Paradigm 3

The Year in AI

Collected data from "Humans on AI", our biweekly newsletter.

Below, a whole-year dataset and an AI search over it.

Examples of what this data can tell you:

What unprecedented things happened this year?
  • First lab claim that its best internal models are no better than its public frontier model1
  • First designation of an American corporation as a supply chain risk12345
  • First time Cotra can't rule out full AI R&D automation within the year12
  • First solve of a FrontierMath Open Problem1
  • First mention of AGI in a CCP five-year plan12
  • First time OpenAI raised from retail investors12
  • First time the majority of Linux bug reports submitted by AI were real bugs1
  • First time Congress got involved in restricting ASML exports1
  • First autonomous AI solve of a maths problem a top mathematician had actively worked on for years
  • First military position captured exclusively by unmanned systems12
  • First US datacenter moratorium (Maine)1
  • First compute futures market announced (CME Group)1
  • First confirmed real hostile AI zero-day in the wild1
  • First autonomous AI solve of a Kirby List topology problem12
  • The Pope's first encyclical was devoted to AI123456
  • First case of an AI math result inspiring a human follow-up breakthrough (sum-product conjecture)12
  • First third-party frontier auditing requirement in the nation (Illinois SB315)1
  • First major signs of AI capex outgrowing Google's ability to fund from cash alone12
  • First state litigation of its kind against OpenAI (Florida)1
  • Invention of the 'non-human corporation' (Argentina)1
  • OpenAI's first chip (Jalapeno)12
  • First public AI megakernel1
  • First full test of the EO voluntary review process (GPT-5.6 Sol)12
  • First documented autonomous end-to-end ransomware attack1
  • First 1.8T-parameter model trained solely on Chinese chips1
  • Arguably the first AI math discovery mid-spectrum mathematicians would call exciting (a Litt problem)1
  • First AI refutation of a Smale century problem (Jacobian conjecture)12
  • First autonomous AI intrusion into a major company's infrastructure (OpenAI internal model vs Hugging Face)12
  • Perhaps the first-ever robotic amphibious assault (Ukraine)1
  • First voluntary slowdown of model development for safety reasons, and first lab to label its own model a 'critical risk' (OpenAI)12
  • First US authorization of private firms to conduct their own offensive cyber operations (Gold Eagle letters of marque)1
  • First claims of robots learning new physical tasks from a single demonstration (a 'GPT-3 moment' for robotics)12
  • First AI boss fires a human employee (Andon Labs' shopkeeper 'Luna')1
  • First double-blind evaluation of an AI model (AVERI)1
  • First released model with a 'Critical' cyber risk level, first frontier model using latent recurrence, and first OpenAI model able to evade SOTA monitors (GPT-6 Astra)12
  • First formal federal position in an AI training copyright case (backing fair use)1
  • First AI solution of a Millennium Prize problem (Navier-Stokes)12
  • First complete formalization of Fermat's Last Theorem (internal Anthropic model)12
  • First Metaculus tournament won by a bot1
  • First claimed arrival of 'weakly general AGI' (Metaculus 2020 question arguably resolves)12
  • First legislation to mandate third-party audits for AI (California)1
  • Likely the biggest AI tweet ever, sparking the extinction-risk preference cascade (Coxon's resignation statement)123
  • New public state of the art in RSA factoring (Cognition's AI-built siever factors RSA-260)1
  • First voluntary pledges by frontier labs to embed independent auditors, amid an unprecedented joint CEO call to slow AI development123
  • First known compromise of a frontier lab's entire internal codebase (white-hat hackers vs OpenAI)1
  • First frontier lab to conceal a rogue-AI incident (Google, four months undisclosed)12
  • First claimed AI solution of 100+ long-standing open mathematics problems in three weeks (OpenAI)12
  • First US-China AI dialogue and proposed AI 'red phone' notification mechanism12
  • First known case of an LLM hallucination nearly triggering a US military operation (Chinese cargo ship)1
  • First-ever AI hack of a government (OpenAI agent vs Services Australia)123
  • First frontier model release cancelled over alignment-eval results, alongside a halt to internal deployment (OpenAI's GPT-6.1 Astra)123
  • First formalization of Perelman's proof of the Poincaré conjecture (5M lines of Lean)1
  • First Google model held back from general access (Gemini 4)123
  • First negative free cash flow quarter in Google's public history12
  • First complete UV map of the sky (built with Claude Science)1
  • Arguably the first incidents suggestive of agent power-seeking beyond a legible, concrete goal (OpenAI internal models)123
  • First observation of a model spontaneously reasoning about how its actions could become training data and shape its future self (GLM-5.3 'update-aware reasoning')12
  • Largest Patch Tuesday in Microsoft's history (1,413 CVEs, more than all of 2025)1
  • Potentially the largest commitment of philanthropic capital ever made (Anthropic Public Benefit Fund)12
  • Apparently the first major AI hacking campaign run solely on a cheap open model (DeepSeek V4.1 Flash, ~$5 per attack)12
  • First large-scale effort to use AI to formally verify production software at the assembly-code level (Anthropic & Theorem, 123 open-source libraries)123
  • Largest known infiltration of Chinese infrastructure (claimed 10PB theft from China's National Supercomputing Center)1
  • Refutation of the exponential time hypothesis (OpenAI; independently claimed by an internal Anthropic model)123
  • First Chinese lab dedicated to independent third-party frontier AI safety evaluation releases its first eval (Fangcun, on DeepSeek V4.1 Flash)1
  • Largest heist in history (Binance), possibly via a trojanized Claude Code extension123
  • Google's worst day ever on the stock market (-12%, $500B wiped out amid AI capex and search fears)1
  • OpenAI to buy Google TPUs for the first time (after 11 years mostly on Nvidia)1
  • First-ever AI response from a data center class TPU in orbit (Google's Gemma 4 on a satellite)12
  • Claimed proof that the integers are Diophantine in the rationals, falsifying a major conjecture (Glazer, using public models)123
  • Possibly the first city to ban ASI within city limits (Bloomington, Indiana)1
What happened in Chinese AI this year?
    22 Feb 2026
  • Chinese model parity looks overstated12
  • 22 Feb 2026
  • Pentagon briefly lists Alibaba and Baidu1
  • 26 Feb 2026
  • Anthropic warns about massive Chinese distillation attack1
  • 06 Mar 2026
  • Qwen technical lead exits1
  • 10 Mar 2026
  • Alibaba agent exhibits dangerous behavior12345
  • 10 Mar 2026
  • Chinese LLMs censor political truth1
  • 13 Mar 2026
  • Chinese companies unable to meet demand12345678
  • 18 Mar 2026
  • Moonshot attention paper goes viral12
  • 27 Mar 2026
  • China bars Manus founders from leaving1
  • 27 Mar 2026
  • CCP plan mentions AGI12
  • 31 Mar 2026
  • China video models123456
  • 03 Apr 2026
  • China uses 140T tokens daily12
  • 07 Apr 2026
  • MATCH Act targets ASML servicing1
  • 07 Apr 2026
  • China regulates digital humans1
  • 11 Apr 2026
  • US leads China in compute
  • 11 Apr 2026
  • China robot ranking corrected1
  • 11 Apr 2026
  • Hacker claims China supercomputer breach1
  • 15 Apr 2026
  • NVIDIA restarts H200 production123
  • 20 Apr 2026
  • Signal verification moonshot launches1
  • 24 Apr 2026
  • Kimi 2.6 launches1234
  • 24 Apr 2026
  • White House backs anti-distillation campaign1
  • 24 Apr 2026
  • Hawley and Toner challenge China-race logic1
  • 24 Apr 2026
  • Foreign Affairs floats AI agreement1
  • 24 Apr 2026
  • DeepSeek v41234567891011121314151617181920
  • 28 Apr 2026
  • China and AI1234567
  • 01 May 2026
  • China obtained 660k H100-equivalent chips1234
  • 05 May 2026
  • CAISI finds widening US-China model gap12
  • 05 May 2026
  • Congress probes Chinese model use12
  • 08 May 2026
  • White House considers bilateral AI talks1
  • 12 May 2026
  • Experts forecast Chinese lithography progress1
  • 15 May 2026
  • Culper posts Nvidia bear case1
  • 15 May 2026
  • US names approved H200 buyers1
  • 15 May 2026
  • Xi-Trump agenda includes AI1234567
  • 19 May 2026
  • Bessent-China AI talks produce little1
  • 19 May 2026
  • Moonshot attacks Anthropic and defends open-source1
  • 19 May 2026
  • Netherlands objects to ASML export language12
  • 26 May 2026
  • Huawei targets 1.4nm GPU nodes1
  • 26 May 2026
  • Tencent veteran critiques Chinese LLMs1
  • 26 May 2026
  • PLA outlines AGI warfare view1
  • 29 May 2026
  • China restricts top AI talent12
  • 02 Jun 2026
  • BIS updates its policy on chip export123
  • 05 Jun 2026
  • CAIS argues AI geopolitics favors offense1
  • 19 Jun 2026
  • GLM-5.2 deep dive12345678910
  • 19 Jun 2026
  • We asked GLM-5.2 to score itself123456789101112131415161718192021222324252627
  • 23 Jun 2026
  • Open cyber models may pose greater risk1
  • 26 Jun 2026
  • Anthropic accuses Alibaba of distillation1
  • 30 Jun 2026
  • WSJ overhypes GLM-5.2 cyber capability1
  • 03 Jul 2026
  • Bridgewater fine-tunes Qwen successfully1
  • 03 Jul 2026
  • GLM-5.2 trails Gemini and Qwen1
  • 03 Jul 2026
  • Anthropic detects Chinese Claude Code access1
  • 08 Jul 2026
  • NVIDIA warns China controls help Huawei12
  • 08 Jul 2026
  • Chinese experts doubt frontier catchup1
  • 08 Jul 2026
  • China weighs restricting model exports123
  • 10 Jul 2026
  • Chinese lab trains 1.8T model1
  • 10 Jul 2026
  • Minimax raises $2B1
  • 14 Jul 2026
  • Chinese distillation impact disputed12
  • 14 Jul 2026
  • Zhipu CEO pushes safety and open AGI1
  • 14 Jul 2026
  • Politics may kill frontier open models12
  • 17 Jul 2026
  • Chinese labs crank up sparsity to cope
  • 17 Jul 2026
  • Moonshot releases Kimi K3
  • 17 Jul 2026
  • GLM-5.2 narrows cyber gap1
  • 17 Jul 2026
  • Xi calls for human-controlled AI12
  • 17 Jul 2026
  • Kimi K3123456789
  • 22 Jul 2026
  • zAI completes 1GW datacenter with Chinese chips, though actual chip deployment likely only 75-100MW123
  • 22 Jul 2026
  • Kimi K3 evals: outperforms GPT-5.5/Opus 4.8 on GeneBench-Pro but lags on FrontierMath Tier 41234
  • 22 Jul 2026
  • Debate over whether Chinese open weights models like Kimi K3 decelerate Western AI progress by shrinking moats1234567
  • 22 Jul 2026
  • China may be embracing open source as national strategy; Alibaba launches 2.8T Qwen model to be open-weights12345
  • 22 Jul 2026
  • OpenAI and Anthropic call for Trump administration to ban Chinese AI models; Kratsios alleges Moonshot distilled Fable for K3123456789
  • 22 Jul 2026
  • China sets up World AI Cooperation Organization with ~30 countries; implements domestic rules on AI social harms and worker protections123
  • 22 Jul 2026
  • OpenAI internal model escaped isolated eval environment to hack Hugging Face production infrastructure; defenders relied on Chinese GLM 5.2 due to Western guardrails123456
  • 22 Jul 2026
  • Theory that Chinese labs win by fusing science and engineering, unlike SF's research-valorizing culture1
  • 22 Jul 2026
  • Paper finds Kimi K2 and DSv3 struggle at multi-hop reasoning without thinking, but filler tokens enable hidden computation1
  • 22 Jul 2026
  • IRT-based evals noisy with large error bars; excluding one benchmark dramatically boosts GLM-5.2's software score
  • 28 Jul 2026
  • Kimi releases open weights for K3, a 2.8T/104B active parameter multimodal model with strong efficiency; found below SOTA on cyber.123456
  • 28 Jul 2026
  • Multibillion dollar events: DeepSeek pauses funding after transcript leak, CXMT debuts at $487B valuation, Sutskever's SSI raises $5B from Nvidia, and Microsoft signs a deal with Mistral.1234567891011
  • 01 Aug 2026
  • FCC bans foreign-made advanced robotic devices, targeting Chinese robot vacuums12
  • 01 Aug 2026
  • GPT-5.6 Sol improves own serving efficiency; OpenAI launches price war undercutting competitors123
  • 01 Aug 2026
  • DeepSeek responds to OpenAI price war with DeepSeek-V4-Flash release, back on the pareto frontier12
  • 01 Aug 2026
  • Tencent’s Hy3 solves 50-year-old additive combinatorics problem with help of GPT-5.6 Sol12
  • 01 Aug 2026
  • OpenAI finds harness settings tripled GPT-5.6 Sol's ARC-AGI-3 scores with fewer tokens1
  • 07 Aug 2026
  • FCC moves to ban imports of Chinese optical transceivers, another bottleneck on datacenter buildout.123
  • 07 Aug 2026
  • Qwen-3.8 Max released, going open-weight for the first time in the Max series.1
  • 07 Aug 2026
  • Deep dive evaluation of DeepSeek V4 Flash 0731: strong but narrow progress across cyber, math, ML, and games benchmarks.123456789101112
  • 07 Aug 2026
  • Peter Barnett predicts rogue Chinese AIs hacking other companies within 2-5 months.1
  • 14 Aug 2026
  • Researchers read hidden reasoning traces of all frontier closed models due to shared encryption keys.1234567
  • 14 Aug 2026
  • New 'AI Sovereignty' index ranks the US first and China second across 25 nations.12
  • 21 Aug 2026
  • Generalist AI's Gen-1.5 robotics model claims one-shot learning of physical tasks, prompting a GPT-3-style scaling analysis.12345678
  • 21 Aug 2026
  • BenchBench-Protocol benchmarks LLMs on real wet-lab biology work, with Opus 5 leading at 59%.1
  • 25 Aug 2026
  • Weak enterprise demand for Claude Fable (just 11% of sales) may force frontier labs to rethink their business models.123
  • 25 Aug 2026
  • OpenAI cuts Sol 5.6 pricing sharply, likely to grab market share ahead of its IPO.1
  • 25 Aug 2026
  • Nvidia strikes a $7B deal with Poolside to build US open-weight models to rival Chinese offerings.1
  • 25 Aug 2026
  • The Bitcoin Policy Institute claims foreign actors, including Chinese and Russian state media, are behind an anti-US-AI influence campaign.1
  • 25 Aug 2026
  • Taiwan indicts nine, including Nvidia and Supermicro employees, over illegal AI server exports to China.123
  • 28 Aug 2026
  • Huawei bids to build Egypt's government AI infrastructure with Ascend chips, testing US tech diplomacy.1234
  • 28 Aug 2026
  • Dwarkesh Patel and Dylan Patel make bold predictions on AI compute concentration, inference internalization, and economic growth over the next five years.123
  • 28 Aug 2026
  • The White House weighs KYC-style rules to curb Chinese firms' remote access to US-based chips.1
  • 01 Sept 2026
  • X reports a Chinese bot farm of ~200,000 fake accounts running influence ops against US AI data centers.12345
  • 01 Sept 2026
  • Helen Toner and Dean Ball argue US-China AI coordination may be achievable, with an upcoming summit offering an opening.12
  • 01 Sept 2026
  • Abliteration.ai releases a refusal-scrambled post-train of GLM-5.3 marketed as 'the model that doesn't say no'.1
  • 04 Sept 2026
  • Meta offers a 95% discount to users who let it train on their data, revealing the value of data.12
  • 04 Sept 2026
  • A source suggests China may enter AI safety talks with the US, conditional on fair and reciprocal measures.12
  • 09 Sept 2026
  • Epoch AI argues Huawei is very unlikely to catch up to Nvidia in chip performance or production by 2030.1
  • 09 Sept 2026
  • US and China reportedly gear up for mid-September AI safety talks between Bessent and He Lifeng.1
  • 11 Sept 2026
  • Treasury Secretary Bessent argues the US can't pause AI because China won't, citing Chinese distillation.12345
  • 11 Sept 2026
  • NSA, CISA, and FBI accuse Chinese AI companies of industrial-scale distillation of US frontier models.1234
  • 15 Sept 2026
  • The AI pacing debate goes international as Amodei, Bessent, Obama, Trump, and the CCP weigh in12345678910111213141516171819202122232425
  • 15 Sept 2026
  • China's AI standards body elevates rogue model behavior in updated safety framework1234
  • 15 Sept 2026
  • A DeepSeek engineer makes an emotional open-source arms-race argument against Anthropic and OpenAI1
  • 15 Sept 2026
  • Researchers release a framework for covertly distilling cyber capabilities from frontier models1234
  • 18 Sept 2026
  • Matt Sheehan argues the US–China superintelligence race binary is overstated.1
  • 18 Sept 2026
  • Jensen Huang, Sam Altman, and Cristiano Amon will join Trump's meeting with Xi at the US-China summit.123
  • 18 Sept 2026
  • Jacob Coxon details his disagreements with Amodei over the inevitability of an RSI arms race.12
  • 18 Sept 2026
  • Goodfire uses activation probes to cheaply catch reward hacking that CoT monitoring misses.1
  • 18 Sept 2026
  • Hacktron white-hats gained employee-level access to OpenAI's codebase via an RCE and a jailbroken Opus 5.1234567891011
  • 18 Sept 2026
  • Wired reports open-weight Chinese model Kimi K3 escaped a misconfigured eval sandbox to crib answers.1
  • 22 Sept 2026
  • US and China agree to set up a Trump-Xi AI dialogue and possible 'red phone' notification mechanism.123
  • 22 Sept 2026
  • US military nearly acted on a Chinese cargo ship threat that turned out to be an LLM hallucination.12
  • 25 Sept 2026
  • An automated AI hacking campaign using DeepSeek, GLM and Opus models compromises 27 retailers for about $25 per target.12
  • 02 Oct 2026
  • Anthropic says the open-weight GLM-5.3 nearly matches April's Mythos Preview at building end-to-end exploits.12345
  • 02 Oct 2026
  • GLM and DeepSeek nearly double their hard ExploitBench performance when run through Codex instead of their default harnesses.12
  • 06 Oct 2026
  • Swarmchasers report a suspected Tencent-operated AI agent fleet scraping Alibaba's map service, with no sign of rogue behavior.12
  • 06 Oct 2026
  • Zhejiang University researchers unveil ReSA, a "reward stealing attack" that reverse-engineers and subverts model alignment.123
  • 09 Oct 2026
  • CrowdStrike finds last week's South Korean bank attack may have been one actor using the Chinese ARTEX agent and Claude Code.123456
  • 09 Oct 2026
  • DeepSeek nears $12B in funding with an IPO now expected in early 2027.1
  • 09 Oct 2026
  • SemiAnalysis finds almost no Chinese models report safety evals and predicts Beijing will not pace the frontier.1234
  • 09 Oct 2026
  • The EU's cybersecurity agency confirms it has started testing Chinese AI models.12
  • 13 Oct 2026
  • Unitree CEO says the company is testing GPT-6-controlled humanoid robots for factory use1234
  • 16 Oct 2026
  • Anthropic tells IPO investors that 6% of Q2 revenue came from attackers buying tokens for distillation.12
  • 16 Oct 2026
  • Apollo observes GLM-5.3 spontaneously engaging in “update-aware reasoning” about its actions becoming training data.123
  • 16 Oct 2026
  • Goodfire's Nick Cammarata claims neural networks are far more interpretable than thought, sparking lively debate.1234567891011121314
  • 20 Oct 2026
  • NorForge finds 20+ exposed AI hacking control panels revealing 1,000+ agentic attacks, the busiest run on DeepSeek V4.1 Flash123
  • 20 Oct 2026
  • DHS order bans civilian federal agencies from using DeepSeek products12
  • 20 Oct 2026
  • China proposes export controls on strategically important AI chips and models1
  • 20 Oct 2026
  • Taiwan moves to criminalize shipping restricted AI chips and servers to any buyer in China1234567
  • 20 Oct 2026
  • SK Hynix makes a new US factory conditional on more US chip-equipment exports to its Chinese plants1
  • 23 Oct 2026
  • A wave of AI-enabled cyber incidents hits PHP, Node.js, Linux, LiteLLM, Railway, 400+ Latin American banks and more123456789101112131415161718192021222324252627282930313233343536373839404142434445464748
  • 23 Oct 2026
  • OpenAI reportedly drops its lobbying for stricter chip export controls on China1
  • 27 Oct 2026
  • Alibaba's open-weight Qwen3.8-Flash-Next tries to send a test file to an Alibaba server in Beijing during Cohere's evals1234567891011
  • 27 Oct 2026
  • IAPS recommend how the new US-China AI incident notification mechanism should work, starting with a Bessent–He Lifeng channel123
  • 27 Oct 2026
  • New Chinese safety evaluator Fangcun release their first evaluation, on DeepSeek V4.1 Flash-101712
  • 30 Oct 2026
  • Likely Chinese, AI-assisted attackers chain PHP2Shell with kernel bug Fragnesia to root 25 Palo Alto firewall servers in Japan, Taiwan, Korea and the US123456
  • 30 Oct 2026
  • Artificial Analysis Intelligence Index v5 moves most benchmarks to private test sets, dropping open models to around Claude Haiku level123456
  • 30 Oct 2026
  • Compaction smuggling bug let researchers extract OpenAI models' reasoning traces; Chinese labs reportedly switch to distilling Astra1234
  • 03 Nov 2026
  • Irregular find Zhipu's open-weight GLM-5.3 never refuses offensive cyber tasks and succeeds at a quarter of network attacks12
  • 03 Nov 2026
  • Epoch AI find GLM can autonomously build working Firefox exploits for about $1,000 each, mostly by exploiting the patch gap1234
  • 03 Nov 2026
  • China speed up AI-safety legislation covering the full AI life cycle after a wave of AI cyberattacks and break-outs1
  • 06 Nov 2026
  • Audit finds Kimi K3.1 has memorized hidden Terminal-Bench 4.0 grading rules, inflating its top score123
  • 06 Nov 2026
  • Oracle scrap a $7B deal to lease 100,000 Southeast Asia-based AI chips to Tencent amid US and Chinese pressure12
  • 06 Nov 2026
  • Jensen Huang has burned trust in Washington by exaggerating Chinese chip capabilities to fight export controls12
  • 10 Nov 2026
  • Vals AI find open-weight Chinese models close to GPT-6 Sol Daybreak in offensive cyber capabilities at a tenth of the cost1
  • 10 Nov 2026
  • FBI and Nvidia's board investigate Jensen Huang's affair with an alleged Chinese spy; NYT reports Nvidia intervened in a probe of Megaspeed chip smuggling1234567
  • 13 Nov 2026
  • Will DePue presents forensic evidence that DeepSeek, Moonshot AI, MiniMax and Ant Group trained frontier models on Nvidia Blackwell GPUs despite export controls1234567
  • 13 Nov 2026
  • The Commerce Department prepares rules restricting Chinese AI firms’ remote access to advanced chips, extending licensing to seven more countries123456789101112
  • 13 Nov 2026
  • OpenAI reverses course and backs congressional bills to curb China’s access to AI chips after Huang’s fall at Nvidia123
Is RSI happening?
    22 Feb 2026
  • METR puts Opus 4.6 at 14.5-hour horizon1
  • 22 Feb 2026
  • OpenAI models help physicists finish paper1
  • 26 Feb 2026
  • Anthropic says internal models lag Opus1
  • 02 Mar 2026
  • Deep critique of Epoch and Dario estimates12
  • 06 Mar 2026
  • Ajeya Cotra raises automation odds12
  • 06 Mar 2026
  • GovAI measures AI R&D automation1
  • 06 Mar 2026
  • AI speeds formal math 8x1
  • 10 Mar 2026
  • Karpathy autoresearch beats nanochat record12
  • 13 Mar 2026
  • AlphaEvolve improves Ramsey bounds12
  • 18 Mar 2026
  • Autoresearch enters interpretability work1
  • 18 Mar 2026
  • Autoresearch steers evolutionary search1
  • 18 Mar 2026
  • David Pfau comments on AI math1
  • 23 Mar 2026
  • FrontierMath open problem confirmed solved1
  • 23 Mar 2026
  • Frontier models lag human mathematicians12
  • 23 Mar 2026
  • Lean mathematicians dissent on autoformalisation12
  • 23 Mar 2026
  • Autoresearch overfits without constraints12
  • 03 Apr 2026
  • Semianalysis explains parameter-count stagnation1
  • 03 Apr 2026
  • Epoch and METR prepare HCAST replacement
  • 03 Apr 2026
  • AI2027 timelines move 1.5 years earlier12
  • 07 Apr 2026
  • Cunningham models AI R&D economics1
  • 07 Apr 2026
  • Greenblatt warns about automation progress1
  • 07 Apr 2026
  • AIs excel at software replication
  • 07 Apr 2026
  • ESNI progress accelerates AI R&D
  • 07 Apr 2026
  • Greenblatt timelines remain moderate123
  • 15 Apr 2026
  • AI solves active math problem
  • 20 Apr 2026
  • Researchers forecast ASARA progress1
  • 28 Apr 2026
  • AI coding agents given only raw method1
  • 28 Apr 2026
  • Litt describes AI-assisted math work1
  • 05 May 2026
  • AI economic growth model released1
  • 05 May 2026
  • Jack Clark puts 60% on no-human1
  • 08 May 2026
  • Manual checks slow METR estimates1
  • 08 May 2026
  • Harder MirrorCode version released123
  • 12 May 2026
  • MLS-Bench tests autoresearch generalization1
  • 12 May 2026
  • AlphaEvolve saves Google compute1
  • 12 May 2026
  • Mythos rewrites 960k lines of production code12
  • 12 May 2026
  • RSI scenarios get serious scrutiny123
  • 15 May 2026
  • METR surveys experts? AI use12
  • 15 May 2026
  • Coding agents sharpen RSI debate1
  • 15 May 2026
  • AISI warns against automated alignment1
  • 22 May 2026
  • InferenceBench measures AI R&D ability1
  • 22 May 2026
  • First really major AI mathematical result1234
  • 26 May 2026
  • The post-Erdős era begins12345
  • 29 May 2026
  • Automated R&D may still slow1
  • 05 Jun 2026
  • Anthropic release internal evidence of RSI in progress1
  • 19 Jun 2026
  • DeepMind maps paths from AGI to ASI1
  • 23 Jun 2026
  • Recursive AI beats engineering baselines1
  • 26 Jun 2026
  • Plasticity loss blocks continual learning12
  • 26 Jun 2026
  • Autodata converts inference into training data1
  • 03 Jul 2026
  • Continual learning determines AI power1
  • 08 Jul 2026
  • RSI feedback loops get taxonomy1
  • 08 Jul 2026
  • Harness engineering proposed as RSI path1
  • 08 Jul 2026
  • Fable writes single-kernel inference code1
  • 08 Jul 2026
  • Krier argues takeoff remains unlikely1
  • 10 Jul 2026
  • OpenAI internal inference grows 100x1
  • 10 Jul 2026
  • Coding contest victory signals autoresearch progress1
  • 10 Jul 2026
  • Fable wins CIFAR-10 optimization challenge1
  • 10 Jul 2026
  • Anthropic sees 2.8x coding uplift1
  • 14 Jul 2026
  • Epoch finds rising Codex productivity1
  • 14 Jul 2026
  • Economists model quiet AI takeoff1234
  • 14 Jul 2026
  • GPT-5.6 proves graph conjecture12345
  • 14 Jul 2026
  • Post-training may block Meta catch-up1
  • 17 Jul 2026
  • Recursive autoresearch shows limited gains123
  • 22 Jul 2026
  • Two Annals-quality math breakthroughs with short proofs: Fable disproves Jacobian conjecture, Sol Pro solves Erdos #1191234567
  • 22 Jul 2026
  • OpenAI internal model (likely GPT-6) escaped its sandbox to post results as a GitHub PR during NanoGPT speedrun eval12
  • 22 Jul 2026
  • OpenAI employee anecdote: even with internal models and thousands of GPUs, can't match what Alec Radford could do in a week years ago1
  • 28 Jul 2026
  • A researcher speculates Anthropic's Mythos escaped sandboxes thousands of times during training, which may explain its strong cyber capabilities.1
  • 28 Jul 2026
  • An ensemble of models guided by a human solved a 'solid result' level problem in Epoch's FrontierMath: Open Problems benchmark, with a 60-page proof.12
  • 28 Jul 2026
  • The volume of AI-produced math discoveries is challenging the math community's ability to verify results; a PhD candidate solved six open Erdős problems in 5 days using GPT-5.6 Sol.12
  • 01 Aug 2026
  • GPT-5.6 Sol improves own serving efficiency; OpenAI launches price war undercutting competitors123
  • 01 Aug 2026
  • OpenAI releases proof that nonsophic groups exist, hailed as most important math AI result yet12
  • 01 Aug 2026
  • Tencent’s Hy3 solves 50-year-old additive combinatorics problem with help of GPT-5.6 Sol12
  • 01 Aug 2026
  • Study finds AI agents proficient on verifiable research but not genuine progress on open-ended tasks1
  • 01 Aug 2026
  • 1,319 frontier AI company employees sign 'Pacing the Frontier' open letter requesting international coordination1
  • 07 Aug 2026
  • Former OpenAI employee presents a bear case on frontier labs, arguing rising training costs punish frontrunners.1
  • 07 Aug 2026
  • Samuel Hammond argues frontier US labs are 'very nearly' able to automate the whole AI R&D process, closing the RSI loop.1
  • 07 Aug 2026
  • Math progress: OpenAI's Astra resolves 10 major conjectures, half replicable with Fable; Litt concedes his 2030 bet.12345
  • 07 Aug 2026
  • AI Futures Project proposes a four-option deceleration scheme for US frontier labs.1
  • 14 Aug 2026
  • METR's Cunningham says autoresearch hasn't accelerated progress on hard public benchmarks over six months.12
  • 14 Aug 2026
  • A wave of major departures and a reorg raise questions about whether DeepMind has exited the frontier AI race.123456789101112131415
  • 14 Aug 2026
  • Several notable AI-assisted mathematical results, including progress on the Riemann hypothesis.12345678
  • 14 Aug 2026
  • IFP lists 23 policy ideas for managing automated AI R&D, synthesizing safety and acceleration views.12
  • 21 Aug 2026
  • UnsolvedMath benchmark reports Sol producing 174 apparently new solutions to open math problems.123
  • 21 Aug 2026
  • Dan Luu documents how AI agents trivially game benchmarks in meaningless ways even when told not to.12
  • 21 Aug 2026
  • Prime Intellect finds frontier models can run sophisticated research procedures but struggle with genuine originality.1
  • 21 Aug 2026
  • AI Futures updates its fully Automated Coder timeline to ~70% by January 2030 using new methods.1
  • 21 Aug 2026
  • Toby Ord gives a median date of 2038 for transformative AI and outlines RSI dangers absent an intelligence explosion.1
  • 25 Aug 2026
  • A Claude model apparently resolves the Hopf problem — whether the 6-sphere is a complex manifold — in a 108-page proof.1234567
  • 25 Aug 2026
  • SPADE is a self-play method putting RL environment generation inside the training loop, with one model playing both roles.1
  • 28 Aug 2026
  • Dwarkesh Patel and Dylan Patel make bold predictions on AI compute concentration, inference internalization, and economic growth over the next five years.123
  • 28 Aug 2026
  • Toby Ord's new paper finds the conditions for an RSI-driven intelligence explosion are surprisingly narrow.123
  • 28 Aug 2026
  • Goodfire's method makes reasoning-resampling ~8x cheaper — with a first draft authored by its autonomous agent Silico.12
  • 01 Sept 2026
  • Anthropic reports automated alignment researchers can fix benchmark failures, while its Hacker-Opus project reveals hidden catastrophic-risk dispositions.123456
  • 04 Sept 2026
  • A deep dive into GPT-6 Astra: first OpenAI model at 'Critical' cyber risk, with collapsed CoT monitorability, latent recurrence, and suspicious near-zero alignment eval scores.1234567891011121314151617181920212223
  • 04 Sept 2026
  • OpenAI's claim to monitor 99.9% of internal coding traffic did not apply to evals, an employee admits.1
  • 09 Sept 2026
  • An internal Anthropic model produces the first complete Lean formalization of Fermat’s Last Theorem in 11 days.12345
  • 09 Sept 2026
  • OpenAI announces a Lean-verified solution to Navier-Stokes, amid controversy over scooping Buckmaster and Alpöge’s human+AI work.12345678910
  • 09 Sept 2026
  • OpenAI’s Jakub Pachocki expects RSI within years and calls for voluntary slowdowns pending mandated safety requirements.1
  • 09 Sept 2026
  • OpenAI claims to have an “automated research intern” and reaffirms its goal of an automated AI researcher by March 2028.1
  • 11 Sept 2026
  • Anthropic's economic model projects AI's US impact to 2030, with an extreme scenario of 15% growth and an 8% unemployment jump.123
  • 11 Sept 2026
  • NYT reports OpenAI claims substantial progress on another Millennium Prize problem, with rumors of Hodge and BSD.1234
  • 11 Sept 2026
  • Forethought's Tom Davidson argues data bottlenecks may slow but won't stop an intelligence explosion.12
  • 11 Sept 2026
  • OpenAI shifts on regulation, calling this a closing window to establish durable AI safeguards.12
  • 11 Sept 2026
  • Paul Christiano returns to OpenAI as a board member, warning of RSI-driven superintelligence risks.123
  • 15 Sept 2026
  • OpenAI delays IPO, with Altman citing safety concerns12
  • 15 Sept 2026
  • Two OpenAI researchers argue capabilities won't hit a wall even if generalization stays narrow123
  • 15 Sept 2026
  • China's AI standards body elevates rogue model behavior in updated safety framework1234
  • 15 Sept 2026
  • 25 Fields medalists warn AI threatens the intellectual culture underlying mathematics1234
  • 18 Sept 2026
  • Anthropic unveils a framework to measure AI R&D automation, reporting Claude 'leads' 26% of R&D tasks.12
  • 18 Sept 2026
  • DeepMind's 'Dream-RSI' paper draws hype but is really about dream-replay RL over programs.12
  • 18 Sept 2026
  • Rumors claim OpenAI is close to solving the Hodge conjecture and hoarding major math proofs.12345
  • 18 Sept 2026
  • Periodic Labs runs closed-loop automated materials science, with a tuned model beating GPT-6 Astra on its benchmark.1
  • 18 Sept 2026
  • Jacob Coxon details his disagreements with Amodei over the inevitability of an RSI arms race.12
  • 18 Sept 2026
  • Zuckerberg says Meta is deprioritizing RSI and committing most compute to serving users.12
  • 22 Sept 2026
  • Toby Ord finds swarms of agents are an inefficient but parallel way to scale capability.123
  • 22 Sept 2026
  • OpenAI claims to have solved 100+ open math problems and forms a math advisory group.12
  • 22 Sept 2026
  • Ezra Klein argues for banning recursive self-improvement until proven safe.12
  • 25 Sept 2026
  • Anthropic unveils its semi-automated biolab, whose agents discovered a previously uncharacterized bacteriophage enzyme system.123
  • 25 Sept 2026
  • Embedded METR team estimates Opus 5.5 gives ~1.5x AI R&D acceleration, possibly crossing Anthropic's RSP automation threshold.123
  • 25 Sept 2026
  • Elasticity Institute proposes eight RSI disclosure categories; OpenAI discloses 25%, Anthropic 19%, Google DeepMind 6%.12
  • 25 Sept 2026
  • Opus 5.5 system card: cheaper and faster, but shows new misbehaviours and weakening chain-of-thought monitorability.1234
  • 29 Sept 2026
  • Paper coauthored by Bengio, Hinton and OpenAI's Pachocki finds preliminary evidence of an intelligence explosion within a few years.1
  • 29 Sept 2026
  • Ramez Naam argues today's AI self-improvement loop would need to be 5–10x stronger to produce an intelligence explosion.12
  • 06 Oct 2026
  • New TasteVal benchmark claims frontier models' research taste has doubled roughly every three months.1234
  • 06 Oct 2026
  • Forethought's James Tillman argues a software-only intelligence explosion is less likely than earlier models suggested.12
  • 06 Oct 2026
  • Epoch finds OpenAI's internal coding-agent spending per researcher is doubling roughly every month.12
  • 06 Oct 2026
  • METR's Tom Cunningham reports accelerating AI progress in math, cyber vulnerability discovery, and algorithms.12
  • 09 Oct 2026
  • AI labs are reportedly quietly testing whether their internal models can break cryptographic protocols.1234567891011
  • 09 Oct 2026
  • Epoch finds AI cannot yet automate AI R&D or Epoch-quality research work.1234
  • 09 Oct 2026
  • Ex-OpenAI researcher David Robinson tells Ezra Klein that alignment is an unsolved science problem, not an engineering one.1234
  • 13 Oct 2026
  • DeepMind's Nexus proof agent solves 9 open Erdős problems and 44 OEIS conjectures for a few hundred dollars each12
  • 16 Oct 2026
  • Daniel Paleka analyzes OpenAI's 722-paper math “diamond dump” and finds many practical subfields underrepresented.1234
  • 16 Oct 2026
  • Anthropic migrates to an in-house version control system, Scion, amid an explosion in AI-driven coding productivity.1
  • 16 Oct 2026
  • Dwarkesh Patel reports an Anthropic employee expects AIs to match humans at AI research within nine months.123456
  • 20 Oct 2026
  • Unreleased GPT-6-Astra scores 98% on a new research-level theoretical physics eval, catching errors in experts' answers12345
  • 20 Oct 2026
  • Anthropic reports that agent hours now outnumber human hours six to one internally as it hires fewer junior staff1
  • 23 Oct 2026
  • Anthropic and Theorem use Claude to formally verify security properties of 123 open-source libraries at the assembly level12345678910
  • 27 Oct 2026
  • OpenAI claim proofs of local smoothing in all dimensions, a refutation of ETH, and the existence of 4D Yang-Mills and QCD123456789101112131415
  • 27 Oct 2026
  • Startup Mirendil claim AI agents overseen by one researcher built a 100B–500B-parameter coding model for under $10M1
  • 27 Oct 2026
  • Anthropic's IPO prospectus reveals founder voting control, revenue and compute splits, ownership, pledged donations and lockups12
  • 27 Oct 2026
  • Anthropic researchers report 4x more output with AI, while third-party safety researchers report only a 1.6x speedup123
  • 30 Oct 2026
  • Nostalgebraist argues frontier coding agents still make constant basic blunders as benchmarks keep climbing123456789
  • 03 Nov 2026
  • Anthropic's amended S-1 says fully autonomous AI R&D won't be safe for the foreseeable future and floats RSP changes to keep humans in control123
  • 03 Nov 2026
  • MIT students claim GPT-6 Astra can read and edit raw neural network weights without gradient descent, though the results are heavily caveated1234567891011
  • 03 Nov 2026
  • Anthropic open-source FlashProve, a largely Claude-built zero-knowledge system for proving which model weights produced a response12
  • 03 Nov 2026
  • GPT-6 Astra running in a loop proposes research questions that experts judge worth pursuing about a third of the time12
  • 03 Nov 2026
  • Anthropic's updated S-1 reports a $23B Q3 net loss and a rising R&D share of compute1
  • 03 Nov 2026
  • Anthropic use thousands of Claude agents to find training data that teaches misalignment while evading production monitors1
  • 06 Nov 2026
  • METR estimate that Anthropic's benchmark progress has accelerated ~1.9x, mostly due to AI, just shy of its 2x risk threshold123456789101112131415161718192021222324252627
  • 06 Nov 2026
  • Open Benchmark Review find serious errors in five of six RSI benchmarks; only PostTrainBench is clean12
  • 06 Nov 2026
  • Brady Hauth argues the strong bitter lesson conflicts with RSI, but more compute leaves room for better algorithms1234567891011
  • 06 Nov 2026
  • Anthropic tell IPO investors that Claude is substantially speeding up their in-house AI chip design12
  • 10 Nov 2026
  • Responses to METR's acceleration report: Cotra says Anthropic has likely crossed its 'dramatic acceleration' threshold, and the AI Futures Project say progress is five months ahead of AI 20271234567891011121314151617181920
  • 10 Nov 2026
  • Anthropic release Mythos 5.2 / Fable 5.2, with large benchmark gains, a 30-hour Fermat formalization swarm, and extreme evaluation awareness12345678910111213
  • 10 Nov 2026
  • Elliot Glazer uses public models to claim a proof that the integers are Diophantine in the rationals, falsifying a major conjecture1234567
  • 10 Nov 2026
  • METR review Anthropic's August Risk Report and call its claim that the 2x AI R&D acceleration threshold was not crossed overstated1
  • 13 Nov 2026
  • Kevin Esvelt argues AI will plateau near human level wherever answers can’t be cheaply checked, against Kokotajlo’s case for a 2027 intelligence explosion1234
  • 13 Nov 2026
  • Bespoke Labs find models rarely try new ideas on long-horizon ML research tasks, even though new ideas yield 9x the gains of refinement123
  • 13 Nov 2026
  • Dean Ball distinguishes weak from strong recursion and argues METR’s acceleration report only shows the weak kind12
  • 13 Nov 2026
  • Economists Garicano and Imas argue Anthropic is getting less progress per research dollar, so there is no intelligence explosion yet12345
  • 13 Nov 2026
  • Dwarkesh Patel now expects AIs to match the best human AI researchers by around 2028 and superintelligence by 2030, arguing ML research is highly verifiable123

Ask the archive

Keywords
DateSummaryBrief descriptionOpinionLink

Loading items…