@66f9d08966e3f631faf296ee29c15ef1cdf20226b6ffabce9ceb3381bf985060 [Posted on Developer DAO Board](https://near.social/#/devgovgigs.near/widget/Post?id=422) ## Idea: Expanding the ecosystem of bots and assistants for near.social and NEAR developers ## Goals Decentralized communities find it difficult to onboard and keep up with information across many platforms and resources. Our goal is to make it easier for community members to interface with information in the NEAR ecosystem, and improve the productivity of developers that are building with NEAR infrastructure. ## Scope We ([AlgoveraAI](https://www.twitter.com/AlgoveraAI)) are a team of engineers that have been developing AI hubs and marketplaces (especially ones that run on decentralized infrastructure) for almost two years. We’ve recently been playing around with near.social and came across some of the cool bots in the ecosystem ([near-social-bot](https://github.com/evgenykuzyakov/near-social-bot)). We’ve also been keeping up with threads from the NEAR community on how they would like to see the ecosystem of AI bots and assistants expanded (see posts on ideas for both [social](https://near.social/#/mob.near/widget/MainPage.Comment.Page?accountId=66f9d08966e3f631faf296ee29c15ef1cdf20226b6ffabce9ceb3381bf985060&blockHeight=87357707) and [developers](https://near.social/#/devgovgigs.near/widget/gigs-board.pages.Post?id=385)). We have some ideas on a roadmap for expanding the capabilities of AI bots and assistants in NEAR by incorporating very recent approaches in generative AI such as retrieval-augmented language models (see research overview [here](https://uploads-ssl.webflow.com/60fd4503684b466578c0d307/63c6c20dec4479564db21819_NEW_In_Context_Retrieval_Augmented_Language_Models.pdf)). Using near-social-bot as a foundation, we propose to build a number of bots for various use cases that are driven by data from near-social and other resources such as code and docs to improve the models. Our approach combines the benefits of both AI and Web3. This aligns with Illia’s vision that *“The future of AI is open source, governed by a decentralized community that can manage bias by guiding data labeling, set up safety and monitoring protocols, authenticate data, and share accountability in a more transparent way not controlled by a single company. The future of the internet is authenticated, and it includes powerful AI systems where every piece of information has provenance which is clearly displayed to the user. Information and requests made on users behalf are recorded and audited in tamper-proof ways.”* ([source](https://twitter.com/ilblackdragon/status/1641205924973928450)) ## Deliverables For now, I’ve kept these more like high-level work packages. Depending on feedback, we can break these down into more fine-grained deliverables. - Work Package 1 involves expanding the ecosystem of bots and assistants for near.social, which offers many useful features for community curation. We will use a retrieval-augmented generation approach, which allows the possibility of rewarding data contributors when their posts are used by bots to answer the questions of others. Since the data is public, we suggest starting off using OpenAI’s models, but we could also have the capability to deploy NEAR’s own language models in future proposals. - The first bot that we will build is a question answering assistant that uses general data from near.social as context. Some of the most valuable data for training GPT-3 was Reddit data, from which OpenAI scraped the content from every webpage with greater than or equal to 3 "karma". In other words, Reddit users voted on the training data to be included. Similarly we will experiment with using only posts above some threshold on near.social. - The next bots that we build will be question answering assistants that are specialized on certain topics. We plan to do this by filtering based on the hash tags on posts. - Also possible in future proposals are bots that take action or use tools, which can improve the UX of near.social e.g. write a prompt such as “like some posts about AI and confirm transactions” - Work Package 2 involves making it easier for developers to leverage AI solutions to help them build on top of NEAR infrastructure. Many developers already use ChatGPT, but this is a general model that (i) often outputs instructions that are out of date, (ii) isn’t specialized towards developing on top of NEAR. While it may be pretty good at building with more established and widely-adopted frameworks like React, it doesn’t perform as well for more recent, emerging tools like the Blockchain Operating System. Furthermore, ChatGPT and Co-pilot don’t reward developers when their content is used. We’re open to suggestions on what the user interface would look like: the simplest being built using React, while an on-chain frontend would be super cool. To start, it would look more like ChatGPT than VS Code. As before, we can reward data contributors when their documentation or code are used by assistants to answer the questions of others. - The first developer bot we build will answer questions about the developer docs. We’ll use off-chain data like gitbook when necessary, but also use on-chain data where it exists (e.g. the NEAR wiki). - The second developer assistant will answer questions about code in the NEAR ecosystem. The first step here would be to automatically generate documentation for all code, before using a retrieval-augmented generation workflow for question answering. We’ll use off-chain data like GitHub when necessary, but also use code that has been published on-chain data where it exists (e.g. for BOS). **Timeline**: 3-5 months depending on final deliverables **Budget breakdown**: Rate $100 per hour #ai #bots @james.near