P2P Bitcoin AI Model

[toribit]

It has came to my attention that there are models of the neuron that are
proof of work compatible.

If true, this could change crypto as we know it, forever.

The duality of machine learning (just like banking) is such that the problem
can be solved in centralized or in a p2p decentralized setting.

In classical machine learning, most neurons train simultaneously using back-prop.
This is not true in the other models of neurons of interest to us, such as
Bloomier Filters or LuT cells.

You may be thinking, has this guy gone nuts? Is he going to scam us using a shit
project that goes nowhere?

Let me explain.

It's trivial to encode the solver's identity inside the neuron. What is also
possible is to make the proof of work chain of the solutions.

Inserting user configurable functions such as attentions has been done before.

What remains to prove, that the p2p network of the solvers can be assembled,
that they can, neuron by neuron, generate a solution and therefore train a
competitive open machine learning model (while earning bitcoin in the process).

[NotATether]

"AI" is not applicable to blockchains.

Sometimes I wish that AI and machine learning models stayed a niche subject for the people who actually use it for productive stuff, instead of the generative slop that's being mass-produced nowadays.

[watashi-kokoto]

"AI" is not applicable to blockchains.

Sometimes I wish that AI and machine learning models stayed a niche subject for the people who actually use it for productive stuff, instead of the generative slop that's being mass-produced nowadays.

Nice GPT answer Mr. smart pants. Have you even read the post? He is not suggesting that AI will be applicable to blockchains, quite the reverse, that a proof of work chain will be applicable to AI training and subsequent payout to trainers.

[BIT-BENDER]

"AI" is not applicable to blockchains.

Sometimes I wish that AI and machine learning models stayed a niche subject for the people who actually use it for productive stuff, instead of the generative slop that's being mass-produced nowadays.

Same here. I believe AI can help better the practice of certain fields, I saw today how surgeries will be carried out using AI, this surgeries will be carried out using AI to make surgeries more direct, safe and more productive but still this won't replace surgeons but rather surgeons will learn to use this to be more productive.

Ai has really been bastardized and all connect I have seen on AI with Bitcoin has been a complete unrelated concept.

[NotATether]

Nice GPT answer Mr. smart pants. Have you even read the post?

Yes.

It's trivial to encode the solver's identity inside the neuron. What is also
possible is to make the proof of work chain of the solutions.

It doesn't take a machine learning undergraduate to understand that the output of the model will be subject to sharp variations in even the smallest changes in data input, and this lack of a deterministic output from the model will ultimately be why we will never see any practical application of AI inside PoW, PoS and similar algorithms.

[toribit]

Although I hold two separate degrees in two different fields, I do not believe
that these help to advance this debate about the role of crypto in machine learning.

It doesn't take a machine learning undergraduate to understand that the output of the model will be subject to sharp variations in even the smallest changes in data input, and this lack of a deterministic output from the model will ultimately be why we will never see any practical application of AI inside PoW, PoS and similar algorithms.

I'm not claiming AI is applied and runs on a blockchain; I'm proposing that
each neuron's training step can serve as proof‑of‑work, with the resulting
weight update recorded on chain. The blockchain only needs to verify
a deterministic proof that an update improves a fixed validation loss, not
the stochastic final output. By rewarding genuine gradient steps, we turn useful
model training into the consensus work that secures the network. This aligns
incentives without sacrificing the randomness inherent in deep learning.

In short, you are wrong.

[NotATether]

I'm not claiming AI is applied and runs on a blockchain; I'm proposing that
each neuron's training step can serve as proof‑of‑work, with the resulting
weight update recorded on chain. The blockchain only needs to verify
a deterministic proof that an update improves a fixed validation loss, not
the stochastic final output. By rewarding genuine gradient steps, we turn useful
model training into the consensus work that secures the network. This aligns
incentives without sacrificing the randomness inherent in deep learning.

In short, you are wrong.

Well, you can only use neuron value updates as Proof of Work if you commit to one weight training algorithm only.

Blockchains will typically not accept a proof of work from two or more possible algorithms, especially unknown (not discovered by researchers yet) training methods. By their very nature, you can't code future yet-to-be-used neural network training methods into a PoW protocol.

And training itself is just a bunch of different stages like the forward pass, the loss calculation, the optimizer, etc. So if any of these algorithms change between two training sessions, you cannot use the new values as PoW for the old "chain", and given the pace of research in ML, the training methods seem to be refined fairly quickly.

TL;DR if you use a specific neural network training as input for PoW, don't expect to be able to use improved versions as PoW values as well.

If you don't believe me, open pytorch on CPU and use any training method on any dataset but just change the activation function, and compare the outputs.  

[Doan9269]

It has came to my attention that there are models of the neuron that are
proof of work compatible.

If true, this could change crypto as we know it, forever.

Let's not take Bitcoin for granted, because before it could stand the test of trials and been successful among other researchers that have attempted something similar, shows how it has been approved for the present and future benefits of what it was being created for, that is why everything in bitcoin deals with codes and these are in consensus with the blockchain technology, we should know more better under the bitcoin network, nodes and the likes, anything that does not aligned with the protocols got rejected because it has been programmed in such a way, therefore bots or any advanced technology cannot be a threat to the bitcoin network for any reason, because they can be in consensus no matter how harder others tried.

[ABCbits]

Your idea isn't new. I have seen few news about the possibility and few discussion on this forum, such as Exploring Tensor-Based Proof-of-Work: Aligning Mining with AI/ML Workloads.

But as layman on AI/ML, i wonder how fast to perform the verification compared with the training? On Bitcoin, you just "need" to perform SHA-256d on block once while miner do SHA-256d billion or more times. In addition, how can node verify the training done properly rather than updating with random number?

[toribit]

Your idea isn't new. I have seen few news about the possibility and few discussion on this forum, such as Exploring Tensor-Based Proof-of-Work: Aligning Mining with AI/ML Workloads.

Thank you, I looked through, and the thing that seems to be dooming those projects is that they either use pointless neuron model (perceptron) or that they try to optimize using proof of work a pointless operation (mat mul).

But as layman on AI/ML, i wonder how fast to perform the verification compared with the training? On Bitcoin, you just "need" to perform SHA-256d on block once while miner do SHA-256d billion or more times. In addition, how can node verify the training done properly rather than updating with random number?

Suppose the ground truth dataset size is O(n). In my system, you need O(log n) operations to validate a solution, which is far less than the full O(n) operations needed to produce a weightless neuron.
This gives verifiers a chance to catch up to the "tip" of the proof of work tree and begin producing solutions on their own.

[memehunter]

You may be thinking, has this guy gone nuts?

Not at all. I was thinking that I am so dumb that I cannot understand what you are trying to convey exactly. Since you posted this in this section, you might need to dumb down it quite a lot OR you can post it in technical section for much specialized discussion.

I believe AI can help better the practice of certain fields, I saw today how surgeries will be carried out using AI, this surgeries will be carried out using AI to make surgeries more direct, safe and more productive but still this won't replace surgeons but rather surgeons will learn to use this to be more productive.

In that sense you always need an observer to observe anything lol. This argument is way too general, and I think AI is already replacing actual people.