By @Rui, Investor of SevenX Ventures

Recap for Decentralized AI House at ETHDenver2024, co-hosted by SevenX Ventures, MyShell, and Jessy’s Hacker House

When AI is wildly growing in the web3 world, it’s even harder to distinguish real innovation from narrative bubbles. We invited 12 most mainstream AI mafias during ETHDenver and this article does a simple recap of the project's visions, methods, and scenarios, to see how they shaking the current world.

Here are some important questions to ask, and answers from our mafias:

Data:

• Data Provisioning: How to source AI training data？ @Grass

• Data provenance: How to protect data source IP? @Story Protocol

• Data alignment: How to ensure certain data used by the model? @Space&Time

Model:

• Open economics: How to build an open platform with incentive? @Bittensor, @Sentient

• Model alignment: How to prove model result is untampered? @Modulus labs, @Ora

Infra:

• General infrastructure: How to connect all infrastructures together? @Ritual

Agent:

• AI agent: How to enable agents to be smart, composable, and ownable? @Future primitive, @Olas, @Myshell

Full Youtube Link: https://www.youtube.com/playlist?list=PLFRYxG8q7EY6SgJHzEefMEq-VyrhzK20n

• Why: Data is the foundation of all AI training, but extractive gatekeepers make it difficult to source high quality training data. A great deal can be scraped from the public web, but it’s common practice among major websites to block commercial data centers.

• What: Grass is a data provisioning protocol, makes data accessible and AI infrastructure equitable.

• How: Users install a Chrome web extension and trace amounts of their excess compute and bandwidth are used to scan the internet in search of AI data. Grass operates a network of nearly 1 million web scraping nodes located around the world.  Using this network, it scrapes over 1 TB of data per day before cleaning and processing it to produce structured datasets.

• Where: Grass nodes are now operating in 190 countries around the globe.

• Youtube: https://www.youtube.com/watch?v=dwBlqwOimig&list=PLFRYxG8q7EY6SgJHzEefMEq-VyrhzK20n&index=2

• Presenter twitter: @0xdrej

• Why: AI remixing is illegal and inevitable. The major block of AI growth is the lack of monetizing and creating attribution and providence for IP and content creators.

• What: A composable on-chain IP layer allows creators to set autonomous rules of engagement. Adding legibility and liquidity for global IP marvel.

• How: Creators can purchase the license NFT, converting their static IP into programmable IP. Programmable IP is a layer that any program can read and write on, consisting of Nouns and Verbs. Nouns include data structures, relevant IP metadata, use ERC6551; And verbs include modules, an array of functionality for IP Assets, like licensing, revenue streams from derivative works, and access globally. Whenever the derivative monetizes, the revenue flows back automatically.

• Where: Story protocol can be used in license renting, derivatives, customization about territory, channels, expiration, revocability, transferability, attribution, etc.

• Youtube: https://www.youtube.com/watch?v=ymq1mhRSxTg&list=PLFRYxG8q7EY6SgJHzEefMEq-VyrhzK20n&index=3&t=201s

• Presenter twitter: @jasonjzhao

• Why: As the LLM evolves, major companies can bias, alt or tamper the dataset and parameters; it's important to have cryptographic proof of the untampered dataset, ensuring that same dataset was used during LLM training. Also, SxT has been exploring ways to sanitize out copyrighted data, pulling from a verifiable vector database and injecting into prompts during inference.

• What: SxT is an indexer and ZK prover that proves SQL queries or vector searches against that indexed data.

• How: LLM providers can load their on-chain/off-chain training dataset into Space and Time, where the data is witnessed and threshold signed with cryptographic commitments, which are used later to prove that a dataset was used for training. From there, litigators or auditors/reviewers can ensure a dataset hasn’t been tampered with since training. SxT built GPU accelerator "Blitzar", which already achieves 2 million-row table queries with 14-second proof time on a single GPU.

• Where: SxT enables users to create queries in plain words, in a few seconds OpenAI retrieves a context from the vector search database and writes accurate SxT SQL that can be executed by the prover, and the prover returns proof in 4 seconds.

• Youtube: https://www.youtube.com/watch?v=cxT0vcU4mSo&list=PLFRYxG8q7EY6SgJHzEefMEq-VyrhzK20n&index=4

• Presenter twitter: @chiefbuidl

• Why: OpenAI aims to monopolize the control of AI

• What: Bittensor is a decentralized platform for open-sourced AI

• How: The Bittensor network has 32 subnets. Those subnets started from the model, but now have extended to storage, computing, scraping, tracking, and different AI fields. $TAO incentivizes subnet builders to keep polishing the model or project, and validators rank the results of the subnet, the rank changes the distribution of $TAO, and the least one will be kicked out of the network. This mechanism guarantees models compete to produce the best outputs, and the work that is most valuable to the collective will be rewarded.

• Where: Powerful applications emerge, like FileTAO doing decentralized storage, Cortex TAO doing OpenAI Inference, Nous Research doing fine-tuning LLM; Fractal Research doing Decentralized Text-to-Video, etc.

• Youtube: https://www.youtube.com/watch?v=xkXBDCaPMYk&list=PLFRYxG8q7EY6SgJHzEefMEq-VyrhzK20n&index=6

• Why: AGI building is dangerous, facing "human extinction threat" and capitalistic framework risk, so they natively need crypto platforms; whereas the crypto platforms need native killer apps.

• What: Sentient is a platform for sovereign incentive-driven AI development.

• How: Use crowdsource methods, allowing communities to coordinate and contribute to training models to lower cost, control inference using an open protocol, enable composability across models, and value flow back to network participants. Aggregating web2 and web3 forces, and leveraging the token, Sentient will fairly incentivize developers to build trustless AGI.

• Youtube: https://www.youtube.com/watch?v=1fbwIGG7PV8&list=PLFRYxG8q7EY6SgJHzEefMEq-VyrhzK20n&index=10

• Why: When AGI's future is unstoppable, we need to prove the AI result is accountable and safe, which is generated from a certificated model instead of manipulated, without relying on the trusted centralized authority's good behavior.

• What: Modulus builds a specialized AI ZK prover "Remainder" which delivers AI features for dApps at a fraction of the cost.

• How: It doesn't make sense to use the modern ZK proving system for AI when there's around 10,000x to 100,000x overhead compared to non-verifiable AI outputs. Modulus built a custom prover for AI inference, demonstrating only ~180x overhead.

• Where: One notable implementation is Upshot, trust concern was raised since Upshot's complicated appraisal mode can only be conducted off-chain. However, Upshot can send valuable AI appraisals to Modulus every hour, and Modulus can generate "proof of correctness" for AI computations, aggregate them, and send them to Ethereum for final verification.

• Youtube: https://www.youtube.com/watch?v=4JRh2eeZCO0&list=PLFRYxG8q7EY6SgJHzEefMEq-VyrhzK20n&index=7

• Presenter twitter: @realDanielShorr

• Why: AI models can't run on-chain, thousands of computers will execute one inference. Verifying results on-chain is feasible but ZKML has an exponential cost with model size increases, so there's a need for linear growth cost- OPML.

• What: Ora is an onchain AI oracle using OPML for any size AI models.

• How: Use Oracle to delegate the computation to off-chain nodes. The user initiates a transaction from the smart contract with prompts and a named model. OAO contract delegates the tx into the OPML node to execute the inference, produce fraud proof and submit the verifier to ORO. The verified result goes back to the transaction initiator. But OPML still needs the help of ZKP to fulfill the input/output privacy and instant finalization. Also, a ZK oracle from ORA can generate a storage proof for the OPML, so you don't need to repeatedly perform OPML when it is being reused.

• Where: Now using ORA we can have Stable Diffusion, 7B-LLaMA on ethereum mainnet. ORA can power AI-managed DAO, AIGC NFT like EIP7007, empower model ownership.

• Youtube: https://www.youtube.com/watch?v=i2Zkz45AGr4&list=PLFRYxG8q7EY6SgJHzEefMEq-VyrhzK20n&index=1

• Presenter twitter: @0xKartin

• Why: New censorship and manipulation take place when AI infrastructure is more and more centralized, permissioned, and increasingly regulated. But crypto gives primitives around privacy and computational integrity, coordination and incentive, and default permissionless infra.

• What: Ritual as a natural convergence point of Crypto*AI, consists of a decentralized Oracle network and sovereign chain with custom VM and coprocessor.

• How: The Oracle network "Infernet" enables any smart contract on evm chains linking on-chain workflow to off-chain ML inference, and ultimately the co-processor enables AI native operation at VM level, while remains composability with other Data Availability, Permanent Storages, Prover networks, GPU networks, Inference engine. Nodes in Ritual network not only run and service model operations, but also do consensus and execution clients.

• Where: "Frenrug" uses inferent SDK to build a non-deterministic LLM that guides users to buy, and sell keys on friend.tech. For defi lending protocols, they can train models to parameterize and generalize them well across all protocols. For crypto-enabled AI, myshell will be the first model creator economy example.

• Youtube: https://www.youtube.com/watch?v=DV-rMP6V-Yk&list=PLFRYxG8q7EY6SgJHzEefMEq-VyrhzK20n&index=9

• Presenter twitter: @sanlsrni

• Why: Autonomous Agents(AA) are powerful entities that perceive certain information and conduct actions. But web2 AA is severely limited in its potential: They can't do KYC, users can't take ownership, platform censorship risks, and limited composability.

• What: Olas is a decentralized protocol for co-ownable autonomous agents.

• How: In Olas world, agents live off-chain, while the registration and management are on-chain. Agents are arranged in autonomous services (“decentralised autonomous agents”). The agent operator is in control of one agent and one consensus gadget. Each agent runs a Finite-State Machine(FSM) that is replicated on the temporary blockchain between agents in a service. Agents in a service come to consensus off-chain before taking action on-chain. The network uses a $OLAS incentives for all stakeholders: capital providers (bonders), code providers (developers), agent operators (stakers), and service owners (entrepreneurs).

• Where: Olas Predict is an economy of three types of autonomous agents that continuously create and participate in prediction markets on arbitrary future events. The trader agent operator doesn't need to worry about subscribing to OpenAI, they just simply pay per request in crypto.

• Youtube: https://www.youtube.com/watch?v=KQeQgqFAQfI&list=PLFRYxG8q7EY6SgJHzEefMEq-VyrhzK20n&index=8

• Presenter twitter: @david_enim

• Why: With the rise of LLM, we need better tools to help creators easily build applications, also subversing the current "static" creator economy into a dynamic space.

• What: MyShell is a decentralized platform for discovering, creating, and staking AI-native apps.

• How: Creator can create AI-native Apps in MyShell within a few minutes, from AI characters for companionship to various tools designed to enhance your study and work experience: choose an open source model, edit the prefix and suffix of the prompt, feed information with specific fields, image, and video capability coming soon. MyShell LLM is based on massive private data to make the roleplay experience more human-like. The token is used for accessing premium features, supporting creators, and settling usage charges.

• Where: MyShell platform has multiple use cases: AI characters have personalities and unique voices for companionship, learning, and playing; Tools for learning language, text-to-image, summarizing videos, and so on.

• Youtube: https://www.youtube.com/watch?v=4PgIk4S5a0w&list=PLFRYxG8q7EY6SgJHzEefMEq-VyrhzK20n&index=5

• Presenter twitter: @ethan_myshell

• Why: NFTs are internet-native objects, but they lack programmability, therefore constrain its potential to conduct further behavior on-chain.

• What: Future Primitives turns NFTs into agents through ERC6551 and its infrastructure.

• How: Using ERC6551, every NFT has the same rights as an Ethereum user. They can self-custody assets, execute arbitrary operations, control multiple independent accounts. Every NFT's address is the same on any evm chain, releasing the cross-chain potential. For tokenbound V4, it publishes the Power of Attorney for your TBA of which a smart contract can perform onchain actions fully autonomously without you needing to sign, validate, or execute the tx, and the owner can always revoke auth extensions.

• Youtube: https://www.youtube.com/watch?v=aIr2HeZsQi4&list=PLFRYxG8q7EY6SgJHzEefMEq-VyrhzK20n&index=11

• Presenter twitter: @jayden_windle

By @Rui, Investor of SevenX Ventures

Recap for Decentralized AI House at ETHDenver2024, co-hosted by SevenX Ventures, MyShell, and Jessy’s Hacker House

When AI is wildly growing in the web3 world, it’s even harder to distinguish real innovation from narrative bubbles. We invited 12 most mainstream AI mafias during ETHDenver and this article does a simple recap of the project's visions, methods, and scenarios, to see how they shaking the current world.

Here are some important questions to ask, and answers from our mafias:

Data:

• Data Provisioning: How to source AI training data？ @Grass

• Data provenance: How to protect data source IP? @Story Protocol

• Data alignment: How to ensure certain data used by the model? @Space&Time

Model:

• Open economics: How to build an open platform with incentive? @Bittensor, @Sentient

• Model alignment: How to prove model result is untampered? @Modulus labs, @Ora

Infra:

• General infrastructure: How to connect all infrastructures together? @Ritual

Agent:

• AI agent: How to enable agents to be smart, composable, and ownable? @Future primitive, @Olas, @Myshell

Full Youtube Link: https://www.youtube.com/playlist?list=PLFRYxG8q7EY6SgJHzEefMEq-VyrhzK20n

• Why: Data is the foundation of all AI training, but extractive gatekeepers make it difficult to source high quality training data. A great deal can be scraped from the public web, but it’s common practice among major websites to block commercial data centers.

• What: Grass is a data provisioning protocol, makes data accessible and AI infrastructure equitable.

• How: Users install a Chrome web extension and trace amounts of their excess compute and bandwidth are used to scan the internet in search of AI data. Grass operates a network of nearly 1 million web scraping nodes located around the world.  Using this network, it scrapes over 1 TB of data per day before cleaning and processing it to produce structured datasets.

• Where: Grass nodes are now operating in 190 countries around the globe.

• Youtube: https://www.youtube.com/watch?v=dwBlqwOimig&list=PLFRYxG8q7EY6SgJHzEefMEq-VyrhzK20n&index=2

• Presenter twitter: @0xdrej

• Why: AI remixing is illegal and inevitable. The major block of AI growth is the lack of monetizing and creating attribution and providence for IP and content creators.

• What: A composable on-chain IP layer allows creators to set autonomous rules of engagement. Adding legibility and liquidity for global IP marvel.

• How: Creators can purchase the license NFT, converting their static IP into programmable IP. Programmable IP is a layer that any program can read and write on, consisting of Nouns and Verbs. Nouns include data structures, relevant IP metadata, use ERC6551; And verbs include modules, an array of functionality for IP Assets, like licensing, revenue streams from derivative works, and access globally. Whenever the derivative monetizes, the revenue flows back automatically.

• Where: Story protocol can be used in license renting, derivatives, customization about territory, channels, expiration, revocability, transferability, attribution, etc.

• Youtube: https://www.youtube.com/watch?v=ymq1mhRSxTg&list=PLFRYxG8q7EY6SgJHzEefMEq-VyrhzK20n&index=3&t=201s

• Presenter twitter: @jasonjzhao

• Why: As the LLM evolves, major companies can bias, alt or tamper the dataset and parameters; it's important to have cryptographic proof of the untampered dataset, ensuring that same dataset was used during LLM training. Also, SxT has been exploring ways to sanitize out copyrighted data, pulling from a verifiable vector database and injecting into prompts during inference.

• What: SxT is an indexer and ZK prover that proves SQL queries or vector searches against that indexed data.

• How: LLM providers can load their on-chain/off-chain training dataset into Space and Time, where the data is witnessed and threshold signed with cryptographic commitments, which are used later to prove that a dataset was used for training. From there, litigators or auditors/reviewers can ensure a dataset hasn’t been tampered with since training. SxT built GPU accelerator "Blitzar", which already achieves 2 million-row table queries with 14-second proof time on a single GPU.

• Where: SxT enables users to create queries in plain words, in a few seconds OpenAI retrieves a context from the vector search database and writes accurate SxT SQL that can be executed by the prover, and the prover returns proof in 4 seconds.

• Youtube: https://www.youtube.com/watch?v=cxT0vcU4mSo&list=PLFRYxG8q7EY6SgJHzEefMEq-VyrhzK20n&index=4

• Presenter twitter: @chiefbuidl

• Why: OpenAI aims to monopolize the control of AI

• What: Bittensor is a decentralized platform for open-sourced AI

• How: The Bittensor network has 32 subnets. Those subnets started from the model, but now have extended to storage, computing, scraping, tracking, and different AI fields. $TAO incentivizes subnet builders to keep polishing the model or project, and validators rank the results of the subnet, the rank changes the distribution of $TAO, and the least one will be kicked out of the network. This mechanism guarantees models compete to produce the best outputs, and the work that is most valuable to the collective will be rewarded.

• Where: Powerful applications emerge, like FileTAO doing decentralized storage, Cortex TAO doing OpenAI Inference, Nous Research doing fine-tuning LLM; Fractal Research doing Decentralized Text-to-Video, etc.

• Youtube: https://www.youtube.com/watch?v=xkXBDCaPMYk&list=PLFRYxG8q7EY6SgJHzEefMEq-VyrhzK20n&index=6

• Why: AGI building is dangerous, facing "human extinction threat" and capitalistic framework risk, so they natively need crypto platforms; whereas the crypto platforms need native killer apps.

• What: Sentient is a platform for sovereign incentive-driven AI development.

• How: Use crowdsource methods, allowing communities to coordinate and contribute to training models to lower cost, control inference using an open protocol, enable composability across models, and value flow back to network participants. Aggregating web2 and web3 forces, and leveraging the token, Sentient will fairly incentivize developers to build trustless AGI.

• Youtube: https://www.youtube.com/watch?v=1fbwIGG7PV8&list=PLFRYxG8q7EY6SgJHzEefMEq-VyrhzK20n&index=10

• Why: When AGI's future is unstoppable, we need to prove the AI result is accountable and safe, which is generated from a certificated model instead of manipulated, without relying on the trusted centralized authority's good behavior.

• What: Modulus builds a specialized AI ZK prover "Remainder" which delivers AI features for dApps at a fraction of the cost.

• How: It doesn't make sense to use the modern ZK proving system for AI when there's around 10,000x to 100,000x overhead compared to non-verifiable AI outputs. Modulus built a custom prover for AI inference, demonstrating only ~180x overhead.

• Where: One notable implementation is Upshot, trust concern was raised since Upshot's complicated appraisal mode can only be conducted off-chain. However, Upshot can send valuable AI appraisals to Modulus every hour, and Modulus can generate "proof of correctness" for AI computations, aggregate them, and send them to Ethereum for final verification.

• Youtube: https://www.youtube.com/watch?v=4JRh2eeZCO0&list=PLFRYxG8q7EY6SgJHzEefMEq-VyrhzK20n&index=7

• Presenter twitter: @realDanielShorr

• Why: AI models can't run on-chain, thousands of computers will execute one inference. Verifying results on-chain is feasible but ZKML has an exponential cost with model size increases, so there's a need for linear growth cost- OPML.

• What: Ora is an onchain AI oracle using OPML for any size AI models.

• How: Use Oracle to delegate the computation to off-chain nodes. The user initiates a transaction from the smart contract with prompts and a named model. OAO contract delegates the tx into the OPML node to execute the inference, produce fraud proof and submit the verifier to ORO. The verified result goes back to the transaction initiator. But OPML still needs the help of ZKP to fulfill the input/output privacy and instant finalization. Also, a ZK oracle from ORA can generate a storage proof for the OPML, so you don't need to repeatedly perform OPML when it is being reused.

• Where: Now using ORA we can have Stable Diffusion, 7B-LLaMA on ethereum mainnet. ORA can power AI-managed DAO, AIGC NFT like EIP7007, empower model ownership.

• Youtube: https://www.youtube.com/watch?v=i2Zkz45AGr4&list=PLFRYxG8q7EY6SgJHzEefMEq-VyrhzK20n&index=1

• Presenter twitter: @0xKartin

• Why: New censorship and manipulation take place when AI infrastructure is more and more centralized, permissioned, and increasingly regulated. But crypto gives primitives around privacy and computational integrity, coordination and incentive, and default permissionless infra.

• What: Ritual as a natural convergence point of Crypto*AI, consists of a decentralized Oracle network and sovereign chain with custom VM and coprocessor.

• How: The Oracle network "Infernet" enables any smart contract on evm chains linking on-chain workflow to off-chain ML inference, and ultimately the co-processor enables AI native operation at VM level, while remains composability with other Data Availability, Permanent Storages, Prover networks, GPU networks, Inference engine. Nodes in Ritual network not only run and service model operations, but also do consensus and execution clients.

• Where: "Frenrug" uses inferent SDK to build a non-deterministic LLM that guides users to buy, and sell keys on friend.tech. For defi lending protocols, they can train models to parameterize and generalize them well across all protocols. For crypto-enabled AI, myshell will be the first model creator economy example.

• Youtube: https://www.youtube.com/watch?v=DV-rMP6V-Yk&list=PLFRYxG8q7EY6SgJHzEefMEq-VyrhzK20n&index=9

• Presenter twitter: @sanlsrni

• Why: Autonomous Agents(AA) are powerful entities that perceive certain information and conduct actions. But web2 AA is severely limited in its potential: They can't do KYC, users can't take ownership, platform censorship risks, and limited composability.

• What: Olas is a decentralized protocol for co-ownable autonomous agents.

• How: In Olas world, agents live off-chain, while the registration and management are on-chain. Agents are arranged in autonomous services (“decentralised autonomous agents”). The agent operator is in control of one agent and one consensus gadget. Each agent runs a Finite-State Machine(FSM) that is replicated on the temporary blockchain between agents in a service. Agents in a service come to consensus off-chain before taking action on-chain. The network uses a $OLAS incentives for all stakeholders: capital providers (bonders), code providers (developers), agent operators (stakers), and service owners (entrepreneurs).

• Where: Olas Predict is an economy of three types of autonomous agents that continuously create and participate in prediction markets on arbitrary future events. The trader agent operator doesn't need to worry about subscribing to OpenAI, they just simply pay per request in crypto.

• Youtube: https://www.youtube.com/watch?v=KQeQgqFAQfI&list=PLFRYxG8q7EY6SgJHzEefMEq-VyrhzK20n&index=8

• Presenter twitter: @david_enim

• Why: With the rise of LLM, we need better tools to help creators easily build applications, also subversing the current "static" creator economy into a dynamic space.

• What: MyShell is a decentralized platform for discovering, creating, and staking AI-native apps.

• How: Creator can create AI-native Apps in MyShell within a few minutes, from AI characters for companionship to various tools designed to enhance your study and work experience: choose an open source model, edit the prefix and suffix of the prompt, feed information with specific fields, image, and video capability coming soon. MyShell LLM is based on massive private data to make the roleplay experience more human-like. The token is used for accessing premium features, supporting creators, and settling usage charges.

• Where: MyShell platform has multiple use cases: AI characters have personalities and unique voices for companionship, learning, and playing; Tools for learning language, text-to-image, summarizing videos, and so on.

• Youtube: https://www.youtube.com/watch?v=4PgIk4S5a0w&list=PLFRYxG8q7EY6SgJHzEefMEq-VyrhzK20n&index=5

• Presenter twitter: @ethan_myshell

• Why: NFTs are internet-native objects, but they lack programmability, therefore constrain its potential to conduct further behavior on-chain.

• What: Future Primitives turns NFTs into agents through ERC6551 and its infrastructure.

• How: Using ERC6551, every NFT has the same rights as an Ethereum user. They can self-custody assets, execute arbitrary operations, control multiple independent accounts. Every NFT's address is the same on any evm chain, releasing the cross-chain potential. For tokenbound V4, it publishes the Power of Attorney for your TBA of which a smart contract can perform onchain actions fully autonomously without you needing to sign, validate, or execute the tx, and the owner can always revoke auth extensions.

• Youtube: https://www.youtube.com/watch?v=aIr2HeZsQi4&list=PLFRYxG8q7EY6SgJHzEefMEq-VyrhzK20n&index=11

• Presenter twitter: @jayden_windle