It's the first time in the history of the world.
After mathematics, AI has invaded the natural sciences!
ClawsGO Science at BitMiracle AI Lab enables a Claude Opus 5.5 to command 136 GPT 6 Astra in 11 days to unravel a chemical conjecture made in 1974.
According to ClawsGO Science, 83 Lean documents have been released to the public.
This is the first year that AI has opened its mouth in mathematics.
Hundreds of mathematical manuscripts were published at a time, and decades of complicated connections were advanced.
A month ago, Xinhua wrote the following sentence:
Physics, chemistry and biology are going to get to this point sooner or later, and every discipline will have a day like this.
After math, it's the natural sciences.Just recently, BitMiracle AI Lab released a 23-page paper that allowed AI to unravel a chemical conjecture put forward half a century ago.It asks the question, which is a common phrase in a chemistry lesson: Reverse reactions ultimately achieve balance.
Does the reversible reaction in a sealed container, no matter how it starts, stop at the same equilibrium point when the sum is the same?
In 1974, the problem was formally posed by the chemist Fritz Horn. For the next half century, it has been on the list of problems in the theory of chemical reaction networks.
This time, the answer is yes.
It's been proven by AI.
This was done by a scientific intelligence body called ClawsGO Science.
Claude Opus 5.5 makes decisions, GPT6 Astra executes, and Lean checks the kernel every step of the way.
A tweet from BitMiracle AI Lab
One sentence in the textbookWhy not a half-century of proof?A and B generate C and C decompose into A and B. The reaction is ongoing and the concentration is changing.
What will happen at the end? Will it continue to oscillate? What kind of matter will be consumed by light? Will it start differently and end differently?
In 1972, Horn and Jackson found a weapon: a quantity similar to nickel. When the reaction takes place, it only decreases. The system is like a ball rolling down the slope of a hill, rolling to the bottom of the valley, which is equilibrium.
They once thought the problem was solved.
But there is a trap on the slope: the concentration of some substance, which may be pointing to zero. There, the slope is almost flattened and looks like the bottom of the valley, but is actually a false endpoint.
To eliminate this trap, chemists have become somewhat stiff-necked over the past half century.
Indicating intent: The gradient moves down the steep slopes, avoiding false ends on the border, and finally stopping at the bottom of the valley.
One Opus, commanding 136 AstraThis question was made in ClawsGO Science. The paper explains the division of labor.
Decision-making body Claude Opus 5.5Scientific research is tasty, responsible for determining the direction: which path to take, which steps to change, and finally hand over the entire proof to the machine for inspection.
GPT6 Astra×136 execution of the smart bodyStrong executive, responsible for implementing: write proof, write Lean code, compile, read errors, repair, and recompile.
Checking the Lean CoreEvery step of the way, the proof is given to it.
The turning point came on day 6.
Dozens of executing smart bodies repeatedly answer the same question: The path in the textbook is too much ground to fill, and the workload is too big.
The decision-making minds didn't harden them, they simply changed their path.
The next day, the step that stuck with the compilation was passed. This step later became the key step in the paper.
This is exactly what the researchers said about the subject, the argument that the machine checks and verifies. The part that people verify is less than 60 lines.
Strictly speaking, the paper proves that this speculation satisfies a kind of delicate equilibrium, taking the first step with the existing theories. From there to the end, the entire chain of proof has passed machine checks.
The division of work at ClawsGO Science: deciding the direction of the intelligent body, implementing the intelligent body, and checking the Lean kernel every step of the way.
AI attacks the natural sciencesWhat's missing is a cloud lab.For AI in mathematics, a system of verified proofs is enough.
The natural sciences are different.
Chemists run GROMACS and LAMMPS, calculating PySCF and CP2K. Biologists use BLAST to compare sequences to check for proteins. One WRF simulation by a meteorologist takes several days.
Each one has to be loaded, fitted, fed, computed. Many of the theme cards have time in the environment, and Bica is long in thought. ClawsGO Science gives each task a cloud workstation.
Lean and its mathematical library have been prepared, compiled, and executed. The intelligent body receives the task and writes directly.
On the same workstation, 533 scientific research software and tools and 91 built-in databases were deployed covering 61 first-class disciplines.
For 11 consecutive days, every command, every output, is kept in the mission log and can be retrieved at any time.
The software, tools and built-in databases preinstalled on cloud workstations, by subject area statistics.
Next, your subject.Mathematics came first, chemistry followed.
It can be used to create a simulation of a series of candidate structures.
If you make a living organism, you can make it have a sequential process of analyzing, running, and checking itself.
To make a drug, it can use RDKit to pump a round of candidate molecules.
If you want to do weather, you can give it a single WRF simulation, and it keeps going until it runs out.
In the past, it took several months for a document to be examined by several experts. Now, it can be accompanied by a machine inspection; the process can be accompanied by a complete record.
After mathematics, it's the entire natural sciences.
This study was completed during the ClawsGO Science inquiry.
Now the official version of ClawsGO Science is online.
clawsgo.ai
ClawsGO Science official release
About the BitMiracle AI LabBitMiracle AI Lab is an independent AI research lab that studies long-range scientific intelligence, verifiable AI reasoning and AI research infrastructure.
The lab is in the angel round of funding, while recruiting researchers and engineers.
This is the first time we've done this.