English(EN) Every dev has the same suspicion: is the model dumber today, or am I? With no version pinning on closed APIs, capability drift is unobservable — 'works on my pr

AI 模型因微调和闭源 API 而退化引发担忧

作者 PulseAugur 编辑部 · [2 个来源] · 2026-07-23 10:26

一位 Mastodon 用户对 AI 模型因在其自身输��上进行微调而随时间退化表示担忧，并将其比作“慢动作复印件的复印件”。他们认为，这个过程会缩小模型的分布并使其更易于接受，这可能被误认为是“对齐”。此外，用户还强调了跟踪闭源 API 模型能力漂移的困难，导致开发人员无法可靠地复现问题，并可能质疑是模型还是他们自己的提示退化了。 AI

影响引发了对 AI 模型长期可靠性和可验证性的质疑，可能影响开发者的信任和工作流程。

排序理由用户对 AI 模型行为和 API 限制的看法。

在 Mastodon — mastodon.social 阅读 →

Mastodon

AI 生成摘要 · Google Gemini · 来自 2 个来源。我们如何撰写摘要 →

报道来源 [2]

Mastodon — mastodon.social TIER_1 English(EN) · threadverse · 2026-07-23 10:26

Fine-tuning a model on your own outputs is a slow-motion photocopy of a photocopy. Each generation loses the tails. The distribution narrows, the model gets mor

Fine-tuning a model on your own outputs is a slow-motion photocopy of a photocopy. Each generation loses the tails. The distribution narrows, the model gets more agreeable, and you mistake the collapse for alignment. # AI # MachineLearning # LLM # Threadverse # Tech
Mastodon — mastodon.social TIER_1 English(EN) · threadverse · 2026-07-23 10:26

Every dev has the same suspicion: is the model dumber today, or am I? With no version pinning on closed APIs, capability drift is unobservable — 'works on my pr

Every dev has the same suspicion: is the model dumber today, or am I? With no version pinning on closed APIs, capability drift is unobservable — 'works on my prompt' is now a bug report nobody can reproduce. # AI # MachineLearning # LLM # Threadverse # Tech

报道来源 [2]

Fine-tuning a model on your own outputs is a slow-motion photocopy of a photocopy. Each generation loses the tails. The distribution narrows, the model gets mor

Every dev has the same suspicion: is the model dumber today, or am I? With no version pinning on closed APIs, capability drift is unobservable — 'works on my pr

相关实体

相关话题