In-depth Analysis of PIN AI, the Latest a16z Investment: Reshaping the AI Landscape with Web3

PIN AI is an open AI network where developers can build useful AI applications.

Written by: Hei Mi

It is an honor to have completed this research under the guidance of Meteorite Labs, based on the experience and communication with hundreds of Web2 AI applications.

PIN AI is a selected project of the a16z Crypto Startup Accelerator Fall Program, with a seed round financing of 10 million USD. In addition to a16z Crypto, well-known VCs include Stanford Blockchain Accelerator, Hack VC, and Foresight Ventures. Angel investors include NEAR Protocol founder Illia Polosukhin, Gitcoin co-founder Scott Moore, Solana Foundation Chair Lily Liu, SUI/Mysten Labs CEO Evan Cheng, Worldcoin research engineer DCBuilder, and others.

I just finished reading an article co-authored by the three co-founders of PIN AI and found it to be the most attractive Web3 AI project to me recently, aside from Sahara, with very interesting application scenarios.

PIN AI is an open AI network where developers can build useful AI applications. "Useful AI Applications"—useful is the product focus. It is somewhat similar to Web2 AI Agents like MultiOn and Jace.ai, aiming to provide users with applications useful for daily life, fulfilling user intentions such as online shopping, travel planning, and investment planning.

Let me briefly introduce Jace.ai, which is an AI Agent capable of autonomously completing browser tasks, based on LLM and its proprietary model AWA-1 (Autonomous Web Agent-1), which supports AI to perform operations on web pages.

Jace's greatest capability is its ability to autonomously plan tasks and perform operations in the browser on behalf of the user.

For example, to understand the application scenario, you can tell Jace, "I'm planning to travel to Beijing for a week on September 20th, with a budget of 5,000 yuan, please help me plan." Jace will automatically create a travel plan, including attractions to visit, hotels to stay at, and food to eat. If I agree to the plan, it will then help me book all the attractions and find the most cost-effective hotel on Meituan and place the order. I only need to enter my personal information and click to pay.

In fact, what PIN AI aims to do is very similar. The biggest difference from generative AI is that this type of AI project mainly focuses on users' daily lives rather than work.

1. Deconstructing PIN AI's Design Concept

Simply put, PIN AI = AI + DePIN

The PIN AI network consists of two types of AI:

• Personal AI: A personalized AI Agent that can adapt to user preferences in real time. It is the connection point between the user and agentic services, somewhat like a coordinator. Users can download it to their mobile phones or computers.
• Agentic Services: AI Agents built on-chain for Web2 platforms, capable of executing tasks on some top Web2 platforms, with the execution process and completion status recorded on the blockchain.
• The official documentation also mentions External AI, which may support interaction with other LLMs or Web2 AI Agents in the future.

Core of the PIN AI Architecture:

PIN Protocol, a DePIN distributed data storage network, allows anyone to connect their devices to the network and share data. It integrates a BERT-based model that anonymizes data at every stage of user data processing, ensuring privacy and compliance with data protection regulations.

Personal AI is built on top of this. On one hand, it provides personalized data for Personal AI, and on the other, it supplies the most relevant data for Agentic Services.

PIN Protocol is built from three components:

1. Private Storage and Computation Layer: Distributed storage of data, securely storing device data shared by users (including photos, videos, etc.), and making the most relevant data available for Personal AI and Agentic Services at any time. Users can connect their devices to the network, provide device data, and receive native token $PIN rewards.

2. Data Connector: Uses zk technology to track and verify user data connected to the network. I believe this is equivalent to the nodes of the PIN network. Node operators need to stake $PIN tokens for validation, and some token holders can stake tokens to nodes. Both can receive staking rewards.

3. Agentic Link: Designed to match Personal AI with Agentic Services. It consists of an agent registry and a transaction mechanism. The former tracks performance metrics, while the latter "thinks" about how to match Personal AI with Agentic Services (based on each agentic service's cost, performance, and completion quality).

User Usage Pattern / Business Logic:

When users make a specific request, PIN AI proceeds as follows:

Step 1: Personal AI — Collects the user's request

The user submits a request to Personal AI, which forwards it to the PIN Protocol.

Step 2: PIN Protocol — Prepares for task execution

Breaks down the user's intent into specific steps, finds the most suitable and cost-effective Agentic Service, retrieves the most relevant data, and provides it to the Agentic Service. (If multiple Web2 platforms are involved, the intent needs to be split among different Agentic Services.)

Step 3: Agentic Service — Executes specific steps

Step 4: PIN Protocol — Feeds back the results to the user

Since most daily life needs involve monetary transactions, in PIN AI, the flow of funds should be as follows:

The user pays a gas fee to the PIN Protocol (I guess this is to activate the intent transaction). Since the PIN Protocol first breaks down the user's intent and then indexes and sends the most relevant data to the Agentic Service, after the Agentic Service completes the task, it will give part of the service fee as a tip to the PIN Protocol.

Therefore, both the PIN Protocol and Agentic Services can take a cut from the service fee paid by the user.

For example:

The user can download Personal AI to their computer or phone, make a request to Personal AI such as "buy the cheapest GTX 3080 graphics card on Amazon," and pay the fee (purchase fee + service fee + PIN Protocol gas fee).

Personal AI forwards this request to the PIN Protocol.

After understanding and breaking down the user's intent, the PIN Protocol will split the intent into specific task steps and send them, along with the most relevant data, to the Agentic Service. There may be dozens of Agentic Services specializing in Amazon shopping, so the PIN Protocol needs to comprehensively consider their cost, performance, and past completion to select the most suitable one.

The Agentic Service finds the most cost-effective GTX 3080 graphics card on Amazon and places the order. After completion, it pays the intent breakdown fee and data call fee to the PIN Protocol. The PIN Protocol and Personal AI feed back the result to the user and send PIN tokens as a reward.

Network Participants

Personal AI users: Install Personal AI on their computer or phone, connect personal data to the PIN Protocol, and can receive PIN token rewards.

Value transfer users: As in the usage pattern above, users who conduct valuable transactions will also receive PIN token rewards.

PIN Protocol nodes: Track and verify user data connected to the network. Operators need to stake, and token holders can stake tokens to nodes. Both can receive staking rewards.

Agentic Services: Developers can earn service fees.

2. Core Development Team

Davide Crapis - Co-founder

Background in blockchain protocol design, with some AI experience.

Previously served as a senior data scientist at Lyft, designing and implementing incentive allocation algorithms that distributed xx USD in growth incentives to passengers and drivers annually. After resigning, he spent some time as an independent researcher, studying incentive schemes and token distribution. Before founding PIN AI, he was a research scientist focusing on "robust incentives" at the Ethereum Foundation.

Developed a machine learning model for "consumer sensitivity to investment/credit product interest rates" and served as a machine learning researcher and mentor at Columbia Business School for four years. Joined the Web2 developer community "South Park" to explore the intersection of large language models and blockchain.

Ben Wu - Co-founder

Operations background, possibly providing strategic direction and AI product ideas.

Graduated from MIT, Y Combinator alumnus. Before founding PIN AI, he was the director of database and operations in Yahoo's Strategic Data Solutions department, responsible for the operation and management of large-scale data projects.

Bill Sun - Co-founder, Chief Scientist

Background in quantitative trading and AI.

Ph.D. in Mathematics from Stanford University, previously conducted AI research at Google DeepMind. Worked as an AI/quantitative trading equity investment manager at a Wall Street asset management firm. Founded the AI research organization AI+Club and the AI technology community AGI House. Angel investor of the a16z scout fund. He is also the founder of Generative Alpha, which provides enterprise-level AI solutions.

3. Reflections and Summary

The first industrial revolution freed our hands with machinery;

The second industrial revolution broke the boundaries of day and night with electricity;

The third industrial revolution merged the boundaries of virtual and reality with the internet.

The emergence of AI is widely regarded as the hallmark of the fourth industrial revolution, and AI Agents are the tickets for this journey of exploration. Each of us can board this giant ship heading toward the future of "human-computer interaction."

Over the past decades, massive activities and data have been generated on the internet every day, yet users do not own these data.

The iPhone 16 has just been released, bringing Apple Intelligence, but PIN AI has the opportunity to build a more open AI Agent ecosystem than Apple Intelligence.

In this ecosystem, developers can earn rewards by developing innovative Web2 platform agentic services, which will give rise to increasingly high-performance and high-quality AI Agents, sparking a wave of innovation.

And billions of mobile users can not only use personalized Personal AI but also share device data to earn rewards.

User data supports the entire PIN AI ecosystem—this is the power of users and the starting point of Web3: decentralization and ownership.

Hopefully, we will soon see the launch of the PIN AI network and whether the incentive mechanism can effectively function, attracting a large number of open-source contributors and creating a larger wave of innovation. The testnet may launch in October, with the mainnet and TGE going live in January next year—something to look forward to.