Decentralized AI
Below are the categories I find intriguing at the intersection of crypto x AI.
I believe openness begets innovation. In recent years, artificial intelligence has crossed the chasm into global utility and impact. Since computational strength grows with resource consolidation, AI naturally fosters centralization, where those with more computing power progressively dominate. This introduces a risk to our rate of innovation. I believe decentralization and Web3 stand as contenders to keep AI open.
This list and these example companies evolve every day. Please don’t take this as a source of truth but rather a snapshot in time. If I missed companies or you think I’m wrong, dm me on Twitter. Would love to debate.
Decentralized Compute for Pre-Training + Fine-tuning
Crowdsourced Compute (CPUs + GPUs)
Argument For: It’s possible the airbnb/uber model of crowdsourcing resources extends to compute and spare compute is aggregated into a marketplace. Problems this may solve: 1) cheaper compute for certain use cases that can handle some downtime/latency and 2) censorship-resistant compute to train models that may become regulated/outlawed in the future.
Argument Against: Crowdsourced compute doesn’t capture economies of scale; most performant GPUs are not owned by consumers. Decentralizing compute is a complete paradox; it’s literally the opposite of performant computing…ask any infra/ML engineer!
Example Projects: Akash, Render, io.net, Ritual, Hyperbolic, Gensyn
Decentralized Inference
Run inference of Open Source models in a decentralized manner
Argument For: Open source (OS) models are approaching parity with closed source in certain ways (1) and gaining adoption. To run inference of OS models, most people use a centralized service like HuggingFace or Replicate which introduces privacy and censorship concerns. One solution is running inference via a decentralized or distributed provider.
Argument Against: There is no need to decentralize inference because local inference is going to win. Specialized chips that can handle inference of 7b+ param models are now being released. Edge computing is our solution for privacy and censorship resistance.
Example Projects: Ritual, gpt4all (hosted), Ollama (web2), Edgellama (Web3, P2P Ollama), Petals
On-Chain AI Agents
On-chain apps that use machine learning
Argument For: AI Agents (apps that use AI) need a coordination layer to transact. It may make sense for AI agents to use crypto for payment as it’s natively digital and obviously agents cannot KYC to open bank accounts. Decentralized AI agents also don’t have platform risk. For example, OpenAI just randomly decided to change their ChatGPT plugin architecture which broke my Talk2Books plugin with no notice. True story. Agents built on-chain don’t have this same platform risk.
Argument Against: Agents are not production ready yet…at all. BabyAGI, AutoGPT etc. are toys! Also, for payments, the entity creating the AI agents can just use the Stripe API, there’s no need for crypto payments. And for the platform risk argument, that’s a trite use case of crypto that we still haven’t seen that play out…why is this time different?
Example Projects: AI Arena, MyShell, Operator.io, Fetch.ai
Data and Model Provenance
Self-sovereign your data and ML models, collect the value it produces
Argument For: Data should be owned by the users who generate the data, not the companies who collect it. Data is the most valuable resource of the digital age, yet its monopolized by big tech and poorly financialized. The hyper personalized web is coming and will require portable data and models. We will bring our data and models from one application to another throughout the internet, just like we bring our crypto wallet from dapp to dapp. Data provenance, especially with how deep fakes are progressing, is a huge problem, even Biden admits this. It’s very possible blockchain architecture is the best solution to solve the data provenance conundrum.
Argument Against: No one cares about owning their data or privacy. We have seen this time and time again through user preferences. Look at Facebook/Instagram registrations! In the end, people will trust OpenAI with their ML data. Let’s be realists.
Example Projects: Vana, Rainfall
Token Incentivized Apps (e.g. companion apps)
Think Character.ai with crypto token rewards
Argument For: Crypto token incentives are highly effective for bootstrapping networks and behavior. We will see AI-centric apps take advantage of this mechanism. One compelling market may be AI companions, which we believe will be a multi-trillion AI-native market. The US spent $130B+ on pets in 2022; AI companions are pets 2.0. We're already seeing AI companion apps reach PMF with Character.ai having an avg session time of >1hr+. We wouldn’t be surprised to see a crypto-incentivized platform take market share here and in other AI app verticals.
Argument Against: This is just an extension of crypto’s speculative mania and doesn’t yield durable usage. Tokens are CAC for Web 3.0. Haven’t we learned our lesson with Axie Infinity?
Example Projects: MyShell, Deva
Token Incentivized MLOps (e.g. Training, RLHF, inference)
Think ScaleAI with crypto token rewards
Argument For: Crypto incentives could be utilized throughout the ML workflow to incentivize behavior such as optimizing weights, finetuning, RLHF -- where humans judge the output of the model to finetune it further.
Argument Against: MLOps is a terrible use case for crypto rewards because the quality is too important. While crypto tokens are good at incentivizing consumer behavior where entropy is okay, they are bad for coordinating behavior where quality and accuracy is essential.
Example Projects: BitTensor, Ritual
On-Chain Verifiability (ZKML)
Prove what model ran efficiently on-chain and plug into cryptoverse
Argument For: Model verifiability on-chain will unlock composability, meaning you can leverage the output throughout DeFi and crypto. In 5 years, when we have agents running doctor models for us instead of going to the physical doctor, we will need some way to verify their knowledge and exactly what models were used in a diagnosis. Model verifiability is akin to reputation for intelligence.
Argument Against: No one needs to verify what model ran. This is the least of our concerns. We are putting the cart before the horse. No one runs llama2 and is scared a different model ran in background. This is crypto technology (zero-knowledge (ZK)) looking for a problem to solve and the aftermath of ZK getting too much hype and venture money.
Example Projects: Modulus Labs, UpShot, EZKL
