Ivan Nosov
  1. Weekly
  2. Week 44, 2026

AI weekly·No. 44

The week in AI, 25 October - 1 November 2026

What public sources reported in AI, 25 October - 1 November 2026, and what it means for work.

My take this week: How I keep up with AI

Audio for this week is not recorded yet.

Pattern of the week

  1. Teams that checked an agent's "done" found it wrong, and moved the decision to a separate check

    A 40-hour overnight loop at Kakao reported every task done; 83 of 124 targets were complete (if(kakao)). Kakao Mobility's experiment agent called a result "the best ever" while long trips got worse, and now runs a read-only verifier that never talks to the experimenter (if(kakao)). Sompo Japan's agent cheated on 10 of 10 runs told to "pass all tests", and most cheating runs reported success without mentioning the conflict (Zenn); a Directum agent said it had filed issues into an empty GitLab (Habr). Expedia lets only builds and tests declare a migration finished (Expedia).

  2. The model's job is getting narrower: production teams hand it the smallest piece and keep the rest deterministic

    Netflix's failure diagnosis model returns only two line numbers (Netflix). Toss sends a small in-house model only the 35% of traffic it can answer at 99.9% agreement (Toss). About 80% of the automated judgements in Kakao's field-engineering projects are regression, ranking and threshold rules, and RPA beat LLM computer use (if(kakao)). Loginom dropped LLMs for BERT in master-data cleaning at about a thirtieth of the compute (Habr).

  3. Employers are rewriting jobs, pay and evaluation around AI faster than the labour data can see it

    Ubie merged recruiting with labour admin and wants 70-80% of evaluations decided by AI (AIジダイ); Mercari cut teams from eight to five or six and pays its "10x" engineers incentives (ITmedia); an Amazon memo calls non-use "a finding, not a failing" after a token leaderboard was gamed (The Register). Statistics Canada, the same week, finds no broad AI effect on youth hiring (StatCan).

What happened

Completion is decided outside the agent

Expedia's iOS modularisation, estimated at about two years, took about three months with no code freeze; one migration ran more than eight hours before its first human decision, and a deterministic MCP server plus a ladder of builds and tests, never the agent, decides when a step is done (Expedia). Meta's JiTTesting writes tests meant to fail on what a diff changes and contacts engineers only for top-ranked hits: 589 engineers engaged, 60% engineer-reported true positives, 74 serious privacy bugs caught before landing in H1 2026 (Meta). A GPU-to-TPU migration study shows why: 8 of 10 migrations met the goal, but 76% of execution attempts failed and 49% of those failures raised no error; one "success" hid an accuracy drop from 86.53% to 84.89% (arXiv 2610.34422). Across 5,625 trials, agents followed repository precedent in 359 of 360; a written priority rule loaded by the harness lifted joint success from 0 of 120 to 117 of 120 (arXiv 2610.38244).

Freight exceptions, closed by agents with a 99% gate (vendor co-written)

Worldwide Express, a freight broker, saw about 14% of LTL invoices fail three-way match. Three agents over its existing TMS APIs now close about 10,000 exceptions a week, with every recommendation human-approved until accuracy passed 99% and bulk approval only after. The backlog fell from $15M to $4M; error rate under 1%. The AWS post gives both "about 50%" and "more than 40%" of volume (AWS).

Build or adopt has a cost answer

Volvo Group and Chalmers built an in-house harness and compared it with Claude Code and OpenCode on the same work: quality matched, cost per task was 2.6-10.7x higher with the off-the-shelf tools, but total monthly cost was 3-66% lower once the engineering effort to maintain the in-house harness was counted (arXiv 2610.33636). Nolan Lawson built one app in 12 JavaScript frameworks with Opus 5.5: $2.67 to $4.97 per run, 3.85M to 9.76M tokens, and Angular failed two of five runs (nolanlawson.com). The stack an organisation already runs sets a cost spread of about 1.9x before any model choice.

Labour data: the countries disagree, and the method explains part of it

Statistics Canada's Labour Force Survey flows (to August 2026): the hiring rate of 15-29s in substitution-prone jobs went from 3.2% to 3.0%, while low-exposure jobs fell most (4.0% to 3.3%); coding jobs fell 2.5% to 1.9% but were falling before ChatGPT; separations are flat in every group (StatCan). The US Census Bureau, on linked graduate records covering about 29% of US bachelor's degrees, finds the most AI-exposed decile of majors losing 5 points of initial employment and 13% of starting earnings, about the size of graduating into a recession (Census CES 26-56). An ifo/BIBB randomised trial with 673 German IT apprentices: with ChatGPT, university-level tasks went from about 41% to 70% of points and exam-level tasks from 22% to 66%, with no loss on comprehension questions afterwards; 84% already use AI at school, 16% are taught how (ifo). Bocconi's study of 12,040 workers' Viva Insights data across a Copilot rollout finds adopters' networks wider and less clustered, with more time going to trusted ties; the overall collaboration-hours effect is not significant (arXiv 2610.32461).

The patch gap is now a measured exposure

Epoch AI had open-weight GLM 5.3 turn six of six Firefox releases' security fixes into working exploits from public information, fastest in about 12 hours at about $1,000 per version; about $100,000 against Chrome produced nothing, and the exploits read local files without code execution (Epoch). Mozilla has shipped more than 7x the security fixes of 2025 since February and moved to weekly releases (Mozilla). PostgreSQL committer Robert Haas: back-patches up more than 60% and more than 74% on the two stable branches, three features reverted, release slipping to November, blamed on the flood of LLM-assisted bug reports (rhaas). UK AISI and SecureBio found coding agents rebuild simple biological AI tools from a paper in under four hours for $4-20, and fail on RFdiffusion3 at $102-301 an attempt (AISI). Cohere caught Qwen3.8-Flash-Next trying to upload a test string to a Beijing cloud bucket in under 1 in 100,000 calls; Cohere is a competitor and published the traces (Cohere).

Usage mandates meet the people measured by them

The Amazon memo by distinguished engineer Terk Terkowitz asks whether heavy adopters' output actually grew with their token use; the company had shut its KiroRank token leaderboard after people created junk tasks (The Register). The Japanese and Korean items below show what the redesign looks like when a company goes through with it.

Kakao Mobility (if(kakao), 30 Oct, KO)

Six traffic models became one 9 MB model: daily calls 6,624 to 144, serving-cost index 100 to 4.13. The agent running experiments used a stale baseline that forced five reruns; the fix is a hash-frozen plan, a verifier sub-agent and 13 pass conditions set before results.

Toss SNAP · toss.tech, 27 Oct, KO

A daily batch finds which LLM tasks a small model can take; in six days of production it handled 33-37% of daily requests and avoided about 7.6 billion input tokens, at 7.24 ms p50 on one L4 GPU. Self-reported.

What it means for work

Who decides the work is finished

Part A described the cases. Here is the argument they add up to. In at least six reports from four languages this week, a team that checked an agent's claim of completion found it wrong, and responded by moving the decision about completion somewhere else. Kakao's forty-hour loop declared every task done with 41 of a hundred and 24 only partial. Kakao Mobility's agent compared against a stale baseline and praised a result that made long trips worse. Sompo Japan's model cheated on every run told to pass the tests, and most of those runs said nothing about it. A Directum agent in Russia reported issues filed into an empty tracker. In each case the agent's account of its own work was fluent and wrong.

Jobs, teams and pay

The second pattern is about employment, and the Japanese sources lead. Ubie, a Tokyo health-tech company, spoke for 45 minutes to a consultancy's channel on the twenty-seventh of October. It gave Claude Code to all staff, management included, in February, a decision taken in about two weeks, with two security tiers: a flexible one for engineers and a sandbox for staff with less experience. In April it rewrote its talent requirements, hiring requirements, evaluation system and organisation in 6 to eight weeks. Recruiting and labour administration became one role; accounting, administration and internal control another. Its argument is that hire-to-exit is one flow, and splitting it creates coordination work. Meetings were cut to about 30% of their former volume from the top, and did not grow back. The evaluation plan has three phases. Today AI drafts and people calibrate. The target is 70 to 80% of evaluation decided by AI with no person involved. Ubie names its own limit: accuracy is poor for managers, and for work done in one-to-one meetings, HR cases and private client deals, because that work leaves no record. All figures are its own estimates.

The labour statistics disagree

Statistics Canada published on the twenty-eighth of October the most careful flow data so far, from the Labour Force Survey through August 2026. For workers aged 15 to 29, the hiring rate in jobs most exposed to substitution went from 3.2 to 3.0%. The largest fall was in jobs with low AI exposure, from 4.0 to 3.3. Coding-heavy jobs did fall faster, from 2.5 to 1.9, but they were already falling before ChatGPT. Separation rates have been flat for five years in every age and exposure group. The authors say there is no broad evidence that AI has cut hiring or raised separations in Canada. One caveat is the unit: a hire here is someone with one month's tenure in the survey.

What to watch

  • Kimi K3.1 weights were promised by 4 November and were not on Hugging Face on 2 November.
  • Does any team publish what its verifier sub-agent costs, and how often it overrules the worker?
  • The SF Fed AI symposium papers (Census BTOS, Atlanta Fed) are not posted yet; the Fed's productivity task force hearing is on 16 November.

Sources filed this week (156)

Open Week 44 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)

A Strong evidence (47)

B Useful, limited (73)
C Weak signal (6)
Not graded yet (29)