Eternity II has an official history (a press release, a prize, a deadline)
and it has a real one, which happened on a mailing list. The
eternity2 group (originally a Yahoo group)
is where the puzzle was analysed before it existed, digitized the day it
launched, declared unsolvable within a fortnight, and where the only prize
money ever paid out was quietly traced to one of the list's own regulars.
This page tells that story from the archive itself: every event links to the
message where it happened. Part I runs from the group's founding to the first
scrutiny date, 2000 → February 2009.
The strangest fact about the Eternity II community is its birthday. Brendan
Owen created the eternity_two group in October 2000, six and a half years
before the puzzle was announced, "because we anticipated a follow up
puzzle" to Eternity I
(msg 404). The earliest
surviving messages guess at what the sequel's pieces might be, betting
wrongly on 3D tetrahedra
(msg 8).
By 2001 the group was already rehearsing the designer's problem rather than
the solver's: Günter Stertenbrink asked how one would build a puzzle with a
huge prize and only a ~1% chance of being solved in ten years
(msg 15), almost exactly the
brief Christopher Monckton's team would later execute. Eternity I veterans
drifted in and waited: Dave Clark, author of the distributed E1 solver
ESolve, rejoined in April 2001
(msg 21). When the first press
reports arrived in December 2005, with Monckton promising the sequel would
need "the lifetime of the Universe" while co-designer Oliver Riordan said
"several years", Stertenbrink drew the obvious conclusion: "So we can
conclude, the end of the universe is in several years."
(msg 34)
The announcement came on 22 January 2007: 256 pieces, US$2,000,000 for the
first correct solution, launch on 28 July, and a London Toy Fair stunt in
which a strongman destroyed "the computer containing a solution"
(msg 36). That computer, a
distributor insider later admitted, contained nothing at all
(msg 148).
The group did not wait for the pieces. Within two days it was counting edge
types on promotional photos, and Owen had derived the expected-solutions
formula that made difficulty a tunable design parameter
(msg 38,
msg 44). He published a
generator for Eternity-II-like puzzles
(msg 41); Alan O'Donnell had the
first working solver a week later
(msg 64). By 8 July,
three weeks before anyone had touched a real piece, anr_56 had estimated
the edge-type counts from launch photos and Owen had run the numbers on that
configuration: about 10^30 times harder than the benchmarks the group was
solving in minutes. "No one will solve this puzzle," he concluded
(msg 665).
Then came the famous 48 hours. Australia got the puzzle first, and Owen was
at Kmart on launch eve, "first one to grab it off their stack"
(msg 1051). He fed the piece
panels through a computer-vision program he had written in advance
(msg 789), reported the edge
distribution "as flat as could be"
(msg 1054), and confirmed the
real parameters: 5 border colours, 17 interior colours, the group's
worst-case scenario (msg 977).
Then, after Dave Clark's "nice chat to Mr Monckton", he published the first
solution-count estimate for the real piece set: roughly 5,930 solutions
with the mandatory clue, against Monckton's own figure of about 5 million
(msg 987). The puzzle was two
days old and its difficulty was already measured.
The same week set the community's norms. To verify piece transcriptions
without sharing copyrighted data, members converged on published CRC
checksums (msg 1063); and
newcomer Sergio Demian Lerner warned about "phantom sets": a stranger's
benchmark could be the real puzzle re-encoded, tricking you into handing
over a $2M solution
(msg 1210).
A week into August, the puzzle's inventor arrived in person; he came to
disqualify someone. Posting from an unverifiable Yahoo address, Christopher
Monckton declared he held copyright on the piece designs and that his
"web-trawling software" had flagged Brendan Owen's website: Owen, who had
been aiming for the lesser prize, was out
(msg 1342,
msg 1352,
msg 1358). It later emerged
the flagged file did not even correspond to the real pieces
(msg 1358). Members protested
that an anonymous forum account could hardly carry official weight; Monckton
doubled down (msg 1374,
msg 1377). The chill was real
and lasting: a year later Owen, as moderator, was still deleting derived
data tables "to be safe"
(msg 5651). It was in this
climate that Owen, showing that E2's flat piece statistics killed the
strategy that had cracked Eternity I, drew his line: "If we cannot talk
about something as fundamental as piece frequencies, then I will give up on
this puzzle now" (msg 1667,
msg 1694).
The theory came fast after that. kubzpa proved by a
parity argument that a score
of exactly 479 (one mismatched edge) is impossible
(msg 1640); hold that thought,
because seventeen months later the argument would meet a counterexample.
Owen then produced the period's signature result: assuming the designers
wanted the hardest possible 16×16, with a flat distribution and about one
expected solution, the interior colour count should be (196! · 4^196)^(1/392) ≈
17.14, hence 17 interior colours and 5 border colours. Exactly the real
puzzle (msg 1947). The 17+5
split was not bad luck; it was a bullseye on the hardness
peak, and the group had reverse-engineered
the targeting.
Theoretical claims were checked against node counts measured on shared
benchmarks. Angel de Vicente, posting as
Txibilis, built the standard suite of E2-like test boards
(msg 1610), and a duel began:
Txibilis's hand-designed fill orders
against the automated strategy-optimizer of doc_s_smith, driving
full-search node counts down by orders of magnitude; one benchmark stood at
89,794 nodes, answered within two days by 85,729
(msg 2896,
msg 2928). Mid-duel, the list
discovered who doc_s_smith was: Dietmar Wolz, finder of most of the known
Eternity I solutions (msg 2972).
The estimates converged too. kubzpa's paper put the solution count near 15
million (msg 3497); Owen
measured the search tree's branching depth by depth, finding the node count
peaks at 161 pieces placed
(msg 3147); and in April 2008
he announced he had "nailed the theory": an exact model of the search tree
whose predictions sat on top of the empirical curves, independently
cross-checked by Louis Verhaard: on the order of 10^27 CPU-years per
solution (msg 5197,
msg 5193,
msg 5209). The group's chief
theorist acted on his own numbers: he was now hunting the highest score, not
480 (msg 4996).
If one computer was hopeless, perhaps thousands were not. Dave Clark's
eternity2.net, a BOINC-based
distributed solver, launched
with the puzzle in July 2007
(msg 756) and had
1,300 members within a month, 160 of them in the USA, where the puzzle had
not even been released
(msg 2122,
msg 2132). The project
submitted a 462-edge partial to Tomy, then 463
(msg 2663), a number that
would function as the community's public score ceiling for over a year.
A prize-sharing "Eternity 2 Syndicate" followed, paying members in
proportion to placements contributed
(msg 3021).
It lasted five months. In December 2007 Clark shut eternity2.net down,
publishing the final accounting: over 1.6 TFlops of aggregate computing,
over 10^19 CPU operations, best scores in the mid-460s. His verdict: a
brute-force solution "was always clearly going to be impossible"
(msg 3511). The list
immediately spawned threads titled "Eternity2 Must Have Been Solved"; it had
not (msg 3554). The project's
files were preserved in the group archive
(msg 3633), and Clark
open-sourced his research solver
(msg 3716). He also left the
archive its best primary source on the puzzle's creation: a phone call with
Monckton, who described judges typing entropy into a generator built by the
Eternity I winners Alex Selby and Oliver Riordan, the solution printed once
and vaulted (msg 4177),
"whilst all parties were out of the room", as the Tomy distributor leaflet
he had posted at launch put it
(msg 901).
The exact-methods flank fared no better. A published
SAT encoding of the
full puzzle came with a call for anyone owning a machine with more than
16 GB of RAM (msg 4084);
integer programming died at 8×8 boards
(msg 5602). Raw speed kept
climbing: istarinz broke 100 million placements per second on multiple
cores (msg 5804), then
measured 558 million on a brand-new Core i7
(msg 6212). But against
search spaces measured in powers of forty, throughput was a rounding error.
The scores had crept up all along: 410 from Pierre Schaus in the first weeks
(msg 1568), 453 from
philippe.dupond (msg 3998),
461 from e2dude (msg 4440),
with 463 as the "current known high"
(msg 5688). The methods
changed character in mid-2008: Schaus posted his constraint-programming
paper, whose key move (remove a set of non-adjacent pieces and re-place
them optimally by solving an assignment problem) became the engine of
antminder's hybrid, averaging a 462 per day
(msg 5589,
msg 5601).
Meanwhile two people had quietly gone past the ceiling. In a long August
exchange, Max reported getting "beyond the 'don't-talk-about-limit' quite
easily"; Louis Verhaard ("Max, you are really a dangerous man!") matched
methods with him and found they had converged on the same heuristic
signature, promising full disclosure "after new-year", that is, after the 31
December scrutiny date
(msg 5767,
msg 5780). Verhaard would
reveal only the geometry: his best high-score fill orders looked like a
"comb": most rows scanned horizontally, the rest vertically
(msg 6112).
Then, on 22 September 2008, Verhaard published
his solver for anyone to run
at fingerboys.se: "This because I am stuck and my only hope to improve my
best score is by using brute force"
(msg 5940). Any prize would be
split 50-50 with the best-scoring user. antminder, whose own program needed
a week to reach 463, was blunt: eii "completely blows it away"
(msg 5950). The prize itself
was still folklore (the rules only promised a discretionary "lesser prize")
until Max traced the $10,000 figure to a Monckton interview on the French
official site (msg 6073).
Verhaard, three months before winning it, claimed not to care: he was in it
for the honour, and only for this year
(msg 6072).
The same autumn the hand-solvers surfaced: Christine Raisin down to 21
pieces left in the box lid
(msg 5935), Verhaard sorting
pieces by pattern with his seven-year-old daughter and running a
solve-by-hand mini-competition, the threads coining nicknames like "pink
swords" and "kipper ties" for the motifs
(msg 5933,
msg 5929,
msg 5958,
msg 5959). The list was never
only a solver bench.
Going into the first scrutiny date, the community was in a holding pattern:
several members refused to commit serious effort until the puzzle proved it
could survive its first deadline
(msg 6216); NickB had £20 on
there being no winner and considered his money safe
(msg 5979).
31 December 2008 came and went. Nothing. Members emailed Tomy's UK and US
addresses and left voicemail, and got no response
(msg 6245); the consumer
Careline knew of "no details as of yet, regarding whether or not there is a
winner" (msg 6295). The rules
said winners would be notified within 14 days and results published in the
London Times and New York Times
(msg 6296,
msg 6336); no such publication
ever visibly appeared. Meanwhile, on 6 January, Verhaard released his
solver's documentation and disclosed its ceiling: users had found 467 more
than forty times (msg 6275).
On 15 January, roughly the last day of the 14-day window, Henk van der
Griendt found a statement behind an unobtrusive link on the UK site only:
hundreds of entries, none complete, the 2Mprizestillopen,anda10,000
runner-up prize to Anna Karlsson of Lund, Sweden, for 467 of 480
(msg 6337). No press release,
no Times, and not a word from Monckton at any point. The list needed about
an hour to decode "Lund + 467": the entry was from Louis Verhaard's
household, submitted in the name of his wife, who had received the
congratulation email days earlier
(msg 6349). Congratulations
poured in from every regular; Max revealed his own best had been 465, each
extra edge costing his program a factor of ~30 in time
(msg 6348). The Swedish press
ran the family photo: over a year of work, 13 "seams" from a full solution,
part of the money going to the World Wildlife Fund
(msg 6374,
msg 6375).
The aftermath had three barbs. A researcher from the Lleida SAT group posted
that reaching 470 was "rather easy" with their methods, a claim never
backed by any submitted entry, as Max pointedly noted
(msg 6306,
msg 6360). Tomy told a member
there would be no further clue puzzles,
which most of the list welcomed as keeping the challenge pure
(msg 6381). And Verhaard
casually settled a seventeen-month-old theorem: 479 is reachable, by
flipping a border piece whose two border edges share a colour in a 480
solution. The 2007 parity proof had missed the outward-facing border edges,
which the score never counts
(msg 6317).
Even the community's cleanest impossibility result had a loophole; the
puzzle, unsolved, kept all of its.
The story continues#
Part II: 2009–2026
The story continues in part II: the
Hungarian "we solved it" rumor resolved, the contest's death with its
solution locked in a safe, the decade of 467, the migration that saved
this archive by days, and the record wave that took the board to 470.