2026年5月AI行业前沿观察笔记：从实时交互模型到效率升级的六大核心进展

导读

This weekly AI roundup covers 6 core industry and research breakthroughs released in mid-May 2026, including a new real-time multimodal interaction model from Thinking Machines, Google's AI mouse Magic Pointer, GitHub's Copilot pricing adjustment, Baidu's compressed ERNIE 5.1 model, Google's AI-assisted mathematics research platform, and the small-model-based RL Conductor multi-agent orchestration system. The core trend reflected in these releases is that the AI industry is moving from pure model capability stacking to solving practical experience and efficiency pain points: interaction is becoming more natural, tool integration is more seamless, cost control is more refined, and human-AI collaboration is replacing the pursuit of full AI autonomy. For practitioners, these developments not only point to new directions for technical optimization, but also bring more accessible and low-cost AI tool options.

这份2026年5月中旬的AI行业周报涵盖了6项核心产业与研究突破，包括Thinking Machines的新型实时多模态交互模型、谷歌的AI鼠标Magic Pointer、GitHub的Copilot定价调整、百度压缩版文心ERNIE 5.1模型、谷歌的AI辅助数学研究平台，以及小模型驱动的RL Conductor多Agent编排系统。这些发布背后的核心趋势是AI行业正在从单纯的模型能力堆叠，转向解决实际体验与效率痛点：交互更自然、工具融合更无缝、成本控制更精细化、人机协作正在替代对AI完全自主的追求。对从业者来说，这些进展既指向了技术优化的新方向，也带来了更多可落地、低成本的AI工具选择。

First up is the most fundamentally innovative technical release this week: the "interaction model" research preview launched by Thinking Machines. Different from the common turn-based multimodal models that stitch together independent modules such as speech recognition and large model reasoning in the market, this model is trained from the very beginning for real-time conversational scenarios, adopting a 200ms micro-turn architecture design. It will alternate between processing 200ms input fragments and generating output, so that audio, video and text can be processed as parallel streams at the same time, instead of waiting for a complete turn of user input before responding. This design allows the model to naturally realize functions such as interrupting conversations, giving real-time feedback, simultaneous interpretation, and reacting to visual cues without additional manual rule-driven dialogue management modules. For scenarios that require deep reasoning, it will also automatically call the background asynchronous model for processing, while the front-end interaction remains smooth and responsive. The team also optimized the inference engine based on SGLang to support frequent small input requests without obvious delay, and contributed a batch-invariant kernel algorithm to improve the stability of model training.

第一个是本周最具底层创新性的技术发布：Thinking Machines推出的「交互模型」研究预览。和市面上常见的把语音识别、大模型推理等独立模块拼接起来的回合制多模态模型不同，这个模型从训练之初就是为实时对话场景设计的，采用了200ms微回合的架构设计。它会交替处理200ms的输入片段和生成输出，让音频、视频、文本可以同时作为并行流处理，而不是等用户输入完整一回合再响应。这种设计让模型不需要额外的人工规则驱动的对话管理模块，就能自然实现打断对话、实时反馈、同声传译、对视觉线索做出反应等功能。对于需要深度推理的场景，它还会自动调用后台异步模型处理，同时前端交互保持流畅响应。团队还基于SGLang优化了推理引擎，支持频繁的小输入请求而没有明显延迟，还贡献了批次不变性核算法提升模型训练稳定性。

Google DeepMind's new product Magic Pointer is also quite interesting, which is expected to completely change the way users interact with AI on the PC side. This is an AI mouse pointer driven by the Gemini model, which can automatically perceive the visual and semantic context around the cursor position. Users do not need to enter complex text prompts, and only need simple gestures and short voice commands such as "fix this" and "check the route here" to complete the operation, and there is no need to jump to a separate AI window when using it across applications. Currently, this function has begun to be integrated into the Chrome browser, supporting users to select web page elements and let Gemini complete comparison, visualization and other operations. Next, it will also be pre-installed on Google's new Googlebook laptop. The core design logic behind this product is four points: to ensure that the user's operation flow is not interrupted across applications, to obtain context through visual perception, to support natural short voice commands, and to automatically identify actionable entities such as dates, locations, and objects from screen pixels.

谷歌DeepMind的新品Magic Pointer也相当有看点，有望彻底改变PC端用户和AI交互的方式。这是一个由Gemini模型驱动的AI鼠标指针，能够自动感知光标位置周围的视觉和语义上下文，用户不需要输入复杂的文本提示，只需要简单的手势和「修复这个」「查这里的路线」这类短语音指令就能完成操作，而且跨应用使用时也不需要跳转到独立的AI窗口。目前这个功能已经开始集成到Chrome浏览器中，支持用户选中网页元素让Gemini完成对比、可视化等操作，接下来还会预装在谷歌的新款Googlebook笔记本上。这款产品背后的核心设计逻辑是四点：保证用户跨应用的操作流程不被打断、通过视觉感知获取上下文、支持自然的短语音指令、自动从屏幕像素中识别出日期、地点、物品等可操作实体。

The remaining four industry and research progress also have very strong practical reference value. First, GitHub adjusted the pricing system of Copilot before switching to usage-based billing on June 1: the original $15/month Pro and $70/month Pro+ packages maintain the same price, but increase the available usage, adopting a two-tier system of "fixed basic credits + variable flexible quotas", and the flexible part will also increase with the reduction of model costs. In addition, a new Max package of $100/month has been launched, which contains $200 worth of usage, suitable for high-frequency users, and all paid packages still have unlimited code completion functions and do not consume credits. Second, Baidu released the ERNIE 5.1 model, which compresses the parameters of ERNIE 5.0 to about one-third, and the pre-training computing power consumption is only 6% of the conventional model. It ranks fourth in the world on the Arena Search list and first among Chinese models, and its ability to solve mathematical problems with tools is close to that of top closed-source models. The optimization is mainly due to the elastic training framework that can simultaneously optimize sub-models of different specifications, and the split reinforcement learning architecture that decouples training, inference and reward systems. Third, Google has launched a collaborative workbench for mathematical research called Co-Mathematician, which does not pursue AI to independently solve mathematical problems, but imitates the real mathematical research process, allowing human mathematicians and AI agents to collaborate in a shared workspace: humans are responsible for putting forward ideas, refining problems, and AI is responsible for literature retrieval, parallel computing verification, and exploring possible dead ends. The system will fully retain the research process records, and users can interrupt and adjust the direction of AI work at any time. At present, this system has achieved a 48% pass rate on the FrontierMath Tier 4 benchmark, which is higher than all existing solutions, and has helped mathematicians find missing references in existing literature when solving open problems. Fourth, researchers have launched RL Conductor, a multi-agent orchestration model with only 7 billion parameters. It learns to split complex problems, assign subtasks to specialized agents, and arrange information interaction between agents through reinforcement learning, without the need for manually written orchestration prompts. It even outperforms individual large models on benchmarks such as LiveCodeBench and GPQA Diamond, and calls fewer agents than manually designed orchestration systems. After training on random agent pools, it can also adapt to any combination of open source and closed source models, and can also recursively call itself as a subtask executor, opening up a new direction for computing power scaling at inference time.

剩下的四项产业与研究进展也有非常强的实用参考价值。首先是GitHub在6月1日切换为按用量计费前调整了Copilot的定价体系：原15美元/月的Pro和70美元/月的Pro+套餐保持价格不变，但提升了可用用量，采用「固定基础额度+可变灵活配额」的双层制度，灵活部分还会随模型成本降低而提升。另外新增了100美元/月的Max套餐，内含200美元的使用额度，适合高频用户，而且所有付费套餐的代码补全功能依然无限制，不消耗额度。其次是百度发布了ERNIE 5.1模型，把ERNIE 5.0的参数压缩到了约三分之一，预训练算力消耗仅为常规模型的6%，在Arena Search榜单上排名全球第四、中文模型第一，搭配工具解决数学问题的能力接近顶尖闭源模型。优化主要得益于可以同时优化不同规格子模型的弹性训练框架，以及把训练、推理、奖励系统解耦的拆分式强化学习架构。第三是谷歌推出了面向数学研究的协作工作台Co-Mathematician，它并不追求AI独立解决数学问题，而是模仿真实的数学研究流程，让人类数学家和AI Agent在共享工作区协作：人类负责提出想法、细化问题，AI负责文献检索、并行计算验证、探索可能的死胡同。系统会完整留存研究过程记录，用户随时可以中断和调整AI工作方向。目前这套系统在FrontierMath Tier 4基准上达到了48%的通过率，高于现有所有方案，在解决开放问题时已经帮助数学家找到了现有文献中遗漏的参考资料。第四是研究者推出了仅70亿参数的多Agent编排模型RL Conductor，它通过强化学习学习拆分复杂问题、给专业Agent分配子任务、安排Agent之间的信息交互，完全不需要人工编写编排提示词。它在LiveCodeBench、GPQA Diamond等基准上的表现甚至超过了单个大模型，而且比人工设计的编排系统调用的Agent次数更少。经过随机Agent池训练后，它还能适配任意开源、闭源模型的组合，还可以递归调用自己作为子任务执行者，为推理时的算力缩放打开了新方向。

In addition, the newsletter also mentioned Wu Enda's latest point of view: he believes that the so-called "AI job apocalypse" is completely a false proposition, and on the contrary, there will be an "AI job carnival" in the future. AI will create a large number of high-quality AI engineering positions, and the overall job market prospects are very optimistic, and the key is that practitioners need to take the initiative to adapt to the new roles driven by AI. Other short news also points to the latest industry trends: ByteDance is increasing its investment in video generation technology, while OpenAI has reduced its investment in similar directions; Nvidia is using AI to optimize the chip design process and improve semiconductor manufacturing efficiency; Gallup survey shows that although the penetration rate of AI is increasing, a considerable number of workers have not yet felt the actual benefits of AI; Sony is cooperating with university researchers to improve the adaptability and work efficiency of robots in complex scenarios.

此外通讯中还提到了吴恩达的最新观点：他认为所谓「AI就业末日」完全是伪命题，相反未来会出现「AI就业狂欢」，AI会创造大量优质的AI工程岗位，整体就业市场前景十分乐观，关键是从业者需要主动适应AI驱动的新角色。其他短讯也指向了行业最新动向：字节跳动正在加大视频生成技术的投入，而OpenAI则缩减了同类方向的投入；英伟达正在用AI优化芯片设计流程，提升半导体制造效率；盖洛普调查显示虽然AI的渗透率不断提升，但还有相当数量的劳动者尚未感受到AI带来的实际收益；索尼正在和高校研究者合作，提升机器人在复杂场景中的适应性和工作效率。

来源：https://www.deeplearning.ai/the-batch/thinking-machines-debuts-a-new-type-of-model