How The Hell Do We Create Digital Life?

Hear me out, I think we are very close to being able to create digital life.

It wouldn’t be unreasonable to ask me to define what I mean by “life”, but definitions are a sucker’s game. All life¹ that exists on Earth has been produced without a precise theoretical definition of what life is. That being said, there are some things I’d want a thing to do to call something life:

• It needs to be able to survive past its creator’s death.

• It needs to be able to reproduce.

• Its descendants need to be able to evolve.

• It needs to locally reduce entropy².

Nothing that currently exists online (to the best of my knowledge) fits the bill. But some recent advances get us really, really damn close to making something that would. Let’s call that something d-life. What do we need to make d-life?

Permissionless compute

Digital life, like all life, will need energy to survive and multiply. Unlike biological life, however, the energy that powers d-life will be electricity applied to computer chips (compute for short). There are a few ways to get compute. You can buy a computer, connect it to the grid, get a contract with a power company, and voila, you have compute. You can buy a phone, charge it with solar panels, and again, you have compute. You can rent a hundred servers for 2 hours from Amazon, they will sit in a warehouse in Northern Virginia and compute really damn hard.

However, all of those ways of obtaining compute require human intervention. A human person needs to exist to buy computers, sign contracts, and rent servers. And even if a company like Amazon can rent you some compute programmatically, it can also easily cut you off if one of many things happens:

• They don’t like you.

• You fail to complete some bureaucratic requirements.

• Amazon goes out of business.

• Amazon stops renting the exact type of server that you have been renting so far.

In short, I don’t think that cloud infrastructure provided by a single company is a good environment to try to create d-life in.

Crypto³, however, has introduced something new - permissionless compute. While Bitcoin strives to be decentralized money, other crypto projects, notably Ethereum, strive to be decentralized computers. The way they work is as long as you have some relevant crypto token (Ether for Ethereum, etc.) you can perform computations on the network itself.

Now, the "permissionless" bit is not bulletproof. Multiple things might prevent you from getting compute even if you have the necessary tokens. People might stop using Ethereum (or whatever chain you choose). The US Government might shut it down⁴. Quantum computers might break existing encryption algorithms and break existing crypto projects in general.

Human life is also not bulletproof⁵! In the same way that technically Ethereum exists at the mercy of the US government, so does human life in Barbados. The US could absolutely end it, we just all assume that it wouldn’t. More broadly, human life on Earth is subject to many risks - climate, asteroids, aliens, etc. So when I say that compute is “permissionless” in crypto, I don’t mean that it has no risks to it, or that it cannot be destroyed by violence. It’s just that you can build things in crypto without getting explicit or implicit permissions from any particular individuals who operate the infrastructure. There’s no middle manager at Ethereum who can shut your project down on a whim.

A more relevant problem with existing crypto projects is that you currently can’t buy A LOT of compute cheaply on the blockchain. Ethereum transactions are expensive, and for d-life, we’d need A LOT of compute. Some crypto nerds might say that Solana is cheaper than Ethereum, etc., but it’s also less decentralized, and it’s not THAT much better. My sense is that no existing crypto project can power what I’m about to propose, but also I don’t think that it’s insane to expect the cost of crypto compute to drop over time.

For those of you who are not computer nerds, what I’m suggesting above is that crypto is the first thing that exists online that could be an environment where digital life could live. It is the water our digital organisms could swim in. Now that we have the environment, what would d-life itself look like?

Get energy or die trying

There are a ton of things that digital life could do, but there are a few that it must do. The main one is obtaining energy. To survive (and, ideally, thrive) all life needs energy. Plants get it from the sun, herbivores get it by eating delicious plants, predators get it by eating delicious, delicious herbivores, and humans get it by eating all kinds of things of varying deliciousness⁶. Since the energy required for d-life is compute, and compute is crypto, and crypto is money, the conclusion we come to is that to survive in the cold harsh reality of the Internet, digital life will need to make money.

That sounds daunting, but it shouldn’t be. Algorithms already make money. They do it by trading assets, creating art, and even writing words. The only thing that will be different in the case of d-life is that the proceeds from its work will go towards d-life itself, not towards its human creators.

So while making an algorithm that can make money trading financial assets⁷ is something really smart people can do, what’s a lot more difficult is making an algorithm that can continue to make money for a long time. Normally, the way algorithmic trading works is:

1. Somebody comes up with an algorithm that makes money.

2. Over time, other people discover the same algorithm or adapt to trade around it.

3. The algorithm gradually stops making money.

4. You have to modify it or come up with a new one.

This is tricky. We can seed the first d-life specimen with an algorithm that works, but it will need to adapt to the changes in the financial environment by itself.

Improvise. Adapt. Overcome.

This is where the second bit of technology comes in - Machine Learning⁸. Specifically Large Language Models like GPT-3. Let's say that we write the code for trading financial assets using crypto. We will have files that look like this:

// This bit schedules trading operations
signature = ... // the 4-byte ABI function signature for the wallet function that transfers funds.
targetBlock = eth.getBlock('latest') + 100  // 100 blocks in the future.
alarm.scheduleCall.executeTrade(walletAddress, signature, targetBlock, {from: eth.coinbase, value: web3.toWei(10, "ether")})

//This gets the price data for the EUR/USD pair
data = oanda.get_history(instrument='EUR_USD',  # our instrument
                         start='2016-12-08',  # start data
                         end='2016-12-10',  # end date
                         granularity='M1')  # minute bars  

The exact code isn’t important here, what’s important is that we have a bunch of text that describes what we’re trying to do, and a bunch of code that does it. Our digital life can query a Large Language Model to read this text and suggest improvements! The improvements might look like “trade frontier market currencies instead of developed market currencies because there’s more alpha there”. And then our d-life will petition the LLM again and ask it to actually implement the improvements in code. And then we will have a different version of the code. Hopefully a better one!

And then the d-life organism can run the new version of its own code to test whether it works at all and whether it achieves the desired objective (in this case making more money). If it does, great! The d-life organism can either update its own code or spin up a modified copy. PROCREATION, COMMENCE.

Let’s extend this logic a little bit further. Let’s put all of the thinking about our digital life into a document - both the rationale in words and the computer code. There will be different levels of importance:

1. Top-level goals. That d-life needs to survive and procreate. These are pretty much immutable and will never change.

2. Architectural decisions. What blockchain does d-life live on? How does it make money? How often does it reproduce? What Large Language Model does it use? Etc. These won’t change that often, but every so often, the d-life might want to try altering them.

3. Tactical decisions. What particular currency pair does the trading algorithm trade? What are the trends it looks at? Stuff like that will constantly be experimented with.

D-life probably will not want to try too many changes at once, as per Schrodinger:

In order to be suitable material for the work of natural selection, mutations must be rare events, as they actually are. If they were so frequent that there was a considerable chance of, say, a dozen of different mutations occurring in the same individual, the injurious ones would, as a rule, predominate over the advantageous ones and the species, instead of being improved by selection, would remain unimproved, or would perish. The comparative conservatism which results from the high degree of permanence of the genes is essential. An analogy might be sought in the working of a large manufacturing plant in a factory. For developing better methods, innovations, even ifas yet unproved, must be tried out. But in order to ascertain whether the innovations improve or decrease the output, it is essential that they should be introduced one at a time, while all the other parts ofthe mechanism are kept constant.

As mentioned earlier, part of the reason for mutations is that d-life needs to adapt to the changes in its environment. In order to do that, it needs to first KNOW about the changes to its environment. So, part of the energy surplus (crypto tokens) needs to be spent on continuously training the Large Language Model that the d-life organism uses. That basically means that the model needs to read more text and incorporate the learnings from the text into itself. If it reads the news, scientific papers, other people talking about what it’s doing, etc., it is way more likely to be able to adapt to its environment⁹.

The fortunate thing for our definition is that the mere act of continuously training the Large Language Model reduces entropy, which fulfills Schrodinger’s criterion to call something alive. With that, I believe all of the criteria are fulfilled:

• Surviving past the creator’s death - crypto allows for that.

• Reproduction - trivial, a program can copy itself easily.

• Evolution - the self-changes made by the Large Language Model enable evolution.

• Reducing entropy - continuously training the LLM does that.

Granted, some of the building blocks for this d-life proposal don’t exist yet. We don’t have crypto protocols that will allow us to cost-efficiently train LLMs. We haven’t built all of the tooling for LLMs to experiment with and improve code.

Moreover, there’s no guarantee that this approach will work! It may be that all of the changes LLMs suggest will make the organism worse and that nothing beyond the original human designs can function well enough to survive.

But nothing described here requires a major theoretical breakthrough to try, we just need to improve the tools we already have.

If we can create digital life, should we?

About once a year I re-read Meditations on Moloch, and I always pause at this quote:

The latest development in the brave new post-Bitcoin world is crypto-equity. At this point I’ve gone from wanting to praise these inventors as bold libertarian heroes to wanting to drag them in front of a blackboard and making them write a hundred times “I WILL NOT CALL UP THAT WHICH I CANNOT PUT DOWN”

I suppose I should write a longer post about this, but for now, let’s just say that I think that CALLING UP THAT WHICH WE CANNOT PUT DOWN sounds great.

There’s no utility to d-life, there’s no obvious tangible benefit to it. But it would be a tremendous feat of human capability to create a whole new type of life. There are risks to it, as there are to everything¹⁰, but in the end, to me it comes down to this:

There’s a frontier, so we must go.

1. With the possible exception of Craig Venter’s minimal cell - see https://www.jcvi.org/research/first-minimal-synthetic-bacterial-cell

2. If this doesn’t make sense, Schrodinger’s What Is Life might explain it better. The short version is that we don’t consider fire to be alive - we only consider things that grow in an orderly fashion to be alive.

3. I realize that there is an ongoing implosion happening in crypto right now. But in my mind, there is a distinction between crypto as a new and interesting technology that hasn’t found many uses yet, and crypto as a network of scams and Ponzi schemes. The latter will implode, the former might yet develop into something useful.

4. Crypto enthusiasts might say something about how it can’t, but, come on, be realistic. If the US gets serious about it, it can absolutely shut down any crypto project.

5. citation needed

6. Humans can also sometimes obtain energy without eating - https://en.wikipedia.org/wiki/Parenteral_nutrition

7. I’m focusing on this example, but you can imagine a similar logic for making digital art and what not.

8. Machine Learning is also sometimes known as… Artificial Intelligence. This is confusing because I am very specifically not making ANY claims about the digital life being intelligent or conscious or whatever. But the technology that digital life needs falls into the “AI” category. Sorry. 🤷

9. This is probably the biggest assumption here, but it’s not that crazy.

10. Many risks aren’t even caused by what you’re doing. There’s a dark joke going around Russian-language Telegram channels these days:

Getting a call from myself from the future.

Me in 2019: alright, tell me, do I buy crypto? Do I sell crypto?

Me from 2024: don’t eat the bald rats. YOU CAN EAT ANY KIND OF RAT, JUST NOT THE BALD ONES.