Sam Altman: GPT-4.5 andGPT5
Here is the translation of the contents within the <document> XML tags:
OpenAI has released a new model specification aimed at guiding its language models (such as ChatGPT) in their behavior and objectives. This specification is based on three core principles: doing the best work, maintaining boundaries, and using an appropriate style. The specification covers aspects ranging from factual accuracy to creativity and programming applications, and emphasizes a balance between user autonomy and preventing harm or abuse.
OpenAI has compiled a set of challenging prompts that were created through a combination of model-generated content and expert manual review, covering everything from typical to complex scenarios. Initial testing results show significant improvement in the model’s adherence to the “model specification” compared to last year’s best models.
OpenAI is conducting a pilot study involving around 1,000 participants who will review the model’s behavior, evaluate proposed rules, and share their thoughts. These studies will help refine the specification further.
Noteably, OpenAI has decided to release the “model specification” under an open knowledge-sharing license, making it available in the public domain for developers and researchers to use freely, modify, and innovate upon.
Translation
OpenAI发布了一个全新的模型规范，旨在指导其语言模型（如ChatGPT）的行为和目标。这份规范基于三个基本原则：做最好的工作、保持边界和使用适当风格。规范涵盖了从事实准确性到创造力和编程应用等方面，并且强调了对用户自主权与预防伤害或滥用的平衡。
OpenAI收集了一系列富有挑战性的提示词，这些提示词是通过模型生成和专家人工审查相结合的方式创建的，以覆盖从典型到复杂的各类场景。初步测试结果显示，与去年五月的最佳模型相比，现在的模型在遵守“模型规范”方面有了显著改进。
OpenAI正在开展一项大约1000人参与的试点研究，参加者需要审查模型的行为、评估提出的规则，并且分享自己的想法。这些研究将帮助进一步完善规范。
值得注意的是，OpenAI决定将“模型规范”以知识共享许可协议发布到公共领域，这对于开发者和研究人员来说是个好消息，他们可以在自己的工作中自由地使用这份规范，还能基于它进行调整和创新开发。
Reference:
https://model-spec.openai.com/2025-02-12.html#overview