A world model for physics.
After two decades of rapid digital innovation, the next revolution will be in physical technologies. Computers can now reason, but to power this revolution, they must also understand and reliably simulate the physical world.
How systems are modelled today.
Hand-derived models
Experts derive equations from physics. Precise and interpretable, but slow to build and hard to scale.
Fitted models
Fit data to known equations. Fast and explainable, but limited to the dynamics those equations can express.
Learned black boxes
Models learn behaviour from data. They handle complex systems, but are hard to inspect and unreliable beyond training data.
Systems as programs.
At the core of Gimle is a one-to-one correspondence between physical systems and typed programs. In this setting, discovering how a system works is equivalent to writing a program. That program isn't then just a representation of the system but it is the system.
Gimle is trained on this programming language and has learned the relationship between observed behaviors of physical systems and their programmatic form.
Given observed behaviour, Gimle finds the equations that produced it. A system that you can read, verify, run, compose and prove things about, as opposed to a black-box prediction you have to trust.
As such, Gimle is a coding model for physics.
When this matters.
Explicit models become particularly important when they must be certified, composed from subsystems, learned from sparse data or used to extrapolate beyond their training conditions.
- Control & autonomous systems
- Mechanical & fluid systems
- Energy & grids
- Finance & risk
- Life sciences & pharma
- Climate & earth systems
A different kind of world model.
Where other models learn how the world behaves from text, pixels or sensors, Gimle learns the actual dynamics of the systems.
With scale, language models progressed from simple mimicry of human language to a deep understanding of human reasoning. The same should hold here: by learning through the language of physics, Gimle should at scale not just replicate dynamical systems but understand deeper connections in the physical world.
The current model is a fifth the size of GPT-1. Everything scale did for language models is still ahead of us.