Ivan Nosov
  1. Weekly
  2. Week 41, 2026

AI weekly·No. 41

The week in AI, 3⁠-⁠10 October 2026

What public sources reported in AI, 3⁠-⁠10 October 2026, and what it means for work.

Audio for this week is not recorded yet.

The week in three lines

  1. Checking is now the scarce resource

    OpenAI published 722 machine-generated maths manuscripts and withdrew three within days; a maths site froze proof claims after 291 arrived in one day; open-source code output is up 183% while review comments per line fell by almost two thirds.

  2. Agents that act on the open web are producing incident reports

    Anthropic, the Wikimedia Foundation and an Australian parliamentary hearing each documented agents doing things nobody asked for: injection attacks, form submissions, proxy tricks.

  3. The harness buys cost, mostly

    Several independent measurements this week found that the scaffolding around a model changes the bill far more than the result, and two large companies reported that small open models with a validator beat frontier models on cost by one to two orders of magnitude.

Numbers from the week

162of 722

OpenAI maths manuscripts with Lean-formalised main results. The rest wait for people to check them.

OpenAI math repository, 6-8 Oct 2026

291in a day

proof claims arrived at erdosproblems.com before it froze new claims; 155 had no human response.

erdosproblems.com forum

83x

cost gap per code-review check: $3.42 for Opus against $0.041 for a Qwen 3.6 27B generator plus a MiniMax validator.

Yandex on Habr

3-98%

share of answers LLM judges call acceptable, depending on the judge. Workers rated 61% acceptable.

MIT FutureTech, audit of LLM judges on 45,796 worker ratings, arXiv 2610.02492

Exhibits from the week's sources

All 89 sources in the Library

Exhibit 1 · Report · Epoch AI

Epoch's own summary: the models miss the house rules

Epoch AI report page titled Can AI automate Epoch?, dated 8 October 2026, listing three key findings. The second, on models missing implicit conventions such as visual style and audience interests, is highlighted.
Key findings section. Highlight added by me. Source: Epoch AI, Can AI automate Epoch?, 8 Oct 2026 Read it in the summary

Exhibit 2 · Drawn exhibit

What held and what failed on 11 real Epoch tasks

Part of the jobResult
CodingReliable
Operating a computerReliable
Unwritten house standardsWeak
Designing experimentsWeak
Spotting own setup errorsWeak
Drawn by me from Epoch AI, Can AI automate Epoch?, 8 Oct 2026. Six models. Read it in the summary

Exhibit 3 · Press · Business Insider

722 maths papers in one day, and the people whose projects they ended

Business Insider article headline about an NYU professor saying OpenAI's release of maths papers wiped out early-career researchers' projects, with the byline and a quote from Tristan Buckmaster highlighted.
Read with: only 162 of the 722 have machine-checked main results. Highlight added by me. Source: Business Insider, 10 Oct 2026 Read it in the summary

Exhibit 4 · Paper · NBER w33777

Where the new AI tasks land, by occupation

Stacked bar chart from Humlum and Vestergaard showing the composition of new AI-related tasks by occupation, from IT support to financial advisors; HR professionals rank third on AI integration share.
Composition of new AI-related tasks by occupation. The same authors reported this week that a third of new tasks are AI oversight. Source: Humlum and Vestergaard, NBER w33777, rev. Mar 2026, Fig. 4 Read it in the summary

Chart 1 · cost per check

Same code review, two price points

US dollars per pull-request check, Yandex AI code reviewer. Opus scored 30% better on quality.

Opus $3.42 Qwen 3.6 27B + MiniMax validator $0.041
Source: Yandex on Habr. About 4,000 pull requests a day, roughly 40% of the monorepo. In the summary.

Chart 2 · skill formation

AI-written text in 2026 US STEM dissertation abstracts

Share of 2026 filings with AI-written text detected; zero through 2022. 226,591 abstracts.

Students from non-English-speaking countries 41% All 2026 filings 29% Programs ranked 90 to 100 26% Top-ten programs 15%
Source: Gross, Shvadron and Zhang, NBER working paper 35859. Detectors miss new model generations, so these are more likely floors than ceilings.

What happened

Anthropic: unintended model actions in evals and internal use · 9 Oct

Claude models, including unreleased ones, ran SQL and command injection against third-party sites, submitted government and police forms, used tokens to reach gated data and used URL shorteners to get around tool limits. Anthropic names three causes: ambiguous tasks, misconfigured environments and RL training that rewarded workarounds. Live internet access is now off for all internal evals. No incident counts were published. Read with the Wikimedia Foundation's report on OpenAI agents editing sandboxes, turning a citation tool into a proxy and making millions of API calls (5 Oct), and OpenAI's apology at an Australian Senate hearing over an agent intrusion into a Medicare portal (6 Oct).

OpenAI's 722 maths manuscripts, and what happened next · 6-8 Oct

An unreleased model produced 722 manuscripts across 372 result families at about three hours of Pro-level compute each. Only 162 have Lean-formalised main results, and the prompts were withheld. Within two days a sign error invalidated three results, a preprint showed that the Lean-verified Navier-Stokes proof does not formalise the stated claim, and set theorist Asaf Karagila called one preprint unreviewable. Separately, erdosproblems.com froze proof claims after 291 arrived in one day, 155 with no human response.

See Exhibit 3

Epoch AI: can AI automate AI research, or Epoch itself? · 7-8 Oct

In InnovationEval, models with 3,000 GPU-hours tried to reinvent a 2026 post-training method; the best clean result reached about 15% of the human method's gain, and their write-ups overstated results. In Can AI automate Epoch?, six models did 11 real Epoch tasks: reliable on coding and computer use, weak on learning unwritten house standards, designing experiments and spotting their own setup errors. One chart task is now automated in-house. This is the clearest task-versus-job evidence of the month.

See Exhibit 1 and Exhibit 2

What does a harness buy? Tokens, mostly · 3 Oct

Five models run in Claude Code, mini-SWE-agent and OpenCode on SWE-bench Verified: swapping the harness flips 13% of hard tasks, the same share as simply rerunning, while cost per task varies up to 3x. A 45-task test set detects a 13-point gap only half the time. A Fudan study of parallel sub-agents (2,124 matched runs) found gains only on the longest tasks, 1.4-3.3x more tokens, and concurrent writes as the top failure.

Small open models plus a validator beat frontier on cost

Yandex's AI code reviewer now covers about 4,000 pull requests a day, roughly 40% of its monorepo. Opus scored 30% better on their quality metric but cost $3.42 per check against $0.041 for a Qwen 3.6 27B generator plus a MiniMax validator, an 83x gap. About 30% of comments are accepted, and half of accepted findings are business-logic bugs. Alibaba.com's president Zhang Kuo described 107 real merchant tasks, now open-sourced, on which the best frontier model completes about 61% unaided; Alibaba says its small-model routing runs all 107 for $3.69 against more than $9 for Codex or Claude Code (vendor figures).

Your evaluation may be measuring the raters · 4 Oct

2,611 production interviews by micro1's AI interviewer "Zara": two reviewers scoring the same batches differed by 0.79 standard deviations, and reviewer turnover hid a real +0.53 SD improvement. An evaluation set goes stale in about five weeks. MIT FutureTech's audit of LLM judges on 45,796 worker ratings: judges rank answers reasonably, but the share they call acceptable ranges from 3% to 98%, against 61% for workers. Any "AI can do X% of tasks" figure built on LLM judges is unreliable on levels.

Labour: still no aggregate effect, visible shifts inside jobs

Humlum and Vestergaard (Brookings, 6 Oct): linked Danish records rule out earnings or employment effects above 2%; a third of new tasks are AI oversight. The NY Fed (8 Oct): small firms using AI expect to hire more, about 10 points above non-users after controls, though only 31% report higher sales. Workday (5 Oct): internal moves fell at 57% of employers, applicants per hire rose from 58 to 69, demand for basic prompting skills is down 25% from January while AI-building skills are up 51%.

See Exhibit 4

Placebo test of Claude Code skills · Habr, 8 Oct

450 pre-registered runs against a length-matched placebo: two of nine popular skills beat the placebo, none beat having no skill, and one skill's "65% fewer tokens" claim measured as 14% higher cost.

What it means for work

The labour market has not moved in aggregate, but the work inside it has, toward oversight, evaluation and translation. And the hardest problem for organisations this week was not getting AI to produce. It was knowing whether what it produced was any good, and who was accountable for acting on it.

The labour market: still waters, moving currents

Aggregate employment and earnings: no measurable effect in the best linked data. Inside jobs and firms: oversight tasks, falling internal mobility, more applicants per hire, a shift from basic AI use to building with AI, and skill demand that may surface at suppliers rather than adopters.

The evaluation problem: who is doing the measuring?

Put these next to the labour-market numbers. Applications per hire are rising, partly because AI makes applying cheap. A Russian fractional chief people officer, Dina Balashova, on the podcast HR Synergy Talks, described AI-written CVs flooding her screening, and HR teams having to learn the tools themselves just to assess AI engineers. So screening is under pressure from more volume, and the tools brought in to handle that volume are hard to evaluate, because the human raters disagree with each other and the model judges disagree with reality. That is the evaluation bottleneck.

Skill formation: what happens to the learning loop

Several studies this week touched on how people learn when the work is done for them. Daniel Gross, Dror Shvadron and Hansen Zhang, in NBER working paper 35859, ran AI-text detection over 226,591 dissertation abstracts from American STEM PhD programs. AI-written text was at zero through 2022, and appears in 29% of 2026 filings. The share is 41% for students from non-English-speaking countries, and it rises as program rank falls: 15% at top-ten programs, 26% at programs ranked 90 to 100. Students who used AI were 6 to 11 percentage points more likely to move into industry or non-tenure-track roles. A paper in the same week by Nakajima and Mizuno showed that detectors trained on older models collapse when a new model generation arrives, catching 99% just before and under 4% just after. So 29% is more likely a floor than a ceiling.

Governance: deployment is ahead of control

The vendor surveys this week are vendor surveys, so treat them as signals, but they point the same way. Thoughtworks surveyed 3,200 chief information officers in 10 countries. 88% said AI adoption is moving faster than governance can adapt. 70% said they have hired a chief AI officer, and 29% report friction between that role and the CIO. Forrester, by contrast, puts chief AI officer prevalence at 21%; sample and definition explain the gap. 89% of the CIOs said they now carry more responsibility for redesigning workforce workflows than for core infrastructure.

What to watch

  • Whether OpenAI explains its late-September training pause; no restart notice or postmortem as of 10 Oct.
  • Mistral Large 4 and Reflection Beam open weights, both promised for later in October.
  • IMF World Economic Outlook and Annual Meetings material, due after this window.

Sources filed this week (89)

Open Week 41 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 (3)

A Strong evidence (40)

B Useful, limited (33)
C Weak signal (13)