Posts

• No AI Without API: How AI Agents Transform Enterprise Workflow and Improve Production Efficiency

  Jul 16, 2026

  AIMachine LearningAutomation

   # No AI Without API: How AI Agents Transform Enterprise Workflow and Improve Production Efficiency In 2025, the McKinsey Global Institute found that generative AI could add up to \$4.4 trillion in annual value to the global economy, with 72% of enterprises now using AI in at least one function ([McKinsey Global Institute](https://www.mckinsey.com/mgi/our-research/generative-ais-impact-on-productivity-at-scale), 2025). Yet the missing link between AI hype and real business outcomes is simple: without robust API infrastructure, AI agents cannot access the data, tools, or systems they need to act autonomously. The agentic AI revolution is fundamentally an API story. AI agents — software systems that perceive environments, make decisions, and take actions to achieve goals — are reshaping how enterprises handle customer service, IT operations, software development, and back-office workflows. This article examines the concrete mechanisms through which API-connected AI agents transform enterprise workflows and deliver measurable efficiency gains. > **Key Takeaways** > - In 2025, 81% of enterprise leaders expect AI agents to be moderately or extensively integrated into company strategy within 12–18 months, and 46% already use agents to fully automate workflows ([Microsoft Work Trend Index](https://www.microsoft.com/en-us/worklab/work-trend-index), 2025). > - McKinsey estimates generative AI could add \$2.6–\$4.4 trillion annually across 63 analyzed use cases, but only 1 in 3 organizations have scaled AI beyond pilot stage ([McKinsey](https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/the-state-of-ai), 2025). > - Klarna's AI assistant handled 2.3 million conversations in its first month — equivalent to 700 full-time agents — reducing resolution time from 11 minutes to under 2 minutes ([Reuters](https://www.reuters.com/technology/klarnas-ai-assistant-handled-two-thirds-customer-service-conversations-2025-02-25/), 2025). > - API-first architecture is the non-negotiable foundation: without programmatic access to enterprise systems, agents cannot execute transactions, query data, or orchestrate multi-step workflows.

• Is Shanghai School District Housing Still Worth It? A Ten-Year Decision Guide for Young Couples (2026-2036)

  Jul 16, 2026

  HousingEducationShanghai

   # Is Shanghai School District Housing Still Worth It? A Ten-Year Decision Guide for Young Couples (2026-2036) If you're a young couple in Shanghai debating whether to drain six wallets to buy a school district home, what you're really facing is a probability problem—not the "buy it and win it" certainty bet of the past. In 2024, Shanghai recorded roughly 112,000 newborns ([Xinhua](https://www.xinhuanet.com/), 2024), nearly halved from the 2016 peak of 218,000. Meanwhile, teacher rotation covered 80% of compulsory education schools by 2025 ([Shanghai Municipal Education Commission](https://edu.sh.gov.cn/), 2025), municipal key high school quotas are allocated to about 60% of non-selective junior highs ([Shanghai Municipal Education Examination Authority](https://www.shmeea.edu.cn/), 2025), and multi-school zoning continues to expand. Under the weight of these three overlapping policies, school district housing is shifting from a "certainty premium" to a "probability premium"—you're paying 100% of the premium but no longer buying 100% of the certainty. This article uses data and trends to answer one question: over the next ten years, is Shanghai school district housing still worth the money for young couples? > **Key Takeaways** > - Shanghai school district housing premiums have narrowed from a peak of 40–50% to 15–35%, and the compression is still underway (Beike Research Institute / CRIC, 2024–2025). > - Primary school enrollment pressure has eased noticeably since 2025—the 2024–25 school year saw roughly 750,000 enrolled students, down 14% from the 2021 peak of 871,000 (Ministry of Education, 2025). > - Teacher rotation covered 80% of compulsory education schools in 2025, with an annual rotation rate of no less than 10–15% and backbone teachers accounting for ≥20% (Shanghai Municipal Education Commission, 2025). > - School district housing yields a gross rental return of just 1.2–1.5%, well below the 1.8–2.2% on non-district newer homes (China Index Academy, 2026). > - Young couples with both owner-occupier and school district needs should prioritize newer homes in mid-tier school districts, reducing the policy and liquidity risks of old, small, run-down units.

• Is The Witcher 3 the Greatest RPG of the 21st Century?

  Jul 16, 2026

  RPGGamingThe Witcher 3

   # Is The Witcher 3 the Greatest RPG of the 21st Century? As of 2026, eleven years after its debut, The Witcher 3: Wild Hunt has sold over 50 million copies worldwide and won more than 260 Game of the Year awards ([CD Projekt investor reports via Wikipedia](https://en.wikipedia.org/wiki/The_Witcher_3:_Wild_Hunt), 2026). Few games stay commercially relevant for a decade; fewer still set the benchmark an entire genre is measured against. So does the numbers back up the "greatest RPG" claim, or is it nostalgia talking? This article weighs the data, the design, and the legacy — and offers a verdict. "Greatest" is a loaded word. So this review scores The Witcher 3 across five concrete dimensions: commercial success, critical acclaim, quest and narrative design, world and replayability, and lasting cultural influence. We also compare it honestly against the strongest rivals of the past decade — Baldur's Gate 3, Elden Ring, Skyrim, and Cyberpunk 2077. > **Key Takeaways** > - In 2026, The Witcher 3 has sold more than 50 million copies — one of the best-selling RPGs ever — and holds more than 260 Game of the Year awards, making it one of the most decorated games in history ([CD Projekt investor reports via Wikipedia](https://en.wikipedia.org/wiki/The_Witcher_3:_Wild_Hunt), 2026). > - It scores 93 on Metacritic (PC), 94 on current-gen consoles, and 95% on OpenCritic — elite marks — while its Blood and Wine expansion alone scores 92, higher than most full games ([Metacritic](https://www.metacritic.com/game/pc/the-witcher-3-wild-hunt/), 2026). > - Its side quests are widely considered the genre's gold standard: choice-driven, morally grey, and written to the same quality as many games' main stories. Personal playthrough confirms it rewards slow, attentive play.

• How to Educate Young Children in the AI Era at Shanghai: A 2026 Step-by-Step Guide

  Jul 15, 2026

  AIEducationShanghai

   # How to Educa

• How to Submit Your First Linux Kernel Patch: A 2026 Step-by-Step Guide

  Jul 14, 2026

  LinuxKernelOpen Source

   # How to Submit Your First Linux Kernel Patch: A 2026 Step-by-Step Guide In 2026, the Linux kernel 7.1 release pulled in 15,849 non-merge changesets from 2,479 developers — and 530 of them were first-time contributors ([LWN.net, "Who Wrote 7.1"](https://lwn.net/Articles/1077425), 2026). The kernel isn't a closed fortress. But its contribution process is famously unforgiving: plain-text email, strict style rules, and a review culture that assumes you've read the manual. If you've ever wanted to contribute but felt lost in the workflow, this guide is your map. You'll go from a fresh clone to a submitted, reviewable patch — the same path those 530 newcomers walked this year.

• How to Choose an Investment Strategy in China for the Next 10 Years: A 2026 Guide for Young Couples in Shanghai

  Jul 14, 2026

  FinanceInvestmentShanghai

  ![Dazzling night view of Shanghai's Pudong skyline with the iconic Oriental Pearl Tower reflecting on the Huangpu River](/posts/how-to-choose-investment-strategy-in-china-for-young-couples-shanghai/i

• How to Choose a Shanghai House-Buying Strategy for Young Couples (2026–2036)

  Jul 14, 2026

  HousingFinanceShanghai

   # How to Choose a Shanghai House-Buying Strategy for Young Couples (2026–2036) If you are a young couple weighing your first home purchase in Shanghai, you are working through the hardest property math in the country. By early 2026, Shanghai's second- hand price index had slipped roughly 8-12% from its 2021-2022 peak even as transaction volumes recovered ([National Bureau of Statistics](https://www.stats.gov.cn), [Shanghai Real Estate Brokers Association](http://www.shfdcrea.com), 2025–2026). Prices are no longer on a one-way bet up, and mortgage underwriting is tighter in real terms than it looks. At the same time, the city's hukou-linked benefits (school districts, housing-fund access, public services) still anchor long-term demand in a way that smaller cities cannot match. The question is no longer "buy everything now" versus "do nothing." The question is which purchase geometry — which district, what policy window, what mortgage structure, what fallback plan — makes sense for the next ten years of your life, starting from where you stand today. > **Key Takeaways** > - Shanghai rental yields hover around 1.4–1.8%, well below what you can earn on a simple money-market fund — so renting is a stronger pure-financial play than many couples assume ([JLL China Residential Report](https://www.jll.com/en-china), 2025). > - The middle ring (Putuo, Hongkou, Yangpu, part of Pudong) offers the best mix of price liquidity, metro connectivity and livability for young couples wary of exurban lock-in. > - Shanghai's first-home mortgage rate was roughly 3.50–3.60% by early 2026 (5Y LPR minus basis points), historically low by Chinese standards and a genuine window for well-qualified buyers. > - The winning strategy for most couples is a 5–10 year committed "buy and live" plan in the middle ring, backed by a joint-housing-fund loan and a cash buffer that survives a six-month job-loss scenario.

• Using ftrace to Selectively Trace Parent Functions

  May 7, 2023

  LinuxDebuggingKernel

  # Using ftrace to Selectively Trace Parent Functions ## Background Taking the code in the memory cgroup subsystem as an example, the `__memory_events_show` function has two call sites, namely `memory_events_show` and `memory_events_local_show`. The code is as follows:

• Linux Kernel Learning — Part 1

  Dec 5, 2021

  LinuxKernelOS

  # Overview 

• Booting a QEMU Linux System Using a Cross-Compiled aarch64 Kernel and BusyBox

  Sep 12, 2021

  LinuxEmulationArm

  ## Compile BusyBox to Generate _install ### Install Build Dependencies bison/flex ```shell sudo apt-get install bison -y sudo apt-get install flex -y ```

• Learning Arm-v8 Assembly — Environment Setup

  May 16, 2021

  ArmAssemblyDebugging

  # Environment Setup - Install the compilation and debugging components - Since I only had an x86 environment with WSL on hand, I needed to use emulation to compile, run, and debug ARM programs - Referencing this [gist article](https://gist.github.com/luk6xff/9f8d2520530a823944355e59343eadc1), it is possible to run armv8 programs on an x86 environment. The steps are as follows: - Install on WSL: - Cross-compilation environment: `sudo apt-get install libc6-dev-arm64-cross gcc-aarch64-linux-gnu` - QEMU emulation environment: `sudo apt-get install qemu qemu-system qemu-user`

• The Page Fault Process After Address Access in Arm-Linux

  Feb 21, 2021

  OSMemoryLinux

  # Address Access Example Taking the access to the user-space **virtual address** 0x0000007000003000 as an example, consider the following code. ```C #include <sys/mman.h> #include <stdio.h> int main() { unsigned long long *x = (unsigned long long)0x0000007000003000; int *p = (int*)mmap(0x0000007000003000, sizeof(unsigned long long) * 10, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_ANONYMOUS, -1, 0); printf("before write, x = %llu\n", *x); *x = 1; printf("after write, x = %llu\n", *x); } ``` If the user-space virtual address `0x0000007000003000` is not mapped in advance via mmap with read and write permissions, a `Segmentation fault (core dumped)` will occur, because `0x0000007000003000` is not managed as a vma (virtual memory area) allocated through mmap or brk in the process's virtual address space, and thus the virtual address is not accessible at this point. The output after compilation is as follows. ``` before write, x = 0 after write, x = 1 ``` As can be seen, a new value was successfully written to the user-space virtual address `0x0000007000003000`.

• Data Mining and Knowledge Discovery on KKBOX Music Data

  Jun 21, 2019

  Data MiningRecommendation

  KKBOX is a leading music streaming service in Asia, boasting the world's most comprehensive collection of Asian pop music. This paper conducts a data mining analysis on the music dataset that KKBOX has provided to the Kaggle community. Through initial cleaning and analysis of the dataset, we propose four data mining–related questions: correlation analysis among features in the dataset, song clustering, user clustering, and predicting whether a user will listen to a particular song repeatedly within a month. We then select appropriate algorithms—such as the K-prototypes clustering algorithm, the t-SNE high-dimensional data visualization algorithm, and the LightGBM algorithm—to address these data mining problems, and finally present corresponding conclusions. In doing so, we explore the application of data mining algorithms to a real-world dataset, transforming knowledge from the data mining classroom into truly practical technology.

• Formal Verification of Smart Contracts

  Apr 22, 2019

  BlockchainSecurityFormal Verification

  A smart contract is a code contract and algorithmic contract that will become a foundational technology of the future digital society. It utilizes protocols and user interfaces to complete all steps of the contractual process. This article summarizes the main technical characteristics of smart contracts as well as existing trustworthiness and security issues, and proposes applying formal methods to the modeling, model checking, and formal verification of smart contracts to support the generation of large-scale smart contracts.

• Transfer Learning for Face Recognition

  Apr 3, 2019

  Deep LearningComputer Vision

  The research history of face recognition is quite long-standing. As early as 1888 and 1910, Galton published two articles in *Nature* on using faces for personal identification, analyzing the human ability of face recognition. However, at that time, the problem of automatic face recognition was beyond reach. In recent years, face recognition research has attracted the attention of many researchers, and a variety of technical methods have emerged. In particular, since 1990, face recognition has made significant progress. Almost all well-known universities of science and engineering and major IT companies have research groups working on related studies. In the early stages, traditional face recognition was usually studied as a general pattern recognition problem. The main technical approaches adopted were geometric feature-based methods. This was largely reflected in the study of profile silhouettes, where a great deal of research was devoted to the extraction and analysis of structural features from facial silhouette curves. Subsequently, appearance-based modeling methods such as Eigenface, Fisherface, and elastic graph matching were continuously proposed. Starting from the late 1990s, researchers began to focus on face recognition under real-world conditions, proposing different face space models, including linear modeling methods represented by Linear Discriminant Analysis, nonlinear modeling methods represented by kernel-based methods, and 3D face recognition methods based on 3D information. New feature representations were proposed, including local descriptors (Gabor Face, LBP Face, etc.) and deep learning methods. Since 2014, deep learning + big data (massive labeled face data) has become the mainstream technical approach in the field of face recognition. Deep neural networks such as VGGFace, DeepFace, and FaceNet have been continuously proposed, and face recognition accuracy has been steadily improving. In 2014, Facebook's work DeepFace, published at CVPR 2014, combined big data (4 million face images) with deep convolutional networks, approaching human-level recognition accuracy on the LFW dataset. Google's work FaceNet, published at CVPR 2015, surpassed human-level recognition accuracy on the LFW dataset by adopting the Triplet Loss function.

• P2P Lending Industry Risk Analysis Platform

  Mar 28, 2019

  Machine LearningFinance

  The goal of this project is to perform risk analysis for the P2P Internet finance industry, establish standardized risk assessment dimensions and risk assessment models, identify high-risk P2P Internet finance enterprises, and ultimately generate automated risk analysis reports for P2P Internet finance enterprises. This project selected over 800 representative enterprises in the P2P Internet finance industry. Based on massive historical data from multiple aspects and dimensions of these enterprises, appropriate risk assessment dimensions were chosen and risk assessment models were built to analyze the risks currently faced by these P2P Internet finance enterprises. The overall risk profile and risk profiles across different domains of each enterprise were evaluated, and risk reports for P2P Internet finance enterprises were generated, providing an intuitive and accurate display of risk data, risk indicators, and risk ratings across all aspects of the enterprises. This meets the regulatory needs of Internet finance supervision and facilitates risk monitoring of Internet finance enterprises.

• Blockchain Network Simulation

  Mar 19, 2019

  BlockchainSecurity

  The simulation is divided into 3 parts: - **P2P Network Simulation** - Use Mininet to build networks with different topologies (such as star, ring, tree, and mesh), and limit the bandwidth of different links. - Each node can generate certain types of data, communicate with other nodes, and build a local database that records the addresses of different nodes and the data types of each node. - Transmit data (of different sizes and in varying quantities) between different nodes. --- - **Blockchain Simulation** - Implement a simulation of the blockchain network. - Simulate transactions (blocks) of the P2P network and the propagation logic. - Implement the PoW algorithm to simulate mining. - Limit link bandwidth and test the network latency and forking conditions under different block sizes and block generation intervals. - Use the PBFT protocol for consensus. --- - **Attack Simulation** - Conduct experiments on the basis of the simulated Bitcoin network. - Run simulation experiments and record the probability of a successful attack (being able to produce a longer chain that invalidates certain existing blocks) under different hashing power levels. - BGP hijacking and Eclipse attacks. The project repository is located at https://github.com/131250106/bitcoin

• Hidden Markov Language Model

  Mar 19, 2019

  NLPMachine Learning

  This post is a summary of the 5th assignment for the Fall 2018 Computational Linguistics course, with the task of using a `Hidden Markov Model` to calculate sentence probabilities.

• Generating Adversarial Input Sequences Based on RNN

  Mar 9, 2019

  Deep LearningNLPSecurity

  This article is a summary of the 6th assignment for the Fall 2018 Computational Linguistics course, with the task of generating adversarial input sequences based on RNN. First, let me give an overall introduction to the concept of adversarial example attacks. Some researchers have found that although deep neural networks achieve very high accuracy, applying small perturbations to the input can cause the model's prediction results to be completely wrong. For images, such perturbations are usually so subtle that they are imperceptible (humans cannot intuitively perceive the perturbation in the image), yet these adversarial input examples can successfully fool deep learning models. This type of attack that deceives neural network models can be broadly divided into two categories: untargeted attacks only need to make the predicted result of the image inconsistent with the original label, while targeted attacks need to misclassify the image as a specific class. It is worth noting that researchers have found that generated adversarial examples are still effective on other models, which is known as the transferability of adversarial examples.

• Multithreaded Concurrency Defect Detection for JVM Applications

  Feb 28, 2019

  ConcurrencyJVMTesting

  Currently, with the rapid development of multi-core computer hardware, multithreaded concurrent programs are gaining increasing popularity and widespread adoption. However, due to the hidden nature of shared memory space access and the randomness of concurrent thread scheduling, multithreaded programs are highly susceptible to concurrency defects, which are often difficult to detect and reproduce. Existing concurrency defect detection methods all have their limitations; for example, methods based on system testing and symbolic execution cannot scale to large concurrent programs due to state space explosion, while methods based on probabilistic scheduling have a very low defect hit rate. To address the enormous challenges faced in concurrent program testing today, the DATE team developed DATE-Confu, a multithreaded concurrency defect detection tool for JVM applications based on dynamic program testing techniques. The tool takes JVM executable programs under test (.class files or .jar files) as input and outputs the set of concurrency defects detected in the program. Currently, the tool can detect six categories of serious concurrency defects, including data races, deadlocks, and null pointer dereferences. Users can select which types of concurrency defects to detect based on their needs. The key innovation of DATE-Confu lies in its effective combination of guided schedule fuzzing techniques and symbolic trace analysis techniques. The figure below shows the architecture diagram of DATE-Confu. The tool uses fuzzing techniques to efficiently traverse the enormous state space of multithreaded programs, while the symbolic trace analysis techniques rapidly discover unexplored control branches in the program under test, thereby providing effective information for iterative fuzzing and enabling DATE-Confu to reach higher coverage faster. DATE-Confu has so far been tested on multiple industrial-level projects. The experiments conducted to date have demonstrated that DATE-Confu achieves high testing efficiency and is capable of uncovering deeper-hidden concurrency defects.

• Personalized Resume Recommendation

  Feb 23, 2019

  Data MiningRecommendation

  This system employs machine learning methods and a big data platform to train on a massive volume of job seekers' resumes. Unlike ordinary conditional filtering, this system can efficiently and accurately provide suitable and reliable talent recommendations from a big data processing perspective for IT positions that recruiters are looking to fill, reducing recruitment costs and using big data technology to bridge the gap between recruiters and outstanding talent. Through research on massive resume data and company recruitment information, this project analyzes the characteristic features of personal resumes, companies, and company positions. From the recruiter's perspective, it conducts exploratory research centered on personalized recommendation technology to help recruiters obtain talent information in a more accurate and efficient manner. Based on this concept, a prototype system is implemented that can automatically recommend more appropriate resumes based on the characteristics of job postings.

• Detecting Improper Sitting Posture with a Laterally Positioned Motion-sensing Camera

  Feb 22, 2019

  Deep LearningComputer VisionHealth

  **Abstract** Sitting posture detection is helpful for preventing musculoskeletal disorders. With the development of motion-sensing cameras and related software development kits (SDKs), it is possible to implement an application using skeleton detection technology. In this paper, a method is introduced to detect sitting posture from a lateral view without disturbing the user. To analyze video stream information, a skeleton thinning algorithm is described, and an averaging process is used to specifically locate the main joints from the lateral side. The results show this method has high accuracy when detecting improper sitting postures. **Keywords** Ergonomics; Motion-sensing Camera; Gesture Recognition; OpenNI