jakiro
Author: Casey, Paradigm Former Investment Partners; Translation: Bit Chain Vision Xiaozou
I believe opening up bring innovation.In recent years, artificial intelligence has achieved leap development and has global utility and influence.As computing power grows with the integration of resources, artificial intelligence will naturally give birth to centralized problems, and the party with stronger computing power will gradually occupy the dominant position.This will hinder our innovative pace.I think decentralization and web3 are powerful competitors to maintain artificial intelligence and openness.
1, Decentralized calculation for pre -training and fine -tuning
Crowdsourcing calculation (CPUS + GPUS.
Support opinion: The crowdfunding resource model of Airbnb/Uber may expand to the calculation field, and idle computing resources will aggregate into a market.This may solve the following problems: providing lower cost computing resources for certain cases (processing certain shutdown/delay faults); training resources with anti -review characteristics to train models that may be supervised or banned in the future.
Opposition: Cross -bag calculation cannot achieve scale economy; most high -performance GPUs are not owned by consumers.Decentralization calculation is a paradox; it actually stands on the opposite side of high -performance computing … If you don’t believe it, you can ask any infrastructure/machine learning engineer!
Project examples: Akash, RENDER, IO.NET, Ritual, Hyperbolic, Gensyn
2, Decentralized reasoning
Run the open source model reasoning in a decentralized manner
Support opinions: The open source (OS) model is getting closer to the closed source model in some aspects, and more and more adoption is obtained.Most people use centralized services to run OS model reasoning with centralized services such as HuggingFace or Replicate, thereby introducing privacy and review issues.There is a solution to operate inference through decentralization or distributed suppliers.
Opposition: There is no need to decentralize the reasoning, and local reasoning will become the final winner.It is now publishing special chips that can process 7B+parameter model reasoning.Edge computing is our solution in privacy and anti -review.
Project examples: Ritual, GPT4ALL (Hosted), Ollama (Web2), EDGELLAMA (Web3, P2P Ollama), PETALS
3On the chainAIIntelligent body
On the chain of machine learningapps
Support opinion: AI Smart (Application of AI) requires a coordinated layer for transactions.For AI intelligence, the use of cryptocurrencies for payment is logical, because it is a digital technology itself, and it is clear that the smart body cannot open a bank account through KYC certification.The decentralized AI smart body does not have platform risks.For example, OpenAI can suddenly decide to change their ChatGPT plug -in architecture, which will destroy my Talk2Books plug -in, but there is no advance notice.This happened.There is no such platform risk created on the chain.
Opposition: The agent has not been prepared for production … No.BABYAGI, AutoGPT, etc. are toys! In addition, for payment, entities that create artificial intelligence agents can use Stripe API without encrypted payment.For the argument of platform risks, this is an old -fashioned case of cryptocurrency. We haven’t seen it played … Why is this different time?
Project examples: AI ARENA, MyShell, Operator.io, FETCH.AI
4, Data and Model Source
Independent management and value collection of data and machine learning models
Support opinion: The ownership of the data should belong to the user of generating data, not a company that collects data.Data is the most valuable resource in the digital era, but it is monopolized by large technology companies and has poor financial performance.The highly personalized network is coming, which requires transplant data and models.We will bring our data and models from one application to another through the Internet, just like we turn our encrypted wallets between different DAPPs.The source of data is a huge problem, especially the phenomenon of fraud is becoming more and more serious, and even Biden has acknowledged this.The blockchain architecture is likely to be the best solution to solve the data source puzzle.
Opposition: No one cares whether to have their own data or privacy.We have seen this from the last time user preference.Take a look at the registration volume of Facebook/Instagram!In the end, people will trust Openai to provide their machine learning data.Let’s face reality.
Project example: VANA, Rainfall
5TokensApps(apps.
ImagineCharal.aiWith cryptocurrency rewards
Support opinions: Cryptoc currency incentives are very effective for starting the guidance network and behavior.We will see a large number of applications centered on artificial intelligence adopt this mechanism.AI partner is a striking market. We believe that the field will be a trillion -dollar AI native market.In 2022, Americans spent more than $ 130 billion in pets; AI companion APPs were pet 2.0.We have seen that the AI companion APP has achieved the product market fit, and the average session time of Character.ai is more than 1 hour.If we see a encrypted incentive platform occupying market share in this field and other AI applications, we will not be surprised.
Opposition: This is just the phenomenon of cryptocurrency speculative enthusiasm and will not last.Tokens is the cost of acquiring web 3.0. Have we not learned lessons from Axie Infinity?
Example items: MyShell, Deva
6, Token incentive machine learning operations (such as training,RLHF,reasoning)
ImagineScaleaiWith cryptocurrency rewards
Support opinions: Encrypture incentives can be used in the entire machine learning work process to motivate behaviors such as optimization of weight, fine -tuning, and RLHF -the output of the human judgment model to further fine -tuning.
Opposition: MLOPS (machine learning operation) is a bad case of cryptocurrency rewards, because quality is too important.Although cryptocurrencies are good at motivating consumer behavior when there is no problem with entropy, they are not conducive to coordinated behaviors when quality and accuracy are crucial.
Project example: Bittensor, Ritual
7, Verified on the chain (Zkml.
Prove which models can be effectively run on the chain and inserted into the encrypted world
Support opinions: The verification of the model on the chain will unlock the combined ability, which means that you can use a combination output in the DEFI and encryption field.Five years later, when we have the intelligent bodies that run the doctor model to check our body for us, we will need some ways to verify their knowledge and what models we use to diagnose them.The verification of the model is like the reputation of intelligence.
Opposition: No one needs to verify what model is running.This is what we don’t care about.We are upside down at the end of this.No one runs LLAMA2 but is afraid of running other models in the background.This is the consequences of the encryption technology (zero -knowledge certificate) intending to find a problem to solve, and the zero -knowledge certificate (ZK) speculation to obtain a large amount of venture capital funds.
Example items: Modulus Labs, UPSHOT, EZKL
