Core Insights - The article highlights significant developments in the AI industry, including new product launches, funding events, and advancements in AI technology. Domestic Developments - DeepSeek released its new model V3.1, enhancing agent capabilities and increasing API prices effective September 6, with input prices rising from 2 to 4 yuan per million tokens and output prices from 8 to 12 yuan [4][6]. - Tesla is integrating advanced AI capabilities into its voice assistant, collaborating with ByteDance's Volcano Engine and DeepSeek [6][7]. - Vivo launched its first mixed reality headset, the Vivo Vision, featuring a lightweight design and high-resolution display, although pricing details remain undisclosed [8]. - Manus reported a revenue run rate of $90 million, indicating strong growth potential in the AI agent platform market [16]. - The Chinese government reported that over 60% of AI model training data is now in Chinese, with some models reaching 80% [15][16]. International Developments - Meta is restructuring its AI division into four groups, focusing on different aspects of AI development [33][38]. - Intel's market value surged by $24 billion, reaching levels not seen since the internet bubble, with a dynamic P/E ratio of 53 [34][37]. - Databricks announced a valuation exceeding $100 billion as it seeks additional funding [34]. - OpenAI has raised $8.3 billion in a recent funding round, with annual recurring revenue projected to exceed $20 billion by year-end [42]. AI Financing Overview - A total of 19 AI financing events were disclosed globally this week, with a total funding amount of 11.59 billion yuan, averaging 828 million yuan per event [46]. - In China, the highest funding was reported for Magic Warehouse Robotics, which completed a multi-million yuan Series A round [53]. - Internationally, Cognition raised $500 million in Series C funding, focusing on AI programming technology [55].

Core Viewpoint - Researchers at Cornell University have developed a revolutionary silicon chip, termed the "microwave brain," which processes information similarly to a living brain using controlled microwave energy pulses instead of traditional clock-driven processors [2][4]. Group 1: Chip Design and Functionality - The "microwave brain" can handle two tasks simultaneously: processing high-speed data streams and enabling wireless communication, all within a compact size comparable to a smartwatch, with a power consumption of only 200 milliwatts [4]. - The chip's design abandons traditional sequential digital methods, utilizing tunable microwave waveguides to transmit information at tens of gigahertz frequencies in real-time, eliminating bottlenecks [5]. - Each waveguide acts like a "physical neuron," where the amplitude, phase, and frequency of microwave signals can be shaped to represent data features, allowing for rich pattern generation before digitization [5]. Group 2: AI Integration and Performance - This chip integrates an artificial intelligence framework directly into the hardware, leveraging the natural behavior of microwaves to process incoming data streams without the need for extensive memory storage or repeated calculations [6]. - In tests, the chip achieved a classification accuracy of 88% or higher for wireless signals, comparable to much larger digital chips, maintaining stable accuracy across both simple and complex tasks without the need for additional circuits [6]. Group 3: Potential Applications and Future Development - The hardware's sensitivity to signal behavior allows for applications beyond AI computing, such as monitoring wireless traffic anomalies, tracking radar targets, and decoding congested radio channels [6]. - The research team, supported by DARPA and NSF, is working on scaling and integrating the chip into existing microwave and digital systems, potentially blurring the lines between computing and communication hardware [7].

美国康奈尔大学研究团队开发出一种低功耗微芯片，被称为"微波大脑"，这是首个能够利用微波物理原 理同时处理超快数据信号和无线通信信号的处理器。相关成果发表在最新一期《自然·电子学》上。 团队强调，在传统数字系统中，随着任务复杂度上升，必须增加电路规模、功耗和纠错机制来维持精 度。而新方法基于概率性计算，能够在不增加系统开销的前提下，在简单和复杂任务中都保持高准确 率。 （文章来源：科技日报） 这一突破性能力源于芯片本身的神经网络架构。受大脑启发，该神经网络通过可调谐波导构建出复杂的 互连模式，具备识别模式和从数据中学习的能力。与依赖时钟同步和逐级执行数字指令的传统神经网络 不同，该芯片在微波频段下以模拟方式运行，利用非线性物理行为直接处理信号，能够应对高达数十千 兆赫兹的数据流，速度远超大多数现有数字芯片。 为了实现这一目标，此次设计中摒弃了许多传统电路设计原则。团队没有试图复制数字神经网络的结 构，而是设计出一种更像受控频率"混沌系统"的结构，最终却能实现高性能的计算功能。 该芯片既能执行基本逻辑运算，也能完成复杂任务，例如识别特定的比特序列或计算高速数据流中的二 进制值。在多种涉及无线信号类型的分类任务中， ...