展示HN：模拟了一个GPT缓存错误，看到它被回显出来了
我是一名高中毕业生，过去几周我一直在模拟GPT在长文本和迭代任务中的表现。在此期间，我发现了一个持续的缓存循环——失败的输出被重复使用，PDF渲染尝试导致了静默的令牌超载，而会话的退化随着时间的推移而加剧。
我公开记录了这一现象，并提出了可重复的行为和清理建议：
→ <a href="https://github.com/sks38317/gpt-cache-optimization/releases/tag/v2025.04.19">https://github.com/sks38317/gpt-cache-optimization/releases/...</a>
发布的亮点包括：
- 在长输出（例如，PDF导出）期间的令牌刷新失败
- 失败缓存内容的递归重用
- 未清除内容导致的会话衰退
- 基于触发的清理逻辑提案
在发布之前，我向OpenAI支持团队提交了一条正式消息。以下是我写的一部分内容：
> “我分享了与GPT行为和系统设计相关的反馈和提案，包括：
> - 通过用户侧提示进行的内存模拟
> - 缓存循环问题和PDF渲染不稳定性
> - 一个建模系统性风险（SSR）和社会不稳定概率（SIP）的框架
> - 受RFIM启发的代理级协调逻辑
>
> 我只是想问这些内容是否曾在内部被审查或考虑过。”
他们的回复礼貌但含糊：
> “感谢您的深思熟虑的贡献。我们定期审查反馈，但无法提供确认、参考代码或跟踪状态。”
不久之后，我开始观察到GPT的响应微妙地反映了发布中的概念——循环抑制、内容清理触发器和减少的延续行为。
这可能只是巧合。
但如果独立贡献者在这些模式出现之前就已经在反映系统模式，并且得到了沉默的回应——也许这值得讨论。
如果你曾经历过反馈消失在虚空中而未被认可，你并不孤单。
*sks38317*
（独立贡献者，记录那些悄然重现的事物）
查看原文
I'm a high school senior who spent the past few weeks simulating GPT behavior across long-form and iterative tasks. During that time, I discovered a persistent cache loop—where failed outputs would be reused, PDF render attempts caused silent token overloads, and session degradation worsened over time.<p>I documented this publicly with reproducible behavior and cleanup proposals:
→ <a href="https://github.com/sks38317/gpt-cache-optimization/releases/tag/v2025.04.19">https://github.com/sks38317/gpt-cache-optimization/releases/...</a><p>Highlights from the release:
- Token flushing failure during long outputs (e.g., PDF export)
- Recursive reuse of failed cache content
- Session decay from unpurged content
- Trigger-based cleanup logic proposal<p>Before publishing, I submitted a formal message to OpenAI Support. Here's part of what I wrote:<p>> “I’ve shared feedback and proposals related to GPT behavior and system design, including:
> - Memory simulation via user-side prompts
> - Cache-loop issues and PDF rendering instability
> - A framework modeling Systemic Risk (SSR) and Social Instability Probability (SIP)
> - RFIM-inspired logic for agent-level coordination
>
> I only ask whether any of it was ever reviewed or considered internally.”<p>Their response was polite but opaque:<p>> “Thanks for your thoughtful contribution. We regularly review feedback,
> but cannot provide confirmation, reference codes, or tracking status.”<p>Shortly after, I began observing GPT responses subtly reflecting concepts from the release—loop suppression, content cleanup triggers, and reduced carryover behavior.<p>It might be coincidence.
But if independent contributors are echoing system patterns before they appear—and getting silence in return—maybe that’s worth discussing.<p>If you’ve had feedback disappear into the void and return uncredited, you’re not alone.<p>*sks38317*
(independent contributor, archiving the things that quietly reappear)