Yuanlong Bao @AcrossClient

@mcgovernmit Language is a way of expressing yourself, but humans don't have just one way to express themselves.

@mrdoubi666 Safety above all!!!

x.com/i/article/2073…

@fchollet Is very little data static? If it is static, how can it obtain generalization and generalized application capabilities? What if it is dynamic? I think our ideas are similar in some aspects.

RIFT: Why Reflection Alone Doesn’t Create Intelligence One observation kept appearing while I was building SelfWeave. People often say AI should “reflect.” But what does reflection actually mean? In many systems today, reflection simply means asking the model another question. “Was your previous answer correct?” “Can you improve it?” This often produces a better response. But I don’t think this is learning. Because once the conversation ends, nothing fundamental has changed. The system may generate a better answer. It rarely becomes a better system. ⸻ This led me to think about a different process. I call it RIFT (Reflective Intelligence Formation Theory). The core idea is simple: Reflection is valuable only when it changes the cognitive structure itself. Experience should not end as another memory. It should become part of the system’s future behavior. A reflection process should continuously ask: What happened? Why did it happen? Which decision was correct? Which strategy failed? What should be reinforced? What should gradually be forgotten? What general principle can be extracted? Only after this process should learning occur. ⸻ RIFT therefore isn’t another prompting technique. It isn’t another reasoning algorithm. It is a theory describing how experiences gradually become intelligence. The process looks something like this: Experience ↓ Reflection ↓ Pattern Extraction ↓ Strategy Reconstruction ↓ Memory Reinforcement ↓ Future Behavior Changes The important part isn’t generating another answer. The important part is changing the system itself. ⸻ This also explains why I think long-term memory alone is insufficient. Memory stores experiences. Reflection interprets experiences. Learning restructures experiences. Only then can intelligence evolve. Without restructuring, memory eventually becomes storage. With restructuring, memory becomes cognition. ⸻ In SelfWeave, this process is closely connected with dynamic memory networks and Hebbian reinforcement. Frequently validated experiences become increasingly connected. Rarely activated memories gradually decay. The system doesn’t simply accumulate knowledge. It continuously reorganizes its own cognitive structure. ⸻ Perhaps intelligence is not defined by how much an AI knows. Perhaps it is defined by how effectively it transforms experience into future capability. That is the question RIFT attempts to explore. #SelfWeave #RIFT #Research #Cognitive

Why Memory Is Not Cognition One thing has been bothering me while building AI systems. Many people believe that if an LLM has enough context, enough memory, or a better RAG pipeline, it will gradually become an intelligent system. I’m no longer convinced that’s true. Memory is important. But memory itself doesn’t produce intelligence. A database remembers everything. A notebook remembers everything. A vector store remembers everything. None of them think. The real question isn’t: “How can an AI remember more?” It’s: “How does an AI change because of what it remembers?” That distinction matters. A cognitive system shouldn’t simply retrieve past experiences. It should reorganize them. Strengthen useful experiences. Weaken outdated ones. Forget irrelevant details. Extract reusable strategies. And eventually transform repeated experiences into stable knowledge. In other words, intelligence isn’t defined by how much you remember. It’s defined by how your memory continuously changes your future behavior. This is why I think the next challenge after LLMs is not larger context windows. It’s cognitive architecture. A system that can continuously organize memory, reflect on experience, update its own strategies, and evolve over time. LLMs have already become remarkably good language models. Perhaps the next question is no longer: “How do we build a bigger model?” Perhaps it’s: “How do we build a system that can continuously become smarter?” I’m exploring this direction through a cognitive architecture project called SelfWeave. There are still many unanswered questions. But I believe cognition is not something we retrieve. It’s something we continuously build. #SelfWeave #RIFT #Research #Cognitive

I have a few questions I want to understand: 1. How is each record stored? Is it stored normally or as vectors? 2. How can multiple conflicting records coexist? 3. In a huge amount of records, how do you find related ones? Using keywords? How do you determine the application of the same keyword in different scenarios? 4. Is there any merging, pruning, or compression between records? 5. Last question, is it static or dynamic?

The biggest limitation of today’s AI isn’t reasoning. It’s the inability to grow. Every conversation produces an answer. Very few produce experience. A human engineer becomes better after every project. Mistakes become intuition. Successful decisions reshape future judgment. Most AI systems don’t. They answer, forget, and start over. We’ve spent years improving reasoning: • Larger models • Longer context • Better prompting • RAG • Agents • Tool calling These advances make AI more capable. But they don’t fundamentally solve one question: How does an AI accumulate experience and let that experience change its future cognition? Memory alone isn’t enough. A vector database stores information. A conversation history stores context. Neither necessarily forms cognition. Real intelligence isn’t just remembering. It’s reinforcing useful knowledge, weakening invalid assumptions, reorganizing internal structure, and continuously improving future decisions. Perhaps the missing layer of AI isn’t another larger model. Perhaps it’s a cognitive architecture that allows experience to become cognition. This is the direction I’m exploring.

新的斗争已经到来🫡🫡🫡