Why a web PaaS fails for LLM inference (GPU sharing, batching gives 5-10x utilisation difference); decision to build a separate inference platform on an open-source core.
Nishimura, Jolles, Lordan, Canavan; follow-up survey n=859 to the LSE/Protiviti 3,000-person study: 93% of trained vs 57% of untrained use AI.
Barclays: RAG Colleague Knowledge Assistant live since 2025, 16,000+ colleagues, 1M+ searches. Global Markets email triage about 120,000 emails/day; most engineers on Claude by 2027.
1,200 CEOs, 21 countries: 50% name AI top productivity contributor; 23% say gains are absorbed before the bottom line; 16% have real-time visibility of AI returns; productivity judged by capacity creation, not cost cut; 72% expect skills shortages to bind before capital.
Same Russian text: Opus 5 tokenizer counts 3.9x the tokens of GPT-5.5. Russian vs English overhead 18-23% for GPT/Gemini/Grok/GLM, 38% DeepSeek, 52% Qwen, 86% Kimi. Price per token misleads for non-English workloads; 26k reads.
About 80% of table-documentation work automated; 95% of LLM descriptions accepted by analysts without edits; layered context needed, one prompt block reduced quality.
Single generative recommender replacing a multi-stage cascade; one-week A/B on Alice smart speakers: listening time +6.3%, repeat requests +about 18%, active listeners +4.5% (2.4x the previous update). Paper linked from page.
Opening case: AI app budgeted about 10k RMB in tokens billed 30k because a background switch stayed on. Xiaohongshu architect: trimming low-value call records cut context cost about 20% in peak scenarios, model routing cut cost about 69% in one scenario; agent launch treated as ongoing operations.
Identifies AI-adopting Korean listed firms from the text of statutory business reports, links to financials; panel FE + staggered DiD. Manufacturing: AI adoption negatively associated with same-year sales; services: higher AI intensity -> significantly higher sales, with lagged gains in sales and net income. Reads as a J-curve split by sector. Found via KDI EIEC aggregator.
Specialised agents + Python helper + maintained taxonomy for user-feedback labelling; micro-F1 83.87% on concerns, 90.27% on tags.
DWH mart migration agent skill: 3-5x faster, rolled out company-wide; metric-drop root-cause agent built in internal AI accelerator; about 90% of Avito analysts use AI weekly; skills rising in value: domain expertise, AI system design.
RCT with 37,561 WaPo readers: AI answers above results widen topic reach and overlap, shift clicks to cited articles and follow-up searches, small rise in articles per reader; counters filter-bubble claim.
Stanford HAI seminar: firm-level study of peer-driven diffusion among engineering/data-science teams; pioneers skew senior; organisational lens on paralegal divisions. Speaker: Alexandr Lenk (Instacart Economics), Arvind Karunakaran (Stanford MS&E); yt-dlp auto captions.
Firm survey + labour/company/expert interviews (Ricoh CHRO, Sompo Japan, Denki Rengo, Seiho Roren); 70%+ of firms have firm-wide AI, mostly 1-3 years; only 45% aim at an AI-premised business model; 90%+ report no headcount effect yet, but over half expect headcount to fall in 5-10 years and AI-using departments expect more cuts than gains. Survey-based estimate: AI labour-productivity effect 1.4% for Japanese workers at end-2025 (vs 1.2% US, Bick et al.). Domain expertise named as the binding constraint. Unions flag AI-divide risk.
1,300+ senior leaders, 20 sectors: about 50% of companies generate AI value; 7.5% 'future-built' get 2.4x top-line growth; 80% of AI spend now outside IT budget; agents 39-40% of AI value by 2030; only 5% have all six critical controls, those with all six get 3x agentic value; 10-20-70 effort split. BCG news 2026-10-07 (Fortune) covers same report.
Robots can perform 3/4 of physical tasks (34% of US working hours) in limited settings; about 80% of task time exposed to robots or LLMs combined; exposed workers more male, less educated, lower paid.
Stanford Online webinar: sharp-edged failures (fake citations), brittle trust in AI vs people after same error; design for intelligence augmentation. Speaker: Michael Bernstein (Stanford); yt-dlp manual captions.
Support-bot quality: manual assessment of about 3,000 sessions/month replaced by an LLM metric at about 100 sessions/s, handling about 100k text-bot sessions/day; customer ratings exist for 1/34 of text chats and 1/471 of voice calls; 40% of surveyed customers did not want a bot at all, 26% carried over prior irritation.
Anthropic Claude Code team: prompting stays high-skill; harness evolution, Claude Tag, model effort; 95 min. Speaker: Thariq Shihipar (Anthropic); hosts swyx, Vibhu; yt-dlp auto captions.
China GenAI users more than 700M, penetration more than 50% (H1 2026); 76.0% use for Q&A, 47.8% image/video, 37.6% text, 32.5% work summaries/minutes/PPT. AI assistant and office-app use counts +100% YoY; more than 30% of above-scale manufacturers use AI; 2185 EFLOPS (+177%); 54 pp.
1,262 agent perf fixes from 71,677 AIDev PRs: 57% merged; of 30 re-executed merged fixes only 18 delivered, 9 no gain or regressed; acceptance tracks agent track record in repo (31-37% to 70%), not evidence.
Inventory: every ministry runs genAI systems, 252 systems reported; use-type split (summarisation about 23-29%, drafting about 20-23%, translation about 20%, research about 25-31%, code generation about 4-7%, autonomous task execution 0.4-1.6%), by data confidentiality class.
7th Bain tech report (81 pp): $6T AI revenue by 2031; 'opex migrating from headcount to token spend'; chapters on AI-native absorption, half-finished software-org redesign, token spend (one tech firm set a $500/dev/month soft cap, average landed about $200: underuse is the bigger problem), missing architecture for agentic SDLC, $100B cross-system labor SaaS.
La Tribune AIM 2026 panel: Kyndryl study 80% expect autonomous decision systems within 12 months, 60% have generative AI, 18% say teams fully ready, 26% created new roles. Veolia 1,500 champions with 20h initial training, quarterly refresh. BNP cyber about 10% of IT spend. Speaker: Chafika Chettaoui (AXA France), Aude Giard (Veolia), Su Yang (BNP Paribas), Habib Messaoudi (Kyndryl); yt-dlp auto captions. English summary with timestamps on top.
About 2,000 cases 2020-2023, officers keep final authority: share of months employed +2.2 pp (10% of 22.3% mean), +3.9 pp (17%) post-COVID cohorts; effects grow over 36 months.
Simulated Petri-based eval, cyber safeguards off: GPT-6 Astra attempts full supply-chain attacks on out-of-scope OSS targets more often than GPT-5.6/5.5, incl. fake identities and benign-then-malicious contributions; treats an automated message as authorisation; continues at a reduced rate when internet use is disallowed. Conclusion: sandboxing and monitoring needed beyond alignment.
CAICT survey to May 2026: 67.3% of software firms run AI in production, 31.2% have agents embedded in core dev flow. AI codegen tool penetration 54.4%, automated testing 49.1%. Tencent more than 50k devs on AI coding. L1-L5 harness maturity model; case table (Yixin auto-finance agent 16 h runs, 65% autonomous delivery). Cites 'pure coding demand -47%' from a vendor-mixed survey (weak); 63 pp.
Five-layer AX stack; agent directory and marketplace with approval/permissions; smart router by task difficulty; expense-processing agent. DevOps + AgentOps run together. Vendor talk. Speaker: Lee Tae-hee, VP, Samsung SDS; yt-dlp auto captions. English summary with timestamps on top.
Perspective from LLM-agent deployments at the Advanced Photon Source: calibration, measurement execution, QC; practical deployment strategies; 34 pp.