🌐 中文 English
Chain of Thought and Reasoning Models
OpenAI’s new model GPT-o1, released on November 23, 2025, introduced the concept of the Chain of Thought. It uses a reinforcement learning game approach, training the model to generate compliant thought chains through rewards and penalties. In 2025, the AI field further discovered that reasoning models are not only powerful themselves but also serve as exceptionally good “teachers.” For instance, DeepSeek achieved high-performance small models by distilling the Qwen model using the DeepSeek-R1 model, breaking through the ceiling of the Scaling Law.
Model Distillation
DeepSeek open-sourced and released a series of small models distilled from the DeepSeek-R1 model, such as DeepSeek-R1-3b and DeepSeek-R1-7b. These models surpassed the performance of the base Qwen and Llama models. During distillation, DeepSeek utilized a large amount of high-quality Chain of Thought (CoT) data generated by the DeepSeek-R1 model. This data contains rich reasoning logic and knowledge, helping smaller models better understand and handle complex tasks. Furthermore, by integrating CoT data for specific verticals, it’s possible to create expert small models with stronger domain-specific capabilities.
RAG (Retrieval-Augmented Generation)
In 2025, RAG technology underwent a paradigm shift from “vector retrieval” to “knowledge association.” GraphRAG (Graph-enhanced Retrieval) became mainstream, enabling models to handle cross-document global problems by integrating knowledge graphs. Additionally, breakthroughs in Long-Context technology freed some scenarios from the hassle of extensive text chunking. Embeddings combined with Late Interaction and Reranking improved matching accuracy, allowing RAG to maintain its dominance in knowledge bases as a low-cost, high-real-time solution with citation traceability.
The Evolution of Mixture of Experts (MoE)
Mixture of Experts (MoE) became a standard feature for all high-performance models. DeepSeek further optimized the combination of Multi-Head Latent Attention (reducing KV Cache) with MoE. By leveraging an ultra-large total parameter count (approximately 671B) paired with an extremely low number of activated parameters (about 37B), it achieved exceptionally high inference efficiency.
MCP: The “Universal Serial Bus” of the AI Era
Introduced by Anthropic in 2025, the Model Context Protocol (MCP) addressed the biggest pain point in the AI toolchain: tool isolation. It provided an interface standard for AI Agents, fueling explosive growth in the AI Agent ecosystem. MCP introduced three key capabilities:
- Resources: Allows models to read external data in a standardized way (e.g., reading local files or database records).
- Tools: Allows models to perform actions (e.g., sending an email or running a Python script locally).
- Prompts: Predefined interaction patterns delivered to the model through a standardized protocol.
AI Agent
If 2024 was about exploring the possibilities of Agents, 2025 marked the stage where Agents truly entered productivity. AI Agents evolved beyond mere conversational bots into complete systems possessing Perception, Planning, Memory, and Action capabilities.
- Maturation of Cognitive Architecture: Reasoning-Driven
Thanks to the maturity of reasoning models (like o1, DeepSeek-R1), Agents gained powerful “slow thinking” capabilities.
- Self-Reflection: While executing tasks, Agents no longer stubbornly follow a single path. They continuously review their own outputs. If code execution throws an error, they automatically iterate based on the error logs.
- Long-term Memory: By combining vector databases with personalized profiles, Agents can remember your preferences for a specific project from three months ago.
- From Single-Agent to Multi-Agent Systems (MAS)
In 2025, the realization dawned that a single, all-powerful Agent is often less effective than a team of specialized Agents.
- Collaborative Work: Similar to how a company operates, one Agent handles architecture design, another writes code, and another performs security audits.
- SOP Automation: Complex enterprise-level processes (like onboarding or expense reimbursement) are broken down into standardized message exchanges between multiple Agents.
- Agentic Workflow: Process as Intelligence The Agentic Workflow concept advocated by Andrew Ng was fully validated in 2025. Rather than pursuing a larger model, smaller models (like the distilled DeepSeek-R1-7B) could surpass top-tier closed-source models on specific tasks through multiple iterations, tool calls, and loop logic.
From AutoGLM to Doubao AI Phone
AI is no longer just a chat window. From Zhipu’s AutoGLM to ByteDance’s deeply integrated “Doubao AI Phone,” the LAM (Large Action Model) concept officially materialized. Models are no longer limited to returning text via API; they gained screen understanding and simulated click capabilities, enabling complex cross-app operations like ordering food or booking a ride. This signifies the official evolution of AI hardware from “AI Inside” to “AI Native.”
Cursor
In 2025, Cursor became a $30 billion unicorn. It’s not just an editor but an “AI pair programming partner.” By indexing the entire codebase and continuously iterating on the Composer mode, developers shifted from “writing code” to “reviewing code.” It proved that the core competency in future software development will shift from syntax proficiency to system architecture and requirement definition capabilities.
Claude Code
If Cursor is the navigator within the IDE, Claude Code was the agent that parachuted directly into developers’ terminals (CLI) in 2025. Unconstrained by the editor’s UI, it obtains system shell permissions to directly execute Git commands, run local tests, and automatically fix build errors. It marks AI’s transition from “suggester” to “executor.”
Vibe Coding: From “Typing Code” to “Conveying the Vibe”
At the end of 2025, the Vibe Coding concept, promoted by tech luminaries like Andrej Karpathy, swept through the developer community. This marks the third stage of software development: from “manually writing syntax” to “AI-assisted completion,” and now to simply providing the “vibe”—intent as code.
Google’s Gemini Architecture
In 2025, Gemini 3 Pro demonstrated the power of a native multimodal hybrid MoE architecture. Unlike other models that attach visual encoders externally, Gemini unified the processing of audio, video, images, and text at the same neuronal layer. This allows the model to excel at multimodal tasks while significantly improving reasoning efficiency.
2025 DeepSeek Experimental Projects: Optical Compression Technology and Manifold-Constrained Hyperlinks
- Optical Compression Technology: Further improved the compression ratio and restoration accuracy of OCR, offering insights for long-context compression.
- Manifold-Constrained Hyperlinks: Built upon hyperlinks by introducing manifold constraints to prevent gradient explosion and vanishing gradient issues.
The 2026 Opening Bombshell: Prompt Skills
At the beginning of 2026, Anthropic introduced Prompt Skills, signaling the demise of traditional “long-form prompts.” Developers no longer write complex prompts. Instead, they encapsulate specific conversational content into “skills” through Q&A with AI, which are then mounted via the MCP protocol. This further reduces the context length overhead associated with prompts. By understanding developer needs through Q&A, developers can focus more on implementing business logic rather than crafting prompts.
2026 Outlook: The Convergence of Agentic AI and Embodied Intelligence
2026 will be the turning point where Agentic AI moves from concept to closed-loop operation:
- From “Conversation” to “Delivery”: AI will no longer be confined to chat windows. It will make autonomous decisions based on preset goals, actively monitoring progress, automatically retrieving documentation upon encountering errors, fixing them, and delivering the final result.
- Scaling of Multi-Agent Systems (MAS): Tens of thousands of micro-agents will collaborate within private networks, forming an automated “virtual factory” handling everything from supply chain optimization to automated marketing.
- Explosion of AI Native Hardware: With the maturation of LAM (Large Action Models), 2026 will see more wearable devices that don’t rely on screens, focusing instead on voice and visual sensing. AI will officially transition from “Inside” to “Native.”
- The Dawn of Embodied Intelligence: Infusing multimodal large models into robotic bodies will grant AI intuition about the physical world, truly entering laboratories and factories.
