AI, Work and the Economy

Latest/Story

AI in science

Eight items, from 25 Sep 2026 to 18 Nov 2026, oldest first.

See also: Point 1 of our view: AI does more office work, cheaper · Point 2 of our view: Fewer office jobs, slowly at first

AI and technologyNew data

Anthropic (25 Sep): Claude computed a nine-loop physics amplitude unaided for $1,000-2,000; ~$100 via bootstrap

Anthropic reports that Claude completed a nine-loop scattering-amplitude calculation in N=4 super Yang-Mills theory two independent ways after a challenge set on 7 Aug, at $1,000-2,000 in total against roughly 96 CPUs for a week; a Chinese Academy of Sciences team reproduced it with GPT-6 within two weeks.

Why it matters. An AI system carried out a frontier physics calculation on its own for about the cost of a laptop, showing expert research work becoming cheap to automate.

Next. Independent checks of the result by physicists in the coming weeks.

Sources: anthropic.com

AI and technologyNew dataUpdate

Anthropic (8 Oct): Claude agents build the first complete ultraviolet sky map in days; traditional route 'weeks'

Anthropic's science team reports that Claude coordinated agents to merge GALEX, Swift and other surveys, correct an artifact across 38,000 observations and predict the unobserved third of the sky to within about 10% in blind tests. The astrophysicist's role was 'limited to guiding them'.

Why it matters. An AI system did most of the work of building a complete ultraviolet map of the sky, a task its author says would have taken weeks of painstaking human effort. If such multi-day research work becomes routine, skilled analytical jobs could shift from doing the work to supervising it, with fewer people needed per project.

Next. Measured evidence on how long and costly a task AI agents can complete unaided, from METR or Anthropic's Economic Index.

Sources: anthropic.com · Point 1 of our view

AI and technologyNew dataUpdate

Anthropic (16 Oct): Claude designs quantum circuits with up to 17x fewer gates than the best published, in a 53-hour run

In eight circuit-design tasks Claude, working in a closed loop with an automated checker, mostly beat the best published circuits; one elliptic-curve design uses 17x fewer gates, a decoder about 12x fewer than Google's. Anthropic says gains that took research groups about eight years came from a single 53-hour run.

Why it matters. An AI model, given plain specifications and an automatic checker, designed quantum circuits that improve on years of published human work, in some cases by an order of magnitude. If such results generalise, specialised research and engineering might be among the skilled work AI can carry out largely unaided.

Next. Outside researchers' checks of the published designs over the coming weeks.

Sources: anthropic.com · Point 1 of our view

AI and technologyNew dataUpdate

Anthropic (20 Oct): Claude designs scorpion-toxin binders largely on its own; 17 of 211 tested designs bind, 11 of 60 single-toxin

An internal Anthropic model ran a protein-design campaign against five scorpion toxins over 64 hours, about 24 of them autonomous, generating 318,484 designs on ~10,000 GPU hours; lab tests found 17 binders among 211 designs, 18% of single-toxin designs, two with single-digit nanomolar affinity. The hit rate matches typical human-led campaigns; none of the 91 multi-toxin designs worked and no neutralisation was tested.

Why it matters. An AI system ran most of a protein-design project itself and reached the success rate of a typical human-led campaign, in under three days. If such results generalise, part of the work of research scientists might be done by machines at a fraction of the time, though this one stopped short of producing a working antivenom.

Next. Tests of whether the designs neutralise the toxins in cells or animals.

Sources: anthropic.com · Point 1 of our view

AI and technologyNew dataUpdate

Anthropic (27 Oct): Claude cuts a drug maker's filing review time 89%; drafts a regulatory document in 4 hours vs ~3 months

Anthropic reported work with six drug developers. At Enveda, Claude found all 11 issues human reviewers caught in 1,061 pages of manufacturing documentation plus seven more, cutting human review from over 25 hours to under three, and drafted a regulatory document in four hours against about three months by hand; Divergence Labs generated about 35,000 antibody designs in a month, three of which beat the standard medication in a lab test.

Why it matters. Anthropic reports that at one drug developer its model checked over a thousand pages of manufacturing records with about a tenth of the human review time, and drafted a regulatory filing in four hours that takes about three months by hand. These are the company's own figures from early-stage work; if they hold more widely, much of the documentation that occupies trained scientists might come to be done by machines rather than with their help.

Next. Whether regulators accept filings drafted this way, and whether the partner firms change their staffing, will show how far this reaches.

Sources: anthropic.com · Point 1 of our view

AI and technologyNew dataUpdate

Anthropic (4 Nov): Claude does most steps of a new protein-structure study in under four months; cryo-EM model in 34 hours

Anthropic's biology team used Claude to find, express and solve the first structure of a vault-like particle from bacteria: it searched about 15 billion predicted proteins, designed DNA for 44 candidates, of which seven formed particles and four were confirmed, and ran the cryo-EM pipeline to a first model in 34 hours. Humans set direction and corrected errors, one of which cost two weeks; Anthropic says most of the work could be done by a typical lab with existing products.

Why it matters. Anthropic reports that its model carried out most of the steps of a new protein-structure study, from searching 15 billion proteins to processing microscope data, under human direction and with errors humans had to correct. If such work becomes routine, research that took a team of specialists many months might need far fewer of them, in science as in office work.

Next. Whether other laboratories report similar AI-run studies, and Anthropic's next data on how its models are used in science.

Sources: anthropic.com · Point 1 of our view

AI and technologyNew dataUpdate

Anthropic (5 Nov): AI could take 27-52% of researchers' time at today's capability; 40% of scientists expect solo AI projects by 2030

Anthropic analysed 207,049 research sessions (Aug-Sep 2026) and surveyed 528 scientists: data analysis and writing dominate use, and use is about twice expected on the tasks AI does best. At current capability, widely adopted, AI could take on 27-52% of research time (median 38%), output 1.35-2.02x; physical lab work (about 38-41% of time) caps lab sciences at 3-4x even with full automation of the rest.

Why it matters. A large sample of scientists' AI use suggests researchers lean on it where it performs best, and that current tools could already take over a third or more of research time if widely used. On this reading, skilled analytical work might be reorganised before the physical parts of science, which remain a limit.

Sources: anthropic.com · Point 2 of our view

AI and technologyNew dataUpdate

Anthropic (17 Nov): Claude Fable 5.2 agents solve two open black-hole physics problems in ~6 weeks, Lean-verified

An Anthropic researcher had Claude list open problems and ran background agents on 22 of them from August; they completed a Hawking-Page calculation set up in 2003 and derived an optimal bound on the lightest black hole's mass in 3D quantum gravity. Results were formally verified in Lean, checked by a physicist, and Claude wrote the paper with minimal guidance.

Why it matters. Anthropic reports that its model, working largely on its own for about six weeks, settled two open questions in theoretical physics, with the results checked by a physicist and by formal proof. If such work generalises, AI could take on a share of research itself, not only of routine office tasks.

Next. Independent physicists' review of the two papers.

Sources: anthropic.com · Point 1 of our view