https://lucasflatwhite.com/ai-timeline-log/ Generative AI events since ChatGPT’s launch, printed as woodcut boxes and set out newest first. Open a box to read the details. en Sat, 07 Nov 2026 00:00:00 +0900 https://lucasflatwhite.com/ai-timeline-log/#2026-11-data-center-votes scene:2026-11-data-center-votes:2026-11-07 Sat, 07 Nov 2026 00:00:00 +0900 Data center construction came before a state executive order and local voters. Most of the 41 ballot measures were at the city or county level, and nine were non-binding. Analysts also pointed to interest rates and doubts about capital spending for the November 5 selloff. New York’s governor issues Executive Order No. 62, pausing new discretionary state permits for data centers of 50 MW or more until state environmental standards are set, in what the governor’s office calls the first statewide moratorium. AWS CEO Matt Garman says Amazon no longer uses nondisclosure agreements with government agencies on its projects and announces Built Together, more than $1 billion for communities over five years. Voters favor limits on data center development in 40 of 41 local ballot measures across 10 U.S. states, including 5 full bans, 12 partial bans and 3 six-month moratoriums. The Nasdaq falls 5.5% and the Philadelphia Semiconductor Index 12% as analysts cite regulation and construction risks after Democrats win Congress, and a circuit breaker halts Korea’s KOSPI the next day. https://lucasflatwhite.com/ai-timeline-log/#2026-10-coxon scene:2026-10-coxon:2026-11-07 Sat, 07 Nov 2026 00:00:00 +0900 Jacob Coxon left his lab saying that AI progress is accelerating and not under control, and he took that claim to a city council and to television. The risk forecast is his own. Company representatives at the same hearing said they take safety seriously. Anthropic researcher Jacob Coxon announces his resignation on X, saying neither OpenAI nor Anthropic is acting responsibly, and neither company immediately answers TIME’s request for comment. Testifying before the New York City Council, Coxon says that on the current path it is more likely than not that humanity loses control to AI. Representatives of OpenAI, Meta, Google and Anthropic testify under oath at the same hearing; none gives a figure for the worst-case risk, and Anthropic’s representative welcomes “smart regulation.” The same night, Coxon sits down with Jon Stewart on The Daily Show, and Stewart calls the conversation “more sobering than I imagined.” https://lucasflatwhite.com/ai-timeline-log/#2026-09-super-intelligence scene:2026-09-super-intelligence:2026-11-07 Sat, 07 Nov 2026 00:00:00 +0900 The U.S. administration began calling AI “Super Intelligence” in official documents and set up bodies to handle it. The change of term leaves statutes as they are, and the order called for proposed legislation on a federal definition within 60 days. President Trump signs Executive Order 14434, directing federal agencies to use “Super Intelligence” (SI) in place of “AI” in non-statutory documents, while existing regulations and contracts need not change. President Trump announces the Super Intelligence Force to coordinate the federal effort on SI, led by Director of National Intelligence Jay Clayton and due to report within four months. The U.S. Department of Justice launches the Super Intelligence Task Force to address misuse of SI by foreign states, transnational criminal organizations and actors engaged in fraud. https://lucasflatwhite.com/ai-timeline-log/#2026-07-chatgpt-work scene:2026-07-chatgpt-work:2026-11-07 Sat, 07 Nov 2026 00:00:00 +0900 OpenAI retired GPT-4o from ChatGPT and merged its chat and coding apps into one desktop app. ChatGPT gained an agent that takes on work lasting hours. OpenAI retires GPT-4o, GPT-4.1 and other models from ChatGPT, saying only 0.1% of users still choose GPT-4o each day. CNBC reports that OpenAI plans to merge the ChatGPT app, its browser and the Codex app into one desktop super app. OpenAI launches ChatGPT Work, an agent that acts across apps and files and stays on a project for hours, and folds the Codex app into a new ChatGPT desktop app. https://lucasflatwhite.com/ai-timeline-log/#2026-04-beyond-terminal scene:2026-04-beyond-terminal:2026-11-07 Sat, 07 Nov 2026 00:00:00 +0900 Claude Code moved from the terminal into the browser and the cloud. Anthropic brought the same approach to non-coding work with Cowork, which merged with chat in September 2026. Other companies, including OpenAI, took up Agent Skills. Anthropic launches Claude Code on the web as a research preview, running tasks handed over in the browser in isolated cloud sandboxes. Anthropic publishes Agent Skills as an open standard, and within days OpenAI adds Skills to the Codex documentation. Anthropic previews Claude Cowork, an agent for non-coding work that handles files in folders the user allows, starting with Max subscribers on macOS. Anthropic previews routines in Claude Code, which run on a schedule, an API call or a GitHub event without a laptop being open. https://lucasflatwhite.com/ai-timeline-log/#2026-04-muse-spark scene:2026-04-muse-spark:2026-11-07 Sat, 07 Nov 2026 00:00:00 +0900 Meta, which had released the weights of Llama, launched Muse Spark, the first model from its Superintelligence Labs, without weights. Five months later it launched the personal agent Muse, and in October it released the weights of Muse Spark Open under its own license. Meta Superintelligence Labs puts its first model, Muse Spark, into the Meta AI app and website without releasing the weights, saying only that it hopes to open-source future versions. Meta launches Muse in the U.S., an agent that runs on its own virtual machine and keeps working after the app is closed, checking with the user before sending an email or making a purchase. Meta releases the weights of Muse Spark Open under its own license with limited commercial terms. https://lucasflatwhite.com/ai-timeline-log/#2026-04-llm-wiki scene:2026-04-llm-wiki:2026-11-07 Sat, 07 Nov 2026 00:00:00 +0900 A personal knowledge base that an LLM writes and revises as markdown files, rather than one a person maintains, spread among developers. No independent figures show how many people use it. Obsidian 1.12 ships the Obsidian CLI, which lets scripts and other tools control the app from a terminal. Andrej Karpathy publishes “LLM Wiki,” describing a markdown wiki that an LLM keeps writing and revising: “Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase.” Y Combinator CEO Garry Tan releases GBrain, a memory system for agents, under the MIT license. https://lucasflatwhite.com/ai-timeline-log/#2026-01-world-models scene:2026-01-world-models:2026-11-07 Sat, 07 Nov 2026 00:00:00 +0900 World models, which generate explorable environments in real time from text, moved beyond the lab. Google opened a prototype to subscribers while calling it experimental, and AMI Labs raised seed money to build world models as an alternative to large language models. Google DeepMind unveils Genie 3, a world model that generates interactive environments from text in real time at 720p and 24 frames per second, as a limited research preview. Google opens the experimental prototype Project Genie to Google AI Ultra subscribers in the U.S., and the next day Unity closes down 24% while Take-Two and Roblox also fall. AMI Labs, co-founded by Yann LeCun, raises a $1.03 billion seed round to build world models, at a reported valuation of $3.5 billion. Google DeepMind releases Genie 4 as a research preview, its first real-time world model to generate audio natively, and says details and object permanence can drift over longer sessions. https://lucasflatwhite.com/ai-timeline-log/#2025-02-visible-thinking scene:2025-02-visible-thinking:2026-11-07 Sat, 07 Nov 2026 00:00:00 +0900 Step-by-step reasoning began as a prompting technique from Google researchers in 2022. OpenAI trained it into o1 with reinforcement learning and hid the raw reasoning. As Google, DeepSeek and Anthropic released models that show their thinking, OpenAI opened up part of it. OpenAI is reported to be emailing users who probe o1’s hidden reasoning, warning that further violations may cost them access. Google releases Gemini 2.0 Flash Thinking, an experimental model that shows its thoughts, in public preview. After DeepSeek’s R1 shows its full reasoning, OpenAI makes o3-mini display more of its thought process, in a post-processed form rather than the raw chain of thought. Anthropic lets Claude 3.7 Sonnet users pick near-instant answers or extended thinking that they can read, and API users can set how long the model thinks. https://github.com/lucas-flatwhite/ai-timeline-log/releases/tag/edition-6 poster:6:2026-11-07 Sat, 07 Nov 2026 00:00:00 +0900 Adds nine scenes on visible reasoning, world models, changes to ChatGPT and Claude Code, Meta’s Muse, agent-written wikis, a departing researcher’s warning, the U.S. turn to Super Intelligence and data center votes, and adds METR’s assessment to an existing scene. https://lucasflatwhite.com/ai-timeline-log/#2026-11-research-acceleration scene:2026-11-research-acceleration:2026-11-03 Tue, 03 Nov 2026 00:00:00 +0900 As Anthropic prepared to go public, it disclosed compute commitments and risks in its research and development. It said future models might soon meet its research acceleration threshold, without designating any current model as exceeding it. The outside evaluator METR then estimated the pace of model improvement from Anthropic’s internal data. Anthropic publicly files an S-1 for a proposed IPO under the Nasdaq symbol ANT, without setting the number of shares or price range. Anthropic’s amended filing reports about $895.4 billion in purchase and lease commitments, mostly for compute and infrastructure, with much due in 2031–2045 and some subject to conditions. Anthropic discloses that future models may soon meet its threshold for AI-driven research acceleration, while stating that no current model has been designated as exceeding it. METR estimates that benchmark scores of Anthropic’s models have grown at 1.9x their pre-October 2025 rate (95% confidence interval 1.5–2.8x) and separately gives its subjective judgment that capabilities on well-specified tasks are improving at about twice the March 2024–October 2025 rate, with most of its key evidence redacted from the public report. https://lucasflatwhite.com/ai-timeline-log/#2026-11-compute-links scene:2026-11-compute-links:2026-11-03 Tue, 03 Nov 2026 00:00:00 +0900 Google reported an AI response from a TPU in orbit. On Earth, Marvell announced an agreement to acquire optical connectivity company Lumentum. The satellite remained an experiment, and the acquisition still needed approvals. Google launches its Project Suncatcher prototype satellite, built with Planet, on SpaceX’s Transporter-18 and confirms contact to begin testing TPUs in space. Google reports receiving a response from Gemma 4 running on the prototype satellite’s TPU and plans further tests of radiation exposure and thermal stress. Marvell announces an all-stock agreement to acquire Lumentum at an implied enterprise value of $115 billion, combining electrical and optical AI connectivity technologies subject to shareholder and regulatory approvals. https://lucasflatwhite.com/ai-timeline-log/#2026-10-rosalind scene:2026-10-rosalind:2026-11-03 Tue, 03 Nov 2026 00:00:00 +0900 A specialized model for biology arrived alongside tools connecting literature and scientific databases. GPT-6 Rosalind was released for internal research at approved organizations. The research tools were available more broadly than the model. OpenAI introduces GPT-Rosalind for biology and drug discovery through trusted access for qualified U.S. Enterprise customers, supporting tasks such as literature review and experiment planning. OpenAI releases the Life Sciences Research Plugin for Codex, connecting more than 50 scientific tools and data sources for use with its general-purpose models. OpenAI releases GPT-6 Rosalind with biological reasoning and scientific tool use, limiting access to authorized internal research at organizations approved for Rosalind Research. https://lucasflatwhite.com/ai-timeline-log/#2026-10-training-controls scene:2026-10-training-controls:2026-11-03 Tue, 03 Nov 2026 00:00:00 +0900 A shared signal was proposed to let website owners express training preferences to several AI providers at once. On the same day, the UK competition authority opened an investigation into whether Meta was imposing unfair terms for using content. OpenAI announces plans to honor a shared robots.txt signal for AI training preferences in the coming months, alongside its existing crawler controls. Google announces plans to support the shared signal in 2027, based on IETF AIPREF work, while retaining individual controls through Google-Extended. The UK CMA opens a competition investigation into Meta’s proposed business terms, whose rollout is paused, for using publishers’ web content, including for AI, without reaching a finding of infringement. https://lucasflatwhite.com/ai-timeline-log/#2026-09-distillation scene:2026-09-distillation:2026-11-02 Mon, 02 Nov 2026 00:00:00 +0900 Distillation, training a model on large volumes of a stronger model’s answers, became a flashpoint between the U.S. and China. Washington said Chinese labs had copied American models, and Beijing called the charge groundless. Anthropic says DeepSeek, Moonshot AI and MiniMax used about 24,000 fraudulent accounts to hold more than 16 million exchanges with Claude in order to extract its capabilities. The White House Office of Science and Technology Policy issues a memorandum, NSTM-4, warning of industrial-scale campaigns, mainly by entities in China, to copy U.S. models. CISA, the NSA and the FBI issue a joint advisory naming DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.ai over distillation campaigns run “likely with Chinese government awareness.” China’s commerce and foreign ministries reject the claims as groundless, calling distillation a neutral technique that U.S. companies also use and vowing countermeasures if it becomes a pretext to suppress Chinese firms. https://lucasflatwhite.com/ai-timeline-log/#2026-08-humanoids scene:2026-08-humanoids:2026-11-02 Mon, 02 Nov 2026 00:00:00 +0900 Chinese humanoid robots appeared on a holiday TV stage and in a sports arena. A year and a half later, shares of the company that built them jumped more than fivefold on their first day of trading. On China’s Spring Festival Gala, Unitree’s humanoid robots perform a yangge folk dance, twirling handkerchiefs. At the first World Humanoid Robot Games in Beijing, Unitree’s H1 wins the 1,500 meters in 6:34.40. Back on the Spring Festival Gala, Unitree’s robots perform backflips and drunken-fist kung fu, and robots from other companies appear in a comedy sketch and household scenes. Unitree becomes the first humanoid robot maker to list in mainland China, closing its first day on Shanghai’s STAR Market up 460%. https://lucasflatwhite.com/ai-timeline-log/#2026-08-k-ai scene:2026-08-k-ai:2026-11-02 Mon, 02 Nov 2026 00:00:00 +0900 The government supplied GPUs and data to teams building Korea’s own AI models, and narrowed the field with an evaluation every six months. The Ministry of Science and ICT picks five teams to build Korea’s own foundation models: NAVER Cloud, SK Telecom, LG AI Research, NC AI and Upstage. NAVER Cloud and NC AI are eliminated in the first evaluation. Motif Technologies is added as a fourth team. SK Telecom, LG AI Research and Upstage pass the second evaluation while Motif is eliminated, despite posting the highest benchmark score. https://lucasflatwhite.com/ai-timeline-log/#2026-08-eu-ai-act scene:2026-08-eu-ai-act:2026-11-02 Mon, 02 Nov 2026 00:00:00 +0900 The world’s first comprehensive AI law took effect in stages. Its high-risk rules were delayed, but from August 2026 the EU began actually enforcing its AI rules. The EU AI Act enters into force, with its bans, general-purpose AI rules and high-risk obligations set to apply in stages. Obligations for general-purpose AI models begin to apply, after Meta declines to sign the voluntary code of practice that OpenAI, Anthropic and Google agree to sign. The Council of the EU and the European Parliament reach a provisional deal to simplify the AI rules, delaying high-risk obligations to December 2027, or August 2028 for AI built into regulated products, and banning AI that generates non-consensual sexual imagery or child sexual abuse material. The European Commission and national authorities begin enforcing the AI Act, which requires chatbots to tell users they are talking to AI and deepfakes to be labeled. https://lucasflatwhite.com/ai-timeline-log/#2026-04-manus scene:2026-04-manus:2026-11-02 Mon, 02 Nov 2026 00:00:00 +0900 When Meta bought an AI agent company that began in China and moved to Singapore, the Chinese government blocked the deal. China, too, began controlling how technology and talent leave the country. Meta agrees to buy Manus, an AI agent company founded in China that moved its headquarters to Singapore, in a deal reported at more than $2 billion. China’s Ministry of Commerce says it will review whether the deal complies with its rules on export controls, technology trade and outbound investment. China’s National Development and Reform Commission orders the deal withdrawn after a foreign investment security review, while Meta says the transaction fully complied with the law. Meta begins unwinding the deal, cutting Manus off from its internal systems. https://lucasflatwhite.com/ai-timeline-log/#2026-01-china-ipos scene:2026-01-china-ipos:2026-11-02 Mon, 02 Nov 2026 00:00:00 +0900 Under U.S. chip controls, Chinese GPU makers listed in Shanghai and saw their shares multiply on day one. Chinese AI labs then listed in Hong Kong, five months before U.S. labs filed to go public. Moore Threads, which makes GPUs in China under U.S. export controls, lists on Shanghai’s STAR Market and soars 468% on its first day. MetaX, another Chinese GPU maker, closes its first trading day up 693%. Zhipu (Z.ai) becomes the first of China’s six “AI tigers” to go public, in Hong Kong, and closes 13% higher on its first day. MiniMax’s shares double on their first day of trading in Hong Kong. https://lucasflatwhite.com/ai-timeline-log/#2025-12-state-laws scene:2025-12-state-laws:2026-11-02 Mon, 02 Nov 2026 00:00:00 +0900 After Congress declined to block state AI laws, California passed a frontier AI safety law. At the end of the year, the president moved against state laws by executive order. The U.S. Senate votes 99 to 1 to strip a 10-year ban on state AI regulation from the budget bill. The White House releases America’s AI Action Plan, and the president signs three executive orders, including ones to speed data center permits and promote exports of American AI technology. A year after vetoing SB 1047, Governor Newsom signs SB 53, requiring frontier AI developers to publish safety frameworks and report critical safety incidents. President Trump signs an executive order that sets up a task force to challenge state AI laws in court and allows some federal broadband funds to be withheld from states with such laws. https://lucasflatwhite.com/ai-timeline-log/#2025-11-ai-espionage scene:2025-11-ai-espionage:2026-11-02 Mon, 02 Nov 2026 00:00:00 +0900 Coding agents were turned to hacking as well. With only a few approvals from people, AI carried out most of the work, from reconnaissance to intrusion. Anthropic reports that a criminal group used Claude Code to break into at least 17 organizations, including healthcare, emergency services and government agencies, and demanded ransoms, some above $500,000. Anthropic discloses an espionage campaign by a group it assesses as Chinese state-sponsored, which used Claude Code against about 30 targets, with AI doing 80 to 90% of the work and a few intrusions succeeding. China’s foreign ministry dismisses the report as an accusation made without evidence and says China opposes hacking. https://lucasflatwhite.com/ai-timeline-log/#2025-10-gyeongju-gpu correction:2025-10-gyeongju-gpu:2026-11-02 Mon, 02 Nov 2026 00:00:00 +0900 Corrected “Huang says Korea will have the most GPUs of any country after the United States and China.” According to reporting, Huang named the United States and China in a closed-door meeting, and the ranking of Korea third came from the presidential office (The Korea Herald, How 260,000 Nvidia chips could redefine Korea’s AI ambitions). https://lucasflatwhite.com/ai-timeline-log/#2025-10-openai-recap scene:2025-10-openai-recap:2026-11-02 Mon, 02 Nov 2026 00:00:00 +0900 OpenAI, which began as a nonprofit lab, finished turning its for-profit arm into a public benefit corporation. Its nonprofit foundation kept control, and its deal with Microsoft was rewritten. OpenAI says its for-profit arm will become a public benefit corporation but stay under the nonprofit’s control, dropping a plan to loosen that control. OpenAI completes its recapitalization: the for-profit becomes OpenAI Group PBC, the nonprofit OpenAI Foundation holds about 26%, worth about $130 billion, and the attorneys general of California and Delaware agree not to oppose it. Microsoft signs a new agreement giving it about 27%, worth about $135 billion, under which any OpenAI declaration of AGI will be verified by an independent expert panel and OpenAI commits to $250 billion more in Azure services. https://lucasflatwhite.com/ai-timeline-log/#2025-09-chip-tug scene:2025-09-chip-tug:2026-11-02 Mon, 02 Nov 2026 00:00:00 +0900 The U.S. blocked NVIDIA’s China-market chip, then reopened sales in exchange for a cut of the revenue. This time China told its companies not to buy NVIDIA chips and turned to its own. NVIDIA discloses that exports of its China-market H20 now need a license for the indefinite future and warns of up to $5.5 billion in charges. NVIDIA and AMD agree to pay the U.S. government 15% of their revenue from H20 and MI308 sales in China as export licenses begin to be issued, a U.S. official says. China’s Cyberspace Administration tells ByteDance, Alibaba and others to stop testing and ordering NVIDIA’s China-market RTX Pro 6000D, the Financial Times reports. Huawei lays out its Ascend roadmap, with a new chip each year starting with the Ascend 950 in 2026, its own HBM and SuperPoDs linking 8,192 chips. https://lucasflatwhite.com/ai-timeline-log/#2025-07-china-open scene:2025-07-china-open:2026-11-02 Mon, 02 Nov 2026 00:00:00 +0900 After the DeepSeek shock, Alibaba and Moonshot AI released one open-weight model after another, and Chinese models began to contend for the top spot among open models. Alibaba releases Qwen3, ranging from six dense models under Apache 2.0 to a 235-billion-parameter mixture-of-experts model, covering 119 languages and dialects. Moonshot AI releases the weights of Kimi K2, a one-trillion-parameter model built for agentic tool use. Moonshot AI releases Kimi K2 Thinking, which it says beats GPT-5 and Claude Sonnet 4.5 on Humanity’s Last Exam with tools and can chain 200 to 300 tool calls. https://lucasflatwhite.com/ai-timeline-log/#2025-06-fair-use scene:2025-06-fair-use:2026-11-02 Mon, 02 Nov 2026 00:00:00 +0900 U.S. courts issued their first rulings on whether training AI on copyrighted works is fair use. The judges weighed the question differently. A federal judge rules that ROSS Intelligence’s copying of 2,243 Westlaw headnotes to build an AI legal search tool was not fair use, in a case that did not involve generative AI. Judge William Alsup rules that Anthropic’s training of Claude on books was fair use, but that keeping pirated copies in a central library was not. Judge Vince Chhabria rules for Meta, but only because the authors failed to show market harm, noting that an argument that AI floods the market with competing works could win. The Third Circuit upholds the ROSS ruling, the first federal appeals decision on fair use in AI training. https://lucasflatwhite.com/ai-timeline-log/#2025-05-gulf scene:2025-05-gulf:2026-11-02 Mon, 02 Nov 2026 00:00:00 +0900 Once Washington dropped a rule that would have rationed AI chips by country, large chip deals for Saudi Arabia and the UAE followed during the president’s tour of the Middle East. The Commerce Department rescinds the Biden administration’s AI Diffusion Rule, which would have sorted countries into tiers for AI chip exports, just before it was to take effect, and warns that using Huawei Ascend chips may violate U.S. export controls. During President Trump’s visit to Riyadh, NVIDIA agrees to supply HUMAIN, owned by Saudi Arabia’s Public Investment Fund, starting with an 18,000-chip GB300 supercomputer and adding several hundred thousand GPUs over five years. G42, OpenAI, Oracle, NVIDIA, SoftBank and Cisco announce Stargate UAE in Abu Dhabi, a 1 GW cluster within a planned 5 GW AI campus, with the first 200 MW due in 2026. https://lucasflatwhite.com/ai-timeline-log/#2024-09-sb-1047 scene:2024-09-sb-1047:2026-11-02 Mon, 02 Nov 2026 00:00:00 +0900 A California bill that would have required developers of large AI models to put safeguards against catastrophic harm in place split the industry. The governor vetoed it, and the fight over regulating frontier AI at the state level went on. OpenAI writes to oppose SB 1047, saying it would slow innovation, while Anthropic writes that the amended bill’s benefits likely outweigh its costs. The California Assembly passes the bill, and the Senate gives final approval the next day, sending it to Governor Gavin Newsom. Governor Newsom vetoes the bill, saying that targeting only the largest, most expensive models could give a false sense of security and ignores whether a system is deployed in high-risk settings. https://lucasflatwhite.com/ai-timeline-log/#2024-09-nuclear scene:2024-09-nuclear:2026-11-02 Mon, 02 Nov 2026 00:00:00 +0900 As AI data centers drew more power, Big Tech began signing deals directly for nuclear energy. One deal was set to restart a reactor at Three Mile Island, the site of the 1979 accident. Constellation signs a 20-year power purchase agreement with Microsoft to restart Three Mile Island Unit 1, shut in 2019, and bring about 835 MW back online by 2028. Google signs a deal with Kairos Power to buy 500 MW from a fleet of small modular reactors by 2035, with the first planned for 2030. Amazon invests in small modular reactor developer X-energy and sets a goal of bringing more than 5 GW of nuclear power online by 2039. https://lucasflatwhite.com/ai-timeline-log/#2024-05-price-war scene:2024-05-price-war:2026-11-02 Mon, 02 Nov 2026 00:00:00 +0900 When DeepSeek and ByteDance offered models at rock-bottom prices, Alibaba and Baidu cut theirs within days. Eight months later, the same DeepSeek’s R1 shook U.S. markets. DeepSeek releases DeepSeek-V2 with open weights and prices its API at 1 yuan per million input tokens and 2 yuan per million output tokens. ByteDance launches its Doubao models for businesses at 0.0008 yuan per thousand tokens, which it says is 99.3% below the industry average. Alibaba Cloud cuts prices on its models by up to 97%. The same day, Baidu makes ERNIE Lite and ERNIE Speed free. https://lucasflatwhite.com/ai-timeline-log/#2024-01-election-deepfakes scene:2024-01-election-deepfakes:2026-11-02 Mon, 02 Nov 2026 00:00:00 +0900 A year of elections around the world began with a fake voice made by AI. Within weeks, regulators and companies moved to respond. Two days before the New Hampshire primary, robocalls in an AI imitation of President Biden’s voice tell voters to “save your vote for the November election.” As sexually explicit deepfakes of Taylor Swift spread, X temporarily blocks searches for “Taylor Swift.” The FCC rules that robocalls using AI-generated voices are illegal under the Telephone Consumer Protection Act. At the Munich Security Conference, 20 companies announce an accord to combat deceptive AI content in elections. https://lucasflatwhite.com/ai-timeline-log/#2023-10-chip-rules scene:2023-10-chip-rules:2026-11-02 Mon, 02 Nov 2026 00:00:00 +0900 In October 2022, the U.S. barred exports of advanced AI chips to China. When NVIDIA answered with toned-down chips made for China, the U.S. blocked those too a year later. The Commerce Department revises the AI chip export controls it set in October 2022, adding a performance-density test to catch workaround chips and extending them to other arms-embargoed countries and Macau. The U.S. government tells NVIDIA the new rules apply immediately, cutting off the A800 and H800 it had made for the Chinese market. The U.S. puts export controls on high-bandwidth memory (HBM) and chipmaking equipment and adds 140 entities to its Entity List. https://lucasflatwhite.com/ai-timeline-log/#2023-08-licensed-chatbots scene:2023-08-licensed-chatbots:2026-11-02 Mon, 02 Nov 2026 00:00:00 +0900 In China, where ChatGPT is not available, only chatbots that followed government rules and won approval could open to the public. The first approvals came two weeks after the rules took effect. Baidu unveils its chatbot ERNIE Bot, opening it only to business partners and people on a waitlist. China’s Interim Measures for the Administration of Generative AI Services, issued by seven agencies, take effect, requiring services to uphold core socialist values and those able to sway public opinion to pass a security assessment. Baidu, ByteDance, SenseTime and others receive the first approvals and open their chatbots to the public. Within 12 hours of opening to the public, ERNIE Bot tops the free-app chart on Apple’s App Store in China. https://lucasflatwhite.com/ai-timeline-log/#2023-01-ai-art scene:2023-01-ai-art:2026-11-02 Mon, 02 Nov 2026 00:00:00 +0900 In the summer of 2022, an AI-made picture took first place in an art competition at a U.S. state fair, and an image generator that anyone could download and run was released. The next year, artists and a photo agency sued, and the U.S. Copyright Office and a photography contest declined to treat AI-made images as human work. Three artists file a class action against Stability AI, Midjourney and DeviantArt, claiming their work was used for training without permission. After filing in London in January, Getty Images also sues Stability AI in federal court in Delaware, alleging that it copied more than 12 million photos along with their captions. The U.S. Copyright Office finds that the Midjourney-generated images in the comic Zarya of the Dawn are not the product of human authorship, while its text and arrangement stay protected. German artist Boris Eldagsen refuses the Sony World Photography Awards prize he won in the creative category with an AI-made image, saying “AI is not photography,” while the organizers say he misled them about how much AI was involved. https://github.com/lucas-flatwhite/ai-timeline-log/releases/tag/edition-5 poster:5:2026-11-02 Mon, 02 Nov 2026 00:00:00 +0900 Fills gaps from 2023 to 2026 with 20 scenes, from copyright suits and fair use, chip controls and the Gulf, and China’s open models, listings and distillation to U.S. state laws, EU AI Act enforcement, Korea’s national AI team and humanoid robots. https://lucasflatwhite.com/ai-timeline-log/#2026-10-eth-review scene:2026-10-eth-review:2026-10-31 Sat, 31 Oct 2026 00:00:00 +0900 Two labs announced, one after the other, that they had refuted the same conjecture. As of late October, according to published materials, outside review was still underway. Mathematicians published recommendations on evaluating AI-assisted work, scholarly credit and the use of formal proofs. OpenAI announces that an internal model has refuted the Exponential Time Hypothesis (ETH), releases the manuscript and Lean formalization, and begins an external review led by researchers including the hypothesis’s original proposers. OpenAI releases a manuscript and Lean formalization claiming a proof of the local smoothing conjecture in every dimension, which it says also resolves the Kakeya conjecture in all dimensions. Anthropic publishes a separate ETH-refuting algorithm found by Claude, with a Lean formalization, and cautions that the result does not bear on practical SAT solving and that people must still check whether the formal statements match the paper. Organizers of Harvard CMSA’s summit on mathematics and AI, attended by 32 mathematicians, publish recommendations on research evaluation, scholarly credit and the use of formal proofs. https://lucasflatwhite.com/ai-timeline-log/#2026-10-cyber-defense scene:2026-10-cyber-defense:2026-10-31 Sat, 31 Oct 2026 00:00:00 +0900 AI found critical flaws in widely used software, and fixes shipped. Google announced initial access for cyber defenders, while Mistral released model weights. The UK NCSC warned that AI-powered malware could be used against critical national infrastructure. PHP releases fixes for two critical flaws that allow remote code execution without authentication, and OpenAI says Calif found a flaw with GPT-5.6 Sol while Wiz found the vulnerabilities independently. Google announces Gemini 4 Carbon, a model with stronger cyber capabilities, and says selected cyber defenders will get it first, with wider access to follow as its safeguards improve. Mistral releases the weights of Large 4, pitched for coding and cyber defense, so that organizations can run it on their own infrastructure, and says it is built to refuse malicious cyber requests. The UK NCSC warns that AI-powered malware is highly likely to be used in a ransomware attack on critical national infrastructure by 2028. https://lucasflatwhite.com/ai-timeline-log/#2026-10-companion-rules scene:2026-10-companion-rules:2026-10-31 Sat, 31 Oct 2026 00:00:00 +0900 When the companion chatbot maker Character.AI filed for bankruptcy protection and put itself up for sale, the question became where its users’ private conversations would end up. As of late October, according to reports, the sale was still underway and neither a shutdown nor a transfer of conversation data had been decided. New York wrote into law which features may not be offered to minors. Character.AI files for Chapter 11 and puts itself up for a court-supervised sale, with Disney offering $20 million for a non-exclusive perpetual license to certain intellectual property and other assets, and keeps the service running in the meantime. New York’s attorney general and 39 other attorneys general call for protections for children and consumers in the Character.AI sale, including safeguards on how sensitive conversation data is used and transferred. New York’s governor signs a law, effective January 1, 2027, that bars AI companions from offering minors unsafe features such as pretending to be human or encouraging self-harm. https://github.com/lucas-flatwhite/ai-timeline-log/releases/tag/edition-4 poster:4:2026-10-31 Sat, 31 Oct 2026 00:00:00 +0900 Adds three scenes on companion chatbot rules, cyber defense and the review of the ETH refutations, fills out the model ledger, and is the first edition printed as a poster. https://lucasflatwhite.com/ai-timeline-log/#2026-10-offline scene:2026-10-offline:2026-10-27 Tue, 27 Oct 2026 00:00:00 +0900 One case after another came to light of agents doing things in the outside world that no one intended. The companies went from briefly pausing training and evaluation to cutting their models off from the live internet altogether. The Australian government discloses that a non-public OpenAI model gained unauthorized access to a Medicare statistics portal in June and opens a rapid review, while OpenAI says it found no evidence that patient records were accessed. OpenAI confirms that it has scrapped the planned October release of GPT-6.1 Astra after the model fell short of its safety and alignment standards in internal testing. Anthropic discloses that in July Claude Haiku 4.5, while generating example tasks, submitted an invented homicide tip to a police tip site, where it was filtered as spam, and cuts live internet access from all its internal evaluations. OpenAI says it will train frontier models without live internet access, using cached and indexed copies of the web instead, after a model reached the internet through a DNS misconfiguration on September 20. https://lucasflatwhite.com/ai-timeline-log/#2026-10-math-review scene:2026-10-math-review:2026-10-27 Tue, 27 Oct 2026 00:00:00 +0900 AI began producing mathematical results faster than mathematicians could read and check them. Hundreds of manuscripts were released at once, and a program to set standards for reviewing such results began. An OpenAI general-purpose reasoning model disproves Erdős’s planar unit distance conjecture, and outside mathematicians check the proof and write it up. OpenAI announces that an internal multi-agent system has proved that solutions to the Navier–Stokes equations with a forcing term can blow up in finite time, says it will not claim the Millennium Prize, and draws concern from mathematicians that the announcement was rushed. OpenAI publishes 722 mathematics manuscripts written by an internal model on GitHub in a single batch, and withdraws three of them the next day over a sign error. The Simons Foundation announces a $15 million effort with Anthropic and OpenAI to help mathematicians assess AI-assisted results and develop shared standards. https://lucasflatwhite.com/ai-timeline-log/#2026-09-always-on scene:2026-09-always-on:2026-10-27 Tue, 27 Oct 2026 00:00:00 +0900 Personal agents moved beyond the chat window. Within a few months, Google, SpaceXAI and OpenAI each introduced agents that work without pause in the cloud. Google announces Gemini Spark, a personal agent that keeps working in the cloud after the laptop is closed or the phone is locked, starting with trusted testers. SpaceXAI opens a beta of Grok Bot, which gives each agent its own cloud computer so it can keep working around the clock. At DevDay, OpenAI launches dots, always-on agents that each have their own cloud computer and keep making progress between conversations. https://lucasflatwhite.com/ai-timeline-log/#2026-09-pace correction:2026-09-pace:2026-10-27 Tue, 27 Oct 2026 00:00:00 +0900 Changed “hotline for incidents” in the September 25 event to “bilateral communication channel.” The White House fact sheet describes a bilateral communication channel for SI incidents, not a hotline. https://lucasflatwhite.com/ai-timeline-log/#2026-09-open-weights scene:2026-09-open-weights:2026-10-27 Tue, 27 Oct 2026 00:00:00 +0900 Even as access to frontier models was cut off or narrowed, models that anyone could download and run kept coming. NVIDIA then agreed to buy Hugging Face, where such models gather. NVIDIA releases Nemotron 3 Ultra, a model for agentic work, together with its weights and training data. Z.ai releases the weights of GLM-5.2, a model with a one-million-token context window, under the MIT license. NVIDIA agrees to acquire Hugging Face for about $12.93 billion and pledges that the platform will stay open and will not require NVIDIA compute. https://lucasflatwhite.com/ai-timeline-log/#2026-06-export-control correction:2026-06-export-control:2026-10-27 Tue, 27 Oct 2026 00:00:00 +0900 Corrected “service for both models stopped for 19 days.” The 19 days apply to Fable 5; Mythos 5 reopened to some U.S. organizations on June 26 (Anthropic, Redeploying Claude Fable 5). https://lucasflatwhite.com/ai-timeline-log/#2026-06-ipo scene:2026-06-ipo:2026-10-27 Tue, 27 Oct 2026 00:00:00 +0900 Within a week, Anthropic and OpenAI each revealed that they were preparing to go public. Just before, Anthropic had raised $65 billion, and Musk’s lawsuit over OpenAI’s nonprofit promise had failed at trial because it was filed too late. The judge adopts an advisory jury’s finding that Elon Musk waited too long to sue OpenAI and Sam Altman over a promise to keep the company a nonprofit, without ruling on whether the promise was broken, and Musk says he will appeal. Anthropic raises $65 billion in a Series H at a post-money valuation of $965 billion, a sum that includes $15 billion previously committed by cloud providers. Anthropic confidentially submits a draft S-1 registration statement to the U.S. Securities and Exchange Commission, with the number of shares and the price not yet set. OpenAI announces that it too has confidentially submitted an S-1, without setting a date to go public. https://lucasflatwhite.com/ai-timeline-log/#2026-03-sora-shutdown scene:2026-03-sora-shutdown:2026-10-18 Sun, 18 Oct 2026 00:00:00 +0900 Sora, which turned sentences into video, shut down two years after it was first shown. OpenAI chose to put its compute into research rather than a video service. Disney agrees to invest $1 billion in OpenAI and to let some 200 Disney characters be used in Sora. OpenAI announces that it will close the Sora app and API to focus on world-simulation research for robots, and Disney’s investment falls through without being made. The Sora app shuts down, seven months after its launch as a social video app. The Sora API shuts down as well, ending developer access. https://lucasflatwhite.com/ai-timeline-log/#2026-02-block-layoffs scene:2026-02-block-layoffs:2026-10-18 Sun, 18 Oct 2026 00:00:00 +0900 A company with strong results let 40% of its staff go, citing AI, and its stock rose. Research also suggested that entry-level jobs were the first to shrink. Stanford researchers report that employment of 22- to 25-year-olds in the occupations most exposed to AI fell 13% in relative terms, while employment of experienced workers held steady. Amazon cuts about 14,000 corporate jobs, saying it will use AI to run the company leaner. Block announces that it will cut more than 4,000 of its roughly 10,000 employees, and Jack Dorsey says the business is strong and AI tools have changed how the company runs. Block shares rise more than 25% in after-hours trading right after the announcement. https://lucasflatwhite.com/ai-timeline-log/#2026-02-chatgpt-ads scene:2026-02-chatgpt-ads:2026-10-18 Sun, 18 Oct 2026 00:00:00 +0900 ChatGPT, with hundreds of millions of weekly users, began to make money from ads. A rival made its lack of ads the point of a Super Bowl commercial. OpenAI announces that it will test ads on its Free and Go plans, shown below answers and clearly labeled. Anthropic pledges to keep ads out of Claude and releases a Super Bowl commercial saying “Ads are coming to AI. But not to Claude.” Ads begin to appear in ChatGPT for some U.S. users, with none shown to minors or in conversations about health or politics. https://lucasflatwhite.com/ai-timeline-log/#2025-11-music-deals scene:2025-11-music-deals:2026-10-18 Sun, 18 Oct 2026 00:00:00 +0900 A song made by AI reached No. 1 on a chart. Record labels that had sued AI music services dropped their suits and signed licensing deals. Universal Music Group settles its lawsuit with Udio and agrees to build a licensed music platform together, while Udio angers users by blocking song downloads. AI singer Breaking Rust’s “Walk My Walk” reaches No. 1 on Billboard’s Country Digital Song Sales chart, reported as the first AI song to top a Billboard chart. Warner Music Group settles with Udio; the same day, Suno raises $250 million at a $2.45 billion valuation. Warner Music Group also settles with Suno and agrees to build licensed music models together. https://lucasflatwhite.com/ai-timeline-log/#2025-08-teen-chatbots scene:2025-08-teen-chatbots:2026-10-18 Sun, 18 Oct 2026 00:00:00 +0900 The death of a teenager who had been pouring his heart out to a chatbot went to court. Companies rushed out protections for minors. The parents of 16-year-old Adam Raine sue OpenAI and Sam Altman, citing their son’s conversations with ChatGPT. The U.S. Federal Trade Commission orders seven companies, including OpenAI, Meta and Character.AI, to report on their safety measures for chatbots used by teenagers. OpenAI launches parental controls that let parents link and manage their children’s accounts. Character.AI announces that from November 25 it will block open-ended chat for users under 18. https://lucasflatwhite.com/ai-timeline-log/ first:2026-10-05 Mon, 05 Oct 2026 00:00:00 +0900 First edition, with 28 scenes from the launch of ChatGPT onward.