2025年12月，OpenRouter发布了基于其平台100万亿Tokens使用数据的实证研究报告，全面揭示了当前真实的AI交互模式。这些发现极具启发性，为数据驱动的LLM系统设计与优化提供了重要参考：报告深入分析了开发者和终端用户在不同任务中调用模型的情况、模型与任务的双向匹配关系、使用模式随地理区域和时间的变化规律，以及定价和新模型发布等外部因素对用户行为的影响。

本文重点分析开源生态、智能体发展趋势和用户留存机制，省略了OpenRouter关于地理区域（因缺少中国样本）和成本定价的分析（后续单独讨论）。

本文基于 SemiAnalysis 2025 年 11 月报告，聚焦谷歌 TPU 的技术升级、商业化进展及产业链布局，分析其对 AI 硬件竞争格局的影响。核心围绕 TPU v7 的性能突破（逼近英伟达 GPU）、成本优势（TCO 更低、利润率更高）、ICI 架构的扩展性创新，结合谷歌与 Anthropic、WULF 等的关键交易，阐述 TPU 从内部使用走向全面商业化的转变。同时梳理了 TPU 产业链生态，对比英伟达 GPU 生态，凸显谷歌在 AI 算力硬件领域的差异化竞争力，预示其将成为英伟达在 AI 训练 / 推理硬件市场的核心竞争对手。

分析当前主流大模型的价格趋势。

关于硅谷投资人 Nathan Benaich 和他创办的 Air Street Capital 所撰写的报告 State of AI 2025 中的一些观点的深度解读。其中包含了一些技术工作，产业实证结论，以及 GW 数据中心的相关盈利研究。

Knative enables serverless workloads to run on Kubernetes clusters, and makes building and orchestrating containers with Kubernetes faster and easier.

Serverless is a hot topic in Cloud, so as Infrastructure as Code(IaC). Infrastructure as code (IaC) tools allow you to manage infrastructure with configuration files rather than through a graphical user interface. IaC allows you to build, change, and manage your infrastructure in a safe, consistent, and repeatable way by defining resource configurations that you can version, reuse, and share.

It’s quite easy to get used to Terraform if you are familiar with CloudFormation as they all being tools to implement infrastructure as code on Cloud Provider, such as AWS. It’s a cornerstone of DevOps, designed to boost the agility, productivity and quality of work within organizations.

This article will cover: the basic components of deploying lambda, and related step to set up a lambda through Terraform.

**Data Science and Machine Learning are surely some fast-moving industries and somewhat need you to study at all times to stay ahead and on top in the industry. But the first step of getting into this area seems dreadfully slow due to widely involved technologies and overwhelming terminologies that scare you out of shit. **

AWS lowers the barrier to entry for companies and organizations looking for solutions of leveraging ML capabilities by offerings more than 20 services including low-level service like SageMaker, which helps build and manage infrastructure for developing environments, as well as high-level systems like Rekognition that come with pre-built Machine Learning models for image recognition.

This blog will go through nearly all the Machine Learning services offered by AWS.

**Data Science and Machine Learning are surely some fast-moving industries and somewhat need you to study at all times to stay ahead and on top in the industry. But the first step of getting into this area seems dreadfully slow due to widely involved technologies and overwhelming terminologies that scare you out of shit. **

AWS lowers the barrier to entry for companies and organizations looking for solutions of leveraging ML capabilities by offerings more than 20 services including low-level service like SageMaker, which helps build and manage infrastructure for developing environments, as well as high-level systems like Rekognition that come with pre-built Machine Learning models for image recognition.

This blog will go through nearly all the Machine Learning services offered by AWS.

In this tutorial, I will walk you through the implementation to reproduce LAMBADA.

From my previous article, which illustrate the basic idea of LAMBADA method that leverage Natural Language Generation(NLG) to boost training set for the Natural Language Understanding(NLU) task including text classification.

What a data scientist to do if they lack sufficient data or suffer from extreme imbalanced dataset to train a deep learning model?

The answer definitely is using IBM’s Lambada AI generates training data for text classifiers. Here is an full implementation of the paper ‘Not Enough Data? Deep Learning to the Rescue!‘ with code.