近日，人工智能领域的又一热点事件引发了广泛关注。一家名为「MultiOn」的初创公司公开宣称，他们操控了一个名为「草莓哥」的社交媒体账号，并在该账号上发布了名为 Agent Q 的智能体。这个智能体被设计用来提供高级推理能力，并且在网络上引发了巨大的关注。
「MultiOn」的创始人Div Garg在斯坦福大学读计算机科学博士期间休学创业，他表示，虽然他们没有等到OpenAI发布其秘密项目「Q*」，但他们已经发布了操控「草莓哥」账号的全新智能体Agent Q。Div Garg还称，Agent Q是一款突破性的AI智能体，它结合了蒙特卡洛树搜索（MCTS）和自我批评，并通过直接偏好优化（DPO）算法来学习人类的反馈。在性能上，Agent Q的零样本性能是LLama 3基线的3.4倍，并且在真实场景任务的评估中，其成功率达到了95.4%。
然而，这一消息并没有得到所有人的认可。虽然Agent Q在技术上取得了显著的进步，但公众的关注点似乎更多地集中在「MultiOn」是否真的利用了「草莓哥」账号进行炒作。有人质疑「MultiOn」的行为是否诚实，甚至有人称他们为无耻的骗子。
尽管如此，「MultiOn」已经公开了Agent Q的相关论文，并表示将在今年晚些时候向开发人员和普通用户开放。论文中详细介绍了Agent Q的主要组件和方法，包括使用MCTS进行引导式搜索、AI自我批评、直接偏好优化等技术细节。
总的来说，Agent Q的发布标志着人工智能在自主规划和自我纠错方面迈出了重要一步。未来，随着技术的不断进步，我们有望看到更多像Agent Q这样的智能体在现实世界中发挥作用。
英语如下：
Title: “AI Hype Unveiled: OpenAI Secret Project Found to be Manipulated by Intelligent Agent”
Keywords: Hype, Traffic, Product
News Content:
A recent development in the field of artificial intelligence has sparked widespread attention. A startup named “MultiOn” has come forward to claim that they operated a social media account named “Strawberry Bro” and released an intelligent agent known as Agent Q on this account. This agent was designed to offer advanced reasoning capabilities and has generated significant buzz online.
Div Garg, the founder of MultiOn, who took a leave of absence from his computer science PhD at Stanford University to start the company, stated that they did not wait for OpenAI to unveil its secret project “Q*” before releasing their new intelligent agent, Agent Q. Garg described Agent Q as a groundbreaking AI agent that combines Monte Carlo Tree Search (MCTS) and self-criticism, and learns from human feedback through the direct preference optimization (DPO) algorithm. In terms of performance, Agent Q’s zero-shot performance is 3.4 times that of Llama 3, and its success rate in real-world task evaluations is 95.4%.
However, this announcement has not been universally accepted. While Agent Q has made significant technological strides, public focus seems to be more on whether MultiOn indeed used the “Strawberry Bro” account for hype manipulation. There are questions about the company’s integrity, with some calling them unscrupulous frauds.
Despite this, MultiOn has published a paper on Agent Q and has stated that it will be open to developers and general users later this year. The paper provides a detailed account of Agent Q’s main components and methods, including guided search using MCTS, AI self-criticism, direct preference optimization, and other technical details.
In summary, the release of Agent Q marks an important step forward in artificial intelligence for autonomous planning and self-correction. As technology continues to evolve, we can expect to see more intelligent agents like Agent Q playing roles in the real world.
【来源】https://www.jiqizhixin.com/articles/2024-08-14-8
Views: 5