Harvard physicist on LLM impedance mismatch

Anthropic publishes Matthew Schwartz’s guest post: the model is strong at code and calculation, weak as a human teammate.

On October 1, 2026, Anthropic posted “Claude-shaped science” on its Science Blog, a guest piece by Matthew Schwartz on the impedance mismatch between what a scientist wants from a human colleague and what current large language models (LLMs, the systems behind those chatbots) actually do well. Schwartz is a physics professor at Harvard and a visiting researcher at the company, in for a stint rather than as a full-time employee. The same day, @AnthropicAI pointed to the post on X.

Labs already talk to the model as if it were another person at the bench. Schwartz says that habit is a bad fit for the work the model already handles well.

Quick recap

Anthropic published the guest post on October 1, 2026, and announced it on X.
Schwartz’s impedance mismatch is scientist versus model, not a call to redo the scientific method.
BootLoops is his open source harness. He maintains it. Anthropic funded it. He and the site say it is not an Anthropic product.
The public numbers sit in NBER Working Paper 35782 (September 2026), which has not gone through journal peer review. Other scientific claims in the blog stay the author’s account.

First: what impedance mismatch means

In physics, an impedance mismatch is what you get when two systems each work, yet they couple badly. The signal dies at the junction. Schwartz borrows that picture for the lab. A scientist wants a colleague who thinks like a person. Today’s LLM writes code and checks numbers across a range of subjects no one person can hold in their head all day. The analogy stops there. This is not a circuit.

In the Outlook of the post, he says he sees neither evidence nor a need to “revisit the scientific method.” The gap he describes is a human workflow against the shape of problem the model already handles well. He writes that Claude and GPT are good at science, but they are not scientists. Treating an LLM as a human collaborator is not, in his words and in the company’s post, “currently the best way to elicit their scientific strengths.”

What the post proposes

The essay picks up “Vibe physics: The AI grad student” (March 23, 2026). There Schwartz described Claude Opus 4.5 as a strong graduate student, only faster, and still a slog: correcting every sentence, pulling the model off useless threads. He used “20 times the speed” as his own analogy, without a published measurement behind it. This time he changes the question. Instead of making the model work like a scientist, he looks for problems already in the shape the agent is good at.

He describes BootLoops, an open source harness: scripts and checks that pin the model to a calculation you can verify, instead of leaving the chat loose. The harness does not think. It only narrows the path. The site is https://www.bootloops.ai/. The code is at BootLoops-ai/bootloops. The guest post and the site say BootLoops is not an Anthropic project. Schwartz owns and maintains it. Anthropic funded it.

Limits, cost, and conflict

The GitHub README includes “Copyright (c) 2026 Anthropic, PBC” and, in the same breath, that it is not an officially supported Anthropic product. This newsroom is not settling that legal tension. The conflict of interest is on the page: the author is a visiting researcher at the company publishing him.

The post does not publish a dollar or token price. RUMOR: beyond Schwartz calling the setup compute- and token-intensive, there is no cost figure. We did not run BootLoops. It is useful if you already have a calculation you can check. It is a poor substitute for a human colleague, and it is not a shortcut around peer review.

What already has a working paper

Outside the blog, the public document is NBER Working Paper 35782 (September 2026), by Matthew Schwartz, Isaiah Andrews, and Jesse M. Shapiro. The National Bureau of Economic Research is a U.S. institute that issues economics working papers, studies that have not yet gone through journal peer review. The authors describe an LLM workflow on 4,452 replication packages from five journals. A replication package is the bundle of files a journal keeps so someone else can redo the math. They report discrepancy flags in 3,460 articles or appendices; in 496 articles, more than a 10-fold cut in calculation time with similar or greater accuracy, they say; in 923, an extension aligned with the original paper’s goals. The NBER notes that Schwartz was an Anthropic contractor on the project and that the results and views are not endorsed by Anthropic. The guest post also mentions about 30,000 routines. That figure is in the blog, not in the NBER abstract.

What we still don’t know

RUMOR: most of the other results Schwartz lists in the October 1, 2026 post (new integrals, ecology, population genetics, 36 manuscripts in 18 fields with 19 coauthors over three months, out of some 400 candidate problems) appear only in the guest post, in the author’s voice. He is the one counting. This newsroom did not find an independent paper or peer review for those items. Schwartz himself marks several as “undergoing further exploration and verification.”

Harvard’s physics department page lists him as Professor of Physics, independent of Anthropic.

The post’s answer to that lab habit is no. Use the model where the problem is already a checkable calculation. Leave the rest of the scientific list on the blog until it shows up as a paper someone else can read.

Sources

Official

Affiliation and working paper

Quick recap

First: what impedance mismatch means

What the post proposes

Limits, cost, and conflict

What already has a working paper

What we still don’t know

Sources

Sources

Markers