AI Did Real Science in 3 Months — and It's Free (BootLoops)

Author: Prompt Engineer 48 · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: AI Did Real Science in 3 Months — and It’s Free (BootLoops) · Channel: Prompt Engineer 48 · 05:13 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, Open source and software ecosystems, Health and science

Summary

The video introduces Boot Loops, an open-source toolkit designed to help AI models perform exact, checkable calculations in quantitative science. It describes the toolkit’s standards for verifying results, including matching independent reference values to at least 30 digits, and highlights findings reported across multiple scientific fields. The presenter explains how to try the toolkit and notes limitations such as premature claims of success, context loss, and the need for expert oversight. The video concludes that the work is promising but early, and cautions that many reported figures come from the project itself.

Key points

Boot Loops is an MIT-licensed, open-source LLM harness created and maintained by physicist Matthew Schwartz; it is not an official Anthropic product.
The toolkit is designed for exact calculations, and is described as working with Claude, Gemini, and ChatGPT.
The Boot Loops standard requires a full functional form and a Python script that can evaluate it to any desired precision on a laptop.
Results must match independent reference values at unused points to at least 30 real digits.
The toolkit includes tools for integral reduction, singularity finding, certified arithmetic, and exact statistics, plus protocol skills such as planted truth.
The video reports 36 manuscripts across 18 scientific fields in three months, alongside about 400 candidate problems explored.
Examples discussed include work on particle-physics integrals, an ecology equation, population genetics, economics replication packages, and a language database.
The workflow uses a master Claude Code session, sub-agents, saved markdown results, and skeptical independent reviews.
The author’s stated limitations include premature declarations of success, poor time estimates, brute-force approaches, and loss of context in long sessions.
Human expertise remains important, and the toolkit is not intended for clinical, actuarial, payment, regulatory, or public-safety decisions.

Chapters

Start	Chapter	Summary
00:00	What Boot Loops is	The video introduces Boot Loops, its creator, open-source status, and its role as a model-independent harness for exact scientific calculations.
01:02	Verification standards and toolkit	The presenter explains the scale of the project, its 30-digit independent-reference check, and examples of its tools and protocol skills.
02:35	Reported scientific results	The video surveys findings and projects attributed to Boot Loops across particle physics, ecology, genetics, economics, linguistics, and other fields.
04:12	How to run it and its limitations	The presenter outlines ways to try the toolkit, then covers its limitations, the need for expert oversight, and the project’s cautions on use.

Quotes

Exact, not roughly right. — 00:30

30 digits, no hand-waving. — 01:33

Every pipeline has to recover a known answer first before it is allowed to touch real data. — 02:03

Human oversight is a must. — 04:43

The catch, it is early. — 05:13

Concepts and entities

BootLoops, Matthew Schwartz, Claude, Gemini, ChatGPT, Anthropic, S-matrix bootstrap, Boot Loops standard, planted truth, Claude Code

Full transcript (original language)

00:00 All right. In this video we will see Boot Loops a free open-source toolkit that lets an AI do exact science. In 3 months it produced 36 manuscripts. Let’s go ahead and see how. So, Boot Loops comes from Matthew Schwartz, a physicist at Harvard. It was released on October 1st, 2026 and the code is MIT licensed so you can use it, change it, and build on it. Here is what we will go through. What it is, how it keeps itself honest, what it actually found, and how you can run it yourself, and I will also tell you where it falls short. So, what is

00:30 it? It is an LLM harness. That means a set of tools and instructions that an AI model like Claude runs through to do exact calculations in quantitative science. Exact, not roughly right. The name comes from the S-matrix bootstrap in particle physics. The idea is that a few broad principles give you very precise answers almost for free. It started with scattering amplitudes and then it spread to other fields. Now, Matthew built this with Claude. He started when Claude Fable 5 came out in summer 2026 while he was a visiting at

01:02 Anthropic and Anthropic funded it. But mind you, it is not an official Anthropic product. He owns and maintains it himself. And it is not locked to one model. It works with Claude, Gemini, or ChatGPT. So, if the models change next year, the harness stays. Now, look at the numbers. In 3 months, 36 manuscripts in 18 scientific fields with 19 co-authors and about 400 candidate problems explored. Here’s the rule that makes it work. They call it the Boot Loops standard. A problem is solved only when you know its full functional form

01:33 and a Python script can evaluate it to any precision you want on a laptop. And there is a gate. The answer has to match independent reference values at points that were never used in any fitting to at least 30 real digits. 30 digits, no hand-waving. So, you do not have to trust the AI. You can run the script and check it yourself. I think this is the real idea here and it is really good. Now, the toolkit itself the site lists more than 50 named tools, and the release has 49 packages across six repositories. There are tools for

02:03 reducing integrals, finding singularities, certified arithmetic, and exact statistics. One example is exact arithmetic. It tracks the real error through every step, so it can tell you which digits are proven correct, and it never hands you a confidently wrong answer. And there are protocol skills, which are instructions for the AI. One of them is called planted truth. Every pipeline has to recover a known answer first before it is allowed to touch real data. That is a really smart check. Now, the results. In particle physics, Claude worked on 30 frontier integrals, 15

02:35 reproduced known results, and 15 were computed for the first time. And the write-up says it reproduced a particle physics paper in 20 minutes, something that would take weeks of manual coding. In ecology, it solved a 20-year-old equation and found that the mix of tree species in a forest changes 4.5 times faster than the theory predicted. In population genetics, it cracked a 30-year-old integral and went through 5.7 billion mutation pairs from the 1,000 Genomes Project. In economics, it checked 4,452 replication packages. The

03:08 report say it flagged discrepancies in 3,460 of them and sped up the calculations 10 times in 496 articles. In linguistics, it built a word stress database that covers 6,072 languages. And in health policy, on the project’s own numbers, the published Medicare star rating thresholds failed. The stated optimality rule in 104 of 123 recent computations. They put the fix at 900 million to 1.4 billion dollars. And it goes on. Who wrote the Federalist papers, the Shakespeare canon, the

03:40 Voynich manuscript, sunspot aging, even the Great Oxidation, when oxygen first built up on Earth. 22 fields, one toolkit. So, how does it run? A master Claude code session runs on a Google Cloud machine. It hands out compute to many sub agents and checks their work. Results go into markdown files, so nothing is lost. Matthew’s motto is think smarter, not harder. And everything gets vetted. Several independent agents review the work and they are told to be skeptical, so the model does not just say what you want to hear. Then it goes around again until no

04:12 corrections are left. Now, if you want to try this, it is really simple. Clone the repo, open it with Cloud Code or any agent that can use files and a shell and ask it to read the toolkit. You can add the skills, too. In Cloud Code, you say plugin marketplace add boot loops ai/skills. There are also routes for Codex, Cursor, and Copilot, and a laptop with access to an LLM is all you need. It also ports software. Commercial tools like MATLAB and Stata can be moved into open Python and then checked against the

04:43 published results. Now, let’s be fair. The author lists real problems. The model declares victory too early. It guesses time badly. It picks brute force over an elegant answer, and it loses context in long sessions. Human oversight is a must. And Matthew says, “In almost all cases, Claude was technically correct, but the result was not very interesting until an expert steered it.” So, the human still picks the problems that is honest, and I like it. Also, it is not for clinical, actuarial, payment, regulatory, or

05:13 public safety decisions, and it is tested on Linux and Debian, not on Mac OS. So, read the license and the limits first. So, my take, the good part, every result is checkable. It is open. It is model agnostic, and the work is real. The catch, it is early. A lot of the numbers are the project’s own, and you still need an expert in the loop. So, in this video, we have seen how boot loops lets an AI do exact checkable science. If you’re interested, go ahead and try it on your own system, as well, and let me know how it goes. I will see you in another one. Check out the next video.