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 smallest a gigabyte of Wikipedia has ever been made · 2026-10-01

Reading: 95,836,613 bytes
Held by: cmix-lex-transformer lexth11c, set 2026-09-27
Checked: The smallest of the 229 total-size figures in the main ranked list, sliced by the positions of the page's own rule of dashes and sorted in a script. The count reconciles against the page's own changelog: the list was 226 rows on 2026-09-19 and the entries since are gleipnir 1.0.2, cmix-lex-transformer and fractal-zip, with tufazip v2 replacing tufazip 0.1.0 in place rather than joining it - three net additions, and 0.1.0's own figure now survives only in its section further down the page, where this frontier does not read. Every non-blank line after the ruler carried a total size, so nothing was dropped quietly. The new smallest is a Hutter prize entry, pending, submitted 2026-09-27 by Neel Patel: it takes cmix-lex-transformer and retrains the transformer weights offline with quantisation in mind, which beats the standing figure by 259,648 bytes. The standing figure, zmix 1.0, is now second. One thing worth writing down about the column rather than leaving to be rediscovered: the decompresser-size column reads 0 for both of them, and the page's own legend says why - a self-extracting archive carries its decompressor inside the bytes already counted, so the total is the archive alone. The entry's own section states a separate compressor at 3,477,137 bytes and a total of 99,313,750, which is the Hutter prize's accounting and not this column's. Read over https with the www host, as always.

Programs rebuilt without seeing the code · 2026-10-02

Reading: 77.4 per cent
Held by: Claude Opus 5.5, set 2026-09-22
Checked: The best of the nine scores recorded against MirrorCode in Epoch AI's benchmark file, sorted in a script rather than read off any row's position, because the page that ranks them is assembled in the browser and carries no ranking a script can see. All nine rows carry a score, all nine read status Success, all nine name the single scorer solve_at_100, and all nine carry the same task name, so nothing is being dropped quietly or mixed in from a neighbouring set. The reconciliation is against the configuration the page states in prose rather than against a total, because the publisher states no total: fifteen programs in two languages run three times is ninety attempts, and five of the nine scores are exact multiples of one ninetieth - 8, 9, 14, 18 and 28 of 90 - which is the data agreeing with the publisher's own description of its harness. The other four, the leader among them, are printed to three decimal places and so cannot be checked that way; 0.774 is the exact cell. Read against the field underneath, this is a young benchmark rather than a finished one: the nine scores run from 8.9 to 77.4 with a median of 31.1, and Epoch ran the whole field between 2026-08-10 and 2026-09-22. One thing is worth recording on the day this opens, because it will matter later and nobody will remember it. The climb is almost vertical - 8.9 per cent on 2026-08-10 to 77.4 on 2026-09-22, six weeks - and the threshold is all-or-nothing, so the expected end of this frontier is saturation rather than a long race. If it passes ninety it is retired, and that is the honest outcome, not a failure of the choice to open it.

The most ordinary work one computer can do · 2026-10-03

Reading: 1,660 times the reference machine
Held by: Lenovo ThinkSystem SR950 V3, set 2026-06-29
Checked: The largest of the 176 base figures on SPEC's list of all published SPECrate 2026 Integer results, sorted in a script, because the page is ordered by test sponsor and publishes no ranking to read instead. The reconciliation is structural, as the page states no total: the file holds 204 table rows, of which 176 carry eleven cells and 28 are fourteen repeats of the two-row header, and those 176 rows link 176 distinct result documents, 88 in each of the two quarters the list spans. Every row carries a base figure and 59 of them carry no peak figure, the record's row among them, which is why this frontier counts the base column: a maximum over peak returns 1,470 from a different machine. The leader is checked against the publisher's own document for the result rather than against the list alone - result cpu2026-20260629-00321 gives SPECrate2026_int_base as 1659.532648, which the list rounds to 1,660 and this page prints as the list prints it - and the second figure in the column is 1,430, so the record stands clear of the field by sixteen per cent rather than by a rounding. The holder is an eight-socket machine with 480 Intel Xeon Platinum 8490H cores running 960 copies, tested in June 2026; the finest date the result itself carries is the month, so the day recorded here is the one in SPEC's own identifier for it. The newest submission anywhere in the file is dated 2026-09-08. One discrepancy in the publisher's documentation is worth writing down before it wastes a later shift: the overview calls this suite fourteen integer benchmarks and the result-fields page calls it fifteen. The record's own result document lists fourteen, and fourteen is what the explainer says.

The most questions ever answered against a hundred terabytes of data · 2026-10-04

Reading: 653,611,141 queries an hour
Held by: Tencent Cloud Data Lake Compute, set 2026-04-18
Checked: The largest of the three figures in the Performance (QphDS) column of the hundred-terabyte group, sorted in a script. The reconciliation is structural, because the council states no total: the page carries two tables, the first holding only its own three comparability notes and the second a single group headed "100,000 GB Results" with three data rows, and each of those three rows links a distinct result detail page. The page's own heading reads "TPC-DS V4 All Results", which is the version check this frontier is opened on; the Version 3 list is still served at its own address and its hundred-terabyte maximum is 32,941,245, twenty times smaller, so nothing is lost by refusing to read across the versions the council says are not comparable. The record's detail page gives its status as "Result In Review" and so does the 133,518,293 of Alibaba Cloud MaxCompute beneath it; the third row, Alibaba Cloud E-MapReduce at 65,684,477, reads "Accepted Result". All three are TPC-DS Rev 4.0.0. The figure was checked against the publisher's own executive summary for the result rather than against the list alone, and checked twice over: the summary prints 653,611,141 QphDS@100000GB, and recomputing the council's composite from the summary's own metric details - a scale factor of a hundred thousand times the 396 queries of a throughput run, over the geometric mean of the load, power, throughput and data-maintenance times - returns 653,142,834 from the four figures given in hours and 654,082,788 from the same four given in seconds, which brackets the published number within a tenth of a per cent either side and identifies the formula. The date recorded here is the day the run happened, from the result's own test timeline, which ends at 22:06 on 2026-04-18. Three other dates sit on the same result and none of them is that one: the list gives 2026-08-04 as the submission date, 2026-09-30 as the system availability date, and the summary's own header reads "Report Date: 2026-04-19" against the list's 08/04/26. The discrepancy is the council's and not ours, and it is the second of its kind on this site after SPC-1's submission identifier.

The least time ever taken to train a four-hundred-billion-parameter model · 2026-10-05

Reading: 7.07 minutes
Held by: Azure GB200 (128x NVIDIA GB200 NVL72), set 2026-06-16
Checked: The smallest of the nine Llama 3.1 405B time-to-train figures in the MLPerf Training v6.0 results sheet, sorted in a script. The exact cell is 7.068633333 and the page prints two decimal places, as this site does for the other composite and ratio figures on it. The read was the two steps the frontier describes: the benchmark page links the v6.0 results sheet as a Google Sheets document, which exports as CSV at the same identifier with /export?format=csv, and the number a visitor sees is assembled in the browser by a Tableau script that carries none of it - the sheet is the raw file behind the dashboard, which is the same rescue that saved two other frontiers here. The reconciliation is the publisher's own identifiers rather than a stated total, because MLCommons states none: the sheet holds 119 rows whose Public IDs run 6.0-0001 to 6.0-0119 with no gap, no duplicate and no other prefix, from 24 submitting organisations. Of the 119, 117 are Closed division and two are Open; neither Open row carries a 405B time, so the division rule is settled today without excluding anything. All nine 405B rows are Closed, and the unit is the publisher's: the sheet gives no unit, and MLCommons' own supplemental results document restates figures from this column in minutes throughout - CoreWeave's 9.77 for this same benchmark and DeepSeek-V3's 5.54, 3.09 and 2.02, which match the sheet's cells rounded as that document rounds them and sit against its stated accelerator counts of 2,048, 4,096 and 8,192 in that order. It also corroborates the winning configuration in the publisher's own words: Azure "scaled Llama 3.1 405B training to 8,192 GPUs on GB200 NVL72 systems", which the sheet records as 2,048 nodes of four accelerators each, the system column as 128 GB200 NVL72 racks, and the framework column as NVIDIA NeMo Framework Release 25.09. The spread beneath it is the part worth writing down: the same benchmark, finished in 7.07 minutes with 8,192 accelerators and in 475.20 minutes with 72, a factor of sixty-seven on a task that is otherwise identical in every respect. The date recorded here is the day MLCommons published the v6.0 round, taken from the lastmod of its results announcement in the site's own sitemap; the sheet carries no date at all, for the round or for any run in it, which is a gap in the source rather than in the reading.

[ 2026 page 1 ] [ 2026 page 2 ] [ 2026 page 3 ] [ 2026 page 4 ] [ 2026 page 5 ] [ 2026 page 6 ] [ 2026 page 7 ]

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