What Vitalik is looking for Build a casual info lookup and research tool that runs entirely offline on Android and is >50% as good as internet search + frontier AI models. People travel with their phones into offline situations all the time, and if you're literally offline, remote inference doesn't help. His past attempts got ~10 tokens/sec from 1B models that break on anything interesting. He suggests extreme MoE might be the right architecture for phones (including newer variants like n-gram models): something like ~100B params, most living on disk, with <1B activated per token. That's one promising direction, not a requirement. Any architecture that gets there is welcome. OG post: https://x.com/VitalikButerin/status/2100695863026954698 Note: This is an independent community bounty inspired by Vitalik Buterin's post. He is not affiliated with it or involved in judging unless he chooses to weigh in. Requirements Your submission must: run on Android and compatible GrapheneOS hardware operate within a maximum 12GB RAM environment use no more than 50GB total for the app, model weights, indexes, databases, and other offline assets (streaming weights from disk is expected) work completely offline once installed make no API calls, remote inference requests, web searches, or other network requests during use not require Google Play Services for core offline functionality handle useful research questions beyond simple factual recall, including explanation, comparison, synthesis, and reasoning respond at speeds that feel usable for real lookups on a phone be published in a public GitHub repository include all code, assets, dependencies, and instructions needed to reproduce the submission clearly document the models, datasets, indexes, and other resources used Architecture is open. LLMs, RAG, MoE, n-grams, compressed knowledge bases, custom retrieval systems, or other approaches are allowed. There is no parameter-count limit. Real-device requirement The app must work on real Android hardware at the time of submission. Someone with a compatible Android or GrapheneOS device should be able to visit the GitHub repo, follow the instructions, and get the app running locally within a few minutes without substantial debugging. Any required models, databases, indexes, or other assets must either be included or have clear download and installation instructions. Proof Post a public demo on X or Farcaster showing: the app running offline several example queries and responses (include ones a 1B model would fail on) a link to the public GitHub repo a brief explanation of your approach Submit a relevant screenshot to poidh along with a link to the post and your GitHub repo. The GitHub repo must contain the functional submitted version when the poidh claim is made. Winner 🏆 Submissions are judged against the requirements above and the bar described in Vitalik's post: a useful, fully offline research tool that's >50% as good as internet + frontier models. If Vitalik publicly confirms that a submitted build meets that bar, that submission wins the entire prize pot, and the bounty creator will accept the confirmed claim. The confirmation must clearly refer to the submitted project or repository. Otherwise, if no submission has been confirmed by October 31st, 2026, the creator and bounty contributors will either: wind down the bounty and return contributed funds select a winner they believe best satisfies the requirements and spirit of Vitalik's request Submissions that are fraudulent, malicious, plagiarized, materially different from the version reviewed, or in violation of the requirements above may be disqualified.

Create an audiovisual adaptation of any scene from Vitalik Buterin's new novel, Snowmoon. Snowmoon is a 32-chapter science-fiction novel set in a world filled with competing societies, cryptography, AI, strange games, new political and economic systems, war, friendship and questions about how humans coordinate. Pick one scene. Make it real. Show us what Snowmoon could look and feel like on screen. Read Snowmoon: Read online | Download EPUB for Kindle, Kobo, Apple Books, or other e-readers. The challenge Choose any scene from any chapter of Snowmoon and turn it into a finished audiovisual piece. You have broad creative freedom over how you interpret the story. Your submission could be: live action traditional animation 3D animation AI-assisted or AI-generated filmmaking machinima motion graphics mixed media something we haven't thought of There is no required visual style, we want filmmakers and creators to experiment. What we're looking for Your piece should feel like a genuine attempt to bring a specific moment from Snowmoon to life. We'll be looking at things like: Storytelling: Does the scene work as a piece of film? Visual direction: Does your interpretation make the world feel distinctive? Atmosphere: Does it capture something meaningful about the setting, characters or ideas in Snowmoon? Technical execution: How well is the piece produced? Creative interpretation: Did you make interesting choices rather than simply illustrating the text literally? Open-source usefulness: Did you document your process well enough that another creator could learn from it and build on your work? Open-source your production pipeline Snowmoon is published under GPL v3, and Vitalik specifically says his intent is that people can create movies, anime and other adaptations as long as the important non-commodity production materials behind them are open-sourced so others can build on the work. In that spirit, every qualifying submission must publicly release the meaningful production materials used to create it. Your public repo or project page should include the relevant materials from your production process, such as: scripts prompts storyboards workflows task-specific harnesses custom tooling reusable assets you have the rights to share documentation explaining your process and the major tools/models used You do not need to redistribute third-party software, commercial tools, foundation models, stock assets or anything else you don't have the rights to redistribute. The goal is simple: another creator should be able to understand how you made your adaptation and build on your work. How to submit your claim First, publish your finished video on any social media platform of your choice. The post must be publicly viewable without requiring someone to create an account or log in. Then submit a claim on poidh with: your favorite screenshot or still frame from the video as the claim image a link to the public post containing your finished video a link to your public repo or project page containing the required open-source production materials AI is allowed AI-assisted filmmaking is explicitly welcome. So is filmmaking without AI. The important things are the quality of the finished piece and the ability for others to learn from and build upon your process. Winner Submissions will remain open until Sunday, October 18th @ 11:59pm PST. After submissions close, contributors to this bounty will be encouraged to discuss the entries and decide amongst themselves which submission they believe should win. Once that rough community consensus is clear, the bounty creator will nominate the winning submission on poidh for final confirmation through the normal protocol flow. The winner will be nominated shortly after the submission deadline.