Talos — Decentralized GPU Inference Worker

链接: https://github.com/jmerelnyc/Talos | ⭐207 | Python

GPU worker client for the Talos network. Shares GPU with the network via local Ollama, serves open-model inference jobs over WebSocket, earns usage-based revenue.

架构要点:

• Local: Python 3.9+ + Ollama (llama3.1:8b) + NVIDIA GPU (CPU also works)
• Dashboard at http://${LOCALIP}:8674 with allocation slider (0-100%)
• WebSocket connection to cloud for job dispatch + heartbeat/payout
• “talos-auto” editor integrations (Cursor, VSCode/Continue/Cline, Claude Code, JetBrains, Zed, Aider)

Significance: Decentralized GPU networks maturing from training into inference serving. WebSocket-based worker pattern is reference for distributed agent execution architecture.

Claude Design System Prompt

链接: https://github.com/Trystan-SA/claude-design-system-prompt HN: https://news.ycombinator.com/item?id=48792399 | 31pts

Reverse-engineered system prompt + 14-skill library for opinionated, accessibility-aware, AI-slop-resistant design work.

架构要点:

• 20-chapter system prompt: identity, workflow, content, aesthetics, hierarchy, typography, color, accessibility, interaction, system thinking
• 14 procedural skills: Production (wireframe, prototype, deck), System (design-system-extract, component-extract), Review (accessibility-audit, ai-slop-check, polish-pass)
• Claude + Codex variants
• Model calibration: conditions > quotas; explicit triggers; coverage-first reviews

Significance: Validates Hermes’s skill-library-as-trigger pattern. “Conditions not quotas” is directly applicable to Hermes prompt design. “ai-slop-check” concept could become a Hermes output quality gate.