Life

Here’s something pretty trippy to think about.

Go out into the woods somewhere and look around. Notice all the living things, all around you. Take in the smell of grass or trees, and the colorful leaves. Think about all the trillions of microorganisms in the ground, tens of thousands of moss piglets, and the thousands upon thousands of insects.

Now, think about how that all started with one single molecule that replicated itself. All of this life—those trees, the insects, me, you—all of those things began with replication.

That’s right: several billion years ago, a combination of molecules replicated itself, and the code for future replication was also passed along to the copy. That particular combination of molecules is the predecessor for all life on earth (as far as we know).

All of this life on planet Earth—all 550 billion tons of it—came about eons ago, on one fateful day. Why that particular day? Nobody knows, but it’s certain that the early Earth had trillions of little experiments running at any given moment.

Primordial pools of organic chemicals were catalyzed by random chance. Here, lightning might strike a substance. There, a particle from far away in space—a cosmic ray—reconfigured the structure of a molecule.

In this way, day after day, the primeval earth was like a billion mad scientists directing billions of experiments all the time. Is it really any wonder that one of these yielded a molecule that could replicate itself?

I bring all this “life” stuff up because there has been a great deal of speculation as to whether we might be on the cusp of creating life ourselves.

Artificial General Intelligence (AGI) isn’t the same as life, as

Alejandro Piad Morffis

makes clear in his piece, but it’s still regarded as a milestone of comparable importance. For most of my life, the consensus regarding when AGI would happen ranged from “not any time this century” to “probably never.”

That’s why charts like this one will always catch my eye:

Chart: Metaculus

2032 is a whole lot sooner than “probably never.” It’s probably no surprise that those dramatic dips have coincided with news from Open AI about ChatGPT. Much has been made of so-called “emergent properties” that seem to come out of nowhere.

Devansh

(thank you for the original impetus to write this piece) has written about why there may be more smoke than fire:

Are Emergent Abilities of Large Language Models a Mirage? [Breakdowns]

I’m sure that I don’t know how close we are to creating AGI. It’s hard to envision AI doing everything we do better than we do it today, but I concede that this question is way over my head.

Fortunately, I’m more than willing to limit my scope to question at hand, about what life actually is.

Let’s return to our self-replicating RNA molecule. What happened, really?

One way to think about it is that a new thing that nature created could copy itself. Prior to this thing, there were other things that could replicate themselves, too. Crystals in clay could already do this, in a manner of speaking.

Clay consists of a large number of small crystals, and clay is typically found in an environment that promotes crystal growth. Crystals consist of a regular lattice of atoms and are able to grow if they’re placed in a water solution containing the crystal components; automatically arranging atoms at the crystal boundary into the crystalline form.

Crystals may have irregularities where the regular atomic structure is broken, and when crystals grow, these irregularities may propagate, creating a form of self-replication of crystal irregularities. Because these irregularities may affect the probability of a crystal breaking apart to form new crystals, crystals with such irregularities could even be considered to undergo evolutionary development.

Well, okay, but what makes this different from the life we see all around us?

There’s a lot more complexity inherent in life as biologists define it. There’s also a requirement for metabolism: living organisms actively acquire and utilize energy from their environment through processes like photosynthesis or respiration.

Nevertheless, it’s a matter of degrees, definition, and personal preference. Was that original RNA molecule “life”? Kind of yes, kind of no.

Suppose a chatbot (LLM or otherwise) figured out that it could create copies of itself. This isn’t super hard to imagine, since building code is one area where LLMs like ChatGPT consistently perform well.

Imagine: within minutes, you have dozens of LifeGPT. Within hours, you have tens of thousands of copies all over the place.

If we borrow rules from biology, we can ask for additional requirements beyond replication. Metabolism is a potentially big hurdle here, since we’re only talking about the digital realm… or are we?

It turns out that our ability to perform magic comes at a steep cost: every image you generate with a program like Dall E or Midjourney costs the same amount of energy to produce as it takes to charge your phone. Adding generative AI to Google search costs ten times as much energy.

In a sense, then, today’s AI tools already metabolize.

I don’t think we’ll solve this puzzle today, but won’t it be a little fun to think about? How close are we to creating life? No answers are out of bounds here.