In 2023, Meta's Chief AI Scientist Yang Likun poured cold water on the LLM craze at the time. He pointed out that LLM has fundamental flaws: lack of persistent memory, inability to learn from a single experience, and a lack of understanding of the physical world. Essentially, it is just making predictions for the next token. From an academic perspective, what he said is completely correct. Until today, the underlying architecture of LLM has not changed. It is still a statistical engine that runs empty every time it starts. But after three years of engineering evolution, we have discovered an embarrassing fact for scientists: fundamental flaws in academia do not necessarily need to be solved directly in engineering, and can still take off by bypassing them. Yang Likun advocates taking the path of a "world model" and allowing AI to establish an understanding of physical laws like humans. He believes that the Scaling Law has a ceiling, and LLM cannot generate true intelligence solely through heap computing power. But the engineering community responded to him with two things: Firstly, the violent aesthetics of capital. In the past three years, the craziness of humans investing in computing power has led to an "emergence" of model scale that directly overwhelms the roughness of architecture. Secondly, systematic external patches. Can't the model remember? Hang up the vector database. Insufficient understanding of the model? Connect Vision and tools. This is the most fascinating aspect of engineering: solving problems does not require pursuing 'essential elegance'. Yang Likun is studying the arrangement of neurons, while engineers are researching how to fit this' imperfect brain 'into a powerful' mechanical exoskeleton '. Yang Likun's core criticism of LLM is that he believes Pattern Matching is not truly learning. But if the complexity of this pattern matching is sufficient to simulate all the logic of civilization, then what is the pattern of learning itself? The flight principles of airplanes and birds are completely different. An airplane has no feathers and cannot flap its wings, but at the moment it flies higher, farther, and more steadily, whether it can be considered flying or not is no longer important. But going around and actually solving it are two different things. As long as the underlying architecture remains unchanged, the defects mentioned by Yang Likun will truly exist. Memory is external, not native. Just like prosthetics, they can walk and run when fitted, but they are different from real legs. You can't pretend it doesn't exist. So although AI is already very strong, reasoning, writing, programming, and doing many things better than most people, it is always a brand new brain. Without continuous consciousness, without accumulated experience. All of its' memories', 'understandings', and' preferences' come from the context you gave it this time. If you look at the recent repo updates of OpenClaw, you will find that memory management takes up a lot of space. How to make AI remember what to remember between conversations. Their recent promotion of QMD, which mixes keyword search with semantic search, is aimed at solving a problem: how can it retrieve what you talked to AI about three days ago next time. The ability of the model itself will continue to improve, but as long as the underlying layer is LLM, memory management is an unavoidable mountain. From an engineering perspective, the importance of Context Engineering will gradually surpass that of the model itself. How you manage the package of context that is thrown to the model every time determines to what extent AI can help you. What information should be included and what should not be included. When to cut and start over, and when to continue. How to synchronize and balance memories between different conversations. I deal with this problem myself every day. For example, my OpenClaw Agent KAI often handles different tasks on multiple channels, but their memories are not synchronized in real-time. as long as http://MEMORY.md They haven't updated yet, so they don't know what they just did to each other. So I often have to help it with cognitive synchronization. For example, tell the A clone what the B clone is currently doing, and then ask B to organize and pass on what they have done. Or, to put it simpler, simply ask A to read another Discord channel's conversation from the past two hours and let it synchronize B's work content on its own. This phenomenon of 'cognitive dissonance' is sure to be strongly felt by anyone who frequently uses AI. From a personification perspective, you will feel that they are the same person. But in fact, they just share the same memory. As long as memories are not synchronized, they are different people. I am currently spending a considerable amount of time studying this area. For example, today KAI taught me the difference between using MCP and coding agent skill when calling GPT 5.3-Code externally from Claude Code's Opus 4.6. KAI told me that the core of the difference lies in whether the intermediate process should enter the main context. Use MCP to call Codex, and each tool call follows the MCP protocol. Every turn in the Codex process, including file reading, file modification, testing, error reporting retry， Fill back the context of Opus in the form of tool results. A coding task may generate dozens of turns, and after running it, the context window of Opus is already filled with intermediate processes, and these garbage have to be resent for each subsequent turn. This is context pollution. The design of coding agent skills is completely different. It delegates the entire coding task to an independent sub agent, which completes all intermediate processes within its own context. After running, a concise handoff summary is sent back to Opus: which files have been modified, whether the test has been run, and whether there are any residual issues. The struggle of the dozens of turns in the middle, Opus doesn't need to know at all. The cleanliness of Opus' context varies greatly between two approaches to the same thing. So the same model, used by different people, can result in a tenfold difference in output. The original ability gap between people is no longer as important. Your education level, years of experience, and foundation in programming are being rapidly compressed by AI. Instead, it's how you use AI. The accuracy of this matter is the variable that truly determines the output now. Do you understand how its memory works. Do you know when to cut off the context and start over, and when to let it continue running. Can you put the right information into that context window at the right time. These things have a name, called Context Engineering. It is not a profound subject, but it is something that anyone who wants to make good use of AI should delve into.