Generative Blockchains
Having spent quite a bit of time researching LLMs and working in the web3 space, I have been looking to answer the question “what is the intersection between AI and blockchains?”
Apologizing in advance
If I were a savvy investor or crypto native, the second I hear the words “AI Blockchain” I would immediately begin getting triggered. Taking two of the most popular fad buzzwords and mashing them together sounds … pretty sketchy. And it has been talked about and tried before.
This is super understandable. And has given me a lot of second thoughts. If I you are going to combine AI and blockchains together, there really needs to be a technical justification.
So is there one?
The answer is yes, there is. In fact there are several potential use cases and research areas which are both technically feasible and potentially in utility or fun. Let’s explore this.
It gets wild.
How LLMs Enable New Forms Of Blockchain
If you are at all familiar with how blockchains work, there are two main flavors of algorithms: Proof of Work and Proof of Stake. There are others, and I am not going to fill up space here outlining or explaining any of them.
Now “AI blockchains” have been talked about before, but in my opinion something has changed: The rise or Large Language Models (LLM).
AI blockchains previously often did not have access to the form of “AI” we have now and relied on outdated algorithms and approaches to AI. LLM research is moving extremely fast, creating a rapidly emerging landscape of new possibilities.
So what are those possibilities?
Well — The most obvious opportunity is a new class of Algorithm: PoG (Proof of Generation).
Here is where things will continue to get weird.
LLMs output as a form of encryption
Under the previous regime of blockchains, Proof of Work (the algorithm used to secure Bitcoin) functioned by solving complex cryptography puzzles across a distributed network of machines to mine new Bitcoins.
That approach has been proven and well established. We know it works and is secure.
But now we have a “new” form of encryption”: LLMs and Prompts. Let’s explore how this can work.
Imagine you have a Large Language Model with millions of parameters. No one knows how it works, LLMs today do not have the characteristic of “Observability” in the same way that human brains likewise don’t.
In the same way that human brains rely on billions of neurons to magically perform calculations when provided with a stimulus, LLMs also come up with results in a hidden way.
You feed a prompt into DALL-E such as “a giraffe wearing a hat,” the LLM “does a lot of secret stuff” and out comes your result.
No one can truly say what happens inside the black box of the LLM, just that it produces an interesting, novel, unique image.
In other words: A sufficiently vast and complex LLM bears striking resemblance to encryption using a Proof of Work algorithm in that you feed in some data, heavy duty processing occurs and then you get a result which is hard to trace back to the source material.
This enables a new class of blockchain algorithms: Generative Blockchains. We are in very early stages, and significant research is required to explore this topic.
A very simple Generative Blockchain
Let’s build a theoretical Generative Blockchain to show how this works.
Imagine you have a network of 1,000 machines each running an LLM designed specifically for the purpose of creating and evolving NFT artwork in random ways.
Rather than 1,000 machines running an algorithm to “mine” crypto on a GPU, you simply swap the application for an LLM running on each machine.
Each machine is running an English language randomizer which picks a Noun, a Verb and another Noun. Example: “Elephant Running Hill.” Or similar the complexity of this prompt generator can be infinite.
Step 1: Start the process with a prompt
The LLM on the first machine runs the prompt generator and gets “Giraffe Wearing Hat.” The LLM produces a result.
The result is converted into a hash and sent to the network. Now, every machine in the network is expected to “guess” how the hash of the image was arrived at by “mining” for the prompt.
That’s right, prompt mining. It works by feeding random nouns and verb sequences into each LLM to try to recreate the original image.
Step 2: All the machines in the network read the hash and begin trying to recreate it
The LLMs “mine” the English language using the prompt mining approach trying to combine words until an image is generated that has a hash equal to the input image.
Step 3: The “winning” machine “mints” the NFT and then uses an English language randomized to produce a new prompt
Eventually, one of the prompt miners recreates the image or something close to it. Then a new prompt sequence is randomly generated: “cheetah eating cheese.”
The process repeats.
Is this secure?
No. It is wildly insecure. And unstable. If my understanding of LLMs is correct, you will generally get random and fuzzy results back from this process.
Even if another machine in the network does create an image resembling the source prompt, it will have differences due to how LLMs work (I need to perform further research to fact check myself here).
Is it an interesting and fun direction to research? Yes.
In my opinion, the stupidity and silliness is going to be a major part of the appeal. Generative Blockchains would have huge appeal to creator communities because, well, they are likely to be pretty creative and silly.
MemeChains
The silliness of this approach to generating a blockchain makes me feel that something like it is certainly feasible. It might make a fun basis for a MemeChain or MemeChains which exist to have fun with what sorts of insane results the community can come up with.
If Bitcoin is the ultra libertarian, serious side of crypto out to disrupt the global banking system and displace gold; and Ethereum is its psychedelics infused successor, MemeChains take things a step further into complete absurdity.
This is a very primitive conceptual model and there are many unsolved problems here, but the general direction of this is “what if we use the output of LLMs running on a network of machines as a new form of PoW” is the gist of it.
I can see communities forming where every new community member “mints” their NFT as part or registration process on their preferred Generative AI chain which has been tuned for their particular community culture.
What is the utility of this?
One observation is that LLMs used in this manner could be used to devise blockchains which generate and mint NFTs as their native form of currency. Or NFT sequences.
This might have interesting applications for games. Or movies.
NFT Sequences
Most NFT collections look like they were generated by AI. If you look at Bored Ape Yacht Club, for example, you can see that we could easily replicate these collections with LLMs.
The novel addition here is to integrate the LLM as a replacement for PoW and make the actual generation of the images, combined with prompt mining in order to unify encryption and minting processes into one.
That’s right, Generative NFT Chains.
Generative Compression Blockchains
Another use case, alluded to by John Carmack might be “compression” blockchains designed to store movies or other assets.
One can imagine storing a “movie” as a sequence of prompts evolving over time.
If you have ever studied stop-motion animation, you will understand the concept of a “Key Frame.” Animators use Key Frames every certain number of frames to capture the state of the characters in an animation sequence and then “tween” between them.
A generative AI blockchain could use an LLM plus a sequence of evolving prompts representing each key frame to store a compressed movie.
That’s right; Movie NFTs.
Mining for MovieChains
This key frame concept gives rise to an exciting idea which is a movie which is generated by a network of machines being fed prompts and perhaps with the input of the community.
Not only would you get a blockchain, you could watch your blockchain in a movie and potentially influence the plot by adding community enabled features to change the script. Or generate the script.
Closing Thoughts
Using LLMs to create blockchains and sequences of NFTs is an emerging theoretical use case. Research will be needed to explore this topic and it’s Potential.
