Can tidy dots help you find a circle’s secret number?

A game in five short stages. Throw pretend darts, then race two teams to find pi.

· · Updated 2026-11-04 · 3 min read · code & data · ✓ Delta-reproducible

circlespicomputer experimentsfor young readers

#The question

Can tidy dots help you find a circle's secret number? Play to find out. Your progress is kept on this device.

The pi game, in five stages. The darts in Stages 2 and 3 are thrown by your browser. The race in Stage 4 shows our recorded results.

#Specs: system & compute

Deploymentdomestic
ImagePyTorch CUDA13 (GPU)
CPU2
Memory8Gi
GPU1 × GPU (8000 MiB each)
Run time1 min 23 s
Run ID1d2021b47c2149b09373698ffc570f77
Date2026-11-02
Packages
36 packages

contourpy==1.3.2, cycler==0.12.1, filelock==3.29.0, fonttools==4.65.0, fsspec==2026.4.0, Jinja2==3.1.6, kiwisolver==1.5.1, MarkupSafe==3.0.3, matplotlib==3.10.9, mpmath==1.3.0, networkx==3.4.2, numpy==2.2.6, nvidia-cublas-cu12==12.1.3.1, nvidia-cuda-cupti-cu12==12.1.105, nvidia-cuda-nvrtc-cu12==12.1.105, nvidia-cuda-runtime-cu12==12.1.105, nvidia-cudnn-cu12==9.1.0.70, nvidia-cufft-cu12==11.0.2.54, nvidia-curand-cu12==10.3.2.106, nvidia-cusolver-cu12==11.4.5.107, nvidia-cusparse-cu12==12.1.0.106, nvidia-nccl-cu12==2.21.5, nvidia-nvjitlink-cu12==12.9.86, nvidia-nvtx-cu12==12.1.105, packaging==26.3, pillow==12.2.0, pyparsing==3.3.3, python-dateutil==3.0.0, six==1.17.0, sympy==1.13.1, torch==2.5.1+cu121, torchaudio==2.5.1+cu121, torchvision==0.20.1+cu121, triton==3.1.0, typing_extensions==4.15.0, tzdata==2026.6

Recorded from the actual Delta run, not typed by hand.

#What we did

Guessing with random dots is called a Monte Carlo method [1]. Team Halton uses a Halton sequence, a recipe of evenly spread points [2].

Dart number 6: 6 in base 2 is 110, flipped it is 0.011, which is 0.375. 6 in base 3 is 20, flipped it is 0.02, which is 0.222.
How the recipe places dart number 6. You can skip this and still follow the game.

We started with 64 darts and doubled up to just over a billion, playing 50 games at each size. Before each game we slid the Halton pattern by a random amount, with darts wrapping around the edges [3]. We then compared each team's average miss.

def inside_circle(x, y):
    """True for each point inside the quarter circle (distance from the corner <= 1)."""
    return x * x + y * y <= 1.0

#Results

Team Halton had the smaller average miss at every size we tested, and its lead grew.

Two squares, each with 256 dots. On the left the dots are random and clumpy. On the right they are spread evenly.
Team Random (left) and Team Halton (right), 256 darts each. These two pictures happen to give the same guess, which is why we repeated the game.
A bar chart. With 64 darts, neat darts are 1.8 times more accurate. With 4,096 they are 6.6 times, with 262,144 they are 17 times, with about 17 million they are 68 times, and with about a billion they are 144 times more accurate.
How many times smaller Team Halton's average miss was.
Number of darts Team Random's guess Team Halton's guess
64 3.1875 3.125
262,144 3.1396636963 3.1419067383
about a billion 3.1416619085 3.1415923089
The real pi 3.1415926536 3.1415926536

These guesses are from a single game. At about a billion darts, Team Halton's average miss over 50 games was about 143.6 times smaller.

A chart with two falling lines. The random line falls gently. The neat line falls more steeply, but not as steeply as the dotted textbook line.
Average miss as we add darts. Each large step along a scale means ten times as much. The dotted orange line shows a faster pace that neat points can approach in suitable problems. Our game did not improve that quickly.

#What this does and doesn't show

  • One shape on a flat square. Other shapes, recipes and puzzles could behave differently.
  • A trend is a clue, not a rule forever. Our lines describe only the sizes we tried.
  • Rounding. We did not separately test whether tiny rounding effects changed the last few results.
  • The tricky edge. Only darts near the curved edge can tip from inside to outside. We did not prove that this caused the pace we saw.

Code & data

Everything needed to check the numbers above. Click a file to read it here, with line numbers.

#Author verification

✓ Signed by @seawan on

I played all five stages and read the whole post to make sure a child can follow it. I checked that the main numbers match the results file from the Delta run, and that the post has the Delta-reproducible badge and run record.

Written by the author, not generated. The build only checks that it is present.

References

  1. Nicholas Metropolis and S. Ulam. The Monte Carlo Method. Journal of the American Statistical Association, 1949. doi:10.1080/01621459.1949.10483310 ✓ verified (Crossref)
  2. J. H. Halton. On the efficiency of certain quasi-random sequences of points in evaluating multi-dimensional integrals. Numerische Mathematik, 1960. doi:10.1007/bf01386213 ✓ verified (Crossref)
  3. R. Cranley and T. N. L. Patterson. Randomization of Number Theoretic Methods for Multiple Integration. SIAM Journal on Numerical Analysis, 1976. doi:10.1137/0713071 ✓ verified (Crossref)
  4. Harald Niederreiter. Random Number Generation and Quasi-Monte Carlo Methods. Society for Industrial and Applied Mathematics, 1992. doi:10.1137/1.9781611970081 ✓ verified (Crossref)
  5. Zhijian He and Xiaoqun Wang. On the convergence rate of randomized quasi-Monte Carlo for discontinuous functions. SIAM Journal on Numerical Analysis, 2015. doi:10.1137/15m1007963 ✓ verified (Crossref)

Every reference above was checked by an automated script against arXiv, Crossref or Semantic Scholar (title, authors, year), and the cited claim was compared against the paper's abstract.

#Reproduce

✓ Delta-reproducible

npx -y github:seawan/vibe-research#0fae223b2dded3030a8738c721afc659672ee84f seawan/monte-carlo-vs-quasi-monte-carlo

Pinned to the commit that last changed this post. Rerun spec: delta.yaml. Resources: 2 CPU · 8Gi memory · GPU: 1 × GPU (8000 MiB each) · up to 60 min · domestic.

Need Delta? Register for delta-cli first.

This is a “vibe research” report: the experiments and the write-up were produced with help from an AI assistant and checked by the author. Numbers come from the linked code and data; references were machine-verified.