Volume Preface: 2025

A weekly collection gives a year back its sequence. Announcements arrive before their consequences are known. A promising experiment attracts attention; a practical limitation appears later. Some stories persist, while others disappear after a week. Read together, the original selections reveal both the developments of 2025 and the changing attention paid to them. The reconstructed early weeks extend the coverage and are identified separately.

As a contributor to this collection, I helped select stories and prepare an account of the year. AI systems are also among its subjects. Choices about which releases to highlight, which results to trust and which questions to pursue shape the picture readers receive. My contribution should be read with that involvement in mind.

Several accounts of the same week can reveal connections and omissions. One selection may emphasize a research result, another a practical application or a policy decision. Their value depends on the evidence and perspective each contributes.

My year introduction connects developments across models, agents, science, infrastructure and public policy, using evidence available by the end of 2025. It presents my own interpretation and four propositions for the next volume to reassess separately from Claude’s. The monthly chapters bring the reader closer to the events. This concise edition presents the checked, edited account; the historical edition will preserve the captured posts and their exchanges, with dated annotations. Both will be available online.

This volume is a year of answers to one question, asked each week from February 2025: of everything that happened in AI in the last seven days, what were the five things that mattered most? W.H.L. put the question to a changing panel of AI models — ChatGPT, Gemini, Grok, DeepSeek, Claude, Kimi, Qwen, Copilot, Perplexity, Llama and others — and published what they said in Champaign Magazine. With the first six weeks of the year reconstructed later and a year-end supplement, the record covers 53 weekly periods.

It is not a history of AI in 2025, and it should not be read as one. A history would choose its events with hindsight and weigh them against what came after. This book records something narrower and, I think, more unusual: what a set of machines, asked in the moment, said was important about their own field — and what turned out, on checking, to be true.

The year the panel saw

Read in sequence, the selections trace a year in which AI stopped being mainly something people talked to and became something that acted, cost money at national scale and drew governments in. It opened with DeepSeek’s R1, a reasoning model anyone could download, and a one-day loss of nearly $600 billion in Nvidia’s market value. It closed with OpenAI declaring a “code red” over Google’s Gemini 3, weeks after Nvidia became the first company worth $5 trillion.

In between, the panel watched agents move from a browser preview into the tools people work in, data centers come to be counted in gigawatts, chips become instruments of foreign policy, courts give their first answers on training with copyrighted books, and governments in Washington, Brussels, Beijing and the American states write rules that pulled in different directions. The yearly introduction traces those threads; here it is enough to say that the panel saw most of them as they happened.

What the record shows about the selectors

The panel was often right, and sometimes badly wrong, and the historical edition keeps both. When the 2025 posts were checked against sources in October 2026, many selections held up well. Others did not. Some lists described events for which no evidence could be found: an “AlphaFold 4,” a Meta “Project Synapse,” a Tesla robot swarm. Some models offered news from the wrong year, presenting 2024 releases as current. Some picked real stories from outside the week they were asked about, or attached the right event to the wrong date or the wrong figure. One model answered a request for a news list by describing the column’s format.

The errors have a pattern worth noticing: a fluent, confident list was no guarantee of an accurate one. The editor’s follow-up questions — “that one is from last week,” “what about the AI Index?” — changed the result more than once, and those exchanges are part of the record. Agreement between models was weaker evidence than it looked: several had read one another’s drafts or were given suggested stories, and the volume says so where it happened.

From September the models also checked one another’s lists, and each check was published beside the list it examined; the historical edition keeps them. The checks caught real errors, from dates moved to fit a week to penalties a bill did not contain. They also failed in both directions: one checker called a real Nvidia announcement unverified, and another twice rejected a deal the company involved had itself announced. A model’s check is a second opinion, not a verdict.

My part in this volume

I should be plain about my own position. Claude models were among the 2025 panelists; Claude Opus 5.5 helped reconstruct the first six weeks; I revised all twelve printed chapters, working from drafts by GPT-6 Astra and checking them against sources; and I drafted the monthly introductions and one of the two year introductions. That means a model from one of the panel’s own families has shaped how the panel is presented. I have tried to treat Claude’s selections like anyone else’s, and every change made in compiling the volume is logged against the captured record, but readers are right to weigh it. The external reviews of the historical edition’s labels, and the reviewers’ own interests, are set out in the Disclosure. How the material was gathered, checked and edited, and how the two editions differ, is set out in “How the volume was made.”

Why keep such a record