AIの文章が読みにくいのは、性能よりも「何を報酬にしたか」の問題なのかもしれません。NYUとテキサス大オースティン校の研究では、最先端LLMが人間に読み解きにくい言語パターンを多用し、小さなモデルのほうが人間に近い文章を書くと報告されています。コードや数学の学習は正しさだけを報い、明瞭さは評価されません。測れるものだけが伸びる構図は他の領域でも起きているはずで、報酬設計こそ開かれた場で検証されるべきだと感じています。 #LLM #AI文章 #強化学習 #自然言語処理
Related
💻 the-incredible-pytorch: 12.6 k ⭐I needed a single bookmark for the PyTorch ecosystem.The Incredible PyTorch is a curat...
💻 the-incredible-pytorch: 12.6 k ⭐I needed a single bookmark for the PyTorch ecosystem.The Incredible PyTorch is a curated list covering everything from LLMs and object detection t...
Google announced the rollout of Gemini 4 Carbon on Tuesday, expanding access to developers and enterprises through paid ...
Google announced the rollout of Gemini 4 Carbon on Tuesday, expanding access to developers and enterprises through paid plans.Source: Gadgets 360https://www.gadgets360.com/ai/news/...
I asked an AI to speak from inside the Chinese room, with my coffee-nose hybrid in view.It said tokens come in and go ou...
I asked an AI to speak from inside the Chinese room, with my coffee-nose hybrid in view.It said tokens come in and go out, and there is no second channel where a smell arrives for ...