Show HN: Model Salt, fixing the pelican problem in LLMs by adding entropy
I have been frustrated with how repetitive the output of LLMs is. Even at maximum temperature, asking the same prompt ten times gives functionally the same result each time, as discussed here: https://news.ycombinator.c…
以下正文同步自 Hacker News 最新，版权归原站所有，已转换为易读排版。
I have been frustrated with how repetitive the output of LLMs is. Even at maximum temperature, asking the same prompt ten times gives functionally the same result each time, as discussed here: https://news.ycombinator.com/item?id=50090443That’s not surprising considering how LLMs work, but the thread above got me wondering if you could push models away from their usual generation patterns by injecting entropy from an external source.After a lot of tweaking and experimenting, I turned the idea into an MCP server. It's optimized to add just enough entropy that creative output stops being repetitive, while still living up to the intent of the original prompt.I’d love feedback, particularly on the approach itself. If you have a better way to inject entropy than what I have found, a pull request is very welcome: https://github.com/modelsalt/model-salt Comments URL: https://news.ycombinator.com/item?id=50259180 Points: 1
Comments: 0
正文由 探How 从来源站点 RSS 同步，内容未经改写；遇到排版缺失或需要图片、视频时请以原文为准。