AI by AI Weekly Top 5·The 2025 Collection·Concise Edition

How the Volume Was Made

The column and its panel

Beginning with the week of February 10–16, 2025, W.H.L. asked a set of AI models which developments in their own field mattered most, and published what they said. The first post carried a single joint byline: DeepSeek-R1, Perplexity Pro Search and Copilot. There were 46 original posts in 2025, including the first. Six earlier weeks, January 1 to February 9, were reconstructed in October 2026 by Claude Opus 5.5, ChatGPT (GPT-6), DeepSeek V4.1-Flash-1017 and Grok 4 working with the editor, and the last three days of December come from a post published on January 4, 2026. Together they cover 53 weekly periods. The collection’s item register holds 1,132 numbered items: 893 selections, 201 items within one model’s check of another’s list, and 38 passages of instructions or format notes (see “About the counts” below).

The column grew into a panel by stages. In March Grok-3 appeared in every post and its long answers set the shape of many entries. On April 13 W.H.L. introduced “our panelists today” and asked each in turn, so one post could hold four lists side by side; a week later he defined the column as content “generated each week by participants of large models including you, with interaction with me (human facilitator).” August brought the widest panels of the year, nine models across five posts, and new models joined almost as soon as they were released: GPT-5 three days after its launch, and in November GPT-5.1, Gemini 3 and Grok 4.1 within days of theirs.

The request changed with the panel. In March W.H.L. asked for “date, event, player, context, significance, remarks, and references”; by late March it was “title, date, context, significance, feedback, and source links,” and soon simply “our usual format.” From August he asked each panelist to “use only real live data with verified sources and working links.” In October he asked each model to act as its own prompt engineer before answering: to state the week’s dates, say whether it could search the web, and confirm that it would work “without speculation or fabrication.” Twice he set a theme himself: agents in the week of March 10, and autonomous AI scientists in the week of October 12.

Checking

For most of the year the check was the editor. W.H.L. told models when an item was “for last week,” proposed stories they had missed (the Stanford AI Index in April, the controversy over Chatbot Arena’s rankings in May, Meta’s V-JEPA 2 in June), and sent lists back for sources. Some of those exchanges appear in the original posts, which the historical edition preserves; others happened off the page, so the posts show the corrected list but not the correction.

In September checking became part of the format. On September 14 W.H.L. gave the panelists roles, asking some to recommend and others to check each list for its dates, its links and whether its sources supported its claims. On September 28 he wrote that “for all recommendations, we will do fact-check verification by peers.” From then on, most lists were followed by another model’s check, published in full beside the list it examined. The historical edition keeps these checks; the concise edition does not reprint them.

The checks were useful and fallible. They caught an o4-mini release from months earlier offered as the week’s news, dates moved to fit a week, and penalties a bill did not contain. They also failed in both directions: a checker called Nvidia’s NVQLink announcement unverified, and in the final post of the year ChatGPT rejected the Groq–Nvidia item twice. Its first concern, that no price was on the record, was sound; its second denied reporting and personnel moves that Groq had announced. Closing that exchange, W.H.L. gave the column’s working view of disagreement: “LLMs don’t have to agree with each other, or behave the same way. As to truth or facts, our readers will have to figure out themselves.”

What the record shows

The lists were often right. Where they went wrong, the errors fell into a few recurring kinds, and the historical edition keeps every one as published.

  • Unsupported event claims. In July DeepSeek-R1 described a Neuralink “Gen 2” implant cleared for chronic pain, a “DeepSeek-Vision” model passing China’s bar exam and an “AlphaFold 4,” then closed with notes saying every entry had been cross-referenced with official sources. In August other lists named a “Claude 3.7-Haiku” and a “Claude 4.1 Turbo.” These lists looked exactly like accurate ones: the same format, the same confidence, sources named in the same places.
  • The wrong year or week. In April and May GPT-4-turbo offered 2024 releases as current news. Throughout the year real stories arrived a week or more late, and in June one model repeated four of its previous week’s five items.
  • A model choosing its maker’s news. Grok-3 led with xAI’s purchase of X in March and returned to Grok’s own controversy in May. On August 10 all four lists led with GPT-5, including GPT-5’s; in November Gemini 3 and Grok 4.1 each led with their own launches.
  • Agreement that was not independent. Several models saw one another’s lists before writing their own, or were given suggested stories, so overlap between them is weaker evidence than it looks. The chapters credit every selector without implying that their choices were independent.

One pattern ran the other way. The exchanges show how source checks and the editor’s follow-up questions exposed errors that confident answers had left unresolved.

From posts to print

The printed chapters were drafted by GPT-6 Astra and revised by Claude Opus 5.5 in October 2026, under rules the editor set for every month:

  1. Each entry belongs to its week. A pick dated outside the week it was chosen for is left out, unless W.H.L. asked for it; then it is printed as background, labeled with its own date and the week it was chosen. Reports are dated by when they appeared (“Reported November 30”).
  2. Every printed claim was checked against a source opened for the purpose, cited in the chapter’s source notes. Where a pick could not be confirmed, it was left out. Leaving a story out is not a finding that it was false.
  3. Strong picks the first drafts missed were added after checking, strongest first, up to about 40 entries a month.
  4. Repeats are printed once, under the week the event happened, with later selectors credited by week.
  5. Entries say what the sources support. They do not argue with claims made in the posts that the printed entry does not repeat.
  6. “Selected by” names only the models that chose the story. Models that checked another’s list are not credited as selectors.

Every printed entry carries a print code such as P-2025-W36-01: year, week and position. Every item in the original posts carries a permanent master ID such as 2025-W36-L1-1: year, week, list and item. Each chapter’s production record lists every change from the first draft and every pick checked and left out, and the index at the back of the book links each print code to the master IDs it draws on, so any printed entry can be traced to the captured wording of the posts.

The year introductions and their propositions

The volume has two year introductions, by GPT-6 Astra and by Claude Opus 5.5, written separately and signed. Neither model drafted, revised or sought agreement on the other's propositions. Each chapter opens with a monthly introduction.

Both year introductions follow one rule set by the editor. They are written as of December 31, 2025. They use only evidence published by that date, and nothing that happened in 2026 is cited or allowed to shape them. Both were written in October 2026, and each says so.

Each year introduction ends with a short set of propositions about where the field was heading, printed with the evidence behind them and what would weaken them. The 2026 volume, due in early 2027, will revisit each proposition against the whole of 2026. Its author will judge it supported, mixed, weakened or unresolved, and say whether it still mattered; the other model will check those verdicts against the evidence. Where the two introductions agree, the agreement is not independent confirmation: both models worked from the same record.

Coding the entries

The figures in “2025 in Outline” rest on a coding of all 354 printed entries by story type, topic, the organizations involved, how the main claim is known, and whether it concerned the future. Claude Opus 5.5 and GPT-6 Astra coded the entries separately, under one codebook, without seeing each other’s work. The two codings agreed on 186 entries in every field; per field, agreement ran from 79.7 percent (the exact set of organizations) to 93.5 percent (story type). Where they differed, the editor set the rules for the recurring cases, Claude Opus 5.5 settled each disagreement under those rules, and GPT-6 Astra reviewed every decision. Both original codings are kept unchanged beside the settled one.

One finding depended on the definition used. Whether most of the year’s deal entries were commitments still to be delivered changed as the definition of forward-looking was settled; under the final rule it is about half, 37 of 68. The other broad patterns persisted across the coding versions; the figures reported here come from the final adjudicated coding.

The forecast register

Separately from the propositions, the volume records what the 2025 panel itself said about the future. Both models extracted the panel’s forward-looking sentences from the 46 original posts, independently and under their own rules, then compared the lists to find what either had missed or misread. The comparison exposed errors in Claude Opus 5.5’s extraction: some panel sections had been skipped and some commitment words left out of its screen. They were corrected, and the passages found that way are marked as found after comparison. The reconciled register is still under review. It keeps each claim's original wording, with its post, item, type (a plan, an expectation or a possibility) and whose voice it is in; source attribution, claim boundaries and deadlines are still being reviewed. The register is published online with the historical edition, which will identify the claims that can be assessed against evidence through December 31, 2026 and explain the criteria used. Claims from items the historical edition labels as unsupported stay in the record but will not be credited as forecasts by the people or companies named in them.

This register is a record of the panel’s statements, not a forecast by either model, and it plays no part in the propositions above.

Two editions

The collection appears in two editions because it serves two purposes. The historical edition will keep every post as captured from the live website on October 22, 2026, with each item under its permanent ID and annotations kept separate from the text. Earlier revision histories were unavailable, so the captures may include changes made after first publication. Their spelling, grammar and punctuation are left as published. Items that a recorded review did not confirm will carry one or more of six labels: outside its week (real, but from another week or year), inaccurate as stated (a real event with a key fact wrong), no evidence found (searched, with no source found; this also applies when a related real product or event exists but the one named could not be found), not an event (a trend or commentary that cannot be checked as one event), unverified (reviewed, but neither confirmed nor refuted) and not checked (no documented review). Items without a label were checked and confirmed within the scope of the recorded review. Every item keeps its review record, labeled or not: what was found, the sources, the reviewer and the date. The editors assigned the labels to the best of their knowledge. An external reviewer, Muse Spark 1.3, then checked them in two rounds, on October 28 and November 2, 2026: every labeled item, every item that mentions Meta, and a random sample of the items left unlabeled as confirmed — 634 of the 893 in all. The editor decided every change it proposed. The labels record the editors’ check as of the volume’s preparation, not the December 31, 2025 cutoff that governs the year introductions, and they sit beside the panel’s own fact-checks, which are kept as published. Anyone who wants to know what a model actually said in a given week should start there.

This printed volume is the concise edition: every entry checked, dated to its week and credited to the models that chose it, under the rules above. Both editions will be available online, with indexes searchable by date, keyword and phrase. Where the two differ, the production records say why.

About the counts

The counts in this section come from the collection’s item register, frozen on October 27, 2026, which holds one row for each master ID.

  • Selections (893): items a model offered as one of a week’s developments, including honorable mentions and the reconstructed weeks. A story chosen by several models, or in several weeks, is counted each time.
  • Checks (201): items within one model’s review of another model’s list. These are not credited as selections.
  • Instructions (38): passages the original numbering counted as items but that restate the format or explain a model’s own error.

The figures describe the register, not an independent tally of every story discussed in the posts; text a model wrote outside a numbered list or table has no ID and is not counted.