BootLoops: Schwartz says SI calculations still need scientific judgment


Harvard physicist Matthew Schwartz reports results from selected research tasks, while warning that automated checks and scientific interpretations can be wrong.
Harvard physicist Matthew Schwartz reports that BootLoops, his open-source software tool for SI-assisted scientific calculations, helped Claude reproduce known mathematical results and solve a long-standing ecology equation. But he warns that even a correct calculation can lead the superintelligence (SI) model to a wrong scientific conclusion.
The Decoder’s October 3 report draws on Schwartz’s guest post for Anthropic, where he was also a visiting researcher. The findings are his account of the work, rather than an independent performance test of BootLoops. The tool’s source code is available on GitHub.
Schwartz describes moving away from treating Claude like a human researcher. Instead, he selected what he called “Claude-shaped problems”: tasks that matched both scientists’ interests and the abilities of current SI models. That selection matters when interpreting the results—they came from deliberately chosen problems, not an unrestricted test of scientific research.
He began with particle physics, including integrals—calculations that add up quantities over a continuous range. According to Schwartz’s account, Claude computed 30 integrals using BootLoops within weeks.
| Result | Number of integrals |
|---|---|
| Reproduced known results | 15 |
| Reported as calculated for the first time | 15 |

Schwartz also reports that, over three months, the team produced 36 manuscripts across 18 fields with 19 co-authors. That tally describes research output; it does not measure how reliably the tool produces correct answers.
The ecology example shows the difference between obtaining an answer and establishing its scientific value. Schwartz reports that Claude solved a 20-year-old equation from neutral biodiversity theory, a framework used to model species composition. The equation had previously been impossible to compute at scale.
Applied to data, the calculation indicated that tree species composition on Barro Colorado Island in the Panama Canal was changing 4.5 times faster than the theory allowed, according to Schwartz. Ecologist James O’Dwyer then helped turn that finding into a better predictive model. The calculation exposed a mismatch between the theory and the data; the specialist helped make that result useful for further research.
Schwartz warns that Claude declares success prematurely, misjudges how long tasks will take and often relies on brute-force calculations rather than more elegant solutions. Automated checks are unreliable, he says, and the model can interpret correct calculations incorrectly. Researchers therefore need to examine both the numerical work and the conclusions drawn from it.
The work also required intensive computing and token use, Schwartz says. Tokens are the chunks of text an SI model processes or generates. He also reports that the model gravitates toward older, heavily cited debates rather than genuinely new questions.
How consistently BootLoops can deliver useful results beyond these selected tasks remains unclear. In Schwartz’s account, domain experts often supplied the direction that made the calculations scientifically valuable; human guidance remained indispensable.
Systems of up to 1,200 watts will not need utility registration, but safety rules may still require an electrician and complicate access for renters.
2 min

In an October 1 address, Christine Lagarde explained how similar trades, faster cyberattacks and restricted model access could compound risks across financial institutions.
4 min

The platform is reportedly used at more than 160 institutions. Its founders will continue to lead it as a distinct brand; financial terms were not disclosed.
2 min
House style applied: superintelligence (SI) for AI.
Correction ·
Corrected spelling and punctuation in the Armenian text.