How does Gemini 3 learn to 'self upgrade'? Interpretation of Google's Latest Paper on "Nested Learning" This may be the path from Google to AGI, perfectly matching the world's dynamic models. This will once again drive demand from storage to communication, further exacerbating the supply-demand tension --Title: The original large language model LLM, after training, seemed to suffer from "amnesia" and was unable to truly absorb new knowledge into long-term memory. This is the dilemma currently faced by large-scale AI such as GPT or early Gemini. The Nested Learning (NL) paradigm proposed by Google researchers in their new paper "Nested Learning: The Illusion of Deep Learning Architects" and the resulting HOPE framework are breaking this deadlock. Some experts believe that the importance of this paper is no less than the well-known pioneering work in the LLM industry, 'Attention is All You Need' It no longer sees AI as a fixed 'knowledge base', but as a dynamic system composed of countless nested 'learners'. HOPE endows AI with the ability to continuously self learn and self correct, enabling models to truly possess the ability for continuous self growth. 1. Theoretical basis: The "human brain model" of AI learning one point one The Illusion of Deep Learning and the Truth of NL Traditional deep learning models, such as Transformers, are like a flat cake, stacked layer by layer. NL believes that this is an illusion that hides the true learning mechanism within the model. The truth of NL: Any complex AI model and its training process can be decomposed into a series of nested, multi-level optimization problems. Optimizers are also memory: even the optimizers we use to train models (such as Adam or Momentum) themselves have been proven by NL to be associative memory modules dedicated to compressing past gradient information. one point two Source of Inspiration: The "Multi time Scale" Updating of the Human Brain NL draws on the mechanism of the human brain: different neural regions in the human brain update information at different frequencies: High frequency layer: fast response, processing instant information (short-term memory). Low frequency layer: slow integration, solidifying short-term information into long-term knowledge (long-term memory). The core goal of the HOPE framework is to introduce this "multi time scale update" mechanism into AI architecture. 2. HOPE's two core competencies: continuous learning and self-improvement The HOPE framework solves the static problem of existing large models through its two innovative modules: 2.1 Continuous Memory System (CMS): Keep Memories Never Stop Addressing 'amnesia': Existing LLMs have fixed parameters (long-term memory) after pre training, and new knowledge can only exist in short-term 'contextual windows'. CMS mechanism: CMS extends the knowledge storage module (FFN) of the model into multiple sets of FFN modules, each with its own update frequency. Continuous solidification: Low frequency modules may only be updated once after processing a large amount of new data, thus slowly but continuously integrating and solidifying new knowledge into the long-term parameters of the model. two point two ️ Self modifying giant: Models learn 'how to learn' HOPE has added a "self modifying giant" module, which is a high-level meta learning mechanism. Meaning: The model no longer relies on external fixed learning rules (such as Adam), but recursively learns how to improve and optimize its own parameter update algorithm through a nested optimization problem. The foundation of Recursive Self Improvement (RSI): This is the key first step in achieving model self upgrading and Recursive Self Improvement (RSI). 3. The Impact of HOPE Framework on the Future HOPE not only outperforms Transformer and Titans in benchmark tests, but more importantly, it further clarifies a possible new direction for AI evolution. 3.1 Accelerating the Evolution of the Big World Model The world is dynamic: the laws and knowledge of the real world are constantly changing. The value of HOPE: The continuous learning ability and multi-scale modeling capability provided by CMS enable the big world model to absorb and simulate the dynamic changes, causal relationships, and long-term patterns of the world in real time. three point two An important step in Recursive Self Improvement (RSI) If recursive self-improvement is a necessary condition for agi implementation, then the Hope framework allows us to see the direction of this path more clearly. It is not an exaggeration to say that HOPE is not only an improvement in the ability of large models, but also a new growth curve of scaling law. Current progress: The HOPE framework has successfully built the infrastructure of RSI, verifying the feasibility of the two core mechanisms of "learning how to update" and "continuously solidifying knowledge". Unrealized parts: Autonomous correction of the objective function, multi-level recursive depth, and breadth improvement of all functional modules of the model have not been implemented yet. Speed up prediction: For narrow AI tasks with clear objective functions (such as optimizing code or strategies), the HOPE mechanism can achieve extremely fast self-improvement. After breakthroughs in multiple narrow AI fields, the capabilities of general AGI may emerge from them. 4. Business and Engineering: Challenges and Opportunities in Storage and Communication The HOPE framework requires higher hardware capabilities, which also creates huge opportunities for related suppliers: four point one Storage increase and memory consumption Static storage (parameter count): CMS expands FFN into multiple modules, which means that the model parameter count may increase by 10% to 50%, requiring more storage space. Dynamic memory (RAM): During runtime, the model requires additional memory to store and manage multi-level gradient flows and memory states across multiple time scales. four point two Comprehensive upgrade of communication network HOPE's requirements for distributed training networks are comprehensive improvements: Inter chip/motherboard communication: requires extremely high bandwidth to manage the memory status and high-frequency layer updates of CMS. Beneficial suppliers: NVIDIA (NVLink), AMD (Infinity Fabric), Samsung/Micron (HBM), Intel (CXL) Inter cabinet/inter cluster communication: The network is required to support non-uniform, multi-scale synchronization, where gradients of different levels are exchanged at different frequencies within the cluster. Impact: This will accelerate the upgrade of data centers to 800G/1.6T optical modules, increase the absolute quantity and value of optical modules, and benefit optical module, InfiniBand, and DPU suppliers. Conclusion: The HOPE framework is not only an innovation in AI architecture, but also another acceleration of scaling law, and a clear rehearsal for future data center infrastructure upgrades and the implementation of universal self improving AI. Accelerate again! (See the link to the paper below)