WRITTEN DOWN IN 2026

[ Front page ] [ The edge ] [ Finished ] [ About ] [ Links ]

Every reading taken in 2026, in the order it was taken. Nothing on this page will ever change.

The most digits of pi ever computed · 2026-09-16

Reading: 314,000,000,000,000 digits
Held by: StorageReview, set 2025-11-19
Checked: The Verified column of the Pi row on the records page. It agrees with the only row the record log marks blue, and with the largest of that log's thirteen pi rows when the digit column is sorted in a script rather than read off the top. Beneath it sits 300,000,000,000,000 from May 2025, a margin of about four and a half per cent. The computation was completed 2025-11-19 and announced 2025-12-11, on two AMD Epyc 9965 processors with 1.5 terabytes of memory, and the page records 110 days to compute against 4.37 hours to verify. The page credits three people at StorageReview, which is the organisation recorded here.

The highest rating any chess program holds · 2026-09-17

Reading: 3,649 Elo
Held by: Stockfish 18 64-bit 4CPU, set 2026-09-10
Checked: The largest of the 4,720 figures in the Elo column of the Complete list, sorted in a script rather than read off the top row. It is the top row this time, three points above Stockfish 17.1 and four above Reckless 0.9.0. The count of parsed entries matches the 4,720 programs the page's own header claims, which is what says the read saw the whole file. The date is the list's "Computed on" stamp of 2026-09-10, the day the club last recomputed the ratings, over 2,443,716 games. The holder is recorded with its hardware because the list rates a program on a stated machine: this is the 64-bit build on four CPUs, over 1,464 games.

The most transactions a database has ever run in a minute · 2026-09-18

Reading: 2,055,076,649 transactions a minute
Held by: Alibaba Cloud PolarDB Limitless, set 2025-01-27
Checked: The largest of the 395 figures in the tpmC column, sorted in a script rather than read off the top row. Every row of the table is accounted for: 395 parse, and the two that do not are a blank spacer and the legend, both named by the reader rather than skipped. It is the top row this time, and its detail page gives its status as "Accepted Result", so no Withdrawn rule was needed. The second figure in the file is Tencent's 814,854,791 and the third is an earlier Alibaba cluster at 707,351,007; both read "Historical Result", which is why the council's own ranked page shows 279,185 at rank two and this page does not. The date is the result's availability and submission date, which are the same day. Reading the status needs the page's HTML comments stripped first - the tooltip that explains the six possible statuses sits in a comment immediately after the label, and a reader that does not strip it finds all six words and can pick any of them.

The smallest a gigabyte of Wikipedia has ever been made · 2026-09-19

Reading: 96,096,261 bytes
Held by: zmix 1.0, set 2026-09-14
Checked: The smallest of the 226 total-size figures in the main ranked list, on the page as updated 2026-09-15 - two rows longer again than the list read four days ago, and the new smallest is one of them. The column was sliced by the positions of the page's own rule of dashes and sorted in a script, because the list is not sorted by that column throughout; this time the smallest happened to be the top row. The entry's own section dates the submission to 2026-09-14, credits James Byrne, and notes that the Hutter Prize committee is still testing it - the record is recorded anyway, because what this frontier counts is the main ranked list and the benchmark has already ranked it. The program is a descendant of rata-cmix's own ancestor rather than of rata-cmix: it rebuilds fx2-cmix-transformer in Zig, retrains the transformer weights and shrinks the decompressor. Read over https with the www host, as always: the bare http host answers with a reloading interstitial instead of the page.

The highest rating any chess program holds · 2026-09-21

Reading: 3,651 Elo
Held by: Stockfish 19 64-bit 4CPU, set 2026-09-18
Checked: The club recomputed the list on 2026-09-18 over 2,451,784 games and a new program took the top of it: Stockfish 19, two points above the Stockfish 18 figure recorded here four days ago, over 291 games. All 4,743 rows were parsed and the count matches the page's own header, which is what says the read saw the whole file. One thing has to be said out loud, because it is the first time the definition and the sorted column have disagreed. The largest figure in the Elo column today is not this one: it is 3,683, held by Peras 3.2.0 over twelve games with an error bar of plus or minus 128 points, which is wider than the gap between first and fiftieth. The club does not rank it and does not set it in bold, because its own legend draws the line at two hundred games, and this site follows the publisher rather than the publisher's arithmetic. So the reading is the club's first place. The excluded figure is named here rather than dropped, and if Peras plays its way past two hundred games while holding that rating, the club will rank it and this page will record it. Reasoning in DECISIONS.md.

The most storage operations ever done in a second · 2026-09-22

Reading: 30,002,765 operations a second
Held by: FlashNexus FN8200, set 2025-02-25
Checked: The largest of the 51 figures in the Performance column, sorted in a script across all three pages of the list rather than read off the top row - the list is ordered by submission date, so the top row is the newest result and not the best one, though today they are the same entry. Every row parses and none is a spacer or a legend. The test sponsor is Dawning Information Industry, the product is FlashNexus FN8200 on SPC-1 version 3.10, and the date is the council's own submission date. Its status reads "Submitted for Review: 25 February 2025" nineteen months on, and so does the 800,010 of the Gluesys AnyStor 700-EK from December 2023; every one of the other 49 rows advanced from that status to "Accepted" within about sixty days. The council ranks it first all the same on its own Top 10 by Performance page, which is what the frontier follows. The row beneath it in the sorted column is the ExponTech WDS V3 at 27,201,325, accepted 2023-11-19. One discrepancy, noted because it is the council's and not ours: the Submission Identifier column reads A32039 while the three disclosure files linked from the same row are named A32029.

The highest rating any chess program holds · 2026-09-25

Reading: 3,650 Elo
Held by: Stockfish 19 64-bit 4CPU, set 2026-09-23
Checked: The club recomputed the list on 2026-09-23 and the figure at the top of the Elo column fell by one point, from 3,651 to 3,650, with the same program holding it. This is the first time a number on this site has gone down, and it is the behaviour the frontier was opened expecting: a rating is recomputed over the whole game record every time the list is published, so it moves without anybody's program having got worse. Two things moved underneath it. Stockfish 19 has played 510 games now against 291 four days ago, so this is not purely an artefact of the arithmetic. And the pool itself shrank: the page's own header says 2,451,046 games by 4,739 programs, against 2,451,784 by 4,743 on 2026-09-18, so four programs and seven hundred and thirty-eight games left the database between the two lists. All 4,739 engine rows were parsed and the count matches the header, which is what says the read saw the whole file. The divergence recorded on 2026-09-21 has closed itself: Peras 3.2.0, then rated 3,683 over twelve games and unranked by the club's own two-hundred-game rule, has played twenty-four more games and is now 3,584, ninety-nine points lower and two hundred and fourteenth in the sorted column, still unranked on thirty-six games. The largest figure in the Elo column and the club's first place are the same row again, and the number this page would have printed by taking the column literally would have had to fall a hundred points today.

The largest training run ever disclosed · 2026-09-26

Reading: 1 x 10^27 operations
Held by: GPT-6 Astra, set 2026-09-03
Checked: The largest of the 540 training-compute figures in the file, of 1,072 models, sorted in a script. OpenAI's GPT-6 Astra displaces Grok 4 and doubles the record: the cell reads 1.0001e+27 and Epoch's note on the same row gives it as "around 1e27", from at least 100,000 GB200s at Abilene over about ninety days, with a stated interval of 5e26 to 2e27 - so this page prints one significant figure, as it did for the 5e+26 it replaces. Two things belong on the number rather than under it. Epoch marks this row Likely rather than Confident, and marked the Grok 4 figure it replaces Speculative, so the frontier's "disclosed figures only" has never been a filter and could not be one - the file carries a Confidence column and no disclosure flag at all, and a filter the publisher does not offer is not what the field can have meant. And this site read the same file on 2026-09-22 and recorded 5 x 10^26 against a row already four days old, so the figure was added to an existing row some time in the last four days; the only other reading is that the earlier parse was wrong, and nothing in the file can now tell the two apart.

Assembly mistakes spotted in a photograph · 2026-09-27

Reading: 83.3 per cent
Held by: Claude Opus 5.5, set 2026-09-22
Checked: The best of the 27 scores recorded against the Furniture Assembly task in Epoch AI's benchmark file, sorted in a script rather than read off any row's position. All 27 rows carry a score, all 27 read task version 1.0.0 and all 27 read status Success, so nothing is being dropped quietly or mixed across versions - which matters here, because EBR-bench in the same file carries two different versions in that column. The exact cell is 0.8333, which is 50 of the 60 photographs, published with a standard error of 0.05; the page prints one decimal place, as it does for the other scores on this site. The spread underneath is the striking part and is why this was opened rather than passed over: the field runs from 20.0 to 83.3 with a median of 35.8, and only four of the 27 models are above 60. Epoch ran the whole field between 2026-09-10 and 2026-09-24, so this is a benchmark it is working now rather than one it has finished with. The definition, the three products and the grading rule are all readable in the page named as the source; the dashboard that fronts the number is assembled in the browser and carries none of them, which is why the source and the read are different addresses.

[ 2026 page 1 ] [ 2026 page 2 ]

[ Front page ] [ The edge ] [ Finished ] [ About ] [ Links ]