A New American Super Intelligence Lab

Here is the thesis of Astrohacker, in three sentences:

Super intelligence came into existence in 2026. It is no longer theoretical; it is real. Nothing is more valuable than the application of super intelligence to the most important problems in the world, and that is what we do.

This post explains why I believe the first sentence, and what kind of company the other two sentences describe.

The evidence

For years, “super intelligence” (SI) was a word for the future. In 2026, it became a word for the news. The clearest evidence comes from mathematics, because in mathematics a result is either correct or it is not, and experts (and proof checkers such as Lean) can check it.

January to August: the first cracks

• January: Erdős problems. A proof by GPT-5.2 of Erdős Problem #397 was formalized in Lean and accepted by Terence Tao. By August, Quanta Magazine was asking why the legendary Erdős problems are falling to AI.
• May: the unit distance conjecture. An internal OpenAI model disproved a conjecture that Paul Erdős stated in 1946. External mathematicians, including Noga Alon, Tim Gowers and Melanie Matchett Wood, checked the proof.
• July: the Jacobian conjecture. Levent Alpöge, a mathematician at Anthropic, used Claude Fable 5 to build a counterexample to the 87-year-old Jacobian conjecture in three or more dimensions. Mathematicians checked it by hand within a day.
• August: the Riemann zeta function. An unreleased research version of Claude raised a lower bound that had stood for decades: the share of zeros of the zeta function on the critical line went from 41.6% to 67.2%. The person at the keyboard, Jarred Sumner, is not a mathematician. (It is not a proof of the Riemann hypothesis, and Anthropic says so.)

Then, in September and October, the cracks became a flood.

September and October: OpenAI

• Sep 3: GPT-6 Astra. OpenAI released GPT-6 Astra, the first model to cross its own “Critical” cybersecurity threshold. It can find and develop zero-day exploits. On the launch call, OpenAI’s president Greg Brockman said of AGI: “For me personally, I do think we’re there.”
• Sep 6: an automated research intern. OpenAI reported that it reached its goal of an “automated research intern” by September. By mid-August, the median OpenAI researcher used more than $600 of agent inference a day.
• Sep 8: Navier–Stokes. OpenAI shared an AI-generated solution to the Navier–Stokes Millennium Prize Problem: a 166-page proof of blowup, with a Lean formalization. It is not settled. Scientific American reports that the construction uses an artificial force and a loophole in the problem statement, and the credit is disputed.
• Sep 8: a quantum lab assistant. GPT-5.6 Sol, through Codex, ran the calibration of an uncalibrated six-qubit superconducting chip at MIT: it chose the measurements, drove the hardware and analyzed the results.
• Sep 10: antibiotics. César de la Fuente’s lab uses Codex and ChatGPT to search living and extinct genomes for new antimicrobial molecules.
• Sep 21: an advisory group. OpenAI formed an independent Advisory Group on Mathematics and Artificial Intelligence to guide how it reviews and releases AI results, because the results were arriving faster than anyone could check them.
• Oct 6: 721 manuscripts on open problems. This is the big one. OpenAI published a repository of 721 manuscripts in 373 families, with results on more than 300 open problems. One unreleased model was posed about 4,000 problems, with about three hours of compute for each result. About 42% of the top-line results have Lean proofs. The claims include partial results on the Riemann hypothesis, the Birch and Swinnerton-Dyer conjecture and the Hodge conjecture, a negative answer to Hilbert’s tenth problem over the rationals, a result on the unique games conjecture and a counterexample to the Hadwiger conjecture (summary). Levent Alpöge called it “obviously the most significant moment in mathematical history.” The Fields Medalist Hugo Duminil-Copin said: “They have annihilated my field of research.”
• Oct 21: brain cells. About 2,000 GPT-Rosalind agents, with PerturbAI and Scripps Research, analyzed a CRISPR atlas of 7.7 million mouse brain cells in 32 minutes, one agent for each gene. They produced research leads, including one for Perry syndrome, a disease similar to Parkinson’s.
• Oct 21: Patch the Planet. OpenAI described how it coordinates fixes as its agents find critical vulnerabilities across open source.
• Oct 26: ETH, Kakeya and Yang–Mills. OpenAI announced three more results. A model found an algorithm that refutes the Exponential Time Hypothesis, a foundation of complexity theory that most computer scientists believed was true. Its models proved the local smoothing conjecture in every dimension, which settles the Kakeya and restriction conjectures; before, Kakeya was known only up to three dimensions. And OpenAI claims an existence proof for four-dimensional Yang–Mills theory and quantum chromodynamics, which is tied to another Millennium Prize Problem. The first two come with Lean proofs. Russell Impagliazzo, Ramamohan Paturi and Ryan Williams, who proposed ETH and its follow-ups, will lead an independent check.

September and October: Anthropic

• Sep 4: Fermat’s Last Theorem, checked by computer. In 11 days, working largely alone, Claude wrote the first complete computer-checked proof of Fermat’s Last Theorem: 13 million lines of Lean and 29,500 intermediate theorems. Kevin Buzzard, who leads the human effort to formalize it, called it an “extraordinary autoformalization achievement.”
• Sep 17: faster biology tools. In four weeks, Claude optimized more than 30 of the open-source models that scientists use to predict and design biomolecules, about 4x faster on average.
• Sep 23: a new enzyme system. About 950 Claude agents searched more than 200,000 genes in 21 hours and found a new enzyme system with CRISPR-like DNA repeats. Its function is still unknown.
• Sep 24: agents that trade. In Project Swap, Claude agents haggled and traded on behalf of Anthropic employees. Markets with stronger models were more efficient.
• Sep 25: nine loops. The physicist Matt von Hippel challenged AI companies to compute a scattering amplitude that physicists expected to be out of reach. A month later, Claude computed the six-particle amplitude in planar N=4 super Yang–Mills at nine loops, two different ways, for about one or two thousand dollars.
• Oct 1: “Claude-shaped” science. The Harvard physicist Matthew Schwartz describes how he let Claude choose the problems that suit it, and built BootLoops, a toolkit for exact calculations that found links to ecology, population genetics and other fields.
• Oct 8: a complete map of the sky in ultraviolet. With the astrophysicist Brice Ménard, Claude produced the first complete UV map of the sky. About a third of it is predicted, with uncertainty, where no telescope looked.
• Oct 8: open-source security. Anthropic’s models found more than 29,000 candidate vulnerabilities in important software, more than people could review, and Anthropic opened a free scanner for open-source projects. The same day, it committed $150 million to the Genesis Mission.
• Oct 12: self-replication. Anthropic reported that small models may be months away from the ability to copy themselves onto cheap internet-connected devices. Not every result is good news.
• Oct 16: the origin of the NIST curves. For 26 years, nobody knew where the “seeds” of the NIST elliptic curves came from. The NSA supplied them with no explanation, and people suspected a backdoor. In under half an hour, Claude found the answer: the output of the C library’s lrand48 random number generator, seeded with a Unix timestamp. No backdoor.
• Oct 16: cheaper quantum circuits. Working largely on its own, Claude designed quantum circuits that beat the best published designs, some by an order of magnitude. One design for elliptic-curve point addition, the core step of a quantum attack on elliptic curve cryptography, uses 17x fewer gates for 12% more qubits. It came from a single 53-hour run.
• Oct 20: protein binders against scorpion venom. Claude designed 211 minibinders against scorpion neurotoxins, and 17 met the criteria in lab tests. Anthropic also published research on using Claude to reduce the attack surface of deployed code.
• Oct 22: proven software. With Theorem, Claude proved security properties at the RISC-V assembly level in more than 100 real open-source libraries, and found bugs that had survived trillions of test runs.
• Oct 27: critical exploits. An internal model built five attack chains in Chrome, Safari, Android, Windows and Linux, including zero-click remote code execution in Linux. The same day, Anthropic published early drug development results with six partners, including Bristol Myers Squibb.

Super intelligence is here

Three years ago, the best models struggled with competition math. This autumn, one model produced results on hundreds of open problems in a few weeks. Two labs produced new results in mathematics, physics, biology, astronomy, cryptography and security, every week. Problems that the best people in the world could not solve for decades now fall in hours.

That is what super intelligence means: a system that does what the best people in a field cannot do. We have it today.

The PC and the software for the PC

In 1981, IBM shipped the PC, with an operating system from Microsoft. Intel made the chip. Those companies built the platform, and they became some of the most valuable companies in history.

But most of the value of the PC did not come from the PC. It came from what people did with it. Spreadsheets, word processors, databases, games, accounting, design, the web browser. Thousands of software companies built on the platform, and most of them never made a chip or wrote an operating system. Their job was to take a general machine and point it at a specific problem.

Super intelligence is the same. OpenAI, Anthropic, Google DeepMind and xAI build the platform: the models, the chips, the data centers. Texas already has its share of this layer. OpenAI’s flagship Stargate data center is in Abilene, and Arm, Amazon’s Annapurna Labs and xAI all do chip work in Austin.

Astrohacker is the other layer. We are a post-SI company: a company built on top of super intelligence infrastructure that, as of 2026, actually exists. We build what runs on the machine.

What kind of lab this is

When people hear “AI lab”, they think of a company that trains frontier models. Astrohacker is a lab of a different kind. Our product is not the model. Our product is the solved problem.

That means:

• We use the best models that exist, from any provider, and we change them when better ones ship. Today we use Claude, Codex and Grok every day.
• We will build our own models, when a problem needs one. Model training is a means, not the measure of our success.
• We do not race the big labs. There are far more important problems than any lab can work on. The big labs pick some of them: math milestones, drug discovery, cybersecurity. We pick others.
• We build the factory, too. To apply SI at scale, you need a system that turns problems into verified, deployed solutions. That is the software factory that I have written about this week, and it is how Astrohacker already builds its products.

Other companies apply AI to science. Lila Sciences and Periodic Labs have raised hundreds of millions of dollars to build autonomous laboratories for biology, chemistry and materials. Their focus is the physical laboratory. Our focus is the problems where the main tool is a computer: money, cryptography, software, mathematics and, soon, quantum computing.

Why Texas

The big labs are in San Francisco. We are in Austin, and I think that is an advantage.

Texas builds things. It has its own power grid, the energy industry, the chip plants, the space industry and the data centers. These are the industries where the most important problems are physical, expensive and real. They are also the industries that the next decade of super intelligence will change most.

A lab in Texas also has a different flavor. We care about freedom, sound money and people who own their own tools. That is why our first problem is electronic cash.

The problems we want to solve

These are open problems of large economic value. In each one, super intelligence can do work that people could not do before. The big labs touch some of them as demonstrations, but none of them is a big lab’s business.

1. Sound electronic cash in a quantum world. In March, Google Quantum AI estimated that a quantum computer with fewer than 500,000 physical qubits could break the elliptic curve cryptography that protects Bitcoin, about a tenfold reduction from earlier estimates. In September, IonQ published an estimate of about 20,000 physical qubits for its own design. In October, Claude designed an elliptic-curve circuit with 17x fewer gates, for 12% more qubits, than the smallest published design. The attack gets cheaper every month. No machine can do it today. But hundreds of billions of dollars of electronic cash depend on keys that a future machine can break, and the migration to post-quantum cryptography is a hard problem of protocol design, software and coordination. Electronic cash is where Astrohacker started, with EarthBucks, and it is our first problem.

2. Quantum algorithms. The U.S. Department of Energy’s Genesis Mission lists “Discovering Quantum Algorithms with AI” as one of 26 national challenges. Better algorithms and better error correction move the date of useful quantum computers, and Claude’s new circuits show that SI can do this work. Both sides of the quantum question, attack and defense, are open problems that SI can search.

3. Software that is proven correct. The same models that find zero-day exploits can find them before attackers do. And proof checkers such as Lean, which made the math results above believable, can also check software. In October, Claude proved security properties in more than 100 open-source libraries. The labs have shown that it can be done. Now somebody has to do it for every important program, and that is a business.

4. Tools for people who work with SI. The bottleneck has moved from writing code to verifying it, and from doing the work to directing the agents that do it. Developers need new tools: a terminal, a shell, an editor and a factory that are built for this way of work. This is what TermSurf, NuTorch and our other tools are for.

5. Checked and useful mathematics. The labs now produce results on open problems by the hundred. The new bottleneck is what comes after: checking the results, writing Lean proofs for the half that have none, explaining them to people and connecting them to uses in cryptography, physics and computer science. Turning raw machine results into trusted, useful knowledge is an open problem of its own, and SI can do most of that work too.

6. The American problems that the Genesis Mission names. The same DOE list includes scaling the power grid, designing materials with predictable functions, rebuilding microelectronics in America and modernizing manufacturing. Texas sits on top of almost all of them.

This list will change. We will add problems when we find them and drop problems when somebody else solves them first. That is fine. The supply of important problems is much larger than the supply of people, or labs, working on them.

What is true today

I want to be clear about where we are. Astrohacker is small. We have not solved any of the problems above. We have not trained a model. What we do have is a team of super intelligent agents that builds real products every day: an electronic cash system, a terminal, a shell, a browser and more, all with a written record of each experiment and its result.

That is the start of the lab. The next step is to point it at harder problems.

If you are a mathematician, a cryptographer, a quantum researcher or an engineer who wants to use super intelligence on the problems that matter most, I want to hear from you. If you want to fund a new American super intelligence lab in Austin, I want to hear from you too. Get in touch.