Author: Casey, Paradigm Investment Partner; Translation: Golden Finance xiaozou
I believe that openness brings innovation. In recent years, artificial intelligence has achieved rapid development and has global utility and influence. Since computing power increases with the integration of resources, artificial intelligence will naturally give rise to centralization problems, and the party with stronger computing power will gradually dominate. This will hinder our pace of innovation. I think decentralization and Web3 are strong contenders to keep AI open.
1, decentralized computing for pre-training and fine-tuning
Crowdsourcing computing (CPUs + GPUs)
Supporting comments: airbnb/uber’s crowdsourcing resource model may be expanded In the computing field, idle computing resources will be aggregated into a market. This may address issues such as providing cheaper compute resources for certain use cases (dealing with certain downtime/latency failures) and using censorship-resistant compute resources to train models that may be regulated or banned in the future.
Objection: Crowdsourced computing cannot achieve economies of scale; most high-performance GPUs are not owned by consumers. Decentralized computing is a complete paradox; it is actually the opposite of high-performance computing... Just ask any infrastructure/machine learning engineer if you don’t believe me!
Project examples: Akash, Render, io.net, Ritual, Hyperbolic, Gensyn
2, Decentralized reasoning
Run open source model reasoning in a decentralized manner
Supporting comments: The open source (OS) model is getting closer to the closed source model in some aspects and is gaining more and more adoption. Most people use centralized services like HuggingFace or Replicate to run OS model inference, which introduces privacy and censorship issues. One solution is to run inference through a decentralized or distributed provider.
Objection: There is no need to decentralize reasoning, local reasoning will be the ultimate winner. Specialized chips that can handle 7b+ parameter model inference are now being released. Edge computing is our solution for privacy and censorship resistance.
Project examples: Ritual, gpt4all (hosted), Ollama (web2), Edgellama (Web3, P2P Ollama), Petals
3, on-chainAIintelligent
Using machine learning On-chainapps
Support comments: AI agents (applications using AI) need a coordination layer to conduct transactions. For AI agents, it is natural to use cryptocurrency for payment, because it is digital technology itself, and obviously agents cannot open bank accounts through KYC certification. Decentralized human AI agents do not yet have platform risks. For example, OpenAI could suddenly decide to change their ChatGPT plugin architecture, which would break my Talk2Books plugin, without any prior notice. This is really happening. Agents created on the chain do not have such platform risks.
Objection: Agents are not production ready... at all. BabyAGI, AutoGPT, etc. are all toys! Additionally, for payments, the entity creating the AI agent can use the Stripe API without encrypted payments. As for the platform risk argument, it’s a well-worn use case for cryptocurrencies that we haven’t seen play out yet… Why is this time different?
Project Examples ：AI Arena, MyShell, Operator.io, Fetch.ai
4, data and model sources
Autonomous management and value collection of data and machine learning models
Supporting comments: The ownership of the data should belong to the person who generated the data users, not the companies collecting the data. Data is the most valuable resource in the digital age, but it is monopolized by large technology companies and has poor financial performance. The highly personalized web is coming, requiring portable data and models. We will move our data and models from one application to another over the Internet, just like we move our crypto wallets between different dapps. Data sources are a huge problem, especially as fraud is becoming more and more serious, and even Biden has acknowledged this. Blockchain architecture may well be the best solution to the data provenance puzzle.
Objection: No one cares about owning their data or privacy. We've seen this time and time again with user preferences. Just look at the number of signups on Facebook/Instagram! Eventually, people will trust OpenAI with their machine learning data. Let's face it.
Project examples: Vana, Rainfall
5, Generation Coin incentivesApps (such as companionapps)
ImagineCharacter.ai >Have crypto token rewards
Supporting comments: Crypto token incentives are very effective for bootstrapping networks and behaviors. We will see a large number of AI-centric applications adopt this mechanism. AI companions are a compelling market, and we believe this area will be a multi-trillion-dollar AI-native market. In 2022, Americans will spend more than 130 billion US dollars on pets; AI companion apps are pets 2.0. We have already seen that AI companion apps have achieved product-market fit, with the average session length on Character.ai being over 1 hour. We wouldn’t be surprised to see a crypto-incentivized platform take market share in this and other AI application verticals.
Objection: This is just an extension of the cryptocurrency speculative mania and will not last. Tokens are the customer acquisition cost of Web 3.0. Haven’t we learned a lesson from Axie Infinity?
Example projects: MyShell, Deva
6, Generation Coin-incentivized machine learning operations (such as training,RLHF, inference)
ImagineScaleAI with cryptographic tokens Coin rewards
Supporting comments: Crypto incentives can be used throughout the machine learning workflow to incentivize behaviors such as optimizing weights, fine-tuning, RLHF, etc. - with human judgment The output of the model for further fine-tuning.
Objection: MLOps (machine learning operations) are a poor use case for cryptocurrency rewards because quality is too important. While cryptographic tokens are good at incentivizing consumer behavior when entropy is no problem, they are not good at coordinating behavior when quality and accuracy are critical.
Project examples: BitTensor, Ritual
7, chain Verifiability on the chain (ZKML)
Prove which models can run effectively on-chain and plug into the crypto world
Support: On-chain model verifiability will unlock composability, which means you can take advantage of combined outputs in the DeFi and crypto fields. In 5 years, when we have agents running doctor models examining our bodies without having to go to the hospital to see a doctor, we will need to have some way of validating their knowledge and what model was specifically used for diagnosis. The verifiability of a model is like the reputation of intelligence.
Objection: No one needs to verify what model is being run. That's the least of our concerns. We are putting the cart before the horse. No one runs llama2 without fear of other models running in the background. This is a consequence of the fact that cryptography (zero-knowledge proofs) deliberately seeks to find a problem to solve, and the hype surrounding zero-knowledge proofs (ZK) garners a lot of venture capital funding.
Example projects: Modulus Labs, UpShot, EZKL
JinseFinance
Beincrypto
Nulltx
Cointelegraph