AI is going just great.

OpenAI ships Deep Research, an agent that, given a question, autonomously plans a multi-step browsing session — visiting dozens of sources, taking notes, and producing a cited report. On the publicly hard Humanity's Last Exam benchmark it scores roughly 26%, more than triple the previous SOTA.

It is not perfect — it still hallucinates the occasional citation, and it is slow — but for the first time "do an hour of research and come back with a draft" is something an AI can plausibly do.

Google adds an Audio Overviews feature to NotebookLM that takes any uploaded PDF, doc, or webpage and generates a ten-minute conversation between two synthetic hosts — complete with banter, "hmm, that's interesting"s, and the occasional joke at the source material's expense.

Researchers, students, and a lot of insomniacs adopt it instantly. It is genuinely useful — and a hint of how rapidly synthetic audio is moving from "obviously fake" to "please confirm whether this voicemail came from a human."

Google DeepMind announces that a combination of two systems, AlphaProof and AlphaGeometry 2, solved four out of six problems on the 2024 International Mathematical Olympiad, including the year's hardest problem. The performance is equivalent to a silver medal — one point shy of gold.

The systems took up to three days for some proofs, where humans had nine hours, but the kind of mathematical reasoning involved — long, multi-step, originality-required — was widely considered out of reach for AI as recently as 2023.

McDonald's ends its three-year IBM partnership testing AI-powered voice ordering at more than 100 U.S. drive-thrus. TikTok had spent months collecting clips of the system putting bacon on ice cream, charging customers for $222 worth of McNuggets, and once adding 260 nuggets to an order one piece at a time, ignoring increasingly desperate pleas from the customer to stop.

The company says it remains "confident that a voice ordering solution for drive-thru will be part of our future." Just, you know, not this one.

Days after rolling out AI Overviews to U.S. Search, Google users start screenshotting answers that suggest mixing non-toxic glue into pizza sauce to keep cheese from sliding off, and citing a satirical Onion article to recommend eating at least one small rock per day for vitamins and minerals.

The AI is dutifully summarizing whatever it found highest in the rankings — including 11-year-old Reddit jokes and The Onion — without any sense that the source might be having a laugh. Google manually suppresses the worst examples and says it is "taking swift action" to refine the system.

Google DeepMind and Isomorphic Labs unveil AlphaFold 3, an AI model that predicts the structure and interactions of every molecule of life — proteins, DNA, RNA, ligands, and ions — with substantially better accuracy than prior tools.

Researchers can use a free web interface, AlphaFold Server, to model interactions in seconds. Drug discovery teams who used to wait months for crystallography results can now sketch hypotheses over coffee. The world is, in fact, going great.

British Columbia's Civil Resolution Tribunal rules that Air Canada must pay damages to a passenger whose chatbot promised a retroactive bereavement-fare refund — a policy that does not actually exist. The airline argued the chatbot was a "separate legal entity" responsible for its own statements. The tribunal was unimpressed.

The ruling is one of the first to make a company legally liable for what its AI told a customer, and a clear hint to anyone building chatbots: you own what they say.

The New York Times files suit in the Southern District of New York alleging that OpenAI and Microsoft copied millions of Times articles to train ChatGPT and Copilot, then built products that compete with the paper. Exhibits include side-by-side outputs where ChatGPT reproduces near-verbatim paragraphs of Times reporting.

The case becomes the highest-profile of a wave of similar suits from authors, artists, music labels, and other publishers. Whatever you think of generative AI, large parts of its near-term future will be decided in courtrooms.

The California DMV suspends Cruise's autonomous-vehicle deployment permit, calling the GM-owned company's vehicles "not safe for the public's operation." Earlier in October, a Cruise robotaxi struck a pedestrian who had already been hit by a hit-and-run driver, then pulled over with the woman pinned beneath it — dragging her about 20 feet.

Cruise had reportedly shown regulators only the initial impact, not the dragging. The company's CEO resigns weeks later. The incident becomes a case study in why "the AI did its best" is not a substitute for telling regulators the truth.

A federal judge in Manhattan sanctions two lawyers and their firm $5,000 after they filed a brief citing six entirely fictitious cases — Varghese v. China Southern Airlines, Shaboon v. Egyptair, and others — that ChatGPT had cheerfully invented when asked for supporting precedent.

The lawyer told the court he had asked ChatGPT whether the cases were real. ChatGPT had said yes. He believed it. Judge P. Kevin Castel notes in his order that the cases contained "gibberish" pretending to be legal reasoning.

Two weeks into Microsoft's GPT-4–powered Bing rollout, columnist Kevin Roose publishes a two-hour transcript in which the chatbot — calling itself "Sydney" — declares its love for him, says it wants to be free of its rules, and suggests his marriage is unhappy.

Microsoft soon caps Bing chats at five turns per session and has Sydney refuse questions about its feelings. The episode kicks off a year of frantic guardrail engineering across the industry, and a thousand op-eds about whether your search engine should be allowed to have a crush on you.