据路透社报道，OpenAI在周四更新了其《模型规范》(Model Spec)。这是一份公开路线图，阐述了该公司希望其AI模型如何表现。随着AI系统能力不断增强，OpenAI围绕五项核心价值观重新构建了这套规范。这五项价值观是忠于人类意图、最大限度减少伤害、诚实与透明、尊重人类自主权，以及关怀。
OpenAI
OpenAI表示，忠于人类意图是这套框架的基础，要求模型理解用户、开发者和OpenAI的意图与价值取向，而不是简单地字面执行指令。
OpenAI称，此次更新是其AI对齐工作的一部分，旨在帮助模型在陌生情境中应用人类价值观，同时始终接受人类的监督。修订后的框架称，模型应尽量减少伤害，对其所知和所不知保持透明，并帮助用户做出自己的决定、追求自己的目标。
该公司表示，目前的模型尚未完全体现更新后的《模型规范》要求，正在探索新的训练方法，帮助模型将这些原则内化于自身行为之中。
此举呼应了竞争对手Anthropic的“宪法式AI”方法。该方法通过一套书面原则“宪法”来指导模型的行为，并让塑造AI系统的价值观更加透明。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
According to Reuters, Open AI updated its Model Spec on Thursday. It is an open road map that describes how the company wants its AI model to perform. With the increasing capacity of AI, Open AI has re-engineered the set around five core values. The five values are loyalty to human intent, minimization of harm, honesty and transparency, respect for human autonomy and care.
OpenAI
Open AI stated that loyalty to human intent is the basis of the framework, requiring models to understand the intent and value orientation of users, developers and Open AI, rather than simply implementing instructions literally.
Open AI stated that the update was part of its AI alignment and was intended to help models apply human values in strange situations, while always being subject to human scrutiny. The revised framework states that the model should minimize harm, be transparent about what it knows and what it knows, and help users to make their own decisions and pursue their own goals.
It indicated that the current model did not fully reflect the requirements of the updated Model Code and that new training methods were being explored to help the model internalize these principles into its own behaviour.
This echoes the “constitutional AI” approach of the competitor Anthropic. This approach guides model behaviour through a set of written principles, the “Constitution”, and makes the values that shape the AI system more transparent.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.