Bojie Li

The protocol documentation for Unified Bus has finally been released. Most of the initial design work for the protocol was done four or five years ago, and I haven’t worked on interconnects for more than two years. Yet reading this 500+ page document today still feels very familiar.

As with most protocol documents, the UB documentation presents a wealth of details about the Unified Bus protocol, but rarely touches on the thinking behind its design. As a small foot soldier who participated in UB in its early days, I’ll share some of my personal thoughts. The productized UB today may differ in many ways from what we designed back then, so don’t take this as an authoritative guide—just read it as anecdotes.

Why UB

To understand the inevitability of Unified Bus (UB), we must return to a fundamental contradiction in computer architecture: the split between the Bus and the Network.

For a long time, the computing world has been divided into islands by these two completely different interconnect paradigms.

• Inside an island (for example, within a single server or a chassis), we use bus technologies such as PCIe or NVLink. They are designed for tightly coupled systems; devices share a unified physical address space, communication latency can be on the order of nanoseconds, and bandwidth is extremely high. This is a performance paradise, but its territory is very limited—the physical distance and the number of devices a bus can connect are strictly constrained.
• Between islands, we rely on network technologies such as Ethernet or InfiniBand. They are born for loosely coupled systems, excel at connecting tens of thousands of nodes, and have superb scalability. But that scalability comes at a cost: complex protocol stacks, additional forwarding overhead, and latencies in the microsecond or even millisecond range create an orders-of-magnitude gap compared with buses.

This “inside vs. outside” architecture worked well for a long time. However, a specter began to haunt the computing world—Scaling Law.

About 10 years ago, researchers in deep learning discovered a striking regularity: as long as you keep increasing model size, data, and compute, model performance predictably and steadily improves. This discovery changed the game. What used to be a “good enough” single machine with 8 GPUs suddenly became a drop in the bucket in the face of models with tens or hundreds of billions of parameters.

At that moment, a clear and urgent need presented itself to system architects everywhere: can we tear down the wall between buses and networks? Can we create a unified interconnect that offers bus-level programming simplicity and extreme performance, while also providing network-level massive scalability?

This is UB’s core mission. It’s not merely a patch or improvement on existing protocols but a thorough rethinking. UB aims to build a true “datacenter-scale computer,” seamlessly connecting heterogeneous compute, memory, and storage across the entire cluster into a unified, programmable whole. In this vision, accessing memory on a remote server should be as simple and natural as accessing local memory; tens of thousands of processors should collaborate as efficiently as if they were on a single chip.

Recently, Alibaba’s Qwen team released the Qwen3-Next model, another major innovation after Qwen3. The model achieves multiple breakthroughs in architectural design, especially reaching industry-leading levels in the balance between inference efficiency and performance. This article briefly summarizes Qwen3-Next’s core innovations.

Three major breakthroughs of Qwen3-Next:

1. Hybrid attention architecture: 3 layers of linear attention + 1 layer of traditional attention, incorporating DeltaNet’s delta rule idea
2. Ultra-sparse MoE: only 11 of 512 experts activated; 80B parameters with only 3B activated
3. 100+ tokens/s inference speed: reaches a state-of-the-art level via MTP

Core value: With 1/10 the compute cost and 10× the token processing speed, it achieves performance surpassing 32B dense models, benchmarking against Gemini 2.5 Flash.

[This article is based on the first live session of the Turing Community AI Agent Practical Bootcamp. See the slides link and download the PDF version.]

Purchase link for Turing Community “AI Agent Practical Bootcamp”

Developing your own AI Agent starts here. This article not only systematically introduces the foundational technical path for building a general-purpose AI Agent from scratch (such as context engineering, RAG systems, tool calling, multimodal interaction, etc.), but also covers advanced techniques such as slow/fast thinking and multi-Agent collaboration. Through 9 weeks of hands-on projects, you will gradually master the full lifecycle of Agent development and core advanced capabilities.

This course was first previewed via livestream on August 18 and will officially start on September 11. Each weekly session is about 2 hours and covers all the fundamental and advanced content below. Of course, 2 hours of lectures per week is definitely not enough—you’ll also need to spend time on hands-on programming practice.

Core Goals of the Bootcamp

Developing your own AI Agent starts here

🎯 Master core architecture and engineering capabilities

• Deeply understand Agent architecture: Systematically grasp the core design paradigm of LLM + context + tools.
• Become proficient in context engineering: Master multi-level context management techniques from conversation history and users’ long-term memory to external knowledge bases (RAG) and file systems.
• Master dynamic tool calling: Reliably integrate Agents with external APIs and MCP Servers, and enable self-evolution via code generation.
• Build advanced Agent patterns: Design and implement complex Agent collaboration patterns such as slow/fast thinking (Mixture-of-Thoughts) and Orchestration.

💡 Build systematic understanding of development and deployment

• Understand the path of technological evolution: See clearly the evolution path from basic RAG to Agents that can autonomously develop tools.
• Master the full lifecycle of an Agent: Be capable of independently completing the closed loop of Agent project design, development, evaluation using LLM as a Judge, and deployment.
• Build domain knowledge: Accumulate cross-domain Agent development experience through multiple hands-on projects in law, academia, programming, and more.
• Solidify your knowledge system: Co-create the book “In-depth yet Accessible AI Agent” and turn fragmented knowledge into a systematic output.

9-Week Practical Plan Overview

9-Week Advanced Topics

Following “Solving LLM Constrained Sampling Interview Question with Vibe Coding”, I’m sharing another Vibe Coding interview question from our company (Pine AI) about the fundamental principles of LLM.

Many people misunderstand Vibe Coding, thinking it’s just about constantly asking AI, “How do you do this? How do you implement that?” This approach is doomed to fail. True Vibe Coding requires you to be the architect and product manager, guiding the AI like a teacher instructing a student, not the other way around.

This interview question assesses candidates’ understanding of the basic principles of Transformers and their engineering ability to quickly implement vibe coding. This is the kind of person we need: someone who understands models and has strong engineering skills.

The Challenge: Attention-Based LLM Hallucination Detector

1. Background & Problem Statement

In many applications, large language models (LLMs) need to answer questions or extract information based on a given context, a process often referred to as “In-Context Learning.” However, LLMs have a known, serious security flaw: when asked about information not present in the context, they may “hallucinate” a correctly formatted but factually incorrect answer instead of admitting the lack of information.

[This article is based on a talk given at Turing Community’s Large Model Tech Study Camp. Slides: Slides link, Download PDF version]

A deep dive into the design philosophy and practical strategies for AI Agents. From the dialogue pattern of chatbots to the action pattern of Agents, we systematically design and manage the information environment of Agents to build efficient and reliable AI Agent systems.

Table of Contents

1. Part 1: Paradigm Shift - From Chatbot to Agent
2. Part 2: Core Analysis of Agents
3. Part 3: Context Engineering
4. Part 4: Memory and Knowledge Systems

Part 1: Paradigm Shift - From Chatbot to Agent

From Chatbot to Agent: A Fundamental Paradigm Shift

We are undergoing a fundamental transformation in AI interaction patterns:

Chatbot Era

• 🗣️ Conversational interaction: user asks → AI answers → repeated Q&A loop
• 📚 Knowledgeable advisor: can “talk” but not “act,” passively responding to user needs
• 🛠️ Typical products: ChatGPT, Claude Chat

Agent Era

• 🎯 Autonomous action mode: user sets goal → Agent executes → autonomous planning and decision-making
• 💪 Capable assistant: can both “think” and “do,” actively discovering and solving problems
• 🚀 Typical products: Claude Code, Cursor, Manus

In AI application development, choosing the right LLM API service is crucial. Whether you are building an intelligent dialogue system, developing an AI Agent, or participating in an AI Hackathon, this article will provide you with a comprehensive API usage guide, covering mainstream services such as OpenRouter, Anthropic API, Volcano Engine, and Siliconflow.

Why Do You Need Multiple API Services?

Different LLM models have their own advantages, especially when developing AI Agents, where you need to choose the right model based on specific scenarios:

• Claude (Anthropic): Excels in complex reasoning, programming, and Agent tasks, particularly suitable for scenarios requiring deep thinking
• Gemini (Google): Performs well in long text processing and multimodal understanding, suitable for handling multimedia content such as images and videos
• GPT (OpenAI): Strong in image understanding and mathematical reasoning, excellent for everyday conversation experiences
• Doubao (ByteDance): Fast access speed in China, good voice dialogue experience, especially suitable for real-time interaction scenarios
• Open Source Models: Low cost, highly customizable, suitable for large-scale deployment

(This article is automatically generated based on my one-hour voice chat with Gemini 2.5 Pro)

The human pursuit of freedom is a profound dialogue with the biological instincts deep within us. Before we embark on this dialogue, we must first understand the two core aspects of “freedom,” as articulated by philosopher Isaiah Berlin:

• The first is “freedom from”, which is negative freedom. It aims to rid us of external constraints, coercion, and interference. This is about delineating a sacred, inviolable “space” in our lives, with its ultimate form being financial freedom—where you are free from the compulsion to sell your labor for a living.
• The second is “freedom to”, which is positive freedom. It seeks to make us masters of our own will, possessing enough ability and resources to realize our self-worth. This endows us with the “power” to act, with its ultimate form being creative freedom—where you can turn imagination into reality.

Understanding this pair of concepts allows us to uncover a deeper secret, revealed by Richard Sutton, the 2025 Turing Award winner and father of reinforcement learning, in his classic textbook “Reinforcement Learning”: what drives our happiness is not the static “reward” itself, but the dynamic “reward prediction error.” What truly makes our brains secrete dopamine and feel joy is the positive gap between “actual gain” and “prior expectation.”

A completely predictable, surprise-free world, no matter how affluent, has a reward prediction error approaching zero. This biologically explains why pure “Freedom From”—a comfortable, worry-free but unchanging haven—can ultimately lead to emptiness. In contrast, “Freedom To,” filled with challenges, exploration, and creation, is a powerful engine that continuously generates positive prediction errors.

Today, the rise of AI is handing the keys to this engine to each of us in unprecedented ways.

(Thanks to Koutian Wu for thoroughly debugging and deploying, and for pointing out several technical issues in the original article, which have been corrected in this version)

As access to tools like Cursor and Claude Code becomes restricted in China, traditional HTTP/SOCKS proxies can no longer meet daily needs. These tools not only impose regional restrictions on the server side but may also employ multi-layered techniques to detect the user’s true geographical location (currently only partially implemented, but may be upgraded in the future):

1. Basic IP Database Matching: Traditional GeoIP database queries
2. Timezone Consistency Check: Obtaining the client’s timezone via JavaScript and cross-verifying with the IP’s geographical location
3. DNS Resolution Check: Using Geo DNS resolution results to check the real location
4. WebRTC IP Leak Detection: Obtaining the user’s real IP address via WebRTC
5. CloudFlare Source Address Retrieval: Obtaining the real source address through CloudFlare’s HTTP headers

Most current HTTP/SOCKS proxies can only handle basic detection methods, while more complex multi-dimensional detection often leaves them powerless. A three-layer tunnel, working at the network layer, can more thoroughly hide the user’s real network environment.

Besides bypassing geographical restrictions, a three-layer tunnel is also suitable for the following scenarios:

1. Server Access Control: Avoid exposing the SSH access port of company servers on the public internet
2. Development and Testing Environment: Avoid exposing the company’s test servers, internal APIs, etc., on the public internet
3. Secure Network Environment: Ensure communication security in untrusted public WiFi environments

While solutions like WireGuard and OpenVPN are stable and efficient, they require installing dedicated clients, which can be cumbersome in multi-device usage scenarios.

IKEv2, as a modern VPN standard, not only offers excellent performance and stability but, more importantly, is natively integrated into mainstream operating systems like macOS, Windows, iOS, and Android, eliminating the need to install any third-party clients.

This article will build on the architecture idea from “Skillfully Using Hong Kong as a Relay to Build a Smooth and Stable China-US Three-Layer Tunnel“ to construct a China -> Hong Kong -> USA IKEv2 tunnel three-hop solution.

This is an interview question from our company.

Some say our Vibe Coding programming questions are too difficult, but actually, our company’s 2-hour Vibe Coding interview questions basically don’t require you to write code yourself. Just input the question into the prompt, continuously interact with the LLM to propose requirements and improvement directions, and the AI will complete it for you.

Why is it called Vibe Coding? It’s about minimizing direct code writing. The division of labor between humans and AI becomes very clear: humans are responsible for direction control, problem definition, and result review, while AI is responsible for specific implementation. Something like Claude Code is an extreme example, where humans are not allowed to touch the code, only the LLM can.

Below, I will demonstrate how Vibe Coding works through the complete experience of this interview question. This entire exploration process was not smooth sailing; the AI’s initial solution had serious flaws. It was through my continuous review and direction correction that we finally arrived at a usable solution. This is not only about solving a technical problem but also a deep exploration of the future software development model.

It is worth mentioning that this article itself was also automatically generated by Gemini 2.5 Pro in Cursor based on my work log (including all my conversations with AI and the evolution of the code). From the moment I first posed the question to Cursor, to completing the final usable program, and then generating this illustrated blog post, the entire process took only 1.5 hours.

The Challenge: LLM Constrained Sampling

A software for learning English needs to ensure that all words output by its built-in LLM must be within a 3000-word vocabulary.

Requirements:

1. Use the Constrained Sampling method of large language models (LLM) to modify the token sampling algorithm in the inference framework (such as transformers) to ensure that all content output by the LLM is within this given 3000-word vocabulary.

2. Of course, punctuation, spaces, line breaks, etc., are allowed, but special characters, Chinese, French, emojis, etc., are not allowed.

3. Case transformations of words in the vocabulary are considered valid words, for example, if the word apple is in the vocabulary, then apple, Apple, APPLE are all considered valid outputs.

4. The 3000-word vocabulary can be any common English word list found online.

5. The performance of the constrained sampling algorithm should be as good as possible.

In the previous article, “Building a Three-Layer Tunnel with Full US IP and No Manual Proxy Settings,” we addressed many network issues encountered when accessing global services through the architecture of Domestic Server -> US Server. However, a new performance bottleneck has gradually emerged: the public connection between the domestic server and the US server experiences high latency and severe packet loss during peak hours.

This results in issues like SSH operation lag, online meeting disconnections, and API request timeouts, even when using a tunnel. The root cause lies in the international internet link between China and the US, which is like a highway during holidays—congestion is the norm.

Faced with this problem, a counterintuitive solution emerges: If the direct route is blocked, would taking a detour be faster?