LLMs are poor RNGs, and low entropy limits ambitious agent systems
austinvhuang · x · 2026-07-26
The quoted discussion makes a simple point: for ambitious LLM systems, low entropy is a real constraint.
It argues that asking a model to “pick a random suffix” is often the wrong tool. If you need stable uniqueness, you should not rely on randomness inside the model or shell-side tricks that may not persist across runs. Instead, uniqueness should come from role identity and coordinated naming, not from synthetic randomness.
The takeaway is broader than naming:
- LLMs are poor random number generators
- practical agent systems need coordinated roles and interfaces
- seemingly small prompt or workflow choices can create reliability problems
The original post links this to a more general limitation in long-horizon use: at normal temperatures, LLMs often produce the same few answers, so the response space can feel too deterministic for more complex projects.
More from coding & agent
- Grok Build’s Plan mode and worktrees are the real leverage points — tetsuoai · 2026-07-26
- Vjeux says coding loops still fail outside training data unless a human babysits them — Vjeux · 2026-07-26
- Hugging Face’s breach analysis hit closed-model guardrails, then worked with an open model — mmitchell_ai · 2026-07-26
- LocalMind brings local AI memory to FastMCP, Ollama, and NixOS — xmrah · 2026-07-26
- Agent harness discussion on Max Agency leaves builders wanting to redesign everything — hwchase17 · 2026-07-26
- Building an MCP Server with 33 Tools: Lessons from Designing for Claude — SimpleSpacer97 · 2026-07-26