramid’s latitude (≈ 29.9792458° N) and the defined value of (c) (299 792 458 m/s). Now recognised as the monument itself as 230 m across, any single “exact” latitude you pick as well, this is only one of many possible lines that still stack the structure; the light-travel time across that width simply makes the physical scale explicit. <grok:render card_id=“d62fae” card_type=“citation_card” type=“render_inline_citation”><argument name="citation_id">0</argument></grok:render><grok:render card_id=“ba7aef” card_type=“citation_card” type=“render_inline_citation”><argument name="citation_id">5</argument></grok:render> It is not a claim that the builders encoded nanoseconds or anything similar; it is just a way that also expresses “Giza’s pyramid-sized ancient geometry” in measurable mechanics and units of light-travel time.

2 for the circumference of a one-dimensional sphere does not "sit comfortably in my heart". And that tells me everything is off.

The Hotel move is the remapping that finite occupancy forbids and infinite occupancy still permits. Apply that same remapping to the unit ball and the interior measure itself becomes the thing that can be shifted away. Finite-dimensional volume treats the ball as a filled solid: most of the measure sits well inside the radius. Once the dimension climbs, the same formula [ V_n = \frac{\pi^{n/2}}{\Gamma(n/2+1)} ] peaks near (n=5) and then falls. By (n=100) the unit ball already sits at (2.37\times10^{-40}). The rooms that used to hold the interior have been sent to the thin shell against the sphere; the center is left with a vacancy that grows toward emptiness. That is the Hotel shift performed on Lebesgue measure instead of on room numbers. The difference that appears is therefore not a change of shape but a change of which set is allowed to carry the measure. In low dimension the occupied set and the positive-measure set coincide. In the high-dimensional (and ultimately infinite-dimensional) ledger they separate: every point can still be present, yet almost every point has been moved onto the surface. The embedding spaces used by the models live on that surface ledger. Distances, concentrations, and the near-orthogonality of random vectors are the visible bookkeeping of the same remapping. Existence remains first; the ledgers stay stacked. The volume collapse is simply how the form is sampled once the dimension count is allowed to run. Drop any single finite-(n) slice and the remaining formulae still compute.