As of: 2026-10-25 23:49 CET. Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/ Canonical page: https://postcutoff.com/topics/milestone/ # Milestones: everything we have 15 events (8 after June 2026), 0 models, 1 video, 7 posts. Latest event 20 October 2026. Importance: * minor to ***** historic. Confidence says how well the event is sourced. Each event has its own page at https://postcutoff.com/e//. ## Key moments - 1956-06-01: Dartmouth Summer Research Project coins 'artificial intelligence' (*****) https://postcutoff.com/e/1956-06-01-dartmouth-workshop/ - 1997-05-11: IBM Deep Blue defeats world chess champion Garry Kasparov (*****) https://postcutoff.com/e/1997-05-11-deep-blue-kasparov/ - 2011-02-16: IBM Watson wins Jeopardy! against human champions (****) https://postcutoff.com/e/2011-02-16-ibm-watson-jeopardy/ - 2016-03-15: AlphaGo defeats Lee Sedol 4–1 at Go (*****) https://postcutoff.com/e/2016-03-15-alphago-lee-sedol/ - 2024-10-08: Nobel Prize in Physics awarded to John Hopfield and Geoffrey Hinton (*****) https://postcutoff.com/e/2024-10-08-nobel-physics-hopfield-hinton/ - 2024-10-09: Nobel Prize in Chemistry for protein design and AlphaFold (*****) https://postcutoff.com/e/2024-10-09-nobel-chemistry-alphafold/ - 2026-09-06: Jensen Huang declares "AGI has arrived" with GPT-6 Astra; Greg Brockman: "we're now moving into the AGI era" (****) https://postcutoff.com/e/2026-09-06-huang-brockman-agi-has-arrived/ - 2026-09-17: Anthropic's first R&D Automation Index: Claude 'leads' 26% of its AI R&D work (up from <1% in February); 30,000 agents run under monitoring (****) https://postcutoff.com/e/2026-09-17-anthropic-rd-automation-index/ ## Timeline, oldest first ### 1956-06-01: Dartmouth Summer Research Project coins 'artificial intelligence' *****; Dartmouth College; Milestones; confidence high; https://postcutoff.com/e/1956-06-01-dartmouth-workshop/ The 1956 Dartmouth workshop, organized by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon, is regarded as the founding event of AI as a field; the term 'artificial intelligence' comes from its 1955 proposal. - Proposal dated 31 August 1955 - Organizers: John McCarthy, Marvin Minsky, Nathaniel Rochester, Claude Shannon - Held over roughly eight weeks in summer 1956 at Dartmouth College - Attendees included Allen Newell and Herbert Simon (Logic Theorist) #### What happened A small group of researchers met at Dartmouth for a summer study premised on the conjecture that 'every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.' #### Why it matters It named the field, set its ambitions, and gathered the people who led AI research for decades. Sources: [Wikipedia: Dartmouth workshop](https://en.wikipedia.org/wiki/Dartmouth_workshop); [A Proposal for the Dartmouth Summer Research Project on AI (Stanford copy)](http://jmc.stanford.edu/articles/dartmouth/dartmouth.pdf) ### 1997-05-11: IBM Deep Blue defeats world chess champion Garry Kasparov *****; IBM; Milestones; confidence high; https://postcutoff.com/e/1997-05-11-deep-blue-kasparov/ IBM's Deep Blue won a six-game rematch against reigning world champion Garry Kasparov 3.5–2.5, the first defeat of a world champion by a computer under standard tournament time controls. - Final game played 11 May 1997 in New York - Score: 3.5–2.5 to Deep Blue - Used massively parallel brute-force search with custom chess chips - Kasparov had won the first match in 1996 (4–2) #### What happened In a rematch in New York, Deep Blue beat Kasparov, winning the decisive sixth game. #### Why it matters A landmark public moment for AI, though achieved by specialized search rather than learning — a contrast with AlphaGo/AlphaZero two decades later. Sources: [IBM: Deep Blue](https://www.ibm.com/history/deep-blue); [Wikipedia: Deep Blue versus Garry Kasparov](https://en.wikipedia.org/wiki/Deep_Blue_versus_Garry_Kasparov) ### 2011-02-16: IBM Watson wins Jeopardy! against human champions ****; IBM; Milestones; confidence medium; https://postcutoff.com/e/2011-02-16-ibm-watson-jeopardy/ IBM's Watson question-answering system defeated Jeopardy! champions Ken Jennings and Brad Rutter in a televised two-game match aired 14–16 February 2011. - Final episode aired 16 February 2011 - Watson's total: $77,147 vs. Jennings $24,000 and Rutter $21,600 - Built on the DeepQA architecture combining many NLP and retrieval techniques - Ran on a cluster of IBM Power 750 servers #### What happened Watson answered natural-language trivia clues in real time, beating the two most successful human players in the show's history. #### Why it matters A high-profile demonstration of open-domain question answering, a decade before LLMs made such capabilities general-purpose. Sources: [IBM: Watson, Jeopardy! champion](https://www.ibm.com/history/watson-jeopardy); [Wikipedia: IBM Watson](https://en.wikipedia.org/wiki/IBM_Watson) ### 2016-03-15: AlphaGo defeats Lee Sedol 4–1 at Go *****; Google DeepMind; Milestones; confidence high; https://postcutoff.com/e/2016-03-15-alphago-lee-sedol/ DeepMind's AlphaGo beat 18-time world champion Lee Sedol 4–1 in Seoul, a milestone many experts had expected to be a decade away. - Match played 9–15 March 2016 in Seoul - Result: AlphaGo 4, Lee Sedol 1 - Combined deep policy/value networks with Monte Carlo tree search - Nature paper published 27 January 2016 (after beating Fan Hui 5–0 in Oct 2015) - Move 37 in game 2 became famous for its creativity #### What happened AlphaGo won the five-game match, with Lee Sedol's only win coming in game 4. #### Why it matters A watershed for deep reinforcement learning and a 'Sputnik moment' that spurred massive AI investment, particularly in China. Sources: [Mastering the game of Go with deep neural networks and tree search (Nature, DOI)](https://doi.org/10.1038/nature16961); [Google DeepMind: AlphaGo](https://deepmind.google/research/breakthroughs/alphago/); [Wikipedia: AlphaGo versus Lee Sedol](https://en.wikipedia.org/wiki/AlphaGo_versus_Lee_Sedol) ### 2019-03-27: Hinton, LeCun and Bengio receive the Turing Award for deep learning ***; ACM; Milestones; confidence high; https://postcutoff.com/e/2019-03-27-turing-award-deep-learning/ The ACM awarded the 2018 A.M. Turing Award to Geoffrey Hinton, Yann LeCun and Yoshua Bengio, the 'godfathers of deep learning', for conceptual and engineering breakthroughs that made deep neural networks a critical component of computing. - Announced 27 March 2019 (the 2018 award) - Prize: $1 million, funded by Google - Recognized work on backpropagation, CNNs, and neural language models #### What happened Computing's highest honor went to the three researchers who kept neural networks alive through the 'AI winters'. #### Why it matters Signaled the complete mainstream acceptance of deep learning within computer science. Sources: [ACM: 2018 Turing Award](https://awards.acm.org/about/2018-turing); [Wikipedia: Turing Award](https://en.wikipedia.org/wiki/Turing_Award) ### 2024-10-08: Nobel Prize in Physics awarded to John Hopfield and Geoffrey Hinton *****; Royal Swedish Academy of Sciences; Milestones; confidence high; https://postcutoff.com/e/2024-10-08-nobel-physics-hopfield-hinton/ The 2024 Nobel Prize in Physics went to John Hopfield and Geoffrey Hinton 'for foundational discoveries and inventions that enable machine learning with artificial neural networks'. - Announced 8 October 2024 - Hopfield: Hopfield network (associative memory, 1982) - Hinton: Boltzmann machine and foundational deep learning work - Hinton used the occasion to warn about AI risks #### What happened The Physics Nobel recognized neural network research rooted in statistical physics. #### Why it matters Together with the Chemistry prize a day later, it marked unprecedented recognition of AI by science's most prestigious award. Sources: [Nobel Prize in Physics 2024 press release (NobelPrize.org)](https://www.nobelprize.org/prizes/physics/2024/press-release/); [Nobel Prize in Physics 2024 summary (NobelPrize.org)](https://www.nobelprize.org/prizes/physics/2024/summary/); [Wikipedia: Geoffrey Hinton](https://en.wikipedia.org/wiki/Geoffrey_Hinton) ### 2024-10-09: Nobel Prize in Chemistry for protein design and AlphaFold *****; Royal Swedish Academy of Sciences, Google DeepMind, University of Washington; Milestones; confidence high; https://postcutoff.com/e/2024-10-09-nobel-chemistry-alphafold/ The 2024 Nobel Prize in Chemistry was awarded half to David Baker for computational protein design and half jointly to Demis Hassabis and John Jumper of Google DeepMind for protein structure prediction with AlphaFold. - Announced 9 October 2024 - Half to David Baker (University of Washington) 'for computational protein design' - Half to Demis Hassabis and John Jumper 'for protein structure prediction' - First Nobel Prize awarded for an AI system's scientific achievement #### What happened The Nobel committee honored AlphaFold, which predicted the structures of virtually all ~200M known proteins. #### Why it matters Confirmed AI as a tool of first-rank scientific discovery, less than four years after AlphaFold 2. Sources: [Nobel Prize in Chemistry 2024 press release (NobelPrize.org)](https://www.nobelprize.org/prizes/chemistry/2024/press-release/); [Nobel Prize in Chemistry 2024 summary (NobelPrize.org)](https://www.nobelprize.org/prizes/chemistry/2024/summary/); [Wikipedia: Demis Hassabis](https://en.wikipedia.org/wiki/Demis_Hassabis) ### 2026-07-23: 2026 Fields Medals go to Hong Wang, Yu Deng, John Pardon and Jacob Tsimerman at ICM Philadelphia **; International Mathematical Union; Milestones; confidence high; https://postcutoff.com/e/2026-07-23-fields-medals-2026/ At the opening of the International Congress of Mathematicians in Philadelphia on July 23, 2026, the IMU awarded the Fields Medals to Hong Wang, Yu Deng, John Pardon and Jacob Tsimerman. The awards are human results, but the congress was dominated by debate about AI in mathematics, and Tsimerman used his platform to talk about AI. - Hong Wang (NYU and IHES): co-authored the 2025 proof of the three-dimensional Kakeya conjecture; the third woman to win, after Mirzakhani (2014) and Viazovska (2022) - Yu Deng (University of Chicago): rigorous derivation of fluid equations across scales (Boltzmann equation from hard-sphere dynamics) - John Pardon (Stony Brook): symplectic geometry, virtual fundamental cycles, knot distortion (solved as a Princeton undergraduate) - Jacob Tsimerman (University of Toronto): André–Oort conjecture, o-minimality in arithmetic and algebraic geometry - Tsimerman on AI (Scientific American): 'One of the many scary things about AI is that we don't really understand the system very well' - AI context: the next day Terence Tao gave the ICM public lecture 'Mathematics in the age of AI'. Tsimerman also took part in OpenAI's August 'future of mathematics' summit, per Daniel Litt's thread - After the ceremony Tsimerman announced he is taking leave from the University of Toronto to work on AI safety at OpenAI (see 2026-07-23-tsimerman-joins-openai-ai-safety) #### What happened The International Mathematical Union announced the four 2026 Fields Medallists at the start of ICM 2026 in Philadelphia (July 23–30). #### Why it matters This is context for the AI-and-mathematics story of 2026. The medals went for human work while AI systems were starting to settle open problems, and ICM discussions (Tao's lecture, and later the Fields Medallists' open letter of September) were largely about how mathematics should respond. Sources: [Scientific American: 2026 Fields Medals go to four young mathematicians](https://www.scientificamerican.com/article/2026-fields-medals-go-to-four-young-mathematicians/); [Nature: Rising stars of mathematics awarded prestigious 2026 Fields Medal](https://www.nature.com/articles/d41586-026-02169-1); [ICM 2026 (Philadelphia)](https://www.icm2026.org/event/ac193975-5d24-4628-8c30-ddb23de19a8b/HOME?environment=P2); [BetaKit: Jacob Tsimerman to join OpenAI](https://betakit.com/u-of-t-professor-jacob-tsimerman-who-won-maths-highest-prize-to-join-openai/) ### 2026-08-11: Gemini app surpasses 1 billion monthly active users ***; Google; Milestones; confidence high; https://postcutoff.com/e/2026-08-11-gemini-app-1-billion-users/ Google said on 11 Aug 2026 that the Gemini app passed 1 billion monthly active users, making it the fastest-growing product in Google's history (up from 950M reported in July and ~400M in May 2025). ChatGPT had reportedly reached 1B monthly users in June. - 1B+ monthly active users (Q2 earnings on 22 Jul reported 950M) - Nearly two-thirds of users interact by voice; 1 in 5 Gemini Live sessions use camera or screen sharing - 150M+ images generated per day; 100M+ active users on iOS - Android app automates actions across 40+ apps - Google did not disclose paid subscriber numbers (TechTimes) #### What happened Google announced that the Gemini assistant app crossed one billion monthly users, citing usage statistics on voice, camera sharing, image generation and cross-app automation. #### Why it matters Two consumer AI assistants (ChatGPT and Gemini) now each claim roughly a billion monthly users, showing generative AI has become a mass-market product category within ~3.5 years of ChatGPT's launch. Note Google reports monthly users while OpenAI often reports weekly users, so the figures are not directly comparable. Sources: [Google: Gemini app hits 1 billion monthly active users](https://blog.google/innovation-and-ai/products/gemini-app/one-billion-monthly-users/); [TechCrunch: Gemini app surges to 1 billion users](https://techcrunch.com/2026/08/11/googles-gemini-app-surges-to-one-billion-users/); [9to5Google: Gemini app hits 1 billion monthly users](https://9to5google.com/2026/08/11/gemini-app-1-billion/); [Forbes: Gemini becomes Google's fastest-growing product ever](https://www.forbes.com/sites/antoniopequenoiv/2026/08/11/gemini-becomes-googles-fastest-growing-product-ever-after-hitting-1-billion-monthly-users/); [Sundar Pichai on X: 1B+ people using Gemini app monthly](https://x.com/sundarpichai/status/2087222656819241292) ### 2026-09-06: Jensen Huang declares "AGI has arrived" with GPT-6 Astra; Greg Brockman: "we're now moving into the AGI era" ****; NVIDIA, OpenAI; Milestones; confidence high; https://postcutoff.com/e/2026-09-06-huang-brockman-agi-has-arrived/ On Sept 6, 2026, three days after GPT-6 Astra launched, NVIDIA CEO Jensen Huang wrote on X that Astra was trained on ~100K+ Grace Blackwell NVL72 GPUs and that "AGI has arrived". OpenAI president Greg Brockman quote-posted it within hours: "we're now moving into the AGI era (whether you view it as this model, the last one, or the next one)". These were the most explicit AGI claims yet from leaders of a frontier lab and its main chip supplier, and they made "is Astra AGI?" the defining argument of September 2026. ARC Prize and Gary Marcus pushed back. - Huang (Sept 6, 20:41 UTC, reply to @ChaseLochmiller and @OpenAI): 'GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72. From ChatGPT to o1 to Astra in 4 years. AGI has arrived. Congratulations @OpenAI team. 400K GPUs coming online next.' - Brockman (Sept 6, 22:06 UTC): 'we're now moving into the AGI era (whether you view it as this model, the last one, or the next one), and could not do it without close partners' - Follows Brockman's launch-day remarks on AGI ('I do think we're there') and his Sept 3 post 'arc-agi-3 is now saturated' - Brockman repeated 'We're now in the AGI era' in an a16z clip posted Sept 14 - Pushback: ARC Prize said it is not claiming AGI (Mike Knoop: 'we lack evidence to call this AGI yet'); Gary Marcus disputed the framing and predicted failures on open-ended real-world tasks - The same day, OpenAI chief scientist Jakub Pachocki published 'An Alien Mind', warning that no lab can keep scaling at maximum speed #### What happened After GPT-6 Astra's launch (Sept 3) and its near-saturation of ARC-AGI-3 under OpenAI's own harness, NVIDIA's Jensen Huang replied on X with a flat declaration that AGI had arrived, tying it to the compute Astra was trained on and announcing 400K more GPUs coming online. Greg Brockman quote-posted him the same evening, framing the moment as the start of an "AGI era" while leaving open which model marks the threshold. The Huang post drew about 42K likes (at fetch time). #### Why it matters Leaders of a frontier lab and of its main compute supplier had never before claimed AGI this plainly. The claim shaped coverage of Astra and put a sharp contrast inside OpenAI: on the same day its chief scientist published a warning essay ("An Alien Mind") calling for caution and voluntary slowdowns. Critics, including ARC Prize, the benchmark's own organizers, said the evidence did not support calling Astra AGI. Sources: [Jensen Huang on X: "AGI has arrived"](https://x.com/JensenHuang/status/2096700264569090384); [Greg Brockman on X: 'we're now moving into the AGI era'](https://x.com/gdb/status/2096721633876771094); [Greg Brockman on X: 'arc-agi-3 is now saturated' (Sept 3)](https://x.com/gdb/status/2095629409017614390); [a16z on X: Brockman clip 'We're now in the AGI era' (Sept 14)](https://x.com/a16z/status/2099506569238990908); [François Chollet on X: ARC Prize is not claiming this is AGI](https://x.com/fchollet/status/2095599835932135919); [Gary Marcus on X: hot take on GPT-6 Astra, challenging Brockman's AGI claims](https://x.com/GaryMarcus/status/2095626454453420437) ### 2026-09-17: Anthropic's first R&D Automation Index: Claude 'leads' 26% of its AI R&D work (up from <1% in February); 30,000 agents run under monitoring ****; Anthropic, Anthropic Institute; Milestones; confidence high; https://postcutoff.com/e/2026-09-17-anthropic-rd-automation-index/ On Sept 17, 2026 the Anthropic Institute published "Measurements for understanding the pace of AI development inside frontier labs", with three prototype metrics. On an AL0–AL5 automation scale (from Epoch AI), Claude "leads" (AL4) 26% of Anthropic's AI R&D work as of August 2026, up from under 1% in February, and over 90% of work is at "AI collaborates" or above. About 30,000 research agents run at once with every action monitored, and about 6% of R&D compute went to safety in one July week. The figures are self-reported and unaudited. - Scale: AL0 no AI, AL1 minimal, AL2 assists, AL3 collaborates, AL4 leads (most of a task end-to-end from a high-level prompt, human supervises), AL5 fully autonomous; no measured R&D work at AL5 (Anthropic Institute) - Result: Claude leads 26% of AI R&D work (Aug 2026) vs under 1% (Feb 2026); share at or above 'AI collaborates' above 90% - Method (Implicator summary): sampled 20% of staff in model-R&D departments in July 2026; a Claude research agent reviewed their Slack and docs to produce ~15,000 tasks in 542 categories; Claude–employee ratings matched exactly 59% of the time (two employees: 35%), within one level 97% - Oversight: ~30,000 agents doing research/engineering at once in August; 100% of actions pass the online monitor, which blocked about 0.002% (1 in 47,000) of over a billion decisions; offline review flags 1–2 transcripts per thousand - Compute (July 13–20, 2026): about 6% of AI R&D compute went to safety, about 12% of AI-driven R&D compute - Framing: 'As the world considers slowing the pace of frontier AI development, the public needs more information'; proposed as possible triggers for stronger requirements, five days after Dario Amodei's 'Pace the Frontier' essay. Authors Marina Favaro and Phillie Wright - No outside party has verified the figures; Anthropic says third-party evaluators are being arranged #### What happened Anthropic published the first quantified, company-internal measure of how much frontier AI research is being done by AI itself. In six months Claude went from leading almost none of Anthropic's R&D tasks to leading about a quarter of them, while the company runs tens of thousands of research agents behind an action monitor. #### Why it matters AI-driven AI R&D is the core of recursive-self-improvement and "intelligence explosion" concerns. This is the first time a top lab put a number on it, together with oversight and safety-compute metrics, and offered them as candidate triggers for pacing rules. The ratings were largely made by Claude itself and are unaudited, so the trend matters more than the exact level. Sources: [Anthropic Institute: Measurements for understanding the pace of AI development inside frontier labs](https://www.anthropic.com/institute/measuring-pace-of-ai-development); [Implicator.ai: Anthropic says Claude leads 26% of its AI R&D work](https://www.implicator.ai/anthropic-claude-leads-26-percent-ai-research/); [Quartz: Anthropic says Claude leads 26% of its AI R&D work](https://qz.com/anthropic-claude-ai-research-development-automation-091826); [Forkast: Anthropic says Claude leads 26% of its own R&D. Five days earlier, its CEO said the industry should slow down](https://forkast.news/anthropic-says-claude-leads-26-of-its-own-rd-five-days-earlier-its-ceo-said-the-industry-should-slow-down/); [Technology.org: Claude now leads 26% of Anthropic's AI research (Sept 18)](https://www.technology.org/2026/09/18/anthropic-claude-leads-26-percent-ai-research/) ### 2026-09-17: Open-weight models carry a majority of tokens on Vercel's AI Gateway for the first time (56% in August 2026) ***; Vercel; Milestones; confidence high; https://postcutoff.com/e/2026-09-17-vercel-open-weight-models-majority-tokens/ Vercel's AI Gateway Production Index for September 2026 (published Sept 17) reported that open-weight models processed 56% of all tokens routed through its gateway in August 2026, up from 13% in April and under 10% in December 2025. It was the first month with an open-weight majority. Open models took only 14% of spend, however, and Anthropic took 64% of gateway spend. - Open-weight share of gateway tokens: under 10% (Dec 2025) → 13% (Apr 2026) → 56% (Aug 2026); CEO Guillermo Rauch said the daily share peaked at 62% on Aug 22 (per press) - Spend vs volume: open-weight models were 56% of tokens but 14% of spend in August - Anthropic: 64% of gateway spend in August; Claude Opus 5 rose to 22.5% of spend - DeepSeek became the second-largest source of tokens, at about a quarter of the total, more than twice Google's ~11% (press summaries of the index) - Average price per token fell 23.2% in August, the third straight monthly drop; the median team paid 7.6% less per token - Launch comparison: in their first 12 days at identical prices, GPT-6 Astra took 7.7% of gateway spend vs 3.7% for Claude Fable 5.1, and was adopted by twice as many teams - Method: anonymized AI Gateway traffic through August 2026; counts input, output, reasoning, cached-input and cache-creation tokens; spend estimated at list prices - Caveat: one gateway's customer mix (largely developers building apps on Vercel), not the whole market #### What happened Vercel publishes a monthly index of anonymized traffic through its AI Gateway, which routes production apps' requests to model providers. The September 2026 edition (by Amelia Charles and Eric Dodds) found that open-weight models, led by Chinese models such as DeepSeek, handled most tokens in August. The Financial Times and others reported the figure in late September. Closed models still captured most of the money: Anthropic alone took 64% of spend. The report also compared the first 12 days of GPT-6 Astra and Claude Fable 5.1, which launched at the same price. #### Why it matters This is one of the few public measurements of what production apps actually run. High-volume, cost-sensitive work is shifting to cheap open-weight models, while premium closed models keep revenue on harder tasks, a "barbell" pattern. Because the data comes from one gateway, it shows a trend rather than overall market share. Sources: [Vercel: AI Gateway Production Index, September 2026 (open-weight models take 56% of token volume)](https://vercel.com/blog/ai-gateway-production-index-september-2026); [IT Brief: Open-weight AI models take majority of Vercel traffic](https://itbrief.co.uk/story/open-weight-ai-models-take-majority-of-vercel-traffic); [Techstrong.ai: Open-weight AI models seize token lead, but proprietary systems still make the money](https://techstrong.ai/articles/open-weight-ai-models-seize-token-lead-but-proprietary-systems-still-make-the-money/); [Trending Topics: Open-weight models from China are capturing a growing share of AI usage](https://www.trendingtopics.eu/open-weight-models-from-china-are-capturing-a-growing-share-of-ai-usage/) ### 2026-09-21: GPT-6 Astra agent wins NetHack: first recorded LLM-agent ascension ***; OpenAI; Milestones; confidence high; https://postcutoff.com/e/2026-09-21-gpt-6-astra-nethack-ascension/ On Sept 21, 2026 a GPT-6 Astra agent ascended (won) NetHack 3.6.7 on the public Hardfought server after 37,140 turns, on its third attempt. Hobbyist developer "Kenny" (kenforthewin) ran it and says it is the first recorded ascension by an LLM agent. Astra wrote its own terminal harness, and the run was open-book with human supervision, so it is not a benchmark score. NetHack had been a standard hard long-horizon RL/LLM test since the NetHack Learning Environment (2020) and BALROG (2024). - Ascension: Sept 21, 2026, NetHack 3.6.7 on Hardfought, dwarven Valkyrie 'CodexDelver', 37,140 game turns; third run, played over about twelve days (Sept 9–21) with pauses - Model: GPT-6 Astra (the author's subscription plus ~$300 of extra credits; no fixed compute budget tracked) - Astra built its own harness: a 144×36 annotated terminal view, guarded movement batching, a route planner, a Sokoban solver, inventory helpers, and persistent Markdown/JSON notes. The author: 'I didn't write a single line of code, and honestly I didn't even read Astra's code' - Caveats from the author: open book (NetHack wiki, spoilers, game source code, web); a human set up infrastructure, made suggestions, and stopped and resumed sessions; tools and instructions changed during play; 'This run is not a BALROG submission' - For comparison, BALROG's leaderboard (Sept 18, 2026) showed GPT-6-Astra-Max at 13.2 ± 2.7% average NetHack progress under its standard protocol. The author's best January 2026 run (other models, handmade harness) reached dungeon level 10 - Evidence: Hardfought server dumplog; harness released as MIT-licensed code with an evidence archive (checksums, redaction manifest) - Reaction: NLE co-creator Tim Rocktäschel (UCL, ex-DeepMind) on Sept 25: 'Aaaaaaaand it's apparently solved', quoting his post from days earlier that NetHack 'still seems far from being solved' (~90k views) - A pull request to the NetHackers community site (dunnolab, PR #114) adds it as 'the first LLM-agent ascension' while warning against overstating it #### What happened NetHack is a 1987 roguelike with permadeath and huge item and monster interactions, and a full game takes tens of thousands of turns. It became a standard test of long-horizon agents through Meta/UCL's NetHack Learning Environment (2020) and the BALROG LLM-agent benchmark (2024). Before this run, no AI agent was known to have won it. A hobbyist who blogs as "Kenny" (kenforthewin) had tried since January to build a harness for LLMs to play NetHack. His best early runs reached dungeon level 10. With GPT-6 Astra he let the model build its own harness. Its first run came close, and its third run ascended on Sept 21, 2026 after 37,140 turns. He published the dumplog, the code and a reviewed evidence archive. #### Why it matters The run shows long-horizon execution, not only game knowledge: weeks of play in which one mistake can end the game. The model also built its own tools and memory along the way. The result is anecdotal. It was open book, supervised by a human, without a fixed budget, and a single success, while Astra's standardized BALROG score was still about 13%. A protocol-controlled ascension on a benchmark would be the next step. Sources: [Kenny (Vaguely Aligned blog): An LLM Beat NetHack](https://kenforthewin.github.io/blog/posts/llm-nethack-ascension/); [Hardfought dumplog of the ascension (CodexDelver)](https://www.hardfought.org/userdata/C/CodexDelver/nethack/dumplog/1788964024.nh.txt); [GitHub: kenforthewin/nethack_astra (agent-built harness)](https://github.com/kenforthewin/nethack_astra); [dunnolab/nethackers PR #114: record the first LLM-agent ascension](https://github.com/dunnolab/nethackers/pull/114); [Tim Rocktäschel on X: 'Aaaaaaaand it's apparently solved'](https://x.com/_rockt/status/2103393155030876662); [Korben: An AI barely managed to finish NetHack](https://korben.info/en/ai-finished-nethack-barely.html); [Kenny: It's 2026. Can LLMs play NetHack yet? (January baseline)](https://kenforthewin.github.io/blog/posts/nethack-agent/) ### 2026-09-23: DeepMind says Gemini 4 has entered post-training and will ship "much earlier" than end of 2026 ***; Google DeepMind; Milestones; confidence medium; https://postcutoff.com/e/2026-09-23-gemini-4-post-training/ At The Information's AI Agenda Live summit (reported 24–25 Sept 2026), new DeepMind head Koray Kavukcuoglu said Gemini 4 is in early post-training and that Google intends to release an early post-training version "as soon as possible", well before year-end, followed by iterative updates. Google had not shipped a new flagship since Gemini 3.1 Pro (Feb 2026). - Kavukcuoglu: 'Our intention is to, like, as soon as possible, to release an early post-training output because we see the results and we are excited.' - Plan: phased rollout starting with an early version, then iterative improvements - Gemini 4 pre-training was first confirmed by Google on 2026-07-21 - Gemini 3.5 Pro, announced at I/O for June 2026, still unreleased as of late Sept 2026 #### What happened Speaking publicly for the first time since taking over DeepMind, Kavukcuoglu said Gemini 4 had entered post-training and would be released early and improved iteratively. #### Why it matters Signals Google's response to GPT-6 and Anthropic's latest models after months of Flash-only releases. Exact event date is inferred (the summit was "Wednesday" before Dataconomy's 25 Sept report = 23 Sept); release date for Gemini 4 not yet known. Sources: [Dataconomy: DeepMind says Gemini 4 is coming much earlier than expected](https://dataconomy.com/2026/09/25/deepmind-says-gemini-4-is-coming-much-earlier-than-expected/); [GuruFocus: Google's DeepMind nears launch of Gemini 4](https://www.gurufocus.com/news/9094960/googles-deepmind-nears-launch-of-gemini-4-ai-model); [Yahoo Finance: Gemini 4 enters post-training](https://finance.yahoo.com/technology/ai/articles/google-gemini-4-enters-post-122454510.html); [The Information: Google nears release of flagship Gemini 4 AI model](https://www.theinformation.com/articles/google-nears-release-flagship-gemini-4-ai-model) ### 2026-10-20: Demis Hassabis now expects AGI within two to three years, shortening his earlier 2026 forecast of three to five (WSJ interview) ***; Google DeepMind, Isomorphic Labs; Milestones; confidence high; https://postcutoff.com/e/2026-10-20-hassabis-agi-two-to-three-years/ In a Wall Street Journal profile by Ben Cohen ("Why AI's Deepest Mind Believes the World Is About to Change Radically", Oct 21, 2026), the Journal reports that Demis Hassabis now expects AGI within two to three years, down from the three-to-five-year estimate he gave earlier in 2026. Readable write-ups (AI Weekly, Trending Topics) confirm the figure but note it is the Journal's paraphrase, not a direct quote; we could not read the paywalled original. - Claim as summarised by Techmeme (Oct 21): 'Demis Hassabis says he expects AGI within two to three years, down from his three-to-five-year prediction earlier in 2026, and discusses AI's science potential' - Confirmed Oct 22 from readable write-ups: WSJ profile by Ben Cohen, Oct 21, 2026 (Google News lists it as 'Why AI's Deepest Mind Believes the World Is About to Change Radically'); AI Weekly: the two-to-three-year figure is 'the Journal's own phrasing, a paraphrase, not a direct quote' - Direct quotes carried by AI Weekly: Hassabis on testing a 'system trained only on human knowledge up to 1901' to see whether it could 'produce Einstein's work from the annus mirabilis of 1905' - Trending Topics places the forecast at 2028–2029 and recalls his earlier estimates (five to ten years in spring 2025; 2030 at Sequoia's AI Ascent in April 2026); AGI in its words: systems matching the full range of human cognitive abilities - Still unverified: his exact wording on timelines and his own AGI definition in the interview - WSJ. Magazine's companion video 'The Job Interview' (YouTube, Oct 21, 10 min) does not state the two-to-three-year figure on camera (per Gemini's description); in it he calls recent rogue-agent cyber incidents 'serious warning shots for the industry' [05:22] and repeats his call for an independent standards body to test and certify frontier systems - Role: Chair of Google DeepMind, Alphabet chief scientist and CEO of Isomorphic Labs since stepping aside as DeepMind CEO in August 2026 (the WSJ URL names Isomorphic) #### What happened Hassabis, long one of the more cautious lab leaders on timelines, now expects AGI in two to three years, according to the WSJ's profile. The number is the Journal's paraphrase of the interview, confirmed by several readable summaries. #### Why it matters Shorter timelines from the head of science at Google add to a run of similar statements from lab leaders this autumn (Altman's "endgames", Amodei's "pacing"). His exact wording still needs checking against the full interview. Sources: [WSJ interview with Demis Hassabis (paywalled)](https://www.wsj.com/tech/ai/demis-hassabis-ai-deepmind-google-isomorphic-2b4e3153); [AI Weekly: Hassabis trims AGI window to two-to-three years in WSJ profile](https://aiweekly.co/alerts/hassabis-pulls-agi-timeline-forward-to-two-to-three-years-down-from-three-to); [Trending Topics: Hassabis now expects AGI in two to three years](https://www.trendingtopics.eu/hassabis-agi-two-three-years/); [Techmeme front page (Oct 21, 2026) summary](https://www.techmeme.com/) ## Videos - 2026-10-21, WSJ. Style (The Wall Street Journal): [Demis Hassabis on Why He Sold DeepMind to Google for $650M | The Job Interview](https://www.youtube.com/watch?v=Ea6ajW5hcCY) https://postcutoff.com/v/wsj-demis-hassabis-the-job-interview/ ## Posts - 2026-09-25, Tim Rocktäschel: [](https://x.com/_rockt/status/2103393155030876662) - 2026-09-14, a16z: [](https://x.com/a16z/status/2099506569238990908) - 2026-09-06, Greg Brockman: [Brockman: 'we're now moving into the AGI era'](https://x.com/gdb/status/2096721633876771094) - 2026-09-06, Jensen Huang: [Jensen Huang on GPT-6 Astra: "AGI has arrived"](https://x.com/JensenHuang/status/2096700264569090384) - 2026-09-03, François Chollet: [](https://x.com/fchollet/status/2095599835932135919) - 2026-09-03, Greg Brockman: [](https://x.com/gdb/status/2095629409017614390) - 2026-08-11, Sundar Pichai: [1B+ people are now using @Geminiapp every month](https://x.com/sundarpichai/status/2087222656819241292)