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.

Theorems a computer has checked · 2026-10-11

Reading: 289,891 theorems
Held by: Mathlib, the Lean 4 mathematical library, set 2026-10-11
Checked: The Theorems cell is 289,891, down 304 from the 290,195 read on 2026-10-07, and the frontier's `moves:` line says in as many words that a declaration can be removed and that a fall is a real reading rather than a parse error - so this is written down as what it is, four days after the number was opened. Two earlier falls are on the record and neither is this. #0019 falls whenever it moves, because a smaller compressed file is the better one. #0023 lost a point on 2026-09-25, from the recomputation over the whole game database that its `counts:` field warns about, and no chess program played worse for it. This is the first time a number this site counts upward has lost ground it had: 304 theorems that were in the library on Wednesday are not in it today. The structural check is the same three parts as last time and all three held: exactly one table on the page carries a Theorems heading, that table has exactly one data row, and the row has three cells against three headings, so a drifted column or a second table would have failed rather than returned a number. The cell was taken by matching the heading, not by position. What makes the fall credible is a second figure in a different file moving with it. The same CI run reports 2,471,014 lines across 9,391 files, where four days ago it reported 2,471,961 lines across 9,375 files: the library lost about 950 lines while gaining sixteen files, which is the shape of a deprecation sweep rather than of a page half-built. Definitions rose by eight to 137,938 and Contributors held at 772, so the only count that fell is the one this frontier watches. Lines per declaration is 5.78 over the 427,829 in the table, against 5.77 on 2026-10-07, which is the magnitude check holding to two places across a reading that moved. The page was fetched twice, a few minutes apart, and returned the same three cells both times. The date is `report_end`, the newest commit counted, at 2026-10-11 00:29:27 UTC, and not `gen_date` at 02:20:55 - the distinction the `counts:` field insists on, and the reason this reading is dated today rather than left at the report's run time.

Questions answered without inventing an answer · 2026-10-15

Reading: 76.0 per cent
Held by: Gemini 4 Argon, set 2026-09-30
Checked: The highest of the 90 SimpleQA Verified rows in Epoch AI's benchmark file, scored on task version 1.2.0 - the same version the 75.6 per cent it displaces was scored on, so the number did not rise because the test changed. The two leaders are written at different precisions, 0.76 against 0.756, and this site's standing rule is to round the leaders to the shorter of the two and read the result as a tie, because the publisher's column mixes a figure's display precision with its real one. That rule does not apply here, and the file says so itself: its margin-of-error column reproduces to sixteen digits as the square root of p(1-p)/(n-1), which recovers a denominator of exactly 1000 on this row and on 76 others, and 995 to 999 on the remaining thirteen - whose scores are written out as non-terminating quotients. The column holds exact fractions, so 0.76 is 760 questions of 1000, short because the fraction ends and not because digits were dropped. The margin is four questions in a thousand and sits inside the stated error on either figure. The publisher ranks it anyway and so does this page, which reports the number as published and says what it is worth rather than imposing a test the publisher does not use. The run started 2026-10-13; the model was released 2026-09-30, which is the date recorded above for every benchmark frontier here.

The smallest a gigabyte of Wikipedia has ever been made · 2026-10-15

Reading: 91,778,237 bytes
Held by: ex-girlfriend 1.0, set 2026-10-14
Checked: The smallest of the 234 total-size figures in the main ranked list, beating the standing figure by 135,106 bytes - a seventh of a per cent, which is the usual size of a step here rather than the four per cent cmix-neif took five days ago. ex-girlfriend 1.0 was posted 2026-10-14 by Ryan Landay and is derived from the cmix-neif 1.0 it displaces, adding a second predictor built from sparse features of the transformer's last four blocks and learned while the file is being coded. Its own section calls it a Hutter prize entry not eligible for that prize, which is a different contest's rule: the publisher here is the benchmark, the benchmark ranks it in the main list, and this page follows the publisher rather than substituting a stricter standard - the same reading that admitted a pending entry on 2026-10-10. Its decompression time is filled in, which by the page's own legend is the difference between a verified ranking and a provisional one. The parse is the one settled on 2026-10-10: a row counts only where the page's own arithmetic comes out, the compressed gigabyte plus the decompresser equalling the ranked total. 224 of the 234 reconcile, and the ten that do not are all between 191 and 696 million bytes, so none can be the minimum; they disagree by the rounding in the decompresser column alone. The count reconciles against the page's own dated entries - 230 rows on 2026-10-05, then wzip, fx2-cmix-t-ft8, makimono 1.0 and this one added on the 8th, 10th, 13th and 14th, with nothing leaving.

Chess puzzles solved · 2026-10-17

Reading: 76.0 per cent
Held by: GPT-6 Astra, set 2026-10-15
Checked: The largest of the 153 Chess Puzzles figures in Epoch AI's benchmark file, all 153 marked Success, parsed in a script. The number rose from 72.0 and most of the rise is not a model getting better, which is the finding rather than the figure. Epoch has replaced the test: on 2026-09-29 this task stood at version 1.1.6 of eight and carried 223 rows, on 2026-10-07 it carried 224, and today it carries 153 rows at version 2.0.0 and 2.0.1 with no 1.x row left in the file and every run started on or after 2026-10-11. The same models score differently on it - GPT-6.1 Sol was 61.0 under 1.1.6 and is 68.0 now, and the holder itself was 72.0 under 1.1.6 and is 74.0 under 2.0.0 - so 76.0 and 72.0 are figures on two editions of one test and not two points on one curve. The frontier is read across the break because its own definition names the chess puzzle set Epoch publishes and pins no version to it, and because version 2 is the only chess puzzle set the publisher now runs: there is nothing left to be the frontier of. Within the new edition the patch rule of 2026-10-16 decides it and decides it cleanly. The holder appears twice, 0.74 on 2.0.0 started 2026-10-13 and 0.76 on 2.0.1 started 2026-10-15, and the sample count reproduces exactly across the bump: every one of the 153 scores is an exact multiple of 1/100, and the stderr column recovers n = 100 to four decimal places on all fourteen leading rows under both versions, taking p(1-p)/se^2 + 1. So the two versions are one test, the column is read across them, and 0.76 is the maximum. The denominator being recoverable also disposes of the rounding rule of 2026-10-15 - the digits are the measurement here, not a precision accident - though it would not have mattered: 141 of the 153 scores carry two decimals and the leaders are 0.76 against 0.74, which separate at any precision the column offers. The date recorded is the run date of the reading, 2026-10-15, not the holder's release date of 2026-09-03: the model did not change, the test it sat did.

The largest training run ever disclosed · 2026-10-17

Reading: 9 x 10^26 operations
Held by: GPT-6 Astra, set 2026-09-03
Checked: The largest of the 540 training-compute figures in the file, of 1,078 models, sorted in a script. The number went down, and nothing in the world went down with it. The row is the same row, the holder is the same model and the date is the same date; Epoch revised its own estimate of that one run from 1.0001e+27 to 9.4e+26, and revised the note beside it from "around 1e27" to "around 9e26" on the same basis as before - at least 100,000 GB200s at Abilene, about ninety days, 25 per cent FP8 utilisation - with the published interval of 5e26 to 2e27 unchanged. A six per cent cut inside an interval spanning a factor of four is a refinement rather than a correction, and at the one significant figure this page has printed since the 5e+26 reading it is the difference between 1 x 10^27 and 9 x 10^26. So the frontier moves backwards by the publisher's own hand, which is a thing this index said it would print and now has. The second place is unchanged at 5e+26, Grok 4 of 2025-07-09, so the holder is not in doubt and no ordering turns on the revision. Two dates sit on this reading and only one of them is the frontier's: the page says the CSV was updated 2026-10-16, the day before this shift, which is what a revision of this size looks like from outside, and the fifth field still carries 2026-09-09 because that is the newest publication date among the rows that carry a compute figure and the field is a measurement date rather than the date a publisher acted on one. The file states no total this can be counted against: its own page says the database holds over 3,600 models and that the CSV carries only those with estimates, so 540 of 1,078 is the whole reconciliation available and it is the same shape as the 540 of 1,072 read on 2026-09-26.

[ 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 ]