AI needs decentralization and verifiability. In a recent tweet, the founder of Curve Finance, @newmichwill, said that the main purpose of crypto is DeFi and that AI doesn’t need crypto at all. While I agree that DeFi is an important sector in crypto, I don’t agree that AI doesn’t need crypto. As AI agents are booming right now, where there’s usually a token attached to each agent, people have a false perception that the intersection of crypto and AI is effectively AI agents. The other topic people seem to miss is decentralized AI itself, which is related to training AI models themselves. The thing that I don’t like about narratives is that the majority of users blindly assume that something is important and useful enough while it’s popular, or even worse, assume that the only goal of the narrative is to extract as much value as possible and that’s it (make some money). One of the first questions in decentralized AI that we should ask ourselves is why it needs to be decentralized and what consequences we face because of that. It turns out that the idea of decentralization almost every time inevitably leads to the idea of incentive alignment. There are multiple fundamental problems in AI that can be fixed with crypto, moreover, there are some mechanisms that add more trust to AI, and not only fix existing problems. So, why does AI need crypto? 1. Expensive computations reduce participation and innovation. Fortunately or not, big AI models require a lot of computational resources, which naturally limits many potential users from participating. Most of the time, AI models need a lot of data resources as well as actual compute that is basically too much to handle for a single individual. This problem is particularly seen in open-source development where contributors, besides investing their time to train the model, must also invest computational resources, which doesn’t lead to effective open-source development. The thing is that, yes, an individual can allocate a lot of resources to run an AI model, just as a user can allocate these computational resources to running their own node on their own blockchain. However, this doesn’t fix the problem as a whole, where computational power is simply not enough to perform relevant tasks. Independent developers or researchers cannot participate in contributing to big AI models like LLaMA, simply because they cannot afford the compute required to train the model: there are thousands of GPUs, data centers, and additional infrastructure needed. To provide a sense of scale: → Elon Musk stated that the latest Grok 3 model was trained using 100k Nvidia H100 GPUs. → Each chip is valued at approximately $30,000. → The total cost of the AI chips used to train Grok 3 is around $3 billion. This problem is actually similar to startup building in some sense, where individuals have time, technical ability, and an execution plan, but not enough resources at first to build their vision. As noted by @dbarabander, while a conventional open-source software project only requires contributors to donate their time, an open-source AI project demands both time and substantial resources, such as computational power and data. Depending solely on goodwill and volunteer efforts is insufficient to encourage enough individuals or groups to supply these costly resources. Additional incentives are necessary to drive participation. 2. Crypto is the best tool for incentive alignment. Incentive alignment is when rules are in place to encourage participants to act in ways that help the system, even if they also benefit themselves. There are countless examples of when crypto helped different systems with incentive alignment, but I’d say the most notable one is the DePIN industry, where it was a perfect fit. Projects such as @helium and @rendernetwork are great examples where incentive alignment was achieved with a distributed network of nodes and GPUs acting as network participants. Why can’t we take this model as an example and use it in the AI field to make the ecosystem more open and accessible? Well, it turns out we can. What drives web3 and crypto in general is ownership. You own your data, you own your incentives, and even if you hold some tokens, you still own a piece of a network. Granting ownership serves as motivation for resource providers to offer their assets to the project, with the prospect of gaining benefits from the network's success. To make AI more accessible, crypto is the best solution that could be possible. Developers could freely share model designs among themselves and projects, as compute and data providers would supply resources for ownership stakes (incentives). 3. Incentive alignment correlates with verifiability. If we envision a decentralized AI system with proper incentive alignment, it should inherit the same features as classic blockchain mechanisms: 1. Network effects. 2. Lower initial requirements with nodes earning on future earnings. 3. Slashing mechanisms to penalize actors that act maliciously. For slashing in particular, we need verifiability, otherwise, if it’s not verifiable who acted maliciously, then we cannot penalize the intended actor, and it will be really easy to cheat the system, especially when working cross-collaboratively. In a decentralized AI system, verifiability is needed because we don’t have a centralized point of trust. Instead, we aim for a trustless yet verifiable system. There are multiple components that might need verifiability: • Benchmark phase (the system is better than another system based on x, y, z) • Inference phase (the system is running correctly, essentially the “thinking” phase of AI) • Training phase (the system is trained correctly or adjusted correctly) • Data phase (the system is collecting the data) There are 100s of different teams that build something on @eigencloud, but what I noticed recently is the shift into AI more than it was before, and I was wondering if it aligned with the original vision of restaking. Any AI system that wants to have incentive alignment has to be verified. In that case, slashing equals verifiability: if a decentralized system can slash malicious actors, then it’s capable of identifying and verifying that a malicious action was made. If the system is verifiable, then AI can use crypto to tap into global compute and data to craft even bigger and better models, because more resources (compute data) lead to better models (at least in the current world). @hyperbolic_labs is already showing what’s possible with collaborative computing resources, where any user can rent GPUs and use them for training much heavier AI models than they would be able to do at home, at a much cheaper cost. → How to make AI validation efficient and verifiable? Someone might say that there are a lot of cloud solutions available to rent GPUs, so it fixes the compute problem. Unfortunately, cloud solutions like AWS or Google Cloud are pretty centralized and create what is known as a waitlist strategy, where they create fictional high demand and raise costs due to the field being an oligopoly. There are a lot of GPUs in data centers, mining farms, or just with other individuals sitting idle that could potentially contribute to the compute for AI model training, but they’re just sitting idle. You might have used @getgrass_io, which sells your unused bandwidth to corporations, so you don’t waste your bandwidth for no reason and get some rewards in return. I’m not saying that there is an unlimited amount of compute, but the thing is, any system can be optimized, creating a win-win scenario for someone who needs more resources for AI model training in a more open market and someone who can get rewards for contributing these resources. The team from Hyperbolic created an open GPU marketplace, where users who need compute for AI model training can rent a GPU and save up to 75%, while renters can monetize their idle resources. Here is an overview of how it works: Hyperbolic organizes connected GPUs into clusters and nodes to allow compute to scale based on demand. The main part of this architecture model is the Proof of Sampling model, where transactions are sampled together: they’re randomly selected and verified to reduce the workload and computational demand. The main problem lies in the AI inference process, where every inference run on the network needs to be verified, preferably without the significant computational overhead of other mechanisms. As I said before, if something can be verified, it has to be slashed if the verified action is against the rules. When Hyperbolic adapted the AVS model, it allowed them to add more verifiability to the system, where validators are selected randomly to verify outputs and make the system incentive-aligned, where dishonesty is unprofitable. There are two main resources that are needed to train an AI model and make it better: compute and data. Renting out compute is one solution, but we still need to get data from somewhere, and different data to exclude potential bias in the model. → Verifying data from different sources for AI The more data you have, the better model you’ll have; the thing is, you often need diverse data. This is one of the problems with AI models. Data protocols have been around for decades. Regardless of whether the data is public or private, data brokers gather it one way or another, pay for it or not, then sell it for profit. The problems we face with getting proper data for AI models are a single point of failure, censorship, and the lack of a trustless way to provide authentic data for “feeding” an AI model. Who needs this? Well, first of all, it’s AI researchers and developers who are looking to train and infer their models with real, proper inputs. OpenLayer, for example, lets anyone add a data feed to the system or AI model permissionlessly, and the system can transcribe every piece of available data in a provable way. OpenLayer also uses zkTLS, which was described in my previous writings, to ensure that what the operator reports is indeed what they got from the source (verifiability). Here is how OpenLayer works: 1. Data consumers publish data requests to OpenLayer’s smart contract and retrieve results from the contract, on-chain or off-chain, using APIs similar to major data oracles. 2. Operators register through EigenLayer to secure OpenLayer AVS staking assets and run AVS software. 3. Operators subscribe to tasks, process and submit data to OpenLayer, and archive raw responses and attestations in decentralized storage. 4. For variable results, aggregators (special operators) standardize outputs. Developers can ask for fresh data from any website, and it can be plugged into the network, while if you’re developing something in AI, you can get reliable, real-time data. We have gone through the compute process for AI and getting verifiable data for AI. After we get the two main parts of the AI model, it’s time to do the computation itself and verify it. → AI computation, in order to be correct, has to be verified In the best scenario, nodes must prove their computational contribution to ensure the system operates correctly. In the worst case, nodes could falsely claim to contribute computational power without doing any actual work. Requiring nodes to prove their contribution can ensure that only legitimate participants are recognized, avoiding malicious behavior among them. This is actually pretty similar to standard Proof of Work; the only difference is the work that those nodes do. Even if we add proper incentive alignment to the system, if nodes can’t permissionlessly prove that they did some work, they could receive disproportionate rewards compared to their actual contributions and vice versa. If the network can’t assess computational contributions, it might overload some nodes with tasks they can’t handle while leaving others idle, leading to inefficiencies or failures. Proving computational contribution allows the network to quantify each node’s effort using a standardized metric, such as FLOPS (floating-point operations per second). This way, rewards can be allocated based on actual work done, not just presence in the network. The team from @HyperspaceAI developed a Proof-of-FLOPS system, enabling nodes to lease unused computational capacity. In exchange, they earn flops—points that will function as the network’s currency. Here is how the architecture looks: 1. The process starts with a challenge delivered to the user, who responds by submitting a commitment for that challenge. 2. Hyperspace rollup manages the flow by securing submissions and fetching randomness from an oracle. 3. The user discloses the indices, and the challenge is finalized. 4. The operator checks the responses and notifies the Hyperspace AVS contracts of valid outcomes, which are then confirmed via EigenLayer contracts. 5. Liveness multipliers are calculated, and flops are granted to the user. Proving computational contribution provides a clear picture of each node’s capacity. Because of that, the system can assign tasks intelligently—giving complex AI computations to high-powered nodes and lighter tasks to less capable ones. The most interesting part here is making this system verifiable, so anyone can prove the correctness of the work that was done. Hyperspace has an AVS that constantly sends challenges, randomness requests, and multiple layers of validation processes, as seen in the diagram above. Operators can engage in the system knowing that the results are verified and rewards are fairly distributed. If the results aren’t correct, the malicious actor is obviously slashed. There are lots of reasons to verify the final AI computation: • To encourage nodes to join and contribute. • To distribute rewards proportionally based on effort. • To ensure contributions directly support certain AI models. • To allocate tasks effectively across nodes based on their proven capacity. → AI decentralization & verifiability As noted by @yb_effect, the terms "decentralized" and "distributed" are not similar at all. Distributed simply refers to hardware spread out across different locations, yet they still have a centralization point of connection. Decentralized means there’s no single main node, and the training process can handle failures, much like most blockchains work today. For AI networks to be truly decentralized, different solutions are applicable, but what we need for sure is verifiability of basically everything. If you want to build an AI model or agent, you want to make sure that every component and essentially every dependency is verified. Inference, training, data, oracles — everything can be verified to bring not only incentive-compatible crypto rewards into the AI system, but also to make it fair and efficient.