A working paper co-authored by Hinton, Bengio and senior figures from OpenAI, Microsoft and Anthropic argues the measurement infrastructure for AI R&D automation doesn’t yet exist — and that’s the first problem policymakers should solve.
Key Numbers From The Working Paper
>80%
Anthropic self-report: AI share of approved code, Jan 2025–May 2026
26%
Anthropic self-report: R&D work autonomously completed (high-level supervision only), Mar–Aug 2026, up from 1%
~4.5 mo
Training compute efficiency doubling interval (software improvements alone), per paper inputs
9–900×
Cost reduction per capability milestone per year — full range, paper’s own figure
A working paper titled “What if automating AI R&D triggers an intelligence explosion?” — Frontier AI Working Paper Series No. 2/2026, published in September 2026 via the Centre for the Study of Existential Risk at Cambridge — was initiated and led by academic and civil society researchers. Its co-authors include Turing Award winners and Nobel laureate Geoffrey Hinton, Yoshua Bengio and Andrew Barto, alongside OpenAI Chief Scientist Jakub Pachocki, Microsoft Chief Scientific Officer Eric Horvitz, and Anthropic co-founder Jack Clark. Co-authorship is not corporate endorsement: the effort’s own framing is that it was driven by academia and civil society, and the executive summary was written by a subset of the authors. Nothing in the paper says OpenAI, Microsoft or Anthropic backs, supports or has accepted any of its recommendations.
The paper’s quantitative case rests on two figures, both Anthropic self-reports cited by the paper rather than independently measured. The first: Anthropic reports that AI systems’ share of approved code rose from low single digits to over 80% between January 2025 and May 2026. The second, measuring a different thing over a different window: the proportion of R&D work autonomously completed with only high-level human supervision rose from 1% to 26% between March and August 2026. These two figures are not combined or presented as a single trend here, because the paper does not present them that way. The OpenAI data point in the same section carries no percentage — it is qualitative, noting that AI assistance is used “in practically all parts of the company” and that systems routinely complete R&D tasks that would take staff days.
The paper does not publish a definition of “AI R&D work” or of “high-level human supervision.” Without those definitions, a move from 1% to 26% is a move on a metric whose denominator is not public. That is a statement about what the document contains, not a suggestion the number is wrong: the figure may be entirely accurate, and it is simply not checkable from outside, which is a different property from being inaccurate.
The key insight: A policy argument whose strongest evidence can only be produced by the party being regulated is, first and foremost, an argument for measurement infrastructure. Visibility is recommendation number one in this paper — not an afterthought — and that ordering is analytically honest.
The quantitative case that AI is automating AI research turns on this series, and on one other figure from the same company.
The Structural Read
The Permission Layer framework names the structural condition precisely: the gap between what regulators can see and what they need to see determines how much any policy can actually do. This paper sits squarely in that gap.
The mechanism the paper describes is parallelism, not raw speed. Its language: “Today, only thousands of researchers work on frontier AI R&D. Because AI systems can be copied and run in parallel, automating this work could add the equivalent of millions more.” From that premise, the paper models consequences under explicit conditions. If r — the returns to research effort — stayed at the levels it describes, and if no other bottlenecks emerged, the pace of AI progress would increase tenfold within about 1.5 years, at which point a year’s worth of progress at today’s pace would take about five weeks. At full automation, even the current pace of efficiency improvements would grow the automated R&D workforce 100-fold over months or years. Both figures are model outputs under those stated conditions, and neither is restated here as a forecast.
The inputs behind those outputs are worth naming in full. Training compute efficiency has been doubling roughly every 4.5 months from software improvements alone. The METR time-horizon metric — measuring the length of tasks AI agents can complete — doubled roughly every 7 months initially, and has accelerated to roughly every 3 months since 2024. Training runs that still take 3 months or more are treated in the paper as a friction on acceleration, not an accelerant. And the cost of reaching a given capability milestone falls by roughly 9- to 900-fold per year, depending on the capability milestone — a range a hundred times wide, and both ends belong in any citation of it.
Frontier AI Working Paper Series No. 2/2026
“Obtaining visibility into AI R&D automation, such as by supervising frontier AI companies through embedded auditors and requiring these companies to report key indicators, including the extent of AI R&D automation.”
The structural point — and it is not a gotcha — is that the paper’s first recommendation is precisely the indicator the argument currently borrows from a single company. That reading runs with the paper rather than against it. The authors plainly know the evidence base and the first ask are the same gap, which is why visibility comes before any speed limit, any air-gap requirement, or any international agreement. You cannot govern what you cannot measure, and this paper is, among other things, a case for building the measurement layer before it is a case for any particular policy response.
Permission Layer — The Structural Condition
The evidence gap is the policy gap
The paper’s empirical case relies on self-reported data from the regulated entity. Its first recommendation is mandatory reporting of that same data. That sequencing — measure first, then govern — is the correct order, and the paper states it explicitly. The Permission Layer cannot function without a data layer beneath it.
Three Implications
IMPLICATION 1 — MEASUREMENT BEFORE MANDATE
Before any speed limit, air-gap requirement or international agreement can be designed, policymakers need a data standard that is independent of company self-reporting. The paper’s embedded-auditor and mandatory-reporting proposals sit logically ahead of everything else it recommends: a speed limit on capabilities growth, monitoring of high-stakes experiments with a pre-built shutdown option, and a requirement that AI R&D run in secure and isolated environments such as air-gapped networks all presuppose that somebody outside the company can see what is happening inside it. Sequencing matters more than any single recommendation.
IMPLICATION 2 — THE PARALLELISM ASYMMETRY
The paper’s mechanism — copying and running AI systems in parallel to multiply the effective research workforce — applies asymmetrically across geography. Nations or organizations that can deploy large compute clusters at scale benefit from the multiplier first. International incident-sharing and verification tools are in the recommendations precisely because the governance problem is cross-border, but verification tools for AI R&D automation do not yet exist in any agreed form.
IMPLICATION 3 — CO-AUTHORSHIP AS SIGNAL, NOT POLICY
The authorship list is the document’s most widely-cited feature, and the most easily misread. Senior figures from OpenAI, Microsoft and Anthropic joining academic co-authors signals that the concern is considered serious enough to associate names with — it does not signal company acceptance of any recommendation. The distinction matters because reading co-authorship as corporate endorsement misreads what has actually been established: three named individuals put their names to a document, and that is a different fact from a company accepting a mandate.
Business Engineer Framework
The Permission Layer
Every AI capability that reaches users passes through a permission layer — government, regulator, platform, or enterprise gate. This working paper is an attempt to redesign that layer before the capability curve outruns the governance curve. Understanding where the Permission Layer sits in the AI stack — and which companies are building toward it, through it, or around it — is the core analytical question for the next three years of AI strategy.
The working paper’s most durable contribution is not its model outputs or its authorship list — it is the honest admission that the evidence needed to justify its policy recommendations is the same evidence that does not yet exist in any auditable form. Some experts cited in the paper believe full AI R&D automation could arrive within the next few years. Whether that timeline is right or wrong, building the measurement infrastructure to detect acceleration is the one action whose value holds across nearly every scenario — and it is the action the paper names first.
Co-authorship is not corporate endorsement. The individuals above are named with their affiliations as the document lists them. The effort was, in its own words, initiated and led by academic and civil society researchers, and the executive summary is written by a subset of the authors. Nothing above says OpenAI, Microsoft or Anthropic endorses, backs, supports or has accepted any recommendation in the paper, and nothing establishes that any company has agreed to embedded auditors or to any other measure described. Both of the paper’s quantitative anchors — the share of approved code and the share of R&D work autonomously completed — are Anthropic self-reports that the paper cites, not independent measurements, and the paper does not publish a definition of AI R&D work or of high-level human supervision. The two figures measure different things across different windows and are not combined above. The cited OpenAI material is qualitative and carries no percentage. The tenfold-in-about-1.5-years figure and the 100-fold workforce figure are outputs of a model under stated assumptions — that the relevant rate stayed at current levels and that no other bottlenecks emerged — and neither is restated above as this publication’s forecast. The cost figure is a range roughly 9- to 900-fold per year depending on the milestone, and is not quoted at one end. The view that AI R&D could be fully automated within a few years is attributed in the summary to some experts rather than asserted by the paper. This is a working paper. Peer-review status, any company’s agreement to any recommendation, any government’s response, the definitions behind the figures, and whether either figure has been independently verified are not established and do not appear — a limit of this reporting rather than evidence that none exist. Nothing above predicts whether an intelligence explosion occurs, places odds on it, or says whether any recommendation should be adopted.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
