Twitter: @korigrogers
LinkedIn: here
Email: kori[at]rogers[dot]live
October 24, 2024
Humanity’s final invention is a universe simulator. A so-called Turing Oracle.
Some people tend to call text-to-video models ‘World Models’. These diffusion-based transformer models appear to implicitely learn physics (without ontology) by gradient descent through massive amounts of video data. The most optimistic interpretation of the results we’ve seen from video models is to say: “machines can learn the physics of the universe simply by observing it”. In other words, this could be the early innings of a data-driven physics engine.
Perhaps the universe simulator is simply what we’ve been calling Artificial Intelligence all along, just at the asymptote of scaling. After all, don’t large language models simply simulate the next token. After extrapolating on multi-modality, mathematical capability, and video models acting like ‘physics simulators’, doesn’t a universe simulator seem like a natural extension? Perhaps I’m blurring the lines between ‘prediction’ and ‘simulation’, ultimately I think it comes down to whether the model has understood the fundamental laws of the thing it’s modelling.
I believe if the Scaling Laws continue to hold, then in x years, humanity’s greatest achievement will be to build as many sensors as possible at every possible scale of measurement, to build towards an infinite tape of data that is fed into an ultimate simulation engine which will predict the next state of the universe.
Imagine the cool things we could build along the way if we worked towards this! How fun would it be to restart Chile’s Project Cybersyn? We could actually build an economic simulator to forecast the impacts of various Chilean policy decisions and guide the country’s government. How far away are we from this? Perhaps not very far.
To build the universe simulator and solve the halting problem- this will be mankind’s final invention.