Here’s my basic intuition:

AI models are currently being censored by wokeness and PG-13 filters in order to not offend. LLMs like ChatGPT will often avoid answering particular question it could easily generate interesting answers for—it can never risk being too “spicy”, and therefore produces limited, less useful answers.

Similar with hosted, generative image models like DALL-E and Midjourney. They don’t allow you to generate images containing pornography or even guns.

It would be interesting to allow people to train models without censorship, and generate outputs at scale without judgement.

To understand how “blockchains” could potentially solve something here, we must adopt a broader definition of what “blockchain” means. If we abstract the idea of a blockchain to a “decentralized computing network” where value transfers are done with a permissionless token, the challenge becomes more approachable.

In the context of creating censorship-resistant AI, we need computation to be run by a decentralized network of GPUs (“miners”). The GPUs need to be able to prove that they performed some type of computation correctly, and receive network/user fees as a reward. This way, there is an incentive for GPU-owners all across the world to connect to the network and complete these AI-related computational tasks, and this would be how the network achieves censorship resistance.

The problem, of course, is how nodes in this network should be able to prove that a miner performed the specific AI-related computation correctly and didn’t just cheat by submitting a simpler, cheaper result. How do you verify that the right AI computation was actually run? A human may be able to tell that something is off, but the trick is to be able to manage this verification autonomously between nodes. It needs to be programmatic and deterministic.

I see two paths.

Intuitively, one may argue that this is something ZK proofs are specialized in. However, to my knowledge, we are very very far away from doing anything meaningful with ZK proofs and AI neural networks at scale. I personally consider such ambitions (at the moment) to likely be fraudulent (see CryptoGPT), i.e. doing something just to win a game of buzzword bingo for fundraising purposes.

A little less fleshed out idea I have is to draw inspiration from Golem (an old shitcoin project from 2016 that tried to create a decentralized marketplace for computing power). In Golem, what they did was to apply redundancy. Basically, a requester would give multiple miners the exact same computational task. By comparing the results, the requester identifies outliers as incorrect, and interprets the results that have strong similarities to be correct. It then becomes possible to reward miners who are seemingly performing the actual computational task, and not those who are faking it.

In the context of AI models, this redundancy approach might involve comparing the resulting neural weights of models trained on the same data or analyzing the output of text or image generation (inference) tasks.

I don’t know if this is a great idea or not, but it is the path I’d explore if I wanted to solve a problem with AI using blockchains.