您正在使用IE低版浏览器，为了您的雷峰网账号安全和更好的产品体验，强烈建议使用更快更安全的浏览器
此为临时链接，仅用于文章预览，将在时失效
0
导语：从实时互动到实时智能，对话 AGI 时代正在到来。
10月23日，iRTE2026实时智能大会在北京举行。声网创始人兼 CEO 赵斌在主论坛发表题为《对话 AGI：实时互动的下一章》的演讲，回顾了过去一年实时互动、对话式 AI 与底层基础设施的关键变化，并提出“五级对话能力”框架，勾勒对话式 AI 从基础应答走向深度理解与关系沟通的演进路径。
赵斌认为，随着 AI 从回答问题走向持续交流、理解上下文并执行任务，实时互动的对象正在从人与人扩展到人与 Agent。实时互动技术的角色也随之延伸：它不仅要让信息传递更稳定、更流畅，还要为 AI 的感知、理解、表达与行动提供实时支撑。
从“问一句答一句”到深度关系沟通：对话 AGI 的五级进化
在赵斌看来，对话式 AI 的进步，不应只用回答速度、语言自然度来衡量。更重要的是，它能否理解对话对象的意图、处境与关系，能否在复杂场景中恰当地表达，并逐步形成长期、深入的理解。
围绕这一判断，赵斌在演讲中提出了“五级对话能力”框架，将对话能力划分为五个层次：
L1：基于自身视角的表达：以自我表达和被动接续为主，实现“问一句、答一句”。
L2：基于互动规律的表达：从被动回应走向主动表达，让对话更自然、连贯。
L3：基于个体认知的沟通：理解个体的身份、意图与处境，实现因人而异的沟通。
L4：基于群体认知的沟通：理解社会关系与角色差异，适应复杂的多人沟通场景。
L5：基于深度关系的沟通：理解长期上下文与关系演变，形成更深层次的理解与思维默契，让沟通从理解当下走向理解一个人。
这五个层次展现的，不只是对话技术的逐级增强，更是 AI 与人之间关系的变化：从完成一轮问答，到理解一个人，再到理解人与人之间的关系，最终形成跨越时间与情境的深度沟通。
赵斌指出，对话式 AI 正在从早期被动接续的交互，向更主动、更自然的全双工对话发展，并开始尝试理解人的个性、处境、情绪与关系。面向未来，AI 还需要突破复杂社会场景下的角色分工与沟通秩序，进一步理解个人长期积累的上下文。这也是对话 AGI 的重要发展方向：AI 不仅能听懂一句话，还能结合完整语境理解一个人的真实意图，并作出恰当回应。
Real-Time AI Infra：支撑对话 AGI 的关键基础设施
对话能力的持续进化，也对底层技术提出了更高要求。
赵斌在演讲中提到，过去一年，OpenAI 等公司对 WebRTC 技术的深入采用，推动实时互动技术在对话式 AI 领域获得更广泛的关注和应用。随着 AI 从单轮问答走向持续交互，实时互动基础设施正在成为连接模型、智能体运行框架与用户的重要技术支撑。
与传统 AI 应用相比，对话场景对延迟、稳定性和双向交互有着更严格的要求。AI 需要一边接收用户输入，一边理解上下文、组织回应，必要时还要随时应对打断，并持续调用工具、推进任务。任何环节出现明显延迟，都可能破坏对话的自然性，甚至影响任务能否完成。
为支撑这些需求，赵斌介绍了面向实时智能交互的基础设施方向：通过全球实时传输网络、分布式云边端联合推理、实时编排与全双工交互能力，让模型和 Agent 能够在真实场景中持续协同。声网对话式 AI 引擎也以全双工交互为核心，连接模型、Agent 运行框架与用户，支撑稳定、同步的实时对话体验。
与此同时，雷峰网(公众号：雷峰网)了解到：声网持续面向开发者完善对话式 AI 开发工具与开发套件，结合 SDK、Skills、CLI 工具及 Recipes 等能力，帮助开发者更高效地构建具备完整逻辑与执行能力的 Agent。
从客服、教育到 AI 陪伴与智能穿戴，赵斌认为，对话式 AI 正在进入更多真实场景。AI 能否真正创造价值，不只取决于模型是否“会说话”，还取决于它能否理解需求、自然沟通，并持续完成任务。实时互动技术将在这一过程中发挥越来越重要的作用。
随后，英伟达首席架构师、AOMedia 视频编解码工作组联合主席 Andrey Norkin 带来了关于 AV2 视频编解码技术的演讲。
圆桌对话：个人智能体，会成为下一代个人入口吗？
在由声网联合创始人 Tony Wang 主持的第一场圆桌中，互动媒体产品经理、科技博主 Eugene Wei，科技自媒体人、播客《乱翻书》主理人潘乱，以及 AIoT 硬件创业者、端侧 AI 自媒体主理人华建新，围绕“个人智能体：从新工具到新入口？”展开讨论。
从手机、眼镜、耳机到更多随身设备，个人智能体正在重新引发行业对下一代人机交互入口的思考。潘乱认为，个人智能体可能带来一场范式变革：AI 产品的价值正在从争夺用户注意力，转向帮助用户更高效地完成任务。未来的个人智能体不仅要响应明确指令，还应主动承担任务，帮助用户处理那些目标明确、却不值得投入大量时间的事务。
对于下一代个人硬件，华建新提出，未来的入口未必由某一种设备独占，眼镜、耳机及其他能够感知环境、执行任务的硬件都有机会。他认为，个人智能体不仅需要理解用户，还需要通过物理设备持续感知现实世界并采取行动，这为不同形态的硬件带来了新的发展空间。
Eugene Wei 则将讨论引向“上下文”的价值：个人智能体能否真正理解用户的兴趣、沟通记录与社会关系，将影响它能否成为有价值的个人入口。在衡量新入口是否成立时，他也提出，可以关注用户每天与 Agent 的交互情况，以及究竟是用户主动找 Agent 更多，还是 Agent 主动找用户更多。圆桌由此进一步讨论了个人智能体的产品价值、硬件形态与未来衡量标准。
圆桌对话：从“能聊天”到“能办事”，对话式 AI 如何创造真实价值？
在由声网首席商业官崔博主持的第二场圆桌中，斜跃智能创始人陈伟、冠客科技创始人耿祥磊、Rokid 乐奇联合创始人兼生态负责人蔡国祥，以及泰坦启明 CEO 吴侃，从具身智能、智能穿戴、企业营销与 AI 招聘等不同场景出发，探讨对话式 AI：驱动下一代的人与 Agent 交互。
陈伟认为，具身智能的关键是让人的意图真正转化为机器人的行动。 机器人不仅要理解语音指令，还需要融合视觉、语音及自身状态等多模态信息，完成对物理环境的感知、规划与决策，并根据行动后的环境变化持续调整任务执行。相比完成预设的演示任务，机器人更重要的是在开放环境中持续理解人的真实需求，让行动始终与人的意图保持一致。斜跃智能正在发展的Duplex Reasoning全新范式的具身基础模型，旨在实现机器人交互和行动的一体化。
蔡国祥则从 AI 眼镜的实践出发，强调对话体验是 AI 穿戴设备的核心竞争力。 他表示，无论硬件和背后的 AI 能力多么强大，如果对话不够自然，用户体验就会受到直接影响。未来，AI 穿戴设备需要通过低延迟、稳定的实时交互，以及对语义和长期上下文的理解，让用户逐渐感受不到自己是在与 AI 对话。他还分享了 Rokid 乐奇 AI 助手采用声网 RTC 链路服务的实践，并看好 AI 眼镜成为 AI 时代重要的个人入口。
在企业应用场景中，耿祥磊指出，AI 语音座席要实现可靠的业务交付，必须解决延迟、打断时机与拟人化等问题，其中延迟甚至可能直接影响任务结果。对于 To B 营销，AI 还需要与行业知识体系及人工协作相结合，才能更好地保障业务结果。吴侃则结合 AI 招聘实践表示，对话式 AI 能帮助招聘业务大规模触达求职者、提升服务效率，而不是简单替代人工。要形成业务闭环，还需沉淀行业数据，以求职者为本体梳理业务，帮助 Agent 建立长期记忆，并围绕到面、入职等实际业务事件，辅助 AI 持续评估和优化服务质量。
从四位嘉宾的分享可以看到，不同行业对对话式 AI 的要求虽有差异，但共同指向一个方向：AI 不仅要能交流，还要能够理解真实场景、融入现有业务流程，并可靠地完成任务。随着实时互动、多模态感知与 Agent 执行能力持续发展，对话式 AI 有望从交互工具进一步成为连接数字世界与物理世界、推动业务运行的重要力量。
从实时互动走向实时智能，开启对话 AGI 的下一章
从赵斌提出的五级对话能力框架，到两场圆桌对个人智能体与真实业务落地的探讨，iRTE2026 主论坛呈现出一个清晰的产业趋势：对话式 AI 的竞争正在从基础模型能力，延伸至上下文理解、实时交互、复杂关系认知与任务执行。
当 AI 不再局限于被动回答，而是能够持续理解人、主动参与交流，并在真实场景中完成任务，实时互动技术也将承担更深层的使命。
面向对话 AGI 的下一阶段，声网将继续与开发者和行业伙伴共同探索实时互动与生成式 AI 的融合，推动人与人、人与 Agent 之间更加自然、高效的实时交互体验。
雷峰网原创文章，未经授权禁止转载。详情见转载须知。
为了您的账户安全，请验证邮箱
请验证您的邮箱
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
You are using a low-end IE browser and strongly recommend a faster and safer browser for your Thunder Peak account security and better product experience.
This is a temporary link, only for article preview, which will expire at the time
0
Introduction: From real-time interaction to real-time intelligence, the AGI era is coming.
On 23 October, the iRE2026 Real-Time Smart Conference was held in Beijing. In a presentation at the main forum entitled Dialogue AGI: The Next Chapter of Real-Time Interaction, the Founder and CEO Zhao Bin of the Voice Network reviewed the key changes in real-time interaction, dialogue AI and lower-level infrastructure over the past year and proposed a “five-level dialogue capacity” framework outlining the evolution of dialogue AI from basic responses to deeper understanding and relationship communication.
Zhao Bin believes that, as AI moves from answering questions to continuous communication, understanding context and carrying out missions, real-time interaction is expanding from people to people and Agent. The role of real-time interactive technology extends: it not only makes the transmission of information more stable and fluid, but also provides real-time support for AI ' s perception, understanding, expression and action.
From Question and Answer to Deep Relationship Communication: Dialogue AGI Five Level Evolution
In Zhao Bin ' s view, progress in the dialogue-like AI should not be measured only by the speed of the answers and the naturality of the language. More importantly, it is able to understand the intentions, circumstances and relationships of its interlocutors, to express them properly in complex contexts and to evolve long-term, in-depth understanding.
In this judgement, Zhao Bin presented a framework for a “five-level capacity for dialogue”, divided into five levels:
L1: Expressions based on one's own perspective: self-expression and passivity lead to “question, answer”.
L2: Expression based on a pattern of interaction: moving from reactive response to proactive expression, making dialogue more natural and coherent.
L3: Communication based on individual knowledge: understanding an individual ' s identity, intentions and circumstances, and achieving individualized communication.
L4: Communication based on group perception: understanding social relationships and role differences and adapting to complex multi-person communication scenarios.
L5: Communication based on deep relationships: understanding the evolution of long-term context and relationships, leading to deeper understanding and a better mindset, allowing communication to move from understanding to understanding a person.
These five levels demonstrate not only the scaling-up of dialogue technologies, but also the evolution of AI ' s relationship with people: from the conclusion of a round of questions and answers to understanding one person, to understanding one person ' s relationship with another, to ultimately creating deep communication across time and context.
Zhao Bin noted that the dialogue-type AI is moving from an early, passive succession of interactions to a more proactive, natural, two-work dialogue and is beginning to try to understand the human personality, circumstances, emotions and relationships. For the future, AI also needs to move beyond the division of roles and communication order in a complex social context to a better understanding of the context accumulated by individuals over time. This is also an important direction for dialogue: AI not only understands a word, but also understands a person ' s true intentions and responds appropriately to it.
Real-Time AI Infra: Key infrastructure supporting dialogue AGI
The continuous evolution of dialogue capacity also places higher demands on bottom technologies.
In his presentation, Zhao Bin mentioned that, over the past year, companies such as Open AI had advanced the use of WebRTC technology to promote greater attention and application of real-time interactive technology in the dialogue-based AI area. As AI moves from a single round of questions and answers to continuous interaction, the real-time interactive infrastructure is becoming an important technical support for connectivity models, AI agent operating frameworks and users.
Dialogue scenes have stricter requirements for delay, stability and two-way interaction than traditional AI applications. AI needs to receive user input while understanding context, organizational responses and responding to interruptions whenever necessary, and to continuously call on tools and advance tasks. Any apparent delay in any chain could undermine the nature of the dialogue and even the fulfilment of the mandate.
To support these needs, Zhao Bin describes the direction of infrastructure oriented towards real-time intelligence interaction: through global real-time transmission networks, distributed cloud-side joint reasoning, real-time organization and full-time interactivity, to allow models and Agent to continue to work together in real-time scenarios. The AAI engine is also centred on a dual-work interaction, linking model, Agent operating framework and users, supporting a stable, synchronized real-time conversation experience.
At the same time, the WeChat Official Account-Ray Peaks Network has learned that it continues to improve its dialogue with developers by developing tools and tools and developing packages that combine the capabilities of SDK, Skills, CLI tools and Recipes to help developers build Agent with complete logic and implementation.
From customer service, education, to AI's company and intelligence, Zhao Bin thinks that the Dialogue AI is entering more real scenes. AI's real value creation depends not only on whether the model is “talking”, but also on its understanding of needs, its natural communication and its continued performance. Real-time interactive technologies will play an increasingly important role in this process.
Subsequently, Andrey Norkin, Chief Architect of In Weida and Co-Chair of the AOMedia Video Decoded Working Group, delivered a presentation on AV2 video decodering techniques.
Round table dialogue: Will a person, AI agent, be an entry point for the next generation?
At the first round table, chaired by Tony Wang, co-founder of the Acoustic Network, interactive media product manager, tech blogger Eugene Wei, tech activist, podcast editor Pancreas, and Aiot hardware entrepreneur, later AI built a new round around “Personal AI agent: From new tools to new entrances?”.
From mobile phones, glasses, headphones to more walk-in devices, individual AI agent is rekindling industry thinking about the next generation of humans. Pan believed that personal AI agent could bring about a paradigm shift: the value of AI products was shifting from competing for user attention to helping users to deliver more efficiently. The future individual AI agent should not only respond to clear instructions, but also take the initiative to help users deal with matters that are clearly targeted but do not deserve to spend a significant amount of time.
For the next generation of personal hardware, China has proposed that future entrances may not necessarily be monopolized by one type of equipment, and that eyeglasses, headphones and other hardware capable of understanding the environment and carrying out the mission have access. In his view, the need for individual AI agent not only to understand users, but also to continuously understand the real world and act through physical equipment, has created new spaces for the development of different forms of hardware.
Eugene Wei will discuss the value that leads to “context”: whether an individual AI agent truly understands user interest, communication records and social relations will affect whether it can be a valuable entry point for individuals. In measuring the establishment of the new portal, he also suggested that attention could be given to the day-to-day interaction of users with Agent and whether it was the user who volunteered to look more for Agent or for Agent. This led to further discussion of the product value, hardware form and future measures of individual AI agent.
Round-table dialogues: from "can talk" to "can do things", how can dialogue AI create real value?
In the second round table, chaired by Choi Bo, Chief Business Officer of the Sound Network, Chen Wei, founders of the Winners Technology, Rokid Lucky, Co-founder and Ecologist Zai Guoxiang, and Titan CEO Wu, from different scenes, such as smartness, smart clothing, corporate marketing and AI recruitment, discussed dialogue AI: driving the next generation to interact with Agent.
In Chen Wei's view, the key to being smart is to actually turn people's intentions into robotic actions. Not only do robots understand speech instructions, but they also need to integrate multi-modular information, such as vision, voice and self-status, to achieve the perception, planning and decision-making of the physical environment, and to continuously adapt mission execution to post-action environmental changes. It is more important for robots to understand the real needs of people in an open environment and to keep their actions in line with their intentions than to complete pre-set demonstration missions. The fundamental model of the new paradigm of Duplex Reasoning, which is being developed by tilting intelligence, aims at the integration of robotic interaction and action.
Cai Guoxiang started with the practice of AI glasses, emphasizing that the experience of dialogue is the core competitiveness of AI equipment. He said that, however powerful the hardware and the AI behind it may be, if the dialogue was not natural enough, user experience would be directly affected. In the future, AI wear devices need to gradually fail to feel that they are talking to AI through low delay, steady real-time interaction and understanding of semantics and long-term context. He also shared the practice of Rokid Lucky AI's assistant using the RTC link service, and looked forward to seeing AI glasses as an important personal entry point for the AI era.
In the context of the enterprise application scenario, it was noted that in order to achieve reliable business delivery, the AI voice booths had to address the issues of delay, interruption and humanization, which could even directly affect the outcome of the mission. For To B marketing, AI also needs to be integrated with industry knowledge systems and manual collaboration to better secure business results. U Wu, in conjunction with AI recruitment practice, suggests that dialogue AI can help recruitment operations to reach job-seekers on a large scale and improve service efficiency rather than simply replacing labour. In order to create a closed business circle, it is also necessary to settle industry data, organize business with a job-seeker approach, assist Agent in establishing long-term memory, and support AI's ongoing assessment and optimization of service quality around actual business events.
As can be seen from the sharing of the four guests, the requirements for dialogue-based AI vary from industry to industry, but it points in a common direction: not only is AI able to communicate, but it is also able to understand the real scene, integrate into existing business processes and perform its tasks reliably. As real-time interaction, multi-modular perception and Agent implementation capacity continues to develop, the Dialogue AI is expected to further become an important force in linking the digital world to the physical world and driving business.
From real-time interaction to real-time intelligence, open the next chapter of AGI
From Zhao Bin ' s five-level dialogue capacity framework to two round tables on individual AI agent and real business landings, the main forum of iRTE 2026 showed a clear industrial trend: dialogue-like AI competition is extending from basic model capabilities to context understanding, real-time interaction, complex relationship awareness and task execution.
When AI is no longer limited to passive answers, it is able to understand people on a continuous basis, engage proactively in communication and perform tasks in real-time settings, and real-time interactive technology will assume a deeper mission.
In the next phase of dialogue-oriented AGI, the AGI will continue to explore with developers and industry partners the integration of real-time interaction and generation of AI to promote more natural and efficient real-time interactions between people, people and Agent.
It's an original Ray Peak article, which is forbidden to reproduce without authorization. For further information, see the information below.
For your account security, please verify the mailbox.
Check your mailbox.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.