Evolution by AI

One of the add-ons for Stable Diffusion is a process called Variate. (Stable Diffusion is kind of late to the game; Midjourney had it from the start.) If you've generated an image that is close to what you want, Variate will create variations that are close to the original. I decided that this could be used to simulate evolution. I generated four 'species' of microbes and then used a random number generator to decide which ones would live after each generation. For each set of four descendants, there was a 40% chance that a single 'species' would survive; a 30% chance that two 'species' would survive, and a 30% chance that all four 'species' would go extinct. (Unfortunately, most of the 'species' that I liked best went extinct.) Then I used the Variate tool to create mutations leading to four new 'species' for each individual that survived.

The results are shown below.

I ran the experiment for 20 generations (G1 to G20). Each row represents a generation. Full-size images of each generation are provided at the end. Lineages that survive are connected by thick green lines; lineages that go extinct are connected by thin red lines.

It looks like the parameters I picked at the beginning worked out pretty well. The experiment never went to complete extinction, but it also didn't grow so large that I couldn't handle it. There were a few close calls: only one 'species' survived G10 (which could have ended things early) and there were six 'species' that survived G17 (which could have ballooned out of control). As it is, it resulted in a total of 264 'species', with only four still alive at the end of the 20th generation.

Now let's dig into this a little more. A groups of 'species' that share a common ancestor is called a clade. Clades within clades are called subclades. Below is a depiction of some notable clades and subclades that arose in this experiment. For each one I'll identify a feature that is unique to that clade (what scientists call a synapomorphy).

Let's start with the Khocho clade (green). The last common ancestor (LCA) of the Khocho clade appeared in G3. This clade gets its name from the fact that these 'species' resemble beetles. This was not a particularly successful innovation since the clade went extinct by G6.

Next is the Mdzivebi clade (purple). The LCA of the Mdzivebi clade appeared in G4. This clade gets its name from the fact that these 'species' became constricted in the middle, which gives them the appearance of beads. This innovation was even less successful than that of the Khocho clade since it arose later and went extinct at the same time, in G6.

The Pariani clade (red) was a little bit more successful. The LCA of the Pariani clade appeared in G2. This clade gets its name from the fact that these 'species' resemble shields. This clade survived until G9, which was more successful than the Khocho or Mdzivebi clades.

That brings us to the Natura clade (blue). The LCA of the Natura clade appeared in G4, alongside the LCA of the Mdzivebi clade. This clade gets its name from the fact that these 'species' resemble lampshades. It quickly rose to dominance by G10 and then experienced a catastrophic bottleneck (G10–G11) where all but one 'species' went extinct. There are several possibilities for why this happened.

• First, naturans may have preyed on the Pariani clade and hunted them to extinction in G9 and most naturans failed to adapt to a new food source (such as other naturans).

• Second, members of the Natura clade may have been dependent on parianis as symbionts and the naturans failed to adapt when the Pariani clade went extinct, possibly due to environmental changes.

• Third, one of the naturans may have drastically changed the environment in a way that was deadly to most other naturans. This is believed to have happened several times on Earth (i.e., the Medea Hypothesis).

• Fourth, an abiotic environmental change may have occurred that was deadly for most naturans. Examples of this include massive volcanic eruptions (called flood basalts), ice ages (which are often caused by plate tectonic activity), asteroid impacts, or radiation from a nearby supernova.

Okay, it's really because the random number generator randomly selected most of the naturans for extinction. But if this were real-world data, any of those explanations could be true.

After the bottleneck, the Natura clade began to diversify again. This led to the evolution of several naturan subclades. The first to arise was the Sakoneli subclade (cerulean). The LCA of the Sakoneli subclade appeared in G11, right after the genetic bottleneck. This subclade gets its name from the fact that these species look like they have four legs (like tetrapod animals). This clade lasted for six generations, finally going extinct in G16.

Next is the Titebi subclade (turquoise). The LCA of the Titebi subclade appeared in G13. This subclade gets its name from the fact that all of its 'species' appear to have fingers. This subclade made it to the end of the experiment, with two 'species' still extant.

Finally there is the Changi subclade (teal). The LCA of the Changi subclade appeared in G13, alongside the Titebi subclade. This subclade gets its name from the fact that all of its 'species' have claws. This subclade also made it to the end of the experiment, with two 'species' still extant.

There's one more point of interest about this experiment: a trait that evolved multiple times. I observed that there were several 'species' (indicated with asterisks) that evolved a filament across the image. All of these 'species' are naturans. If you look back at the LCA of the Natura clade, you'll notice that that image has a diffuse glow behind the microbe. That glow could be considered a precursor mutation or potentiating mutation that made the later evolution of the filament possible.

Now all that's left is to give you the full images for all these 'species'. Here, they are, organized by generation.