The Case For Decentralized AI: Blockchain Based AI Infrastructure & Applications
Much like the internet, AI is an exponential technology that is poised to transform almost every industry in the world. Similarly, crypto (what I collectively call bitcoin, crypto, NFTs and web3) is also going to transform almost every industry in the world.
For the first time, these two technologies are intersecting and creating an exponential tornado of new experimentation, use cases, addressable markets and new technologies. The OpenAI coup last year made it clear that control and monopoly centralization over AI is a huge deal. Others like Balaji are taking a stand that AI must be decentralized
We have just begun to scratch the surface of what is possible and how these two technologies will interact. The following are some categories of overlap we are excited about at Stateless Ventures.
CRYPTO x AI USE CASES:
- Open Source Models
- Decentralized Resource Sharing
- Verified Compute/On-chain Inference
- Decentralized AI Agents
- AI Blockspace (Ex: Bittensor)
- Decentralized Tech Stack (Ex: file storage, decentralized GPU networks, decentralized RPCs, and AI social graphs)
First: Underestimating the Total Addressable Market for Crypto
Crypto is often cited as having 100m users in total — underwhelming for the hype and market caps attributed to the top blockchain protocols. However, unlike social apps, online marketplaces (Uber, Airbnb), crypto networks can onboard and service AI/agent users. If we begin to calculate the total daily active users (DAUs) and total addressable market (TAM) to include AI and AI agents, the figures start to look massive. At an average of 10 agents for every human, thats 80 billion potential users. Combined with the speed of distribution and these applications can grow exceptionally fast. (For example, OpenAI reached 100m users within two months after launch).
OUR CORE FOCUS AREAS OF DECENTRALIZED AI
1. OPEN SOURCE MODELS
Decentralized AI would not be possible without the creation of open source large language model (LLM) AI models. Much like Bitcoin could not exist without open source codebase running on multiple open source technologies such as SHA 256, consensus layer, Merkle root trees, etc, Decentralized AI cannot run without open source AI models such as Llama 2 (Meta), Mistral, and Falcon.
We expect many more new models to be made open source or even leaked from top AI companies in the interest of advancing humanity.
2. DECENTRALIZED HOSTING & RESOURCE SHARING
Decentralized GPU networks such as Render, Akash, and Nosana. These networks onboard the long tail of idle GPUs and CPUs around the world and harness it into a usable GPU compute as a service network for AI inference, models and applications to run on.
Three benefits:
- Cheaper cost of compute: by leveraged a decentralized incentive based network, the cost of traditional compute can be reduced significantly.
- Unstoppable, un-censorable hosting for AI models and uptime (the antithesis to OpenAI on AWS servers).
- Balancing the load of the centralized GPU infrastructure.
3. VERIFIED COMPUTE
Verified compute is a new category of computation that allows one to verify a computation has been done correctly off chain and verify it on-chain. Often via a ZK proof, this allows for almost infinite scaling of applications on chain. The ability to use the speed and compute power of off-chain computers/GPUs and have the output be a trusted verified output enables all existing web applications to be run “on chain.” This means you can run an AI model such as Llama 2 or even chatGPT and trust that the network running the model ran the correct LLM to generate your answer.
4. ON-CHAIN INFERENCE
In AI, inference is the act of asking an LLM a question and receiving an answer. Using ChatGPT you are generating an inference, asking it questions and receiving the inference as an answer. Unlike with model training, these LLM’s have already been built.
Corcel, a clone of ChatGPT running on the Bittensor network is one of the best examples of a decentralized AI inference application at the moment.
5. DECENTRALIZED AI AGENTS
As AI and AI agents proliferate, there will be a deep need to track and audit how AI agents are acting on our (humans) behalf. In the not-too-distant future, there will be multiples more AI agents than human beings in the world. We estimate there will be at minimum 10 active AI agents at any given time for every human. With many AI agents being spun up and shut down repeatedly resulting in a never ending supply of AI agents.
AI agents need crypto to transact and interact on chain. AI’s will need a way to hold money and transact online. The crypto industry has been building decentralized financial and legal (contracts) infrastructure for oover a decade. Crypto assets will be the preferred method of conducting business online for Agents.
AI agents are extremely early in their development. There are projects (such as ChainML) that are building AI agents to help humans in a number of use cases.
Today, DeFi apps are already primitive versions of autonomous AI agents. — on chain smart contracts that perform financial services without the need for humans. (Ex. Aave, Uniswap, Link etc). This is a very primitive form of an agent. You ask Aave to lend out your money, it sets the interest rate variably based on demand, liquidates users who over-leverage and accrues fees to Aave token holders.
Future AI agents include projects like Autonolas who are building on chain agents to go out and read, analyze real world data and perform actions on the web and on chain for its owners.
There have also been exciting experiments in AI agents as social profiles/artists and we see future AI agents as in-game NFTs or non-player characters (NPCs).
6. AI BLOCKSPACE
AI blockspace is a new concept that is an evolution of the blockchain based virtual machine.
First we had bitcoin miners creating a very simplistic decentralized triple entry accounting ledger. Maintaining account balances by providing hashing compute to secure the bitcoin network.
Second, Ethereum created the Ethereum Virtual Machine (EVM) where developers could build Turing complete applications comprised of a series of interconnected smart contracts.
Third, we argue, Bittensor has created a generalized AI blockchain creating a new type of block-space resource that allows anybody to build and contribute to the proliferation of AI. This includes AI inference, apps, train new models, prediction (subnet 8), 3D rendering, machine to machine translation and audio generation. All of the use cases above can be built out in a subnet on the Bittensor network. The native token of Bittensor (TAO) is used to manage and create a market for the resources on the Bittensor Network.
Subnet owners pay TAO in rent/fees to register and operate their subnet. Miners receive block rewards in TAO for providing the best output for the requested subnet task. Validators determine the best output for each subnet task and allocate the miner reward. Validators receive a % fee for providing the validation.
For more on Bittensor read the messari report:
7. OTHER DECENTRALIZED AI INFRASTRUCTURE (SOVEREIGN DIGITAL INFRASTRUCTURE)
These are the different layers of the tech stack needed to power the decentralized AI. Creating a decentralized tech stack is important because it reduces the power (and potential harm) any centralized party has over the system, while aligning incentives across a broad base of contributors in a democratic fashion.
Examples include file storage (Filecoin, Arweave, and Shadow), decentralized RPCs (Pocket Network), web hosting, and AI social graphs.
BONUS: Where do we go from here?
The intersection of Crypto and AI has just begun. We can’t imagine the things that will be built over the next decade. Some ideas I think will likely get built:
- GPU Futures
- AI Owned Crypto Games and NPCs
- AR DeGINs: (Decentralized Gamified Incentive Networks):
- Social Networks for AI’s
- AI Celebrities
- AI Agent Website logins
More on the list above in a future post. The reality is most of the best break out use cases we can’t imagine in January of 2024. We will have to watch this space closely to see what ignites.
We are so excited to be able to participate and play a part in bringing about this future.
— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — —
Resources:
https://a16z.com/what-builders-talk-about-when-they-talk-about-ai/
https://a16z.com/the-economic-case-for-generative-ai-and-foundation-models/
https://www.plaintextcapital.com/blog/aiandsolana/
Sequoia: https://www.sequoiacap.com/article/generative-ai-a-creative-new-world/
https://arxiv.org/abs/1706.03762
https://messari.io/report/decentralizing-machine-learning
________________________________________________________________
The information herein was prepared by Stateless Ventures LLC (“Stateless”) and is believed by Stateless to be reliable and has been obtained from public sources believed to be reliable. Stateless makes no representation as to the accuracy or completeness of such information. Opinions, estimates and projections in this presentation constitute the current judgment of Stateless and are subject to change without notice. This is not an offering or the solicitation of an offer to purchase an interest in any fund managed by Stateless (collectively, the “Fund”). Any such offer or solicitation will only be made to qualified investors by means of a confidential private placement memorandum and only in those jurisdictions where permitted by law. The identified investment does not represent or approximate all of the investments purchased, sold, or recommended for the Fund. It should not be assumed that recommendations made in the future will be profitable or will equal the performance of the identified investment. An investment in the Fund is speculative and involves a high degree of risk. Opportunities for withdrawal, redemption, and transferability of interests are restricted, so investors may not have access to capital when it is needed. There is no secondary market for the interests, and none is expected to develop. The fees and expenses charged in connection with this investment may be higher than the fees and expenses of other investment alternatives and may offset profits. No assurance can be given that the investment objective will be achieved or that an investor will receive a return of all or part of his or her investment. Investment results may vary substantially over any given time period.
