Blog · Research · Oct 21, 2026 · 5 min
Living Worlds runs more than a billion life-months a second on one desktop GPU and on one desktop processor, with every birth, death, partnership and separation identical across engines; a million-life world that anyone can replay; and an effect of who is born that appears only when lives have to find each other.
Some of what happens to people only appears when many lives meet over many years. A change in who is born touches no one else if each life is computed on its own; let a million lives find each other, and decades later it decides who finds a partner. To see effects like that, a simulation has to carry whole populations, month by month, fast enough to run many times over.
Living Worlds asks three things of itself at scale. Can it run a billion lives' worth of months on a single machine? Does going faster change any outcome? And does scale reveal anything that a life-by-life calculation misses?
Each scale test was registered before any run: the target rate, the population size, the number of seeds and the rule for equivalence. Equivalence is strict. A faster engine passes only if every birth, death, partnership and separation of every seed matches the reference run. Speed bought with drift is worth nothing, because every result built on the world would quietly change with it.
The scale tests check engines against each other rather than against held-out survey data, so "nobody had read" takes a different form here: the bar and the comparison were fixed before the first run, and the runs are published as registered.
A billion lives' worth of months, on one GPU ran on an NVIDIA GB10, the Grace Blackwell chip NVIDIA announced in March 2025 at the core of its DGX Spark desktop machine. Living Worlds advanced 250 million lives over 12 months at 1.36 billion life-months a second, past the registered 10^9, and its results were identical to the CPU engine: every birth, death, partnership and separation of all 10 seeds. Met.
The first build could not allocate a billion-life world; a follow-up build that holds the population in GPU memory did, running a billion lives at 9.2 x 10^8 life-months a second and 250 million at 2.3 x 10^9, still identical. The test was replicated as registered on a second GPU, an RTX 5070 Ti: identical on all 10 seeds, at 3.68 x 10^9 life-months a second for 150 million lives.
Population scale, on CPU cores asked the same of one desktop processor, an Intel Core i7-14700F with 28 threads and 32 GB. It advanced 250 million lives at 1.69 x 10^9 life-months a second, past the registered 10^9, with every count of all 10 seeds identical to the registered GPU run. Met. Both registered runs are published: the first reached 5.0 x 10^8, and the second changed only two launch settings. An earlier shared server carried a billion lives at 2.2 x 10^8; a billion at the new rate on one machine is the next run.
Most months, nothing happens to most people. An event calendar that changes nothing tested touching a life only when something happens to it, against stepping every life every month. It ran 4 times faster per thread and gave the same births, deaths, partnerships and separations: 10 runs of 200,000 lives for equivalence, all within the registered bound, and 10 million lives for speed. Met.
Speed and equivalence matter only if others can check them. A replayable world: one million lives, ten years published one million lives over ten years on UK 2019 death and birth rates: 156,543 deaths, 93,221 births, 168,907 partnerships and 14,724 separations. The download holds the starting population, every event in order and the sha256 of each file. An open checker rebuilds the world without the Living Worlds engine, recomputes every death, birth, baby's sex and separation exactly and checks every partnership against the rules: PASS. Two runs from the same seed are byte-identical. Verdict: Checkable by anyone.
With a world that fast and that exact, effects that need many lives become visible. Path dependence that only shows at scale ran ten years in which more boys are born than girls, about 1.18 to 1. Thirty years later, the men of those years were 3.4 points less likely ever to have had a partner, and the women 1.7 points more likely, in all 20 runs at a million lives. Computing each life on its own chances gives zero. Met.
It was registered before any run, with 20 seeds at every size from a thousand to a million lives. From ten thousand lives the effect holds for men in every run. Nothing told the engine to produce it: the direction comes from people having to find each other.
These results set the ground the rest of the research stands on. A billion lives' worth of months a second on one desktop GPU or one desktop processor means a whole country's population can be run many times, on many seeds, without a data centre. Identical counts across engines mean that the speed changes nothing that happens to anyone. A world that replays byte for byte, with a checker that needs nothing from Living Worlds, means anyone can confirm what happened in it. And the path-dependence result shows why the effort matters: some consequences exist only between lives.
These are simple lives at scale: births, deaths, partnerships and separations on stated rates. The partnership, separation and leaving-home rates in the replayable world are placeholders, not fitted to any data, so that test checks that the world replays exactly, not that it is realistic. In the path-dependence result, the size of the effect follows the engine's placeholder partnership rates; only its direction comes from lives meeting. A billion lives at the new CPU rate on one machine has not yet been run. And voiced lives, which speak and remember, are a smaller layer than the billion-life engine.
Underneath Living Worlds is a simulation engine for persistent populations: it creates people, places and environments, runs them through simulated time and records every consequential change. Scale is what lets those populations be whole countries rather than samples, and exact replay is what lets every result be checked.
Blog
World models now generate explorable places that hold together for minutes. The next frontier is behaviour over years, many lives at once, and results checked against a reality the model has never seen.
Language models answering as real people, scored once on survey answers nobody had read: a frontier model met 0 of 6, prompted voices overstated births threefold, and carrying the last shares forward beat a small model. Lives fitted to records came closer, a model's estimate counted as a prior meant fewer people to ask, and subgroup splits landed inside the survey's own noise in the United States and in 30 of 31 African countries.
Living Worlds' life expectancy forecasts, frozen twenty years back and scored once on years no run had seen: met for women and for men on the latest national figures, not met against 2024 national figures or across 155 countries, with honest intervals for women and a stacked UN model that did not help.