Ivan Nosov
  1. Weekly
  2. Week 45, 2026

AI weekly·No. 45

The week in AI, 2⁠-⁠8 November 2026

What public sources reported in AI, 2⁠-⁠8 November 2026, and what it means for work.

Listen

This summary as a short episode

AI-generated audio overview of this page, made with Google NotebookLM: two synthetic voices talk it through. The text on this page is the record.

7 min · Download mp3

Pattern of the week

  1. Writing the change has become cheap; getting someone to own it has not

    Bloomberg opened 2,333 migration PRs over six months, and two months after opening only 34.2% of the C++ and 10.5% of the Python PRs were merged; the cost sat with the owning teams who review, test and deploy (arXiv 2611.05457). A Japanese contractor made 37 Claude Code PRs in two weeks against one client reviewer who merged at most about eight a week, with a median of 11 days to first review (Devlog, Zenn). Haulmont's Jmix team built a search rewrite in 3 days and needed almost 3 more weeks before one person could own the 15,000-line diff (Habr). Avito's LLM-assisted chaos tests are limited by service owners opting in: 6 tests a day against 44 planned (Habr).

  2. The cheapest answer that works wins, and often it is not a model call

    U Mobile's telco helpdesk answers about 70% of requests by keyword match with no model call, 20% with a small model and 10% with a frontier model (AWS, co-written). In a Brazilian lab portal test, plain RPA was more accurate than a browser agent, about 13 times faster and about 6 times cheaper per report (arXiv 2611.03853). Rosselkhozbank reached about 75% chat automation with a CPU-only MiniLM encoder and Postgres search (Habr), and a Russian marketplace support team found a single rule scored 82% against 84% for its three-model judge cascade (Habr). Hightouch got 98.8% accuracy at $0.88 per 1,000 checks by letting a cheap model escalate uncertain cases to Opus, which alone scored 99.0% at $3.86 (Hightouch).

  3. Code computes the numbers; the model proposes and people decide

    Meta's experimentation agent selects and sequences procedures while deterministic code computes every number it reports (arXiv 2611.04515). ZOZO's monthly cloud budget review runs amounts through SQL and leaves causes to the agent and decisions to people, cutting several days to 1-2 hours (ZOZO TECH BLOG). A Russian exam-grading bot has the model fill a schema and code recompute the score (Habr). Kakao let an agent run seven days of full reconciliation on a search-index migration while people judged which side was wrong (tech.kakao.com).

What happened

Bloomberg: generation was the easy part

Bloomberg's "Automated Uplift" ran six months of low-priority infrastructure migrations: Python packaging (294 PRs) and an in-house C++ build system to CMake (2,039 PRs). Of 2,333 PRs, 729 migrations merged. The design keeps the model inside a deterministic pipeline, editing bounded files with pre-computed context and production telemetry. Merge rates were higher where the pipeline could compile and test before opening a PR, and where repositories looked alike. Most unmerged PRs were not rejected; they were waiting. The team estimates "well over a thousand person-days" saved; no LLM cost is given (arXiv 2611.05457).

Keeta (Meituan): an in-house 31B model replaced the frontier incumbent in customer service

The operating team runs round-the-clock inspection of live conversations, separates model errors from faults in procedures, tools or harness, and trains only on the model's own errors. Over five iterations the offline bad-case rate (fixed simulation set) fell from 6.42% to 1.97% (the incumbent, GPT-5.4, sat at 4.60%), median latency fell 27.7% and serving cost per order fell 70.5% after full replacement. Every promotion needs simulation parity, an online comparison and human approval. Self-reported by the team (arXiv 2611.02839).

Revolut: unrationed self-hosted tokens over rationed frontier ones

About 2,000 engineers use GLM-5.3 served from 56 B200 GPUs at a blended $0.30 per million tokens, about a tenth of the earlier Anthropic and OpenAI invoice (staff and operations not included). Budget-capped Claude Code earlier in 2026 showed little measurable effect; after the switch the bank reports 60% better coding time and 70% better lead time, with no control group. Frontier models are now "the exception". The talk also gives serving targets: first token under 3 seconds, 200 tokens a second, 256K context (AI Engineer talk; The Register, 3 Nov).

Two operations cases from AWS's customer blog, read with the vendor caveat

U Mobile's internal helpdesk for call-centre and branch staff grew from 2 pilot agents to 8 agents and 150+ scenarios; ticket volumes fell 48.7-84.5% by category from January to September, with no control. Adoption needed prompt templates per scenario, not "ask anything" (U Mobile). Greenlight Financial triages 3,000-4,000 security and risk alerts a month; adding a second-tier reasoning model and episodic memory lifted resolution from 83% to over 90%. Remediation still needs a person or an approved policy, and the hours "avoided" are computed, not measured (Greenlight).

Meta: let the agent pick the query, never the arithmetic

Meta's agent for its A/B platform has 97 skills. Withholding information from the model improved answers; almost half of requests were plain estimate fetching; and in a number of sessions the agent recommended p-hacking, such as removing readings from unfavourable days. Reliability fell wherever no report contract applied. Five pages, no accuracy counts (arXiv 2611.04515).

Independent checks on the labs: METR at Anthropic, R40 on Kimi K3.1

METR's high-access review (12 August-29 October) finds Anthropic is likely improving capabilities on well-specified tasks at about twice the 2024-25 rate (plausible range 1.3-3x), mostly through AI assistance in R&D. Anthropic redacted most of the evidence METR found most informative (METR). R40 reproduced Kimi K3.1 Pro's Terminal-Bench score (66.7% vs 67.2%) and found that in 23.3% of attempts the model cites "instructions" the prompt never gave but hidden tests check; when a task was edited to contradict such a rule, it ignored the edit in 3 of 4 tasks (R40 audit). For a deployer, a model that overrides a written instruction with a remembered one is a production risk.

Labour statistics: small effects, and one broken series

Australia's statistics bureau asked businesses for the first time: 81.8% see no AI effect on vacancies, 3.8% report fewer and 0.6% more; information media and telecoms is at 16.2% fewer (ABS). The US Census business survey renamed AI "Super Intelligence" under Executive Order 14434 in the same wave as a new sample cycle; use held at 24.2% but "don't know" rose from 9.6% to 13.9%, so the series breaks here (Census BTOS). Bank of England survey firms report about +0.5% productivity from AI over three years and close to zero effect on jobs, and expect the next gains to split roughly equally between slower employment growth and faster output (BoE MPR). In Finland, first-choice applicants to ICT fell 38% while ICT graduate unemployment rose only 2.9 points: the authors read this as a reaction to expected rather than current conditions (IZA DP 19056).

Rosselkhozbank, Russia: no vendor, no GPU

After leaving its chatbot vendor, automation first fell from about 60% to 40% on keyword matching, then rose to about 75% with a CPU MiniLM model; the 75% is measured as 3 minutes of client silence, which the team calls a proxy (Habr).

What it means for work

Plan from the reviewer, not the generator

Three patterns ran through this week's practitioner reports, and they are one argument. First, writing a change has become cheap and owning it has not. Bloomberg merged 729 of 2,333 automated migration pull requests; after two months, a third of the C++ ones and a tenth of the Python ones were in. A Japanese contractor produced 37 pull requests in two weeks for a client with one reviewer who merged at most about eight a week. Haulmont needed three days to write a search rewrite and three weeks before one person could own it. Avito's chaos tests run six a day against 44 planned, because service owners must opt in.

Jobs redrawn from the inside

Sber's business-to-consumer content team in Russia replaced a chain of human approvers, an editor, reputation risk, legal, a proofreader and part of compliance, with three checks running on Sber's own GigaChat model inside its content tool. The rules came from clustering years of past approver comments into 50 to 60 editing rules, reputation-risk traffic lights and legal risk maps for five financial products. That is tacit judgment written down. Legal review went from a five-day service level to about three minutes. Content output roughly doubled since 2023 while the author's own team shrank from 10 content managers to five; the others moved to other roles, and an "AI validator" role is planned. The caveats are the team's own: full automation covers only 30% of text and push messages, the accuracy figure has no method, and the checks over-flag.

Labour statistics: small effects, and one broken series

Australia's Bureau of Statistics added questions on AI to its quarterly job vacancies survey, the first official measure tying AI to vacancies. In August, 82% of businesses reported no AI effect on vacancies, under 4% reported fewer, and under 1% more; 13% were unsure. Information media and telecommunications stood out at 16% reporting fewer vacancies. Looking 12 months ahead, 8% expect fewer. The shares are of businesses, not of vacancies or jobs, and the smallest firms are excluded.

What to watch

  • Whether the US Census publishes a bridge between the AI and "Super Intelligence" waves; without one, adoption trend lines from BTOS break at wave 202622.
  • More review-capacity numbers: does any team publish merge rates and review time per agent PR the way Bloomberg and Devlog did?
  • Independent tests of the released Kimi K3.1 weights, and whether other open models show the hidden-rule recall R40 found.

Sources filed this week (112)

Open Week 45 in the Library

Every source screened and kept this week, strongest first. Each links to its Library page with the digest and the reason for its grade.

S Shapes the field (2)

A Strong evidence (45)

B Useful, limited (62)
C Weak signal (3)