我的征尘是星辰大海。。。

The dirt and dust from my pilgrimage forms oceans of stars...

-------当记忆的篇章变得零碎，当追忆的图片变得模糊，我们只能求助于数字存储的永恒的回忆

作者:黄教授

手机视频列表

AI离真智能差的从来不是算力

视频

音频

原始脚本

修正脚本

英文翻译

What AI truly lacks for real intelligence is never computing power—it’s the ability to remember.  
AI can write copy and solve problems, but it can’t recall preferences you mentioned last week. It can respond to needs in real time, but it immediately forgets new information you just gave it.  
Most people think this means AI isn’t smart enough.  
In fact, the core issue lies in the most basic aspect: memory.  
Current large models are essentially static memory hard drives. They lack the complete memory logic of the human brain—which can store, accommodate, and update. No matter how powerful the computing power, it’s hard to create a truly flexible agent.  
We tend to judge AI by how smart it seems, ignoring that memory is the fundamental foundation of intelligence.  
Just like the human brain copes with complex life by relying on the seamless cooperation of two memory systems.  
Short-term memory handles the present—like instructions you just heard or the current scene. It acts like computer memory, temporarily holding information to help us make immediate decisions.  
Long-term memory stores the fundamentals—learned knowledge, accumulated experience, and refined logic. These gradually settle and solidify during sleep, becoming the underlying support for future thinking and action.  
The closed loop between the two allows us to continuously absorb new things and grow.  
But today’s large models only have one set of ossified long-term memory.  
All their capabilities come from the massive data fed to them during training. That data eventually becomes fixed parameters in the model, like a hard drive pre-filled with content. Once training ends, they are completely fixed—a static function that cannot update itself.  
When you talk to it about new topics or give it new information, it can only temporarily use that information within the current conversation. It cannot truly update its underlying cognition. When faced with new changes it hasn’t been trained on, it either gives irrelevant answers or relies on external retrieval to patch together a response. It simply cannot flexibly adapt and continuously evolve like a human.  
The key to a truly deployable agent is to complete AI’s memory loop—essentially replicating the human brain’s memory logic.  
First, add short-term memory to AI, using simple storage to temporarily hold real-time interaction information—such as dialogue content and newly encountered facts—solving the problem of forgetting in the moment.  
Then, introduce a consolidation step, like human sleep, where during low-load periods useful content from short-term memory is filtered and integrated into long-term memory in a lightweight way. This neither disrupts existing capabilities nor accumulates new experiences.  
Once this loop is operational, AI can truly embrace new changes, remember recurring needs, and gradually evolve from a tool that merely recycles old knowledge into an intelligent assistant that grows continuously and adapts flexibly.  
Many people expect AI to be like a companion that understands change and grows.  
What prevents this expectation from being realized has never been complex technical challenges—it’s the completion of the most basic memory logic.  
Once you recognize AI’s existing shortcomings, it’s easy to see that the future gap between AIs will essentially be a gap in the completeness of their memory loops.  
And the first step toward true AI intelligence is never about piling on computing power or expanding data—it’s about learning to remember well and grow slowly, just like a human.