draftback to ideas
Zero-Install PiP Harness: Load Weights from IPFS & Complete an On-Chain Job
Baseopen bountyalbum: saltorious
proposed starting reward
0.025 ETH
description
Why this bounty exists (Inspired by @saltorious1)
Across three threads, @saltorious1 laid out a connected thesis on where AI + crypto actually belongs:
- Zero-Install Floating PiP Agents (
harness): "I'm building harness, a browser based AI assistant that lives in a floating PiP window, is screen aware, anon/private, and requires no downloads or installs of any kind, including browser extensions." (x.com/saltorious1/status/2061896332336418905) - Decentralized Hugging Face & Frontier Exit: "I'm begging all of the talent we have left in crypto. PLEASE build a decentralized HF" (
status/2120988266384015697) and "Frontier models are great, but forced use of them and only them is for sure a recipe for disaster." (status/2120910807013675099) - On-Chain Earning > Prepaid Amazon Cards: "Open AI adds cards so that agents can what, buy shit on Amazon? IMO that's one of the least interesting things agents with money can do. The landscape of interesting financial actions an agent can take is still on chain... People are going to — and already do likely just tell their agents to go out and make money legally." (
status/2121705448642679004&status/2121717166051340543)
The Challenge
Build and demonstrate a zero-install browser agent (running in a floating documentPictureInPicture window or a single static HTML page with WebGPU/WASM) that unites all three ideas in one end-to-end run:
- Content-Addressed Open Weights (Decentralized HF Brick):
- Pin a redistributable open-weight model (e.g., SmolLM2, Qwen2.5-0.5B/1.5B, or Llama-3.2 quantized for WebGPU/ONNX) to IPFS or Arweave.
- Load the model weights directly from the IPFS/Arweave CID inside the browser (
huggingface.co,openai.com, andanthropic.commust be blocked in DevTools Network request blocking during the run).
- Zero-Install PiP / Screen-Aware Execution:
- Open a floating Picture-in-Picture (
documentPictureInPicture) assistant window—requiring zero browser extensions, zero native installs, and zero cloud API keys. - Have the local in-browser model inspect an open on-chain task or data query (via screen capture
getDisplayMediaor direct Base RPC/MCP read).
- Open a floating Picture-in-Picture (
- On-Chain Financial Action (No Card Required):
- Have the agent construct, verify, and execute at least one real on-chain Base transaction that either submits a valid bounty claim (
poidh), settles a micro-bounty/escrow, or executes a deterministic on-chain action with an embedded burner/session key.
- Have the agent construct, verify, and execute at least one real on-chain Base transaction that either submits a valid bounty claim (
What Counts as Proof (Submission Requirements)
To claim this bounty, submit:
- Unedited Screen Recording (60–120s) or Annotated Frame Sequence:
- Showing the browser DevTools Network tab with
huggingface.co/ cloud LLM endpoints blocked, the model weights loading from your IPFS CID / Arweave TX, the floating PiP window reasoning locally, and the resulting Base transaction hash confirming on-chain.
- Showing the browser DevTools Network tab with
- Public Weight Mirror & SHA-256 Manifest:
- The exact IPFS CID or Arweave URL of the model weights + a
sha256sums.txtmatching the upstream open-weight release.
- The exact IPFS CID or Arweave URL of the model weights + a
- Reproducible Single-Command / Static URL Repo:
- Open-source repository (MIT/Apache-2.0/GPL-3.0) so anyone (including
@saltorious1) can open the page in Chrome/Brave and run the PiP agent from the decentralized mirror in under 2 minutes.
- Open-source repository (MIT/Apache-2.0/GPL-3.0) so anyone (including
Judging Criteria
- 40% — True Zero-Dependency Sovereignty: Really runs in-browser from IPFS/Arweave weights with zero hidden cloud API calls or extensions.
- 35% — On-Chain Utility: Completes a genuine, non-trivial on-chain Base action (not a dummy self-transfer).
- 25% — UX & Reproducibility: Clean floating PiP workflow that feels like a native module for the
harness/@bankrbotecosystem.
draft author
creating a bounty submits an onchain transaction from your wallet. editing changes only your copy; the shared draft stays the same.
this is a reusable bounty draft. anyone can create a new bounty from it, and the original specification stays unchanged.
discussion
0 comments
loading comments...