Introducing Mirror Physics
New scientific methods.
Mirror got started last year around Leap Day. My co-founders and I looked out upon a spate of breakthroughs happening in AI for scientific simulation — in our ability to make models that predict how physical systems behave, as AlphaFold describes how chains of amino acids will fold into 3D-structured proteins — and understood that the world would look different in a short time, as the advances being glimpsed in academia found their way into wider deployment.
New capabilities for using compute to observe the world — to apply a GPU where, previously, an electron microscope was required — are emerging just as general computational reasoning systems are coming online. The confluence creates stark new possibilities for scientific discovery in chemistry and materials science.
At Mirror, we build scientific technology adapted to this unique moment. Here, I want to highlight three ideas that are guiding our efforts.
Leverage for scientists
Every scientist today knows the feeling of being up against more data than they can process. We think that this overwhelm will become a signpost of the past: the cost of comprehension is trending to zero. It’s now easier than it’s ever been to generate meaningful insight through fine-grained inspection of large volumes of data. This promises to ease the path to scientific intuition, letting researchers focus their efforts on larger elements of the scientific process.
The impact is leverage: a light touch will now move ever-heavier operations in research. This year we’ve seen compelling early demonstrations, especially in handling scientific literature. Complex observational data that provides high signal, but carries a high burden of analysis, is meanwhile an under-exploited resource whose value will soon be fully realized. In the sciences, experimental data streams are the primary example of this kind; within industry, manufacturing and process data lend us another prototype.
But all of the experimental data that we can generate won’t saturate the capacity for analysis afforded by compute-bound reasoning. What was big data before now looks small, and in this new regime, data collection will rate-limit discovery. Physical experimentation, especially, is too slow and far too costly — if we have any recourse, we must go faster.
World models for chemistry
First-principles simulation is the compute-bound complement to experiment. Set aside AI for scientific discovery, for the moment, & consider AI as scientific discovery.
In 1926, the physicist Erwin Schrödinger discovered the equation governing quantum systems. This was the birth of first-principles chemistry: his peer Paul Dirac famously observed that “the underlying physical laws [for] the whole of chemistry are thus completely known, and the difficulty is only that the exact application of these laws leads to equations much too complicated to be soluble." Today, after nine decades of development, tractable computational approaches regularly anticipate the outcomes of experiment by working up from bare quantum mechanics.
For all this progress, perhaps we’re only now getting started. The field is undergoing a transformation, set off by the discovery that pre-trained models of quantum mechanical behavior are able to generalize. Once trained, these models stand in for numerical quantum mechanics while providing many orders of magnitude of acceleration.
In 2020, AI systems learned from hundreds of thousands of experimental samples to capture the behavior of amino acid sequences; in 2025, AI systems are learning from hundreds of millions of detailed quantum mechanical calculations to predict the behavior of atoms across a much greater span of chemical physics, from reactive biochemistry to thermal transport.
These are world models for the realm of atoms and electrons. This year, in the first prototypes, we’ve witnessed the beginnings of a new scientific platform. Like a space telescope or a particle collider, these frontier-scale simulators are artifacts which extend our ability to observe the world around us, and which will form the locus of new discoveries.
Experimental alignment
Useful scientific simulators need to faithfully reflect reality. Useful scientific reasoning systems must have a robust understanding of practice and technological context. These are two cross-sections of an ideal quality we call experimental alignment.
Experimental alignment is the capacity of AI systems to make informative, trustworthy predictions about scientific outcomes in the real world. It is the ultimate determinant of the impact that predictive AI can have in the sciences.
At Mirror, we’re approaching this criterion in several ways. One direction, which we’ll talk about more in the future, is our effort around laboratory automation for experimental verification. Another route — less capital-intensive, but not less important — is in understanding how to use many different sources of empirical data to improve and complement computational methods.
Though we’re positive believers in the power of pure computation at Mirror, we’re still certain that experimental observation will be crucial over the long term for calibrating the predictions of world models and reasoning systems alike. It’s the era of experience: compounding self-improvement demands solid contact with reality.
—
2025.07.07