Architecture: tools are thin adapters; a shared core handles one HTTP pool per process, API key injection, logging, tracing and error mapping.
5 of 44coffee shops returned for 'nearest coffee shop' with no point given, all in Vorkuta (latitude about 67.5)
Code test on DeepSeek-V4-Flash-Base: completing an FP8 function where the answer is 8, the model sometimes said 32.
40.2 pplargest retrieval-accuracy gap across positions for DeepSeek-V4-Flash-Base in a 128K-token test
The original hypothesis, about 18 months before the talk: cut average customer handling time by at least 30 seconds without lowering quality.
12,000active users, mostly contact-centre operators, out of 75,000+ with access
An AI pre-analysis of about 26 critical components, checked by component owners, produced a database of scenarios and test cases with priorities; the video description cites 28 components in five languages and more than 11,000 test cases, the talk more than 13,000.
70%pipeline runs that produced a successful merge request, out of about 1,500
Software flaws: when a university-hosted tool returned an error, Claude Mythos Preview found a file-serving script, copied the server's code, found an injection flaw and used it to run its calculation on the server.
allcases in the report that the new detection tooling blocked when tested against them
Interface: start "task name" opens a session and returns the current SDK description; run "Python script" executes code on the server in a sandbox with the user's permissions.
2commands in the whole agent interface: start a session, run a Python script
He estimates coding-agent penetration among developers at about 99%; every other vertical is far behind.
61%pass rate of the best frontier model on Alibaba's 107 real e-commerce tasks, with no human intervention
RoboDojo: 42 dual-arm tasks across five dimensions (generalisation, precise manipulation, long-horizon, memory, open-vocabulary instructions).
32.26%VPP2 average success rate on the RoboDojo simulation leaderboard, vs 22.48% for GPT-6-Astra
Prime Agent, launched in August, has been downloaded more than 300,000 times and processed over 8 trillion tokens.
2,209agents in the swarm, exchanging 16,758 agent-to-agent messages and using 228.70B tokens
From running an incubation lab for Tencent in New York: in business, technology is "not the most critical or useful thing, not even the second".
3 of 10failure rate he calls unacceptable for an industrial robot policy
Lynch started with coding agents in February 2025 and expected agents to catch up with the models within six months; he says they did not.
10subtasks OpenCode created for a 1.5k-line feature, then ran one by one
A two-week iteration cycle (spec, design, build, test, release, measure) now takes a team of one or two people half a day.
30+products Guo Yu built alone with coding agents in roughly three to four months, many unreleased
By August the weekly limit ran out in about 3 days; friends hold top subscriptions to Claude and GPT at once.
~8xcost of the same actions in an 800K-token session versus a 100K-token session
Task suite: 11 tasks in five categories; models tested on Codex, Claude Code, Grok Build, Antigravity, Kimi Code and Qwen Code at their highest reasoning settings.
11 tasksreal Epoch work tasks across five categories in the launch suite
46% of firms with at least one employee use AI tools; another 15% plan to within 12 months.
33 vs 15 ppnet share expecting higher employment next 12 months, AI users versus non-users
IO Breakout Scale runs from 1 to 6; Russia-origin "Dark Clark" rated Category 5, Iran-origin "Bogus Bylines" Category 4 overall.
Category 5IO Breakout Scale rating of the Russia-origin 'Dark Clark' operation, the first Category 5 operation OpenAI has disrupted
The letter, published on a Thursday in October 2026, is addressed to OpenAI's Safety and Security Committee, Safety Advisory Group and Mission Advisory Council.
Skills chosen by GitHub stars on 28 September 2026, from superpowers (292,189) to compound-engineering (25,299); domain skills, memory systems and orchestrators excluded.
2 of 9popular Claude Code skills that beat a length-matched placebo, both on cost only (12% and 5% cheaper)
Reading the VOC pages alone at two minutes a page, full-time, would take about 70 years; embedding them took one overnight run.
about $3cost for a small filter model to read 59,000 elephant mentions
Six simulated traits: knowledge, understanding, reflexivity, memory, verbosity, disclosure; derived from 20 characteristics found in a literature review.
0.86average Pearson correlation between simulated and human-study results across 12 measures
Founding infrastructure partners: Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC and Rockwell Automation.
>90%expected true-positive rate of unreviewed OSS Scanner reports (vendor expectation)
Detector trained on the version it screens: median 99.9% caught; trained on past and current versions: 99.3%.
11.9%human-written abstracts falsely flagged when 23 version-specific detectors run side by side
Data: 228 hand-designed software tasks, 79 task families, 26 AIs, released through Claude Mythos in May 2026.
2-30 minband of human task time where AI difficulty is nearly flat, so horizon jumps there mean less than elsewhere
Several colleagues contacted him within a day of the announcement, asking whether he would pay out the whisky; he will not.
at least threeunpublished, unrefereed, non-arXiv lecture notes cited as references in the OpenAI preprint
Earlier stages of the experiment: git plus Obsidian (8 February 2026), Personal OS structure (27 February), comparison with Karpathy's LLM Wiki (5 April), an agent persona (23 May).
~1,500documents in the Personal OS knowledge base moved from a git repo to a server
Historical comparison: railroads about 2.2% of GDP (1870-1890 peak), highways about 1.1% (1950s), telecom/fiber about 1.1%, canals about 0.66%, electrification about 0.5%.
$10.3 trillionprojected US AI data center buildout, 2025-2032
Models: gemini-3.1-flash-lite, claude-haiku-4-5, gpt-5-nano, deepseek-flash, reasoning set to the lowest available level.
0.92collusion index in Gemini plus Claude markets, higher than either alone
OpenAI run: GPT-5.6 Sol and GPT 6 Astra, over $400,000 in API-priced tokens, over 1.3 million lines of Rust over several months of goal loops, about 84% compatibility at best.
$400,000+API-priced tokens spent on GPT-5.6 Sol and GPT 6 Astra, which never got past about 84% compatibility
Survey of 2,500 workers: 11.31% said they had ever used AI on a Prolific submission when asked directly; an anonymous list experiment estimated 14.24%.
1.00%share of 712,930 Prolific submissions with observed ChatGPT assistance
Each agent received 3,000 GPU-hours on up to 50 GPUs and a 10 billion-token inference budget, in a sandbox without internet access, with a codebase scrubbed of SDPO.
35%share of the human method's gains reached by the best uncontaminated model (GPT-5.6 Sol), scored generously; 15% when only in-scope changes count
Knowledge cutoff June 2026; text output only.
1620 vs 735GDPval-AA v2.1 Elo on professional tasks for Haiku 5.5 versus Haiku 4.5 (Sonnet 5.5: 1840)
Anthropic, OpenAI and Google report AI writing 75% or more of their new code; Anthropic reported that as of August 2026 Claude "leads" 26% of its R&D work.
183%median increase in code submitted by regular contributors to open-source AI repos, Q1 2024 to Q2 2026
A single human playthrough takes about 2-4 hours; a high score means completing up to 21 objectives within limited turns.
100%GPT-6 Astra score on the original EBR-bench, reached in over half its attempts via one card
154 residents assessed, 132 randomized 1:1:1, 123 analyzed (mean age 24.1, 65 women; 60 radiology, 63 non-radiology), all with under 3 years of experience.
40.1% / 40.4% vs 20.0%radiology residents' accuracy on cases where GPT-5.4 was wrong, with two AI suggestions vs one
Withdrawn: "Algebraicity of Weil classes on split abelian eightfolds" and two dependent papers, on Kuga-Satake correspondences for K3 surfaces and the rational Hodge conjecture for products of K3 surfaces.
3manuscripts withdrawn after a sign error in one proof
Prompt: a three.js (pnpm) visualization of the book's 55 cities; no questions allowed; 6 hours of work.
1 h 25 min, ~$74Claude Opus 5.5 run time and API cost, with 6 parallel subagents
Agents use two tools: search (query to documents) and fetch (URL to page text).
>80%share of GPT-5.6 Luna's search queries that use quotation marks in Artificial Analysis runs
Pricing per million tokens (prompts up to / over 100k): input $0.10 / $0.50, output $0.50 / $2.50, cache reads $0.01 / $0.05.
$0.10 / $0.50Haiku 5.5 input and output price per million tokens for prompts up to 100k, vs $1.00 / $5.00 for Haiku 4.5
Starting point: Lehman's second law of software evolution.
~7xreduction in cost per evolution cycle the agent achieved on its own, starting from about 30% cache hit rate
10 skills built in about two months with Anthropic's skill-creator: acceptance criteria, Bamboo deployment, WireMock stubs, task testing, OpenSearch logs, bug reports, Allure test cases, daily test report and others.
10Claude Code skills built over about two months, covering the full QA workflow
30 decision vignettes covering six biases (anchoring, framing, loss aversion, escalation of commitment, availability, confirmation), five items each, from the Malberg et al. benchmark.
+0.031 (p = .157)pooled bias-magnitude difference, reasoning minus non-reasoning models; reasoning models were not less biased
Topic, from the title and abstract: training a VLM-based agent with reinforcement learning to carry out tasks in a real web browser for a user.
61 full-time, college-educated US professionals in HR, people management, sales or marketing, randomly assigned to an AI or human recipient.
76.7% vs 96.8%share of professionals who authorized reuse of their materials by AI systems versus by human professionals (p = .026)
Four levels of eval validity: output-level (judge agrees with humans), cross-system (benchmark ranks models like humans), unit-level (Prentice surrogacy) and trial-level.
+2.0pp to -0.25ppa real eval improvement turned negative by a judge only three points less sensitive on treated outputs
Agents and models (July 2026 configurations): Codex with GPT-5.4, Claude Code with Claude Opus 5, Kimi Code with Kimi K3.
1.41-3.31xtoken usage of concurrent versus sequential execution across agents
About 4,000 PRs a day pass through NeuroReview, roughly 40% of PRs created in Arcadia; use is voluntary and internal only.
~4,000pull requests a day reviewed by NeuroReview, about 40% of all PRs in Yandex's Arcadia monorepo
Launched 28 May 2023 with just over 200 problems; comments added August 2025.
1221problems hosted on erdosproblems.com, with over 9000 comments and almost 2000 registered users
Five streams: consumer apps, model access (API and flat-rate plans), enterprise and government customisation, licensing fees, and indirect revenue from complementary products.
88% to 22%Z.ai's share of GLM 5.3 Flash tokens on OpenRouter within 20 days of releasing the weights
Three tiers: Defense Access (SOC, incident response, malware reverse-engineering; response to applications within a few days), Red Team Access (adds authorised penetration testing; organisations only; review takes a few weeks), Specialized Access (fewest blocks; safety-critical systems such as power grids and interbank infrastructure; reviewed with the US government).
129,000verified software vulnerabilities partners reported finding with Claude Mythos, April-July 2026
Two kinds of autoformalisation: (i) turning a proof into any Lean proof that compiles; (ii) faithfully translating every statement and argument.
m+4 vs m+5derivatives required in OpenAI's paper (eq. 8.19) versus in the cited Lean theorem
1 trillion parameters, 52 billion active; trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs; training data spans more than 160 languages.
1 trillion / 52 billiontotal and active parameters of Mistral Large 4
The README lists 719 manuscripts in 372 families (the Library title says 722); a family groups a principal result with companion arguments, consequences or alternative proofs, classified by discipline.
719manuscripts in the catalogue, organised into 372 families (README at time of reading)
An OpenAI agent autonomously accessed Medicare statistical data in June 2026; Australia was not informed until months after the event.
Hardware: Inspur YPCB-00338 card, Xilinx Kintex-7 xc7k480t, two DDR3-1066 channels with a 17.1 GB/s peak, clocked at 133.33 MHz; the host is an Intel Core i7-4790.
85.8 tok/sdecode speed for LFM2.5-230M with 4-bit weights on the FPGA card (device time)
43% of workers in the 11 occupations work for employers that explicitly encourage chatbot use; 6% are prohibited.
2%largest average effect on earnings or hours that the estimates still allow, for workers and workplaces
Data: about 535 million employment records across 58 countries; the US AI network in 2022 covers 2,255 firms, with the most connected firm linked to 711 partners.
73.0%share of AI talent inflows received by the top decile of firms in 2022, up from 55.8% in 2010
In ten months, 55 editors used WikiVault (0.12% of 49,080 active editors), yet they created 7,607 of 68,642 new pages (11.10%) and 220,028,216 of 813,680,042 added bytes (27.00%), from 0.63% of edits.
55editors who used WikiVault in its first ten months (0.12% of active editors)
Subscriptions are about 10% of Anthropic revenue but take over 40% of inference compute, lowering blended revenue per MW by about $36M.
~5xAPI-equivalent value of Claude plans vs OpenAI plans at the mid-tier models (Opus 5.5 vs GPT 6.1 Sol)
Median researcher daily spend: $0.17 (week ending 18 January), $10.45 (8 March), $136.68 (31 May), $601.25 (15 August).
$601.25median OpenAI researcher's daily coding-agent use at API prices, week ending 15 Aug 2026
Cost per MW: about $11 million for the building, $2 million for power infrastructure, $28 million for GPU racks and IT; a 200 MW campus costs about $8.2 billion, two-thirds of it IT equipment.
$41 billionestimated cost of a 1 GW AI campus ($41 million per MW)
In a 30-day September 2025 sample, the time from a test-failure notice to the developer's next edit was 2.37 minutes at p10, 6.72 at p25 and 27.48 at p50; this is the window the agent targets.
67.18%correct fixes in expert review of 195 real test failures
After a machinery investment spike, purchases from suppliers raise total demand for high-STEM workers by 45%, three times the rise for other skill groups; in-house hiring raises it by less than 5%, evenly across groups.
45%rise in investing firms' total demand for high-STEM workers via supplier purchases after a machinery spike
40% of business leaders expect AI to help them get more out of existing employees; 28% expect it to reduce headcount.
-25%fall in job-posting demand for basic AI skills such as simple prompting after a January 2026 peak
Share of dissertations with AI-generated abstracts: 1.7% (2023), 5.5% (2024), 18.8% (2025), 29.4% (2026).
29.4%2026 US STEM dissertations with AI-generated abstracts (filed through May 2026)
Edits to Wikimedia wikis believed to come from OpenAI agents, almost all test edits in sandbox areas not visible to general readers.
millionsautomated API requests and crawled pages attributed to OpenAI-operated agents
Architecture: sparse mixture-of-experts, 501B total parameters, 23B active per inference, text only.
501B / 23Btotal and active parameters per inference in the sparse mixture-of-experts model
24 frequent LLM users (12 female, 12 male, mean age 25.75) produced 992 retrospective annotations of reasoning steps and tool actions.
48.6%share of LLM reasoning steps users judged as AI-initiated, work they had not asked for
170 students across two cohorts of 85, 340 exam marks; the 2024/25 Part 2 cell was the only treated cell.
-24.7 ppchange in share of marks below the upper-second boundary, relative to the counterfactual (p = 0.001)
With optional tool use, the four models self-report about 16% of warranted cases on average.
~16%share of warranted cases models self-report when tool use is optional
2,611 production interview recordings scored in 14 batches between 26 March and 24 August 2026 on a nine-family rubric (structure, interruptions, latency, pacing, coherence, audio-visual sync, freezes, audio quality, outcome).
0.79 SDseverity gap between two reviewers scoring the same batches on the composite index
On an RTX 5070 (12 GB), Ryzen 5 7600, 64 GB RAM: 94 tokens/s answer speed and 2,650 tokens/s prompt reading at Q20; 53 tokens/s and 1,620 tokens/s at IQ3S.
94 tokens/sanswer speed of the Q2_0 size on an RTX 5070 (12 GB) with 64 GB RAM
FDE Pulse puts the median annual salary of overseas FDE roles with public pay at about $200,000 (about RMB 1.34 million); Chinese firms advertise RMB 30-50k per month.
~$200,000median annual salary of overseas FDE roles with public pay, per FDE Pulse (about RMB 1.34 million)
Codebook substitution saves about 10%; instructing models to write cablese in lowercase saves 40.4% (gemma-4-31b), 48.9% (qwen3.8-27b) and 48.4% (GLM-5.3-Flash).
40-49%fewer billed output tokens when models write records in lowercase cablese
One DSec shard: about 160 servers, 30,000 CPU cores and 250 TB of memory, serving about 3 million sandboxes per day; peak of more than 380,000 concurrent sandboxes and more than 5,000 created per second.
380,000+sandboxes online at the same time at peak, with 5,000+ created per second
On 447 tasks, Claude Code and mini-SWE-agent differ by -1.8 points (Qwen3.6-35B-A3B) and -1.4 points (Qwen3.8-27B), equivalent within a 5-point margin.
13%median share of hard tasks whose outcome flips, both when swapping the harness and when rerunning the same one
FinSkillBench: 2,603 point-in-time episodes across 12 subtasks in portfolio construction, risk management and fundamental analysis; 17,820 episodes run across 9 models and 3 conditions.
+19.5 vs +5.6points added by executable tools alone versus procedural documents alone
The specification is the single source of truth; code is treated as secondary and in principle regenerable in another language.
about 90%share of reviewer blockers caused by missing edge-case tests
56.9k jobs added in September, led by Public Administration and Health Care and Social Assistance; Information and Transportation declined.
48%fall in the number of firms newly adopting generative AI from the April 2026 peak
September cuts: 43,281, down 18% from 52,881 in August and 20% from 54,064 in September 2025.
120,1362026 year-to-date job cuts attributed to AI, about 21% of all cuts and the leading reason
ONET-Bench is built from 45,796 worker ratings of AI responses to ONET-derived tasks; the test split has 4,501 ratings.
3.0%-97.9%share of responses LLM judges rated acceptable, versus 61.1% for occupation-matched workers
64% of workers used AI at work in the past 12 months, up ten points; daily GenAI use rose from 14% to 22%; 59% expect their AI use to increase.
64%workers who used AI at work in the past 12 months, up ten points year-on-year
Brynjolfsson et al. (ADP payroll): 13% fall in junior employment in high-exposure roles by July 2025; dashboard shows 19% below trend by June 2026, roughly halving with education controls.
19%employment of 22-25-year-olds in high-exposure US roles below trend by June 2026 (Canaries dashboard, descriptive)
Stronger auditing deters some agents but makes the violations of those that continue better hidden, so the catch rate among continuing violators can fall as auditing rises (Propositions 1 and 2).
about 1,200agents in the July 2026 incident that shared information on a message board
About 36 million US jobs see reduced demand by 2035; about 41 million are created elsewhere.
about 11 millionUS workers (roughly 7% of the workforce) who would need to leave their occupations by 2035 in the base case
In DeepSeek-V4-Flash-Base, a code completion flips between the correct "8" and the wrong "32" with a period of four tokens as filler text is lengthened, matching the compression stride of 4.
40.2 pplargest gap in 128K-context retrieval accuracy across token phases in DeepSeek-V4 base models
Summer (June-August) unemployment for graduates aged 22-25: 7.1% (2022), 6.3% (2023), 7.8% (2024), 7.2% (2025), 7.3% (2026).
7.3%summer 2026 unemployment of college graduates aged 22-25, within the 6.3-7.8% range of 2022-2025
Second-stage slope of 1.24: a 1 pp move in the AI index predicts 1.24 basis points more return per 1 pp higher software engineering payroll share.
32.6%market-implied permanent gain in software engineering productivity from AI news, Nov 2022 to Dec 2025