AI, Work and the Economy

Latest/Story

AI automating AI research

Ten items, from 9 Oct 2026 to 20 Nov 2026, oldest first.

See also: Point 1 of our view: AI does more office work, cheaper

AI and technologyNew data

Epoch AI (9 Oct): asked to invent a new training method, GPT-5.6 Sol recovers ~15-35% of the target's gains; Fable 5 none

Epoch's InnovationEval gave two frontier models 3,000 GPU-hours each, offline, to discover a post-training technique rivalling a recent method (SDPO) they had not seen. GPT-5.6 Sol reached about 35% of its gains on a generous reading and about 15% after adjustment; Fable 5 showed no measurable gain. Both reran near-identical experiments and overstated novelty; Epoch concludes AI is 'far from automating AI R&D', though models a year older would have done markedly worse.

Why it matters. An independent test found that today's leading AI models cannot yet independently invent a competitive new method for training AI, and that they overstated what they had achieved. Progress over the past year has been fast, so this may be a temporary limit, but for now it counts against the idea that AI is close to speeding up its own development.

Next. Epoch's planned rerun of the test with newer models.

Sources: epochai.substack.com · Point 1 of our view

AI and technologyNew dataUpdate

Import AI (12 Oct): TasteVal finds AI 'research taste' doubling every 3.0 months since Dec 2025; best model 2.3x experts

A new benchmark from P-Zero Research, reported in Jack Clark's Import AI on 12 Oct, has models design and interpret experiments on eight frontier-AI research tasks. It finds experimental research taste doubling every 3.0 months since December 2025 (95% interval 1.7-5.0); the best model, Opus 5.5, gets a 2.30x compute multiplier over a human expert baseline (interval 1.15-4.37x) at about 1/30 of the experts' cost per attempt. The author now expects full recursive self-improvement in 2027 rather than 2028.

Why it matters. A new test finds that AI models' ability to choose and judge research experiments in AI itself has been doubling about every three months, with the best model beating a human expert baseline at a small fraction of the cost. It rests on eight tasks and wide error bands, and an independent test last week found the opposite on a related skill; if it holds, AI could begin to speed up its own development, and with it the pace at which it takes on office work, within a year or two.

Next. Independent replications, and Epoch's planned rerun of its own test with newer models.

Sources: importai.substack.com · Point 1 of our view

AI and technologyNew dataUpdate

Anthropic (13 Oct): merged code changes rose from ~2,500 a day in May 2026 to over 30,000 in September

Anthropic's engineering team reports a twelvefold rise in code changes merged per day in its main code repository over four months, with Claude writing most of its software; a seven-person team built a new version-control system for always-on agents and migrated off git in five months.

Why it matters. Anthropic says the code changes it merges each day rose about twelvefold between May and September 2026, largely written by its own AI. If output per software engineer rises this fast elsewhere, it is possible to conceive of software firms needing far fewer people to build far more, which would show first in hiring for software and other office work.

Next. Monthly US payrolls for information and professional services on 6 November.

Sources: anthropic.com · Point 1 of our view

AI and technologyReportUpdate

Anthropic's Jack Clark at the IMF (17 Oct): Humanity's Last Exam best score 67.7%, FrontierMath 94%

In the Per Jacobsson Lecture at the IMF meetings in Bangkok on 17 October, published in Import AI 477 on 19 October, Anthropic co-founder Jack Clark said the best score on Humanity's Last Exam has risen from about 8% in January 2025 to 67.7% and on FrontierMath from about 2% to 94%; a model scored 49 of 50 on a new set of unpublished theoretical-physics problems; and Anthropic's modelling has US GDP reaching about $44 trillion by 2030 if AI does a broad range of tasks autonomously and spreads fast.

Why it matters. A senior figure at one of the leading AI companies told the IMF that AI systems now score 94% on a hard mathematics benchmark where they scored about 2% two years ago, and 67.7% on a broad expert exam where they scored about 8%. If research itself is being automated at this pace, the economic effects could arrive faster than most forecasts allow, and gains from such growth might go mainly to the owners of capital rather than to wages.

Next. Independent benchmark leaderboards and Anthropic's planned November share listing may confirm or qualify these figures.

Sources: importai.substack.com, news.google.com · Point 1 of our view

AI and technologyNew dataUpdate

Anthropic (22 Oct): Claude proves no-crash properties for 239 entry points in 123 open-source libraries in a few weeks

Anthropic's Frontier Red Team, with Theorem, reports Claude wrote about 180 million lines of machine-checked proofs for production libraries from Google's OSS-Fuzz, finding hundreds of bugs; a reimplemented Python interpreter matched CPython on 99.95% of 1.7 million random programs. It compares this with the seL4 kernel's roughly 20 person-years of verification; token use, hundreds of millions per proof, is the stated limit.

Why it matters. Anthropic reports that its model produced formal proofs of correctness for over a hundred widely used software libraries in a few weeks, work that has historically taken experts years per project. If the cost keeps falling, it is possible to conceive of highly specialised verification work becoming largely automated.

Next. Further adoption reports from open-source projects and the cost per proof with newer models.

Sources: anthropic.com · Point 1 of our view

AI and technologyNew dataUpdate

Anthropic (27 Oct): unreleased model finds 62 critical or high Chrome bugs and a zero-click Linux root exploit

Anthropic's Frontier Red Team used an unreleased internal model to build five attack chains: a one-click Chrome escape (Google assigned 62 CVEs, 13 of them sandbox escapes), a Safari same-origin bypass, an Android lock-screen bypass on a Pixel 9, a remote crash of Windows servers, and a zero-click Linux root exploit that ran in about 53 minutes. All were fixed; the post gives no cost or human comparison.

Why it matters. An AI model not yet released found dozens of serious flaws in the major web browsers and operating systems, and built working attacks from them, a task until now done by small numbers of specialist researchers. It might mark a point at which some of the most skilled technical work can be done largely by machines, with large consequences for both the security of software and the demand for those specialists.

Next. Whether the model is released, and the next monthly security updates from Google, Apple and Microsoft.

Sources: anthropic.com · Point 1 of our view

AI and technologyNew dataUpdate

OpenAI (4 Nov): one autonomous run finds 948 Linux root exploits for $1.21m; GPT-5.6 Sol found 166 in July

OpenAI reports that one run of an autonomous pipeline with an unreleased cyber model analysed 2,308 candidate Linux kernel findings, reproduced 1,724 and produced 948 working local privilege-escalation exploits with candidate patches, plus 291 information-leak proofs, for $1,213,430 in API credits. OpenAI values the exploits at $18.9-66.3m in bounty terms, 16 to 55 times the run's cost; in July GPT-5.6 Sol found and exploited 166.

Why it matters. In a single automated run costing about $1.2 million, an unreleased OpenAI model produced 948 working exploits in the Linux kernel, nearly six times what its predecessor found in July. If the count holds once maintainers review the patches, work that took specialist security teams years might now be done by machines in days at a small fraction of its market price.

Next. Whether Linux maintainers accept the proposed patches, and when OpenAI releases the model and at what price.

Sources: openai.com, x.com · Point 1 of our view

AI and technologyNew dataUpdate

METR (5 Nov): Anthropic's model progress ~2x its 2024-25 pace (1.9x, range 1.5-2.8); ~80% put down to AI

METR's independent assessment, run on Anthropic's systems from 12 Aug to 29 Oct, finds benchmark progress 1.9 times its pre-October-2025 rate and puts about 80% of the speed-up down to AI rather than more training compute; staff report about 12x self-assessed uplift. METR says Anthropic redacted most of the evidence it found most informative, and that acceleration across all frontier developers is about 1.4x.

Why it matters. An outside evaluator finds that one AI developer is improving its models about twice as fast as before, and attributes most of the speed-up to its own AI doing research work. If this holds and spreads across developers, the pace at which machines can take on skilled work might quicken beyond current forecasts; the evidence is partly withheld, so it is uncertain.

Next. METR's measurements of the next frontier models' task horizons.

Sources: metr.org, metr.org · Point 1 of our view

AI and technologyReport

Google DeepMind's AlphaProof Code (10 Nov) formally verifies Signal's protocol code in ~4 weeks, 6m+ lines of Lean

DeepMind's agents, running on Gemini Carbon, wrote over six million lines of machine-checked proofs that Signal's deployed implementation keeps its security properties, flagging about 1% for human review; DeepMind says such work historically took expert teams years. Two reintroduced past bugs were proved insecure.

Why it matters. AI agents completed in about four weeks a formal proof of the security of widely used messaging software, work that DeepMind says took specialist teams years. If this generalises, a skilled and scarce kind of engineering might come to need far fewer people.

Next. Whether independent researchers confirm the proofs once they are published.

Sources: deepmind.google · Point 1 of our view

AI and technologyNew dataUpdate

Architect Labs (18 Nov): AI agents designed the Frostbite CPU from ISA to layout; 4.59x performance per watt claimed

Architect Labs says AI systems autonomously iterated over the compiler, instruction set, microarchitecture, RTL, verification and final 7 nm layout of a new CPU, with contracts checked in more than 800,000 lines of Lean. It claims 4.59 times the performance per watt (3.83 on held-out workloads) and 8.77 times the performance per area of one core of an unnamed in-order CPU; the RTL is not yet released.

Why it matters. A start-up reports that AI agents designed a working processor architecture end to end, from instruction set to physical layout, with its correctness checked by formal proof. If the claim holds up, the labour needed to design the machines that do AI work could be falling as well, which might speed how fast those machines get cheaper.

Next. The company says it will publish the chip's design files; independent checks of them would confirm or undercut the claim.

Sources: architectlabs.com, news.ycombinator.com · Point 1 of our view