July 31, 2025

From sand to artificial universes, and back

tl;dr: Reality is fundamentally more compressible and structured than previously assumed, as demonstrated by AI models intuitively capturing complex physical phenomena. Thus, building hyper-realistic artificial universes may require significantly less computation and energy than traditionally thought, lending plausibility to the simulation hypothesis.

One of the biggest philosophical implications of AI is that there seems to be a lot more latent structure to reality than we previously thought [1]. AI models have proven they can intuitively pick these up - be it in language, cognition, or physics.

The latest video generation models are a great example of this, where models don’t need to solve Navier-Stokes equations to predict the behaviour of fluids, but rather “just know” how water spills onto the floor after watching millions of YouTube videos.

Remarkably, all this intuitive understanding emerges from a computational substrate made essentially out of sand (silicon wafers arranged precisely). This observation strengthens the intuitive plausibility of the simulation hypothesis. Perhaps creating hyper-realistic universes like our own requires far less compute and energy than we previously assumed, given that reality appears much more compressible than traditional physics might suggest. Or, if significant compute and energy are still required, at least we’ve begun to grasp exactly how to create these artificial realities.

Think about it for a moment. We have taken sand and, within roughly 80 years since the invention of transistors, turned it into a thinking machine that can predict the next few seconds/tokens/frames of aspects of reality we previously thought would require insane amounts of compute to predict. And we’ve used only the tiniest subset of the data available in our universe to build these models. This should be an eyebrow-raising moment even for the most cynical.

We are glimpsing the early steps toward creating hyper-realistic artificial universes in silico; each potentially developing their own physics, evolutionary pathways, and life forms. Whether our own universe is itself artificial in the same sense remains to be seen.

So perhaps we can imagine a future, say in 2040, where you can prompt a ChatGPT-like system to generate an artificial universe and “run” it for billions of years, and perhaps you can put on your VR goggles and jump into any point of this universe at any time.

[1] Demis Hassabis elaborates on this early in his latest interview with Lex Fridman: YouTube

Previous post The Three Bricklayers of AI Over the years, several mental models have been used for predicting the impact of AI on jobs: Back in 2016, Andrew Ng’s 1-second rule stated that