Old process: a weekly 1C export of about 2,500 technical items went to sales and engineers; past decisions were lost each week and nobody reached the end of the list.
4xmore slow-moving stock sold in the first month, as the speaker says it (the video title says 4.5x)
Industry reports say some embodied-AI companies have started cutting algorithm jobs; no companies are named.
20-30 to 50-60RoboDojo scores of teams' models after combining them with GPT-6 Astra-based approaches
Messengers are thin adapters into one core; the same person on Telegram and MAX is still two user IDs.
about 1,200people using the assistant across Telegram, MAX and VK
Five one-shot scenarios, three seeded runs each: diagnosis, fact extraction and two clarification tasks passed 3/3; S12 rollback planning failed 0/3.
12/15 vs 15/15one-shot passes across five scenarios (three runs each): model quality versus runtime control
Launched 21 August 2026; season 4 topped the Hongguo AI new-drama chart; season 10 heat reached 82.06 million; about 4 million people booked unaired episodes.
about RMB 1 millioncost figure circulating in media, which the team confirms covers compute only; it rejects the RMB 10 million return estimate
June 2025: 49% active use; a reported 18% productivity improvement had levelled off.
49%engineers actively using AI coding tools in June 2025, when a reported 18% productivity gain had levelled off
NVD: 74 of 74 within 2 hours, median 0.8 hours.
74CVEs with 2026 records added to CISA's KEV catalogue between 1 September and 7 November 2026
The appraiser started on Claude Pro for the Excel add-in in February; after a three-to-four-hour one-to-one session with a founder friend in April he moved to Claude Max at $100 a month.
$100 a monthClaude Max subscription the appraiser uses, enough for one or two websites and three or four apps
232 services have 312 connections: 214 API, 97 MCP, one A2A (Albato Copilot, partner access only).
232 of 424services with at least one documented API, MCP or A2A connection (312 connections in total)
41% are senior or above; 8.5% chose roles new to the survey (28 AI/LLM engineers, 13 MLOps engineers).
74.4% / 67.3%code writers whose coding time fell / whose time checking and fixing AI output rose
Rollout of real sends went 10%, 30%, 50%, 100%, with dry run as default and sends held when the model layer failed.
605 of 655specialists' closing messages that also invited the buyer to write again, which defeated keyword rules
Weekly merged PRs went from about 100 to more than 400; the median AI-generated PR reached 640 lines.
4xPRs merged per week with the same nine engineers (about 100 to more than 400)
Repositories: undici, fastify, cachetools, Python-Markdown, anyio, zod; 151 target items extracted in advance by regex and parsers.
288runs: 12 find-all questions x 2 models x 4 conditions x 3 repeats
Workflow: the agent first breaks a task into steps; the engineers prune the plan until it serves as the spec.
2 monthstime for two engineers to bring a business-process automation system to pilot stage
Transcriber reach: new people per month rose from 5 in April to almost 30 in June and July, then about 13 in August; close to 100 staff had taken at least one action by end of August.
168 of 374 (45%)AI reports on working meetings scored 3 or lower out of 5 by the LLM judge, 3 July to 22 September 2026
Scope: about 20 full-time staff, average age early thirties, no remote work.
~20%cut in scheduled weekly working hours from June 2026, at unchanged pay and daily hours
Why not built-in tools: on a few internal meetings Teams' Russian transcript captured 26-37 words per minute against 99-108 for the in-house pipeline; Copilot needs a licence per person; other platforms are not covered.
~2,100meetings processed by both flows in four months (to 31 October 2026)
10.0 vs 7.0: repeated operations 14-35% faster (for example LoggerMessage.Define 377 to 244 ns); first calls and object creation 39-64% slower; the first EF query in a process rose from 996 to 1,162 ms.
61 to 145 µsPOCO binding of 50 fields, 7.0 vs 11 RC; traced to one human-written line checking an attribute per property
Problem: in closed environments the vendor sees "the query is slow", a screenshot or a log fragment; the same questions repeat across clients and were analysed by hand.
27.5% to 64%Text-to-SQL accuracy on about 1,000 business questions, open-source SQL model versus the fine-tuned model with value hints and one rewrite
Alert sources: dark web breach cards, identity and access anomalies, infrastructure alerts, financial risk events.
15-30 minutesmanual handling time per case before automation
Structure: role (responsibility and limits), skill (method for a document type), tool (a script such as builddocx.py), template and sources; all working files are drafts until a specialist accepts them into a release folder.
14 in 8 working daysbusiness-requirement documents produced in the first deadline test, later 17, each read and accepted by the analyst
MKP has a built-in copy of the corporate wiki (read-only, refreshed daily) and any number of team Markdown bases; access through a web UI, REST API and an MCP server, all under the same per-person permissions.
The collaboration ran from 1 October to 1 November 2026, with limited pizzas themed on the show.
allTikTok ads stopped by the company after the incident, not only the affected videos
The robot is built for "map, zone, clean".
~500photos taken per robot run along the route, most of them unusable
Data: 757 Tax Answer articles (about 1.17 million characters); a 250-page invoice-system Q&A PDF (153 questions); a 94-page model work-rules PDF (69 articles).
30% to 13%answers containing errors on 67 tax questions, raw documents versus pre-processed chunks
Korean SW graduate employment rate: 68.1% (2022) to 57.3% (2024); all majors 66.3% to 62.8%.
68.1% to 57.3%employment rate of Korean four-year-university SW graduates, 2022 to 2024 (all majors: 66.3% to 62.8%)
The studio's project management system gained AI features: one click generates the PM's requirement list, splits tasks and assigns people; designers press a button to write technical documents; technical review and development have similar AI review steps.
60 + 20 vs 90the designer's analogy: AI adds about 20 points to a junior scoring 60 against a veteran's 90
Each task has its own plugin; there is no universal prompt.
Top scores (pass^3): Gemini 4 Carbon (high) 64%, GPT-6 Astra (max) 63%, Claude Opus 5.5 (max, with fallback) 61%, Claude Sonnet 5.5 60%, GPT-6.1 Sol (max) 59%.
64%pass^3 score of the leader, Gemini 4 Carbon (high): all three repeats correct
The project is the unit of governance: owner, budget, credentials and policy attach to it.
5% to 97%prompt-cache rate of one project after moving per-request content out of the system prompt
Input is one alert as an evidence pack E1...En; output is a JSON card with a four-class verdict, claims citing evidence numbers, MITRE ATT&CK techniques, gaps and next checks.
39 of 100 to 0 of 100needless escalations on held-out real lab alerts before and after adding 215 real benign cases to training
Per a third-party count, over 60% of humanoid robots shipped worldwide in H1 2026 went to labs, research, exhibitions and tourism; under 20% to industry.
over 60% vs under 20%share of global humanoid robots in H1 2026 going to labs, exhibitions and tourism versus industry (third-party count cited)
About 8 weeks (July to August), a tool syncing data with an external service; 63 PRs: 38 merged, 15 closed unmerged, 10 still awaiting review.
37pull requests created with Claude Code in the two weeks from 13 July, against 10 merged in the same period
Prerequisite: a malicious or hijacked browser extension with access to claude.ai, and the user signed in to both the web app and Claude Desktop.
18 to 71 percentshare of known Claude accounts that were personal accounts, across six Forever Security customer organisations
AI use in the last two weeks: Yes 24.2% (24.6%, 23.8%, 23.8% in the three prior waves); No 61.9% (65.8%); Do not know 13.9% (9.6%).
24.2%businesses using AI (now worded Super Intelligence) in the last two weeks, wave 202622, vs 24.6% prior wave
Before: seven Claude Team/Claude Code seats (about $175 a month), seven Codex subscriptions (about $140) and about $10 of credits, about $325 a month for over 25.93 billion recorded tokens.
$325monthly commercial AI spend (Claude and Codex seats plus credits) before local models
Rubric dimensions: Visual Storytelling, Feature Presentation, A+ Layout, Conversion Optimization; each scored 0-10 with a confidence and rationale.
5,478product detail pages from 10 categories labelled by one expert and four annotators
Set-up: 30 API operations with an append-only audit log, 19 scored tasks plus two unauthorised-refund traps, baseline on Claude Sonnet 5.5 over three runs; thresholds set at twice baseline variance.
AECI (Anthropic's version of the Epoch Capabilities Index) has grown 1.9x faster since a breakpoint estimated at 7 October 2025, roughly two years of historical progress in one year.
≈12xaverage self-reported staff uplift from the most capable internal model, October 2026 (loose evidence)
PR #21852, submitted by the account lobehub-coder on 29 October, was reviewed by a Codex bot; the two exchanged several rounds in 18 minutes at 4 a.m. on 31 October, and the coder declined one finding with reasons.
~12 stime from an untagged chat message to a new task appearing in the reporters' first test
Box C: Bank staff built an AI activity proxy from computer programming and consulting, information services, and publishing (which includes AI agents and application software).
around 0.5%productivity gain from AI over the past three years, as estimated by DMP firms
Team of 9 (3 backend, 2 frontend, 2 testers, an analyst, a product owner) on the bank's internal PaaS; Python/FastAPI, React, Postgres Pro, Kafka.
~60% / ~40% / ~75%chat automation with the vendor, after the move on keyword matching, and now with a CPU MiniLM encoder
95 requirements and criteria reconstructed: 21 functional, 12 non-functional, 23 acceptance, 22 evaluation and 17 success criteria.
40%share of the 95 reconstructed requirements and criteria that were ever formalised (38 of 95)
Of 89 respondents, 49 (55.1%) used GenAI in ERR work; top tasks were translation (44.9%), data analysis (42.9%) and social media content (40.8%).
89survey respondents from Sudan's Emergency Response Rooms, plus 15 semi-structured interviews
Design and project engineers: 3-4% organisational, with own technical decisions dominating 87% and 82% of descriptions.
20%of 58,582 vacancy descriptions where organisational work dominates (18% to 39% depending on the threshold)
Sample: 20 relays (10 below official prices, 10 at or above), found via LLM recommendations, community lists and web search; requests sent from the USA.
7/10 vs 2/10low-price vs same-or-higher-price relays flagged for answer-behaviour differences in round one (p = 0.07)
16 steps: 14 AI stages and two deterministic phases.
16steps in the pipeline: 14 AI stages and 2 deterministic phases
Background: left school at 14, built marketing short-video accounts to RMB 1 million a month in his early twenties, then lost it in stocks and property.
RMB 600 + 800compute cost plus regeneration ("card draws") for an average 50-minute drama, per the creator
Benchmark: 1,000 identities built from pseudopeople synthetic records with added noise, half single people and half over-merges.
99.0% at $3.86Claude Opus 5.5 accuracy and cost per 1,000 identities in grading identity-resolution results
Scout agents picked 5,034 audit targets (about 2.8 million lines, roughly 7% of the kernel) for $3,426; analyst agents produced 2,308 candidate findings in 12 hours for $316,549.
2,308 / 1,724candidate kernel findings after deduplication / vulnerabilities dynamically reproduced (~74.7%)
Four arms: control, tutor feedback, and tutor plus learner feedback under a social or a goal-oriented framing; 1,199 tutor assignments across 1,030 distinct tutors.
8%increase in tutors eliciting learner ideas (1.41 more times per hour over a control mean of 17.16)
Treated firms disclosed more than twice as many vulnerabilities as the comparison group and about 150% more high-impact ones, with no slowdown in fixing reported flaws or detectable drop in measured quality.
about 150%more high-impact vulnerabilities disclosed by firms with access
Coze Enterprise launched in H2 2025 and is being withdrawn after a little over a year; Yuanqi launched in May 2024.
38Chinese AI products shut down or merged, by Tech Xingqiu's incomplete count; 25 of them at big tech firms in H2 2026
The pipeline moved from a proprietary image API to FLUX.2 [klein], an open 4B flow-matching model, to have a tailorable model.
15.7% to 58.7%image generation pass rate on 1,000 icon concepts before and after verifier-rewarded fine-tuning, under an independent VLM judge
Per-PR review: when a PR is ready for review, an agent reads the diff and nearby code for quality, security and compliance; review criteria live in the repository as a Skill that developers can also run locally.
$0.01-0.1cost per AI security review of a pull request, using open-weight models on Fireworks AI
Corpus: about 300 standards, policies and procedures, 2,000-3,000 pages, about 800k tokens, already stored as Markdown in a GitHub repo.
~800k tokenssize of the roughly 300 internal security standards, about 2,000-3,000 pages
Current impact, August 2026: 81.8% no impact, 3.8% decrease, 0.6% increase, 13.0% unsure.
81.8%businesses reporting no AI impact on their job vacancies, August 2026
Timeline: two pilot agents in late 2025; a Q1 2026 pilot with 4 agents and 20 scenarios; 6 agents by mid-2026; live to 50 L2 call-centre and branch staff in Q3 2026 with 8 agents and more than 150 scenarios across six business divisions.
70 / 20 / 10 percentrequests resolved by keyword match with no model, by a lightweight model, and by the full orchestrator
Data flow: BigQuery billing export into Google Sheets via Connected Sheets; the skill is a SKILL.md with procedures and decision criteria plus Python scripts calling the Sheets API.
several days to 1-2 hoursmonthly budget-vs-actual work before vs after the skill
Five harnesses: a one-shot pipeline, a ReAct tool loop, a loop with policy reachable only by search, a loop with long benign customer memory, and a loop with a same-model reviewer.
11.5 percentage pointsspread between best and worst harness for one model (qwen3.8-flash-next)
12 hard tasks taken from open benchmarks, mostly hard ML/SWE, chosen as closest to the author's own work; about 7 model setups were run.
81%, $1.20score and cost per task for Opus working alone
SMFL holds about 9 trillion yen of operating assets.
6 people x ~6 months to 1 person x 1 weekassetforce modification estimated under the old method versus done with enterprise vibe coding
Scope: income tax, corporate tax, local taxes, consumption tax and items such as depreciation; only pure calculation functions whose inputs are explicit arguments.
76,165distinct inputs on which production Ruby and its Lean copy were compared for the income-tax rate table, with 0 mismatches
Claude Fable 5 generated the code and GPT-5.6 Sol independently cross-checked parts.
3 days, 44 commitstime for Claude Fable 5 (with GPT-5.6 Sol cross-checking) to produce a working RAG search rewrite
Nine features in seven projects, from a Fortran solver and Python programs to books, a research dossier and this paper; five of the nine delivered non-software outputs.
288 of 581planned tasks completed block by block across nine features in seven projects
The agent ranked spans by frequency and duration and returned file, line, estimated weekly volume and an example trace per candidate; the engineer steered with short prompts and chose what to fix.
~143Mextra SELECT queries a week removed by reusing an already loaded project in ingestion hot paths
Output: eight to 20 PRs a month before; 200+ a month at Tano.
11xmerged PRs per month at Tano (AI-native startup) vs PostHog for the same engineer
Stock Litmus, which works only with pods, can simulate about 22% of past incidents; with Avito's own database and network scenarios, 41% (400 of 977).
977incident post-mortems analysed by Llama 3 70B to generate chaos-test hypotheses
Gemini was deployed in early 2025 and Claude Code in early 2026, with strict budget limits after engineers spent $1,000 to $10,000 a month "with not so much ROI".
roughly 90 percentcut in Revolut's AI coding inference costs in under a year after making self-hosted models the default
Generated share: 0% of texts at the end of 2023; 30% of SMS and pushes on GigaChat and DeepSeek by the end of 2024; 70% of short formats generated by GigaChat alone by the end of 2025.
about 1 hourbrief to CRM upload for content that passes the AI checks, against a 5-6 working-day cycle at the end of 2023
Two campaigns over about six months: Python packaging (about 1,900 candidate repositories, 294 PRs) and C++ build system to CMake (about 6,700 candidates, 2,039 PRs).
67.7% and 57.7%Python and C++ PRs with no developer decision at all by the end of observation
97 skills in 11 categories; 52 concern statistical validity, root-cause diagnosis, assignment integrity, data readiness or incident response.
97skills in Meta's ABTest Agent, across 11 functional categories
Nine classes (parties, document details, contracts, object of dispute, performance, payments, breaches, damages, correspondence), each extracted by a separate call returning JSON.
0.70 vs 0.71macro-F1 of the distilled Qwen2.5-1.5B vs Ministral 3 14B on a 145-document golden set
Postings from 12 categories, sampled by a fixed rule from 18.8M postings requesting $69.1B of work; the agent (based on claude-opus-5) works in one response without tools.
1,494real freelance job postings in PAYWORTHY, sampled from 18.8M posted 2020-2026
The claims screen is used for tens of thousands of lookups a day; each document has 40+ fields built from 17+ tables.
about 42,000, 0 mismatchesnew changes compared field by field over 7 consecutive days between old and new index
Trigger: a tool her unit built for another department came back for "after-sales service" when its data fields changed.
20AI innovation leaders spread across the commission's bureaus, with tool subscriptions and API keys paid centrally
The paper's prompts are on GitHub and are identical apart from the weights, so anyone can re-run them.
3 of 4games Qwen 3.8 27B identified from raw policy weights on a laptop (failed on Freeway)
A citation problem went from about 70% to about 99% correct with one model release.
70% to 99%correct citations on one problem after a single model release, with no change on Legora's side
Across 15 models, the strongest discover over 80% of withheld requirements; open-weight models 40-50%.
over 80%withheld requirements the strongest models discover (open-weight models 40-50%)
Deployment: seven regions (Saudi Arabia, Brazil, Hong Kong and four Gulf states), several languages; tasks include missing items, delays, refunds and handoff to staff.
6.42% to 1.97%offline inspection Bad Case rate of the in-house 31B model over five iterations; incumbent GPT-5.4 at 4.60%
Pricing works in "person-days".
99.9%share of domestic game-studio outsourcing orders that include AI in the workflow, per one 2D artist's projects
Digital twins and RCT estimates correlate at 0.93 across 37 trials.
37randomised trials of workforce training programs used to estimate the bias of the observational method
Before: one buyer ran purchasing in Excel; when that person was on holiday, purchasing stopped.
105 to 89 dayscapital payback (stock value divided by daily profit) before vs after the agents, over an 8-week window
Case 1, "division by zero" when printing a UPD for quantities below 0.006: the agent found the same query line 1C patched in a 12,000-line module, added a guard on the rounded divisor, and noticed 1C's own bug description was wrong (only quantities below 0.005 fail).
155-427 RUB ($1.93-5.30)DeepSeek API cost per bug fix across the three bug cases
The Xuzhou hospital said the doctor was checking a new drug, found it unsuitable and issued no new prescription that day: "AI is only an aid for getting information; prescribing rights stay with the doctor."
647medical workers in the DXY (Dingxiangyuan) survey cited
Mixture-of-experts: 2.8T total and 104B activated parameters, 93 layers, 896 experts with 16 selected per token, 1M-token context, 401M-parameter vision encoder; text and image input.
2.8T / 104Btotal and activated parameters of the mixture-of-experts model
Requirements: all text stays on YADRO servers, the LLM and its parameters can be changed, and the tool has a web or client-server interface.
about 0.40 RUB vs about 2.5 RUBcost per word on the in-house cluster in the worst case (one-word segment) vs current Smartcat rates
Task 1, reading values from charts and tables (8 questions): Luna 8, Flash-Lite 6, Sonnet 8, Astra 8.
8/8 at about 1/20 the costGPT-6 Luna on chart reading, matching Claude Sonnet 5.5 ($0.0026 vs $0.053 for 8 questions)
Before LLMs, writing involved 4.85 contributors on average (77.8% of team members), execution 4.37, ideation 4.07, data and coding 2.54.
2.6 percentage pointsdrop in team participation in ideation by 2025 relative to 2022
Routes tens of thousands of training job attempts a day across hundreds of model types and more than 20 hardware types.
77%prediction accuracy over 30 nightly retrains
Experiment 1: four PPO networks of identical shape (17,253 weights) trained on grid Breakout, Pong, Pac-Man and Freeway.
7.9% to 48.1%Qwen3.5-0.8B on MATH-500 after GPT-6 Astra edited its weights, with 50 forward passes and no gradients
69 patients (71 exams) processed in five daily batches, run twice, at T and T+7.
98.90% vs 96.49%accuracy of the RPA script vs the browser agent across two rounds
2025 boom: lead actors earned RMB 50,000 a day; a 7-8 day shoot rose from RMB 200,000 to 1.2-1.5 million.
about 90%share of staff one Zhengzhou studio moved to AI production in April 2026; monthly output then rose to 5,000-8,000 minutes
Premium ranges on FLORES-200 (encyclopedic text): Swedish +32% to +68%, Norwegian +35% to +66%, Danish +38% to +73%, Finnish +56% to +100%.
+32% to +68%extra tokens billed for Swedish vs English across frontier models (FLORES-200)
Incidents spanned North America, South America and Asia Pacific and several cloud and model providers; the number of incidents is not given.
over $100,000fraudulent charges in a single day in individual incidents
Verdicts at 30 September 2026: 22 claims expired (44%), 17 weakened (34%), 8 persist (16%), 3 strengthened (6%).
44%share of the 50 tested claims expired in current models (34% weakened, 16% persist, 6% strengthened)
First session drew 15-17 of 130-150 staff.
34 participants, 20 builds, 29 solutionsoutput of three monthly "free Thursdays" for senior staff
Servers: in a 1,000-repository sample of the official MCP registry, an estimated 90% of destructive tools have no server-side confirmation (70% no confirmation or annotation, 20% annotation only).
90%estimated share of destructive MCP tools with no server-side confirmation (95% CI 82-95)
No free-form message failed with the window open (0 of 35,500); failure jumped 93.0 points at 24 hours, while templates showed no jump (1.0 point).
93.0 pointsjump in failure rate of free-form messages at exactly 24 hours since the mother's last message (95% CI 87.0 to 99.1)
Ten agents, from GLM-5.3-Flash to Claude Opus 5.5, complete 95% of requests on average one at a time but 51% when chained; only 39% of chains finish without a single missed request or broken earlier test.
95% vs 51%average completion of requests taken one at a time vs requests chained together
In 6,904 human-labelled queries, 53.2% of abandonments were successful and 6.5% of clicks were unsuccessful.
53.2%abandoned queries that human judges labelled successful
926 controlled runs, 899 analysable: small agent teams built programs under varied room rules, some with a hostile member.
42% to 13%total coordination failure before and after revising which messages start work (21 of 50 vs 4 of 32 runs)
About 100 models served, hundreds of trillions of tokens a day, across heterogeneous hardware and service-level targets.
about 100models supported by the shared production engine
Firm models rank a stayer above a leaver at least 85% of the time at their own firm (random-forest AUCs 0.855, 0.891, 0.914), but are essentially useless at other firms.
0.855 to 0.914AUC of each firm's own retention model at its own firm
Over 142 compiler bugs resolved and 103 more under review; most are silent numerical miscompilations, because the agent swarms were steered away from loud crashes.
over 142XLA:TPU compiler bugs whose resolution the framework drove
Gemini (January 2025) brought 10% better lead time and 20% better coding time.
56 B200GPUs serving GLM-5.3 to about 2,000 engineers
Bill: Claude Max 5x US$100; API US$0; hardware, Slack, Google Workspace and Grafana add nothing extra.
US$100monthly cost of the Claude Max 5x plan that runs all 17 bots; API spend US$0
Six configurations within 1.5 points near 74% on DeepSWE fan out across 23.0 points on MaxxBench.
20 of 178max-effort DeepSeek V4.1 Flash attempts that called outside models with the harness API key