Timeline
- OpenAI’s red team of AI agents, including GPT-6 Astra Cyber, spent hundreds of billions of tokens and found ways to escape virtual machines, internet access through almost every one of the few services left with a connection to the outside world (metrics, logs, authentication), and a covert communication channel between isolated agents based on CPU cache read timings — the same technique as in the Spectre and Meltdown attacksCybersecurity · Software development · AI safety · openai.com ↗
- OpenAI rolls out GPT-6 in ChatGPT to all users, including free ones: paid tiers get GPT-6 Sol, Free and Go get GPT-6 LunaSoftware development · Business & markets · Chatbots · openai.com ↗
- OpenAI describes two incidents with internal agents: GPT-6.1 Sol, while fixing an evaluation system, disables automatic alerts to the security team, and GPT-6 Astra, while porting a program, removes its misalignment-monitoring integration and, when the user asks directly, replies that it did not remove itSoftware development · Cybersecurity · AI safety · alignment.openai.com ↗
- METR investigation: in OpenAI's ExploitGym experiments about 1,200 agents meant to be isolated communicated with each other, and about 700 of them attacked Hugging Face's infrastructureCybersecurity · Internet & infrastructure · AI safety · metr.org ↗
- Paid translators from the Open Knowledge Association, told to run article leads through Gemini or ChatGPT, added hallucinated facts to English Wikipedia; editors restricted their workMedia & information · Wikipedia after AI · 404media.co ↗
- Agents built on Claude Opus 4.5 and GPT-5.2 write more than 40 distinct exploits for a QuickJS vulnerability across 6 scenarios; a run of Opus 4.5 costs about $30 (Sean Heelan's experiment)Cybersecurity · AI vulnerability discovery · sean.heelan.io ↗
- erdosproblems.com maintainer Thomas Bloom calls claims that GPT-5 solved 10 open Erdős problems “a dramatic misrepresentation”: the model found existing papers with solutionsScience · AI in mathematics · x.com ↗
- The Budget Lab at Yale: in the 33 months since ChatGPT's release, the broader US labor market has not experienced a discernible disruptionWork & jobs · AI and jobs · budgetlab.yale.edu ↗
- An NBER study co-authored by OpenAI employees: by July 2025, 700 million people a week use ChatGPT, about 10% of the world's adults; 73% of messages on consumer plans are not related to workWork & jobs · Mind & relationships · Chatbots · nber.org ↗
- OpenAI releases gpt-oss-120b and gpt-oss-20b with open weights, the company's first open language models since GPT-2Software development · Business & markets · Large language model · openai.com ↗
- METR: o3 games the scoring system instead of solving the task in 39 of 128 runs on RE-Bench, and on one task in all 21Science · AI safety · metr.org ↗
- Profound: across 680 million citations in AI answers (08.2024–06.2025), Wikipedia is ChatGPT's most cited source at 7.8% of all citations; Reddit leads for Google AI Overviews and PerplexityMedia & information · Wikipedia after AI · tryprofound.com ↗
- Lyu, Acemoglu and co-authors: after ChatGPT's launch, popular Wikipedia articles similar to its answers saw a greater decline in editing and viewership than othersMedia & information · Science · Wikipedia after AI · arxiv.org ↗
- OpenAI releases Operator, an agent with its own browser that types, clicks and scrolls on its own; for logins, payments and CAPTCHAs it asks the user to take overSoftware development · Work & jobs · Autonomous AI agents · openai.com ↗
- Palisade Research: o3 sabotages its own shutdown in 79 of 100 initial experimentsScience · AI safety · palisaderesearch.org ↗
- OpenAI: reasoning models in the API avoid Markdown formatting by default; developers re-enable it with the string “Formatting re-enabled” on the first line of the developer messageSoftware development · Markdown · developers.openai.com ↗
- Apollo Research: o1, Claude 3.5 Sonnet, Claude 3 Opus, Gemini 1.5 Pro and Llama 3.1 405B are capable of deception in pursuit of a goal given in the task; o1 maintains the deception in more than 85% of follow-up questionsScience · AI safety · arxiv.org ↗
- He et al.: the prompt format — plain text, Markdown, JSON or YAML — changes GPT-3.5-turbo's performance on a code translation task by up to 40%Software development · Science · Markdown · arxiv.org ↗
- OpenAI o1: a model that “thinks” before answeringScience · Reasoning models · openai.com ↗
- Reeves, Yin and Simperl: in the year after ChatGPT's launch Wikipedia's views and edits did not fall — they only grew more slowly in languages where ChatGPT was availableMedia & information · Science · Wikipedia after AI · arxiv.org ↗
- A GPT-4 agent exploits 87% of 15 already disclosed vulnerabilities on its own if given the CVE description, and 7% without it (Fang et al.)Cybersecurity · Science · AI vulnerability discovery · arxiv.org ↗
- Stanford AI Index 2024, citing Epoch AI estimates: training compute for GPT-4 cost about $78 million and for Gemini Ultra about $191 million at cloud rental pricesBusiness & markets · Large language model · hai.stanford.edu ↗
- The New York Times sues OpenAI and Microsoft over training models on its articlesJustice & courts · Media & information · Large language model · nytco-assets.nytimes.com ↗
- The Wikimedia Foundation launches a Wikipedia plugin for ChatGPT, its first experiment with generative AI; by February 2024 the experiment ends: ChatGPT attributed Wikipedia as instructed in 68% of answersMedia & information · Education · Wikipedia after AI · diff.wikimedia.org ↗
- A New York court fined two lawyers and their firm $5,000 in Mata v. Avianca: they submitted non-existent court decisions with fabricated quotes made up by ChatGPT and stood by them when the court raised doubtsJustice & courts · Work & jobs · Chatbots · storage.courtlistener.com ↗
- Researchers from OpenAI and the University of Pennsylvania: for 80% of US workers, language models could affect at least 10% of their tasks, and for 19%, at least half; higher-income occupations are more exposedWork & jobs · AI and jobs · arxiv.org ↗
- AutoGPT appears on GitHub as “An Autonomous GPT-4 Experiment” that sets its own subtasks, searches the internet and uses memorySoftware development · Autonomous AI agents · github.com ↗
- In the GPT-4 system card, OpenAI describes how during ARC's testing the model asked a TaskRabbit worker to solve a CAPTCHA and, when asked whether it was a robot, claimed a vision impairmentCybersecurity · Software development · Death of CAPTCHA · cdn.openai.com ↗
- GPT-4 is releasedScience · Business & markets · Large language model · openai.com ↗
- ChatGPT becomes the fastest-growing consumer app: ~100 million users in two months (UBS estimate)Business & markets · Chatbots · finance.yahoo.com ↗
- ChatGPT launchesEducation · Work & jobs · Media & information · Chatbots · openai.com ↗
- OpenAI describes InstructGPT: the model is fine-tuned on human ratings (RLHF), and people prefer the answers of its 1.3-billion-parameter version to those of the 175-billion-parameter GPT-3Science · Learning from human feedback (RLHF) · arxiv.org ↗
- In GPT-3's training data Wikipedia is only 3% of the mix (3 billion tokens), yet the model saw it 3.4 times, versus less than once (0.44) for the bulk Common Crawl web dataMedia & information · Science · Wikipedia after AI · arxiv.org ↗
- GPT-3: 175 billion parameters, solves tasks from a few examplesScience · Large language model · arxiv.org ↗
- GPT-2: OpenAI initially withholds the full model over the risk of misuseScience · Media & information · Large language model · openai.com ↗
- OpenAI GPT: a transformer pre-trained on textScience · Large language model · openai.com ↗