我的征尘是星辰大海。。。
The dirt and dust from my pilgrimage forms oceans of stars...
-------当记忆的篇章变得零碎,当追忆的图片变得模糊,我们只能求助于数字存储的永恒的回忆
作者:黄教授
手机视频列表
领悟到大模型公司的两个浅显的道理
视频
音频
原始脚本
今天我们聊透的核心本质是 AI 技术选择的底层逻辑,永远跟着应用需求走。 不是技术越高级越好,而是技术能不能最高效的满足场景需要。 我们可以把这些关键认知梳理成两个清晰的核心结论,帮你更直观的沉淀。 核心结论一,为什么 GPU 成了多场景通用选择?不是它在单场景最优,而是综合性价比最高。 企业选 GPU 本质是在避免重复投入和覆盖未来需求之间做平衡。 对既要做大语言模型、文字,又要做多模态图像视频的公司,比如 OpenAI、国内的字节、阿里。 TPU、NPU 确实在纯文字推理上更省能耗,但要额外处理图像,就需要再买图像专用芯片。 比如英伟达的 A100本身就擅长图像渲染。 与其买文字芯片加图像芯片两套设备,不如直接用用 GPU,一套硬件覆盖所有需求,反而省了硬件采 采购机房维护的成本,综合性价比更高。 对只做纯视觉的公司,比如特斯拉,做安防识别的海康。 GPU 的图像优化能力,比如并行处理像素,快速渲染多视角图,本就是为视觉设计的,比为文字优化的 TPU、NPU 更适配。 而且这些场景不需要文字语义,直接用 GPU 处理图像。 还能省掉文字绑定的额外计算开销,效率更直接。 核心结论二,AI 分纯感知派和语言加感知派,本质是场景要不要抽象语义。 两类技术路径没有高低之分,只有场景适配之分。 纯感知派跳过语言,只玩图像、声音,适配不需要理解意义,只需要识别生成感知规律的场景。 比如特斯拉自动驾驶,只需要识别前方是车、人、红绿灯的视觉特征,不用懂车这个词的语义。 安防抓逃犯,只需要比对人脸的视觉特征,不用绑定逃犯名字。 工业质检,只需要看产品有没有划痕的视觉特征,不用懂划痕的文字定义。 这类 场景的优势是链路短、开销小,不需要大语言模型的抽象能力,直接用视觉模型就能落地。 语言加感知派,文字和图像绑定,做多模态,适配需要抽象语义来联动感知的场景,比如 文生图,要理解一只会飞的猪的文字语义,再生成对应的图像。 图生文,要识别猫在追老鼠的图像,再用文字描述出来。 智能客服要同时理解用户的文字提问和发来的故障图片,再给解决方案。 这类场景需要语言作为桥梁来实现跨模态联动,所以必须绑定大语言模型,走更复杂但更智能的路径。 简单说,今天我们聊透的底层逻辑就一句话,AI 选什么技术,用什么硬件,从来不是看技术多先进,而是看场景要什么。 需要跨模态联动,就走语言加感知,只需要感知规律,就纯视觉学直脊,需要覆盖多场景,就选通用 GPU。 只聚焦单一场景,再选专用芯片。 这些看似粗浅的结论,恰恰是业界做技术落地时最核心的判断标准。
修正脚本
今天我们聊透的核心本质是 AI 技术选择的底层逻辑,永远跟着应用需求走。 不是技术越高级越好,而是技术能不能最高效的满足场景需要。 我们可以把这些关键认知梳理成两个清晰的核心结论,帮你更直观的沉淀。 核心结论一,为什么 GPU 成了多场景通用选择?不是它在单场景最优,而是综合性价比最高。 企业选 GPU 本质是在避免重复投入和覆盖未来需求之间做平衡。 对既要做大语言模型、文字,又要做多模态图像视频的公司,比如 OpenAI、国内的字节、阿里。 TPU、NPU 确实在纯文字推理上更省能耗,但要额外处理图像,就需要再买图像专用芯片。 比如英伟达的 A100本身就擅长图像渲染。 与其买文字芯片加图像芯片两套设备,不如直接用 GPU,一套硬件覆盖所有需求,反而省了硬件采购、机房维护的成本,综合性价比更高。 对只做纯视觉的公司,比如特斯拉,做安防识别的海康。 GPU 的图像优化能力,比如并行处理像素,快速渲染多视角图,本就是为视觉设计的,比为文字优化的 TPU、NPU 更适配。 而且这些场景不需要文字语义,直接用 GPU 处理图像。 还能省掉文字绑定的额外计算开销,效率更直接。 核心结论二,AI 分纯感知派和语言加感知派,本质是场景要不要抽象语义。 两类技术路径没有高低之分,只有场景适配之分。 纯感知派跳过语言,只玩图像、声音,适配不需要理解意义,只需要识别生成感知规律的场景。 比如特斯拉自动驾驶,只需要识别前方是车、人、红绿灯的视觉特征,不用懂车这个词的语义。 安防抓逃犯,只需要比对人脸的视觉特征,不用绑定逃犯名字。 工业质检,只需要看产品有没有划痕的视觉特征,不用懂划痕的文字定义。 这类场景的优势是链路短、开销小,不需要大语言模型的抽象能力,直接用视觉模型就能落地。 语言加感知派,文字和图像绑定,做多模态,适配需要抽象语义来联动感知的场景,比如文生图,要理解一只会飞的猪的文字语义,再生成对应的图像。 图生文,要识别猫在追老鼠的图像,再用文字描述出来。 智能客服要同时理解用户的文字提问和发来的故障图片,再给解决方案。 这类场景需要语言作为桥梁来实现跨模态联动,所以必须绑定大语言模型,走更复杂但更智能的路径。 简单说,今天我们聊透的底层逻辑就一句话,AI 选什么技术,用什么硬件,从来不是看技术多先进,而是看场景要什么。 需要跨模态联动,就走语言加感知,只需要感知规律,就纯视觉最直接,需要覆盖多场景,就选通用 GPU。 只聚焦单一场景,再选专用芯片。 这些看似粗浅的结论,恰恰是业界做技术落地时最核心的判断标准。
英文翻译
The core essence we thoroughly discussed today is that the underlying logic of AI technology selection always follows application demands. It's not about choosing the most advanced technology, but about whether the technology can most efficiently meet the needs of the scenario. We can distill these key insights into two clear core conclusions to help you internalize them more intuitively. **Core Conclusion 1:** Why has GPU become a universal choice across multiple scenarios? Not because it is optimal for a single scenario, but because it offers the best overall cost-effectiveness. When companies choose GPU, they are essentially balancing the avoidance of redundant investment with covering future needs. For companies that need to handle both large language models (text) and multimodal images and videos—such as OpenAI, ByteDance, and Alibaba in China— TPUs and NPUs are indeed more energy-efficient for pure text reasoning, but to additionally process images, they would need to purchase a dedicated image chip. For example, NVIDIA’s A100 is already inherently good at image rendering. Rather than buying two separate sets of hardware (a text chip plus an image chip), it’s more efficient to directly use a GPU. One set of hardware covers all needs, thereby saving hardware procurement and data center maintenance costs, resulting in higher overall cost-effectiveness. For companies that focus purely on vision—such as Tesla in autonomous driving or Hikvision in security surveillance— The image optimization capabilities of GPU (e.g., parallel pixel processing, rapid rendering of multi-perspective images) are inherently designed for vision, making them more suitable than TPUs or NPUs optimized for text. Moreover, these scenarios do not require text semantics; they directly use GPU to process images, thus saving the extra computational overhead tied to text, achieving more direct efficiency. **Core Conclusion 2:** AI is divided into the pure perception school and the language-plus-perception school. The essence is whether the scenario requires abstract semantics. There is no hierarchy between these two technical paths; they are simply suited to different scenarios. The pure perception school skips language and works only with images and sound. It is suitable for scenarios that do not require understanding meaning but only need to recognize and generate perceptual patterns. For example: - Tesla’s autonomous driving only needs to recognize the visual features of vehicles, pedestrians, and traffic lights ahead, without understanding the semantics of the word "car." - Security surveillance for catching fugitives only requires matching facial visual features, without binding the fugitive's name. - Industrial quality inspection only needs to check whether a product has visual features like scratches, without understanding the textual definition of "scratch." The advantage of such scenarios is a shorter pipeline and lower overhead. They do not require the abstract capabilities of large language models and can be implemented directly with vision models. The language-plus-perception school binds text with images to achieve multimodality. It is suitable for scenarios that require abstract semantics to link perception, such as: - Text-to-image generation: It must understand the textual semantics of "a flying pig" and then generate the corresponding image. - Image-to-text generation: It must recognize an image of a cat chasing a mouse and then describe it in text. - Intelligent customer service: It must simultaneously understand the user's text query and the attached faulty image, then provide a solution. These scenarios require language as a bridge to achieve cross-modal linkage, so they must incorporate large language models, following a more complex but more intelligent path. In short, the underlying logic we thoroughly discussed today boils down to one sentence: What technology and hardware AI chooses is never about how advanced the technology is, but about what the scenario requires. - If cross-modal linkage is needed, go with language plus perception. - If only perceptual patterns are needed, pure vision is the most direct. - If multiple scenarios need to be covered, choose a general-purpose GPU. - If only a single scenario is targeted, then choose a specialized chip. These seemingly superficial conclusions are precisely the core judgment criteria used in the industry when implementing technology.
back to top