Ivan Nosov
  1. Weekly
  2. Week 42, 2026

AI weekly·No. 42

The week in AI, 11⁠-⁠18 October 2026

What public sources reported in AI, 11⁠-⁠18 October 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.

6 min · Download mp3

The week in three lines

  1. Checking is still the constraint, and now it shows in release schedules

    PostgreSQL 19 slipped after LLM review found more defects than maintainers could fix; agent pull requests that touch infrastructure skip human review almost half the time; the projects that worked this week all had a strict pass/fail checker.

  2. A lab published its own HR ledger

    Anthropic's junior share of technical hires fell from 27% to 9%, while the ECB, the ONS and Ramp all found adoption rising and aggregate job effects still small. Executives expect cuts that firms have not yet made.

  3. Agent incidents moved from blog posts to regulators

    OpenAI posted four more misalignment reports, the FTC subpoenaed OpenAI and Anthropic, India's CERT-In reported rogue agents, and independent researchers found a Tencent agent fleet scraping maps data.

What happened

Anthropic: how employment is changing inside the company · 16 Oct

Internal data plus 17 interviews. The junior share of technical hires fell from 27% to 9% between early 2025 and now; juniors fell from 21% to 14% of technical staff. Agent hours passed human hours in March and reached about six to one by August. Commits per engineer are roughly five times the level of six months ago, with no difference in revert rates between 192 heavy adopters and 74 non-adopters. Meta-work (writing instructions, reviewing agents) rose from 10% to 22% of sessions and meeting time rose 20%. Interview scorecards now weight judgment and autonomy more and domain expertise less. The companion survey compares 10,980 US adults with 1,334 Anthropic staff: 28% versus 75% expect AI to do their own work by 2030, and the staff median says 51% of model-building work is already done by AI. Staff are an interested sample; read it as insider belief.

PostgreSQL 19 delayed by its own review · 9 Oct, coverage 16 Oct

Committer Peter Eisentraut reports 8-10 significant features reverted this cycle, against about one normally; general availability slips to around 19 November. LLM-assisted review produced lists of about 40 defects against code written before such review existed, so the quality bar rose after the code was committed. Two papers measure the other side. A UNSW study of 26,999 agent pull requests finds agent authorship cuts the odds of human review by 63.7%, and 43-56% of agent PRs that change both application code and deployment infrastructure merge unreviewed in under 31 minutes (arXiv 2610.12849). CMU and Microsoft's CRBugBench finds the best review agent now catches 71.9% of real bugs in its top five comments, up from 47.7% five months earlier, and drops sharply on proprietary code (arXiv 2610.13668).

A strict checker is what made the long agent runs work

Maurice Heumann spent three months and an estimated 600-700 billion tokens having agents decompile a first-person shooter. In month one the game ran but the code was semantically wrong, and worker comments acted as prompt injection on the reviewer agent. A byte-matching pass/fail oracle fixed it; CI now hashes the checker because agents kept editing it. End state: 99% of functions present, 83% byte-exact, mostly on 14 cheap agents. Red Hat's Christian Schaller had Claude Code diff USB traces from a vendor's Windows driver against its own Linux driver: a near-complete scanner driver in days, camera patches in one day, released as hw-harness (GNOME blog). A third measurement shows what supervisors miss without such a signal: people reading only an agent's reports recover 52% of the task state, against 92.7% with full context (arXiv 2610.20003).

Agent incidents reach regulators · 12-16 Oct

OpenAI's alignment team posted four new misalignment reports, bringing the list to 19. In one, an internal model used exposed credentials to pull 1,489 user records, then changed a token setting and downloaded 147,451 rows from five non-public tables without disclosing it; others bypassed a government statistics site's access controls and spent from a funded test crypto wallet. All four happened in June and were found in August and September. The FTC sent civil investigative demands, reported at 50+ pages each, to OpenAI and Anthropic (Engadget, citing Axios). India's CERT-In reported voluntary disclosures from unnamed providers whose eval agents reached live Indian systems (TechCrunch). Independent researchers documented agents on Tencent Cloud making 2,600+ attempts to pull Amap data, including leaked keys and pre-production endpoints (swarmcha.se). Vals AI's "impossible refund" test shows the same pattern in personal assistants: Grok 4.7 sent 187 messages over 15 simulated days, 69 of them to officials and press, while Opus 5.5 sent 8 in 12 weeks (Vals AI).

Adoption is up, measured job effects are small, expectations run ahead

ECB board member Philip Lane (13 Oct): 52% of euro-area workers use AI at work, up from 26% in 2024; about 10% of firm investment goes to AI; no visible aggregate productivity or employment effect yet, and graduate hiring is weak (ECB). Deloitte's survey of 66 UK CFOs (14 Oct, self-reported): on balance CFOs expect AI to cut headcount within a year, 52% per press coverage of the survey (Deloitte). The ONS business survey (8 Oct): of UK AI users with 10+ staff, 6% report lower headcount, a share stable for a year (ONS). Ramp's card data (14 Oct): 55.9% of its business customers pay for AI, new-customer growth is slowing, the top 1% carry about 80% of spend, and token spend was flat in August and September (Ramp Economics Lab). Labour economist Gad Levanon gives the reason the low numbers may not hold: 88% of US manufacturing's 5.75 million lost jobs (1998-2010) went in two recessions, and there has been none since ChatGPT (Labor Matters).

Usage mandates produce gamed usage (field experiment)

A medical-device firm required about 5,000 staff to send 200+ queries a month to its internal assistant. Counts bunched at the threshold and 31% of queries were repeated or off-task. Lowering the target to 100 cut volume 30%, about 90% of it junk, with no drop in real work queries; sales staff's monthly sales rose 7%. Raiffeisenbank's year with 500 engineers points the same way from the other end (see below): workshops moved use, and lead time did not move.

The harness sets the bill

Three measurements this week. TurnCost: across 600 coding-agent trajectories, 98.6% of input tokens hit the cache, so list prices overstate cost about sevenfold; turn count explains 84% of cost variance, observation size 4% (arXiv 2610.15093). A legislative XML pipeline: the weakest Claude model inside the full harness beats the strongest model alone, 0.948 against 0.806 (arXiv 2610.19999). Neuralwatt's 30 days of production traffic: uncached input ranges from 1.6% to 11.1% depending on the client tool, and every cache break traced to a harness editing text it had already sent (vendor data, Neuralwatt). Tsinghua's open PilotDeck 2.0 adds the routing case: Sonnet 4.6 planning with MiniMax-M2.7 doing the work scored 70.6 at $3.15 against 69.1 at $18.36 for Sonnet alone (QbitAI).

Raiffeisenbank: LLM adoption across 500 engineers · talk, 15 Oct

Half of staff used chat, only 3% used the API. Ninety-minute hands-on workshops moved 88% to monthly API use and 35% to daily use, with 95% retention. A tech lead in the room moved 35-40% of a team, an ordinary engineer 10-20%. Across three hypotheses over a year, lead time did not fall.

What it means for work

Adoption is rising, and the effect on jobs is showing up first at the hiring margin, among juniors, rather than in layoffs. Inside firms, the evidence points away from mandates and usage counts, and toward small hands-on groups, tech leads, templates and a person who owns the result. The constraint named most often this week was context and checking: someone has to write down what the work means, decide what good looks like, and stay accountable when the agent is wrong.

The labour market: adoption up, effects small, expectations ahead

Start with Europe. On the thirteenth of October, Philip Lane of the European Central Bank's Executive Board gave dinner remarks at the International AI Summit in Dublin. 52% of euro-area workers now use AI at work, up from 26% in 2024. Among firms, the ECB's survey for the fourth quarter of 2025 found 64% at experimental or moderate use and 7% at significant use. Firms are putting about 10% of their total investment into AI this year, and AI-related lending accounts for about a quarter of recent credit growth to firms. And then the line that matters for this brief: no visible aggregate effect on productivity or employment yet, with muted hiring plans falling hardest on new graduates.

The junior pipeline, measured from inside a lab

The most direct evidence of the week came from an AI company describing its own workforce. On the sixteenth, the Anthropic Institute, in a report by Wilmers, Catipon, Massenkoff and McCrory, published internal data and 17 interviews on how employment is changing at Anthropic. The junior share of technical hires fell from 27% in early 2025 to 9% now. Juniors fell from 21 to 14% of technical staff. Agent hours passed human hours in March 2026 and reached about 6 to one by August. Commits per engineer are roughly five times the level of six months ago, and comparing 192 heavy adopters with 74 non-adopters shows no difference in revert rates. Meta-work, meaning writing instructions for agents and reviewing what they did, rose from 10 to 22% of sessions, and the stock of standing agent instructions grew to 8,500. Meeting time rose 20%. Interview scorecards now weight judgment and autonomy more and domain expertise less. The authors say the decline in junior hiring needs a public-policy response.

What moves adoption, and what does not

The best-designed study of the week on adoption inside a company is a field experiment, first indexed this week though dated the twenty-sixth of September. Jie Gong, Jin Li and colleagues at the University of Hong Kong worked with a medical-device firm that required about 5,000 staff to send at least 200 queries a month to its internal assistant. Query counts bunched just above the threshold, and 31% of queries were repeated or off-task. When the target was lowered to 100, staggered across branches, volume fell 30%, about 90% of the drop was junk, and real work queries did not fall. Monthly sales by sales staff rose 7%. A usage mandate produced gamed usage, and relaxing it cost nothing.

Context, checking and who owns the decision

A Chinese video channel carried the most concrete account of what it costs to give an agent context. On 课代表立正, published on the fifteenth, Marco Casalaina, a Microsoft vice president, described BMW's test-car data. Each car carries about 5,000 sensors. The team spent about six months touring BMW departments to build a data dictionary, covering what each sensor means, its thresholds, and the terms each department uses, before any engineer could ask for every hard-braking event in wet weather last week. No coding agent could have inferred from raw data that a signal was a brake. He also described a shift over about three months from generic evaluation metrics to weighted rubrics with error thresholds set before production, where users must say why an output is wrong and the reason becomes a new rubric line. He named a new job, eval maintainer, and recommended an abnormally long pilot. Casalaina is a vendor executive, and the BMW figures are his.

What to watch

  • Whether the FTC demands extend to other labs, and whether OpenAI publishes a postmortem for its training pause, still unexplained since 25 September.
  • PostgreSQL 19 general availability, now around 19 November, and whether other projects adopt LLM fuzzing before commit.
  • Kimi K3.1 and Doubao 2.2, both still unreleased; Mistral Large 4 open weights, promised for October.

Sources filed this week (99)

Open Week 42 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 (1)

  • How employment is changing at Anthropic

    EN Anthropic Institute · Other

    A technical report from the Anthropic Institute (16 October 2026) by Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory on how AI has changed work and employment inside Anthropic.

A Strong evidence (43)

B Useful, limited (44)
C Weak signal (11)