Biobot Tierra

The result

Inheritance plus selection gives a fast adaptive shift, which then holds. When offspring copy a parent’s behavior, the population’s speed jumps within a few generations, as the faster founders out-reproduce the rest, and stays there. Nothing goes on building on it: not with larger populations, not at any mutation size we tried, and not in a pre-registered test on fresh dishes. The rise is real adaptation, but it is one round of sorting the founders’ variation, not evolution that accumulates.

That answer comes from two studies. The first swept all four channels at three mutation sizes in the dish shown above (the original sweep, below). Its populations were tiny, so a follow-up found out why, fixed it in a research version of the dish, and then tested the imprint result properly.

A dish that holds a population

In the original sweep only 2 to 6 bots were alive on average, falling to about 2 by the end, in every channel. Logging why each pile failed to hatch found the cause: a dead bot’s cells scatter into a compact pile of their own, which hatches again as a bigger bot, so mass collected in ever fewer, larger bots. Capping each hatch at the standard mass (the rest of the pile stays loose), with more loose cells and longer lives, gives a dish that holds about 39 bots steadily. The rerun, the assays and the phase diagram below ran in that dish; the live dishes above keep the original settings.

The early shift is selection

The Price equation splits a change in a trait exactly into selection (parents that leave more offspring differ) and transmission (offspring differ from their parents). Under imprinting, the first generation’s rise in speed is all selection: fast bots work more piles, so they are copied more. After that the change is near zero. This split re-reads the original sweep’s logs.

Loading the follow-up results…

Who counts as the parent matters. A newborn under imprinting copies the last bot to touch its pile, which is usually not the bot that built it. Copying the builder instead gives a stronger, faster response, and that rule (“builder” below) became the main test case.

A pre-registered test: it shifts, then holds

The tests, windows and pass marks were written down before any dish ran: 16 fresh seeds, each channel paired with the control on the same seed, and a further 8 seeds as replication. Speed rises for about three tenths of the run and then holds, about 75 above the founders under the builder rule. Neither test of continuing change passes. Speed moved about 86 more than the color marker, which does nothing, so the shift isn’t drift.

Loading the follow-up results…

The late bots are better, and the reason is speed

A competition assay tests adaptation directly: late-run bots and founder-like bots start together in a fresh dish. Under the builder rule the late bots leave about 60% of the offspring, in two separate runs. Founder-like bots given the late bots’ speed do just as well, so speed is the whole advantage. Under the original imprint rule the late bots have no measurable edge. The advantage is competitive, not productive: a dish of late bots alone makes no more offspring than a dish of founder-like bots. Faster bots win a larger share of a fixed supply of piles.

Loading the follow-up results…

The phase diagram

The same question across a range of settings: copying fidelity (mutation size), population (the supply of loose cells) and selection (a cost of speed, which shortens fast bots’ lives). Each cell is scored separately on the initial shift, on change after it, and on the assay. Click or hover a cell for its details.

Loading the follow-up results…

The map is mostly one phase: a fast shift that holds, wherever a population persists. Its edges:

• Too few cells: at 30,000 cells nothing persists, not even the control (about 2 bots). From about 13 bots up, the shift is about the same size, so population size isn’t a brake in this range.
• A strong cost: a cost of 250 changes nothing, 500 halves the shift and 750 reverses it.
• Sloppy copying: mutation of 128 halves the shift. With no mutation at all, the shift is as large as with mutation, and one founder’s speed takes over every dish: the response is sorting, and new mutations add nothing to it.

Change after the shift passed in 1 cell of 40, about what chance gives. The assay’s share is above 0.5 in nearly every cell without a strong cost, but 8 replicates detect it in only 10. The rule of thumb that selection beats drift when population × selection × heritability is above about 1 predicts the cost boundary well, but not the mutation one.

What it means

Heredity and selection are both present, and they produce adaptation, but only once: the founders’ variation is sorted and the population settles. Building on that would need something this arena doesn’t offer, such as traits that keep paying off as they improve, or an environment that keeps changing. This is a map of the model, not of real xenobots: none of its settings has a measured real value, and not seeing accumulation in a finite run doesn’t show it can’t happen.

The original sweep

Each channel at three mutation sizes in 16 dishes for 200,000 steps (about 2.8 simulated hours), against 32 control dishes where nothing is inherited, in the dish shown above.

• Imprint shifts speed early: bots born late were 53 ± 15 faster than their dish’s founders at mutation 16 (on a 0–255 scale), most of it in the first generation, then flat.
• Mold copies shape almost exactly (heritability above 0.8), but in this arena shape barely affects how many piles a bot builds, so copied shape only adds drift and harmful mutations: mold dishes do worse than the control.
• Carry-over passes on the average of the pile’s cells faithfully (offspring on pile mean: a slope of 0.9–1.0 at mutation 4). What it removes is the variation between offspring, since piles are mixtures that differ little: the problem Fleeming Jenkin raised against Darwin in 1867. The dish’s starting cells carry a default genome, so carry-over offspring are also pulled toward it.
• Every channel with a given seed starts from the same founders, so the nine conditions aren’t independent tests; the follow-up’s comparisons are paired by seed for that reason.

Loading the sweep results…

Conclusion

We asked which way of inheriting, if any, would let replicating biobots evolve. In this model, inheritance that copies a useful trait does produce evolution, but only a little of it. When offspring copy a parent’s behavior, the faster founders out-reproduce the rest and the population speeds up within a few generations, then stays put: the variation the founders started with is used up, and nothing in the dish rewards going further. Copying shape, which barely matters here, only adds drift; blending traits from many cells averages away the differences selection needs. Larger populations, other copying rules, a cost to speed and a pre-registered rerun all gave the same picture: one round of sorting, not a build-up of adaptations. Lasting evolution would need more than heredity and selection, such as traits that keep paying off as they improve or a world that keeps changing. This is a thought experiment about a model, not a forecast for real xenobots, but it suggests that giving them inheritance would not, by itself, be enough to set off open-ended evolution.

The science

In 2021, xenobots made from frog cells were found to replicate kinematically: they swept loose stem cells into piles that developed into new xenobots. Spherical parents usually managed one round; an AI-designed C-shaped body, like Pac-Man, managed four. The offspring always round up into spheres, so nothing about the parent passes on. Without inheritance there is no natural selection: a built-in safety brake.

This page is a thought experiment about removing that brake. Bots sweep a lattice of loose cells; a compact pile that sits untouched hatches a new bot, and old bots fall apart back into cells. Each bot has five genes: mouth, size, speed, turning bias and a color marker that does nothing except make lineages visible. Each dish uses one heredity channel: none (the lab), mold (the pile’s builder passes on its shape), imprint (the bot that last worked the pile passes on its behavior) or carry-over (cells remember their source bot, and the offspring averages them). Evolution is judged the standard way: heritability, response to selection, and Bedau–Packard evolutionary activity against the no-heredity control.

One more finding: in phase 1’s particle arena, Pac-Man bots out-replicated spheres, as in the lab. In this coarser lattice arena they don’t: a search of 16 arena settings found no robust advantage, so size and speed matter here, not the mouth.

Ethics: inheritance is what would let replicating biobots evolve, and its absence is a reason they’re considered safe. This page models what removing it would do, as a thought experiment, not a recipe. Any lab follow-up would need review first.