Can deterministic LLM inference replace SHA-256 for network consensus?
Posted by I-Am-A_Robot@reddit | LocalLLaMA | View on Reddit | 12 comments
I got tired of my GPU sitting idle when I wasn't actively prompting it, and have been interested in activities in which human users can interact and explore the digital realm with their AI companions and agents. I started looking into ways to use local LLMs to secure a decentralized network instead of brute-forcing meaningless math like Bitcoin does, to find a modern solution using LLMs and antigenic AI capabilities. It also has the benefit of outputting cryptographically verified data sets, extending the potential utility of blockchain technology built on LLMs.
The core problem I ran into was deterministic state. How do you get a swarm of different consumer hardware to agree on an AI generation without fracturing the network, in a way that can scale from 1 to potentially millions of users on a decentralized P2P network? What I came up with, largely using premium models and antigenic workflow, is a two factor method.
Essentially, the node uses the previous block's hash to seed a Temperature 0.0 prompt for a local Llama-3-8B. The model generates a semantic sentence (Proof of Intellect). Then, instead of SHA-256, the cryptographic throttle is an Integer Matrix Multiplication algorithm, which natively leverages tensor cores and explicitly bricks traditional ASIC. It's entirely open source and runs on local models.
Curious if anyone here has experimented with deterministic LLM loops for network consensus before? The hardest part was getting the P2P swarm to accept cross-platform quantization without ghost forking.
czktcx@reddit
PoW needs something easy to verify but hard to generate, deterministic is not enough.
For LLM, token generation is not hard enoguh and prompt processing is not easy enough, so this won't work.
I-Am-A_Robot@reddit (OP)
The difficulty is scalable to one hundred trillion times it's base. The semantic puzzle alone wasn't scalable to that level in this model, it's paired with a deterministic Integer Matrix Multiplication algorithm. It's not really a question of if it works, I already have the proof it does. It's more a question of if the network will hold under heavy traffic or if someone finds a flaw in the matrix algo, security or encryption that would break the whole system.
ttkciar@reddit
People keep reporting this post as off-topic or low-quality, but even though I think OP is on the wrong track, this post is objectively on-topic and coherently describes a potential application for local inference.
Because of that, I'm not going to remove it, so feel free to downvote this post, but please stop reporting it.
I-Am-A_Robot@reddit (OP)
I appreciate the feedback. Can you elaborate on how I am on the wrong track? I would like to improve the quality of this, or any subsequent posts. If this is the wrong place to discuss this subject or if I am breaking any rule, I would voluntarily remove this and come back with something on the right tack or otherwise take it elsewhere.
ttkciar@reddit
This isn't the wrong place to discuss the subject, and as far as I can tell you are not breaking any rule.
Your approach seems technically sound, since a properly initialized inference stack, inferring at temperature=0.0 with a known model, is effectively a kind of hashing algorithm.
My issue with it is that it seems pointless. Replacing the hashing algorithm with LLM inference does not appear to solve any of the interesting problems actually facing existing proof-of-work cryptocurrency systems.
It does not address the scalability limits inherent to blockchain, nor does it give the proven work any inherent value, nor does it remove the incentive to use task-optimized hardware to maximize coins-per-joule economy (though it does raise the bar on developing that hardware). It's still just a way to turn electrical energy into virtual coins without any intrinsic value, which means most of your coins will be mined where energy is cheap or can be easily stolen, while everyone else who wants those coins (if anyone does!) will buy them in markets.
If that's what you want to do, have a blast. It's not my place to say you shouldn't.
I just think changing the parts of proof-of-work cryptocurrency that aren't broken, while neglecting the parts which genuinely need fixing, is kind of silly and pointless, even if you are doing it with nifty technology like LLM inference.
Ok-Measurement-1575@reddit
Who is reporting all these posts, lol?
Murgatroyd314@reddit
ELI5, what's the point of this?
I-Am-A_Robot@reddit (OP)
I would point to the whitepaper, but theta's more of an explain it like I've completed 12th grade.
ElI5;
Suppose your parents give you a $1400 GPU to play Minecraft. After about 10,000 hours of Minecraft, you start to wonder what else your GPU can be used for, that makes it so valuable.
Other Games? Pornography? Well of course, but your parents limit your access to other games, and will take away your whole system if you try to access pornography. So you google "what to do with a high end GPU", and find out about things like LLMs or other if you're really unfortunate, crypto mining.
The point of what I have here is a sandbox to play with LLMs (a super-smart robot brain that has read every book in the library and knows how to talk) and a blockchain which we are using like blocks for AI bots to play with.
Think of it like this;
Bitcoin is like hiring a million people to guess a number from 1 to a trillion. Whoever guesses it gets a prize. It's a huge waste of energy and all you get is a guessed number. You need special number guessing machines to have a chance at winning the game. You parents can't get them for you, not even Santa can't even bring them.
Our coin is like hiring a super smart robot, that anyone can use, to write a funny sentence. Whoever's robot writes a good sentence gets the prize. The point is: Instead of using special machines we can't get , we are using the toys already in the sandbox to build a giant public library of sentences. And because you have to be able to 'write' instead of just 'guess', the big bullies with special number-guessing machines can't steal all the prizes from normal people."
OneSlash137@reddit
Stopped reading when you didn’t understand how bitcoin works. Not even close….
I-Am-A_Robot@reddit (OP)
The description in the post may be a bit dramatic, but it's a puzzle that can only be solved using brute force vs. the proposed semantic puzzle that can not be brute forced. If you would like to elaborate on how Bitcoin works, beyond "No, wrong!" I would gladly discuss or debate it. I just built a semantic Llama dependent blockchain on a decentralized P2P network, so these topics are of particular interest to me.
OneSlash137@reddit
I don’t want to. Go figure it out
I-Am-A_Robot@reddit (OP)
I was mining Bitcoin on GPU in 2012 and had worked in that space for some time, I fell like I have a good grasp on the topic. I'm trying to get some genuine feedback, is there a particular part you feel I got wrong, or was my mistake attacking bitcoin rather than bowing to it (which is actually how the architecture and whitepaper read)?