projects

desert is.land

Jul 2026

Every Desert Island Discs castaway names the eight records they would take to the island. desertis.land reads your Spotify top artists and tells you whose eight sit closest to yours: your castaway twin.

The match is rarity-weighted, so sharing an obscure rockabilly singer counts for more than sharing The Beatles; everyone shares The Beatles. What comes out is a sentence — my castaway twin is Alfred Hitchcock — and a page you can send to the group chat, card and all.

Behind the party trick is the substance: since 1942 the BBC has been asking famous guests the same question. If you were cast away alone on a desert island, what would you take? What eight songs; what book; what luxury. desertis.land is every episode over the decades with the items and reasons extracted, searchable by the half-remembered impression rather than the name.

The desertis.land landing page: a mid-century tuning dial beside the headline 'Eight records, a book, and one luxury', with presenter eras marked along a long-wave scale
The wireless:

The five presenters sit as stations along the long-wave scale.

What they choose

Tally every request since 1942 and the winner is a song about regretting nothing.

castaways
  1. 1String Quintet in C major, D. 956 Franz Schubert64
  2. 2St Matthew Passion, BWV 244 Johann Sebastian Bach61
  3. 3Cello Concerto in E minor, Op. 85 Edward Elgar46
  4. 4Piano Concerto No. 2 in C minor, Op. 18 Sergei Rachmaninoff45
  5. 5Nimrod (Enigma Variations) Edward Elgar39
  6. 6Ode to Joy (Symphony No. 9 in D minor, Op. 125 'Choral') Ludwig van Beethoven38
  7. 7Soave sia il vento (Così fan tutte) Wolfgang Amadeus Mozart37
  8. 8Non, je ne regrette rien Édith Piaf36
  9. 9Mass in B minor, BWV 232 Johann Sebastian Bach34
  10. 10The Four Seasons Antonio Vivaldi34

The records age slowly; the luxuries are carbon-datable. Pianos and golf clubs belong to one era, guitars to another, and nobody has asked for painting materials since the sixties. My favourite oddity: nine separate castaways asked for a Michelangelo sculpture, most of them in the sixties, as though the decade briefly agreed marble was a reasonable thing to want on an island.

Pianohalf as popular since the 2010s
1950s — 4.0% of episodes1960s — 4.0% of episodes1970s — 3.3% of episodes1980s — 4.3% of episodes1990s — 5.7% of episodes2000s — 5.9% of episodes2010s — 2.3% of episodes2020s — 3.0% of episodes5.9%1950s2020s
Guitarclimbing since the fifties
1950s — 0.9% of episodes1960s — 1.6% of episodes1970s — 0.2% of episodes1980s — 1.5% of episodes1990s — 1.4% of episodes2000s — 1.7% of episodes2010s — 2.7% of episodes2020s — 2.4% of episodes2.7%1950s2020s
Golf clubsa sixties fixture, fading since
1950s — 1.7% of episodes1960s — 2.2% of episodes1970s — 0.6% of episodes1980s — 0.4% of episodes1990s — 0.5% of episodes2000s — 0.7% of episodes2010s — 0.5% of episodes2020s — 0.3% of episodes2.2%1950s2020s
Painting materialsnot chosen since the sixties
1950s — 2.0% of episodes1960s — 1.8% of episodes1980s — 0.9% of episodes1990s — 0.2% of episodes2000s — 0.7% of episodes2020s — 0.3% of episodes2.0%1950s2020s
Michelangelo's sculpturenine requests, mostly the sixties
1960s — 0.8% of episodes1970s — 0.6% of episodes1990s — 0.2% of episodes2010s — 0.2% of episodes0.8%1950s2020s
Cigarettesnot chosen since the nineties
1950s — 0.9% of episodes1960s — 0.6% of episodes1990s — 0.2% of episodes0.9%1950s2020s

Weighing the evidence

A raw tally is unfair to anything young. Elgar’s Cello Concerto has had every episode since 1942 to be chosen in; a song released in 2005 has had a few hundred. So the records chart now ranks by rate rather than count: picks per eligible episode, where eligibility starts the year the record existed.1 That means composition year for the classical canon and first release for everything else, checked against an unforgiving constraint: nothing may be picked before it exists.

Small counts lie, though. Two picks in forty episodes is a wonderful rate and terrible evidence. Every rate is therefore shrunk towards the archive’s average (a Gamma prior worth a few hundred imaginary episodes, if you want the incantation), and the chart sorts by the estimate’s cautious lower bound rather than its middle; read each entry as “we’re three-quarters sure it is at least this popular”.2 Fix You has 8 picks in the 886 episodes since 2005. The naive rate is nine per thousand; the shrunk figure says six; the gap is the model declining to be excited by thin data. The warhorses barely move under any of this. The action is in the mid-table, where Eminem now sits at #40 on nine picks.

The same machinery, pointed at time, finds the risers and fallers. The archive splits at 2000, and each item is measured against its own era’s base rate, so a growing catalogue doesn’t make every old record look like a faller. The chart then ranks by how confidently the pick rate moved between the halves.3 Confidence scales with evidence (going from 3 picks to 6 outranks going from 1 to 3, identical ratios notwithstanding), which is why the captions promise “at least”: cigarettes fell at least ninety-fold, and the drum kit rose from nowhere.

Recovering the run

The BBC’s own programme listing only carries episodes with surviving audio. For the full run you need BBC Genome, the digitised Radio Times listings. Several of the early decades the BBC taped over.

The result is 3,431 episodes and 3,293 castaways, including hundreds of broadcasts nobody can listen to any more. For those the site can only show the listing: a name, a date, no audio.

Episode pages play the original recording with a transcript that scrolls in time with it; go and find your twin.

Section titled Footnotes

Footnotes

  1. For record rr with eligibility year trt_r, exposure NrN_r counts disc-extracted, dated episodes broadcast in or after trt_r (2,558 all-time). Eligibility years cover 9,880 of 11,368 canonical songs, derived by a two-pass LLM sweep (bulk pass, then an independent blind re-derivation of a 10% sample; disagreements beyond ±2 years re-done at higher effort) and validated against first-pick causality: any claimed trt_r later than the record’s first pick is demoted to undated. Undated records take the full-archive exposure, the same treatment as a pre-1942 work. ↩

  2. Gamma–Poisson empirical Bayes: kr∣λr∼Poisson(λrNr)k_r \mid \lambda_r \sim \text{Poisson}(\lambda_r N_r) with λr∼Gamma(a,b)\lambda_r \sim \text{Gamma}(a, b). The prior is fitted by method of moments over all qualifying records (≥2 choosers): prior mean a/b=∑k/∑Na/b = \sum k / \sum N; strength bb from the excess of Var(k/N)\text{Var}(k/N) over the expected Poisson sampling noise, clamped to [50, 5000] pseudo-episodes. The fitted strength is then deliberately softened by 4 (fitted b≈666b \approx 666, used b≈166b \approx 166): a parameter sweep showed the top of the chart is insensitive to bb, while newcomers were over-shrunk. Posterior λr∣kr∼Gamma(a+kr,b+Nr)\lambda_r \mid k_r \sim \text{Gamma}(a + k_r, b + N_r); the display shows the posterior mean per 1,000 episodes and the ordering uses the posterior 25th percentile, computed with the Wilson–Hilferty cube-root normal approximation to keep the API scipy-free. ↩

  3. Windows split at 2000-01-01; tracks clip both windows by eligibility year as well, while luxuries and books use plain calendar windows. Priors are fitted separately per window (unsoftened), so each rate is measured relative to its era’s base rate and catalogue-growth dilution cancels. The shift is ρ=log⁡2(λnow/λthen)\rho = \log_2(\lambda_{\text{now}} / \lambda_{\text{then}}) with posterior quantiles from 20,000 Monte-Carlo draws per item (fixed seed, vectorised); risers rank by P10(ρ)P_{10}(\rho) descending, fallers by P90(ρ)P_{90}(\rho) ascending. A measurability floor requires ≥200 episodes of exposure in each window and ≥4 picks in total. Captions display 2P102^{P_{10}} (or 2P902^{P_{90}}) as an “at least” bound rather than the median, which is prior-dominated whenever a window holds zero picks. ↩