Reddit data DAO and everything to know about Gen AI model training

The Cryptonomist · 2024-07-20T07:11:14.000Z

The Cryptonomist interviewed Anna Kazlauskas, CEO and Co-founder of Vana’s, which Reddit Data DAO, which in the first week saw 140k users sign up with verified Reddit accounts. Anna now is working with developers to build Data DAOs for other platforms, like LinkedIn and ChatGPT. In addition to DAOs, they have other outlets for users to pool their data into datasets that can then be used for GenAI model training, such as creating portraits or avatars. Beyond what Vana is doing, with Anna we talked about the growth of the decentralized AI space, as platforms help people use and monetize their data for new applications. Can you provide an overview of Vana and its mission in decentralized AI space? Vana is a user-owned AI platform powered by user-owned data. Our mission is for users to own their data and the value it creates through AI models. There’s a growing need for more training data to improve AI model performance, as ultimately AI models are only as good as their data. For example, LLaMA 3 is trained on about 15 trillion words, which is roughly the amount of data available on the public internet. Companies are now trying to acquire more data, sometimes paying hundreds of millions of dollars for it. Major tech platforms are hoarding valuable user data and building new technologies without considering user permissions, which is holding back innovation. At Vana, we’re liberating data from these walled gardens by putting it under user control. We allow users to directly contribute to AI models, choose how their data is used, and how AI is used. We believe we can actually outperform leading models if we can access the very best data–beating the performance of models like GPT-6 by accessing data only available directly from users. Vana is architected as a layer 1 blockchain designed from the ground up for private, user-owned data. The Reddit Data DAO saw 140k users sign up in its first week. What do you think drove this rapid adoption, and what lessons did you learn from this launch? The Reddit Data DAO was an incredible success from an adoption perspective, with over 140k users signing up in the first week. This level of adoption is unusual for DAOs–it is now the largest data DAO in history. One of the things that drove rapid adoption is that so much of the story had already been set out, as users are becoming more and more aware of the value of their data through press coverage of data sales. Realizing that Reddit is selling your data for $200M or that Apple is buying up data for $50M makes you much more aware of its value. There’s also a strong appetite for user-owned products built in web3 that move beyond familiar DeFi products to a new frontier of ownership. We’re seeing this trend in projects like Farcaster, DePIN networks, and data DAOs built on Vana, which represent a new wave of user-owned products. One important lesson was the need for proof of contribution requirements. Over a million people tried to join the Reddit Data DAO, but many didn’t meet the criteria of having a Reddit account that’s been around for a certain time and has a minimum amount of data. This highlights the importance of having mechanisms to ensure quality contributions. You mentioned plans to create Data DAOs for platforms like LinkedIn and Chat GPT. What unique challenges and opportunities do you see in expanding to these platforms? Vana is a peer to peer network for user-owned data, and builders have created various data DAOs like the Reddit Data DAO, LinkedIn Data DAO, and ChatGPT Data DAO. These different data sources are incredibly valuable for training AI models, but they’re currently locked away in walled gardens. Each of these platforms can be tricky to get data out of, but it’s always possible because of data regulation. How does Vana empower users to monetize their data, and what are some examples of how users have benefited from this? Our goal is to help users monetize and protect their data simultaneously. For example, with the Reddit Data DAO, they’re now training a user-owned model (mostly focused on shitposting at this stage, but it’s a start). Users get paid every time the model is used, creating an economic incentive for joint ownership of the model. And user data stays fully private – rather than selling data, the data is just “rented” where the underlying data never leaves the secure environment. With the growing concern around data privacy, how does Vana ensure that user data is secure and used ethically within Data DAOs? Data privacy has shifted from being just an ideological or preference question to an economic one. If someone has your data, they can potentially create an AI version of you that is economically valuable, earning revenue and potentially competing with you. That’s why privacy is so important and core to Vana. We invented a concept called “non-custodial data”, which is similar to a non-custodial wallet but for your personal data. It keeps your data under your full control, permissioned by your private key. This allows your data to be portable across applications and adds a native financial layer on top, enabling things like data DAOs to be built. How do the datasets created through Vana’s Data DAOs enhance the training of generative AI models, and what advantages do they offer over traditional datasets? Typically, AI models are trained with data scraped from the public internet – data that’s available without logging in anywhere. But if you think about it from the perspective of teaching a child about the world, you wouldn’t want them just wandering the public internet randomly. You’d want to give them high-quality information that might not be publicly available – things like high-quality writing, thought processes, or messages. AI is primarily trained on public data, but it really needs private data to push the frontiers. This is what data DAOs enable: users contributing their private data to create user-owned AI. We believe AI should be created more like open source software, by a community. Our goal is to give researchers access to the best datasets that are currently held captive inside walled gardens to push the frontiers of AI performance. What trends do you foresee in the decentralized AI space over the next 5-10 years, and how is Vana positioning itself to lead in this evolving landscape? The decentralized AI space has really accelerated over the past year. For example, at EthCC this year, there was a decentralized AI event almost every day, compared to none last year. People are figuring out how to apply sovereign technologies that have worked well for finance to the AI space. At Vana, we believe that the core foundation of all this is data. To build user-owned AI and sovereign AI, you need user-owned data, so our focus is on that data piece. In the next 5-10 years, I’m excited about a few milestones: 1) A user-owned foundation model collectively owned by 100 million people. 2) More autonomous AI agents that can earn on their own, and ensuring those agents are truly owned by the users who contributed to training them. As AI plays a more and more important economic role, ensuring that power is widely distributed from both a technical and social perspective. Can you share more about your collaboration with developers to build Data DAOs? What are some of the innovative projects currently in the pipeline? Vana is a permissionless network, so anyone can build a data DAO. It is a layer one blockchain designed from the ground up for private, user-owned data. There are over 100 data DAOs deployed on the Satori testnet today. Many of the builders are early participants in the Bittensor ecosystem who deeply understand the intersection of crypto and AI. Some notable projects include the Twitter Data DAO, LinkedIn Data DAO, and GitHub Data DAO. We’re also partnering with projects in the ZK space and DAO tooling space to make data DAOs even easier to create and manage. What ethical considerations are most pressing in the development of decentralized AI, and how does Vana address these issues? I think one of the biggest questions in AI today is on who should own models and decide what data goes into them. As we start to rely on AI more and more for information, they become our source of truth. Whoever decides what goes into the AI is essentially deciding the truth. It’s scary to have a single entity controlling this. Our view at Vana is that the community, not a single company, should make these decisions. One other question that comes up in decentralized AI is: if the AI is fully decentralized, then what if the AI goes rogue and there’s no off button? The way we approach this at Vana is that AI models are ultimately owned by the users who have contributed to them, so they always stay in full control. What advice would you give to aspiring entrepreneurs looking to enter the decentralized AI space, based on your experiences with Vana and Data DAOs? It’s a great time to start building in the decentralized AI space. There is a lot of opportunity to apply some of the crypto economic primitives that have worked well for DeFi to the new emerging category of decentralized data and AI. I’d also recommend spending some time diving into the non crypto, open source AI space to learn about some of the approaches people are taking outside a crypto context. I would dive in hands onto some of the existing projects to see what sort of primitives are available to build with, including trying out starting a data DAO on Vana.

Cryptonomist 採訪了 Vana 的首席執行官兼聯合創始人 Anna Kazlauskas，該公司是 Reddit 數據 DAO，在第一週就有 14 萬名用戶使用經過驗證的 Reddit 帳戶註冊。Anna 現在正在與開發人員合作爲其他平臺構建數據 DAO，例如 LinkedIn 和 ChatGPT。
除了 DAO 之外，他們還有其他渠道供用戶將他們的數據彙集到數據集中，然後可用於 GenAI 模型訓練，例如創建肖像或頭像。
除了 Vana 所做的事情之外，我們還與 Anna 討論了去中心化人工智能領域的發展，因爲平臺可以幫助人們使用數據並將其貨幣化以用於新的應用程序。
您能否概述一下 Vana 及其在去中心化 AI 領域的使命？
Vana 是一個由用戶擁有的數據驅動的用戶擁有的 AI 平臺。我們的使命是讓用戶擁有自己的數據以及通過 AI 模型創造的價值。人們越來越需要更多的訓練數據來提高 AI 模型的性能，因爲最終 AI 模型的好壞取決於它們的數據。
例如，LLaMA 3 是在大約 15 萬億個單詞上進行訓練的，這大致相當於公共互聯網上可用的數據量。現在，公司正試圖獲取更多數據，有時甚至爲此支付數億美元。各大科技平臺都在囤積有價值的用戶數據，並在不考慮用戶權限的情況下構建新技術，這阻礙了創新。
在 Vana，我們將數據從這些圍牆花園中解放出來，將其置於用戶的控制之下。我們允許用戶直接爲人工智能模型做出貢獻，選擇如何使用他們的數據以及如何使用人工智能。我們相信，如果我們能夠訪問最好的數據，我們實際上可以超越領先的模型——通過訪問只能直接從用戶那裏獲得的數據來超越 GPT-6 等模型的性能。Vana 的架構是第 1 層區塊鏈，從頭開始設計用於私有的、用戶擁有的數據。
Reddit Data DAO 上線第一週就吸引了 14 萬名用戶註冊。您認爲是什麼推動了它的快速普及？您從這次上線中學到了什麼經驗教訓？
從採用的角度來看，Reddit Data DAO 取得了令人難以置信的成功，第一週就有超過 14 萬名用戶註冊。這種採用水平對於 DAO 來說是不尋常的——它現在是歷史上最大的數據 DAO。
推動其快速採用的因素之一是，隨着新聞報道數據銷售，用戶越來越意識到其數據的價值，故事的大部分內容已經被闡述。意識到 Reddit 以 2 億美元的價格出售你的數據，或者 Apple 以 5000 萬美元的價格購買數據，會讓你更加意識到它的價值。
此外，人們對基於 web3 構建的用戶自有產品有着強烈的需求，這些產品超越了人們熟悉的 DeFi 產品，進入了新的所有權領域。我們在 Farcaster、DePIN 網絡和基於 Vana 構建的數據 DAO 等項目中看到了這種趨勢，它們代表了新一波用戶自有產品。
一個重要的教訓是需要證明貢獻要求。超過一百萬人試圖加入 Reddit Data DAO，但許多人不符合擁有 Reddit 帳戶、存在一定時間且擁有最低數據量的標準。這凸顯了建立機制以確保高質量貢獻的重要性。
您提到了爲 LinkedIn 和 Chat GPT 等平臺創建數據 DAO 的計劃。您認爲擴展到這些平臺有哪些獨特的挑戰和機遇？
Vana 是一個針對用戶擁有數據的點對點網絡，構建者創建了各種數據 DAO，例如 Reddit 數據 DAO、LinkedIn 數據 DAO 和 ChatGPT 數據 DAO。
這些不同的數據源對於訓練人工智能模型非常有價值，但它們目前被鎖在圍牆花園中。從每個平臺獲取數據都很棘手，但由於數據監管，這總是可能的。
Vana 如何幫助用戶將其數據貨幣化，您可以舉例說明用戶從中受益嗎？
我們的目標是幫助用戶同時將數據變現和保護。例如，藉助 Reddit Data DAO，他們現在正在訓練一個用戶擁有的模型（現階段主要專注於發垃圾帖子，但這只是一個開始）。每次使用該模型時，用戶都會獲得報酬，從而爲模型的共同所有權創造了經濟激勵。
而且用戶數據保持完全私密——數據不是出售，而是被“出租”，底層數據永遠不會離開安全環境。
隨着人們對數據隱私的關注日益增加，Vana 如何確保用戶數據的安全性並在數據 DAO 中合乎道德地使用？
數據隱私已經從意識形態或偏好問題轉變爲經濟問題。如果有人擁有你的數據，他們就有可能創造出具有經濟價值的人工智能版本，賺取收入並可能與你競爭。這就是爲什麼隱私對 Vana 如此重要和核心。
我們發明了一個概念，稱爲“非託管數據”，它類似於非託管錢包，但用於存儲您的個人數據。它使您的數據完全受您控制，並由您的私鑰授權。這使您的數據可以在應用程序之間移植，並在其上添加一個原生金融層，從而可以構建數據 DAO 之類的東西。
通過 Vana 的數據 DAO 創建的數據集如何增強生成式 AI 模型的訓練，以及與傳統數據集相比它們有哪些優勢？
通常，人工智能模型是用從公共互聯網上抓取的數據進行訓練的——這些數據無需登錄任何地方即可獲得。但如果從教孩子瞭解世界的角度來考慮，你不會希望他們只是隨意在公共互聯網上閒逛。你會想給他們提供一些可能不會公開的高質量信息——比如高質量的寫作、思維過程或信息。人工智能主要在公共數據上進行訓練，但它確實需要私人數據來推動前沿發展。這就是數據 DAO 所實現的：用戶貢獻他們的私人數據來創建用戶擁有的人工智能。
我們認爲，人工智能應該像開源軟件一樣由社區創建。我們的目標是讓研究人員能夠訪問目前被禁錮在封閉環境中的最佳數據集，以推動人工智能性能的進步。
您預見到未來 5 到 10 年去中心化 AI 領域會出現哪些趨勢，Vana 如何定位自己以引領這一不斷髮展的領域？
過去一年，去中心化 AI 領域發展確實加速。例如，今年在 EthCC 上，幾乎每天都有去中心化 AI 活動，而去年沒有。人們正在研究如何將對金融領域效果良好的自主技術應用到 AI 領域。在 Vana，我們認爲所有這一切的核心基礎是數據。要構建用戶擁有的 AI 和自主 AI，您需要用戶擁有的數據，因此我們的重點是該數據部分。
在未來的 5-10 年裏，我對幾個里程碑感到興奮：1）由 1 億人共同擁有的用戶擁有的基礎模型。2）更多可以自行賺錢的自主 AI 代理，並確保這些代理真正由爲訓練它們做出貢獻的用戶擁有。
隨着人工智能發揮越來越重要的經濟作用，確保從技術和社會角度的權力得到廣泛分配。
您能否分享更多與開發人員合作構建數據 DAO 的情況？目前正在籌備哪些創新項目？
Vana 是一個無需許可的網絡，因此任何人都可以構建數據 DAO。它是專爲私有、用戶擁有的數據而設計的第一層區塊鏈。目前，Satori 測試網上部署了 100 多個數據 DAO。許多構建者都是 Bittensor 生態系統的早期參與者，他們深刻理解加密和 AI 的交集。一些值得注意的項目包括 Twitter Data DAO、LinkedIn Data DAO 和 GitHub Data DAO。我們還與 ZK 領域和 DAO 工具領域的項目合作，使數據 DAO 更易於創建和管理。
在去中心化人工智能的發展中，最緊迫的道德考慮是什麼？Vana 如何解決這些問題？
我認爲，當今人工智能領域最大的問題之一是誰應該擁有模型並決定將哪些數據輸入其中。隨着我們開始越來越多地依賴人工智能獲取信息，它們成爲我們的真相來源。決定將哪些數據輸入人工智能的人實際上就是在決定真相。由一個實體控制這一切是可怕的。在 Vana，我們的觀點是，應該由社區而不是一家公司來做出這些決定。
去中心化 AI 的另一個問題是：如果 AI 完全去中心化，那麼如果 AI 失控並且沒有關閉按鈕怎麼辦？我們在 Vana 處理這個問題的方式是，AI 模型最終歸爲其做出貢獻的用戶所有，因此他們始終保持完全控制權。
根據您在 Vana 和 Data DAO 方面的經驗，您會給那些想要進入去中心化 AI 領域的有抱負的企業傢什麼建議？
現在是開始在去中心化 AI 領域進行構建的好時機。有很多機會將一些在 DeFi 中效果良好的加密經濟原語應用於新興的去中心化數據和 AI 類別。我還建議花一些時間深入研究非加密、開源 AI 領域，以瞭解人們在加密環境之外採取的一些方法。我將親自研究一些現有項目，看看可以使用哪些類型的原語進行構建，包括嘗試在 Vana 上啓動數據 DAO。

創作者的更多內容

實時新聞

創作者的更多內容

實時新聞

熱門文章