Acceleration is being accelerated: when AI begins to evolve on its own The so-called foam is just a linear thinking creature unable to understand the illusion of exponential curve thinking; And RSI is the force that further changes the shape of the curve. ---Title: The recently released Cloud Opus 4.5 is less than a week after Google released Gemini3, Nested Learning (NL) paradigm, and Hope framework. The model also exhibits some characteristics of recursive self evolution, such as: Call the tool yourself to check your answers Generate a test set for yourself Generate a better answer template for your future Generate a 'stronger strategy for the next answer' for oneself Of course, this is not truly recursive self-improvement, but it exhibits a series of "RSI like" features in reasoning, self reflection, tool invocation, and other aspects This is a subtle yet profound moment. The model did not improve by exponentially increasing the number of parameters or computing power, but relied on internal structural adjustments, resulting in a significant synergistic enhancement of reasoning, tool calling, code generation, and other abilities. This is the first time we have seen a large system showing a tendency towards self-improvement in certain parts. This may not be AGI, but it seems like its mathematical premise. This may not be a singularity, but it looks like its accelerating orbit. 1、 The concept of RSI: The improvement is improving The so-called RSI (Recursive Self Improvement) refers to a system that can continuously improve its own abilities based on its performance feedback in the environment. Furthermore, when a system improves its own capabilities, it also increases the efficiency of the next round of improvement. This means that the source of improvement is shifting from external to internal. The system no longer relies solely on new data or inputs, but begins to rely on its own structural efficiency to obtain new data or inputs and other necessary elements to achieve continuous self-improvement. Once RSI is activated, it will continuously accelerate. 2、 The true meaning of domain specific RSI: Cloud Opus 4.5 On Cloud Opus 4.5, we first see signs of localized and authentic RSI, which can be referred to as domain specific RSI or narrow domain RSI, mainly coming from three directions: 1. Self enhancement at the cognitive level (reasoning → task planning → data generation) The stronger the reasoning ability, the better the model can break down tasks, generate higher quality data, and provide a more accurate "learning roadmap" for the next round of training. This continuously improves the efficiency of the improvement itself. 2. Self enhancement at the engineering level (code generation → toolchain optimization) The improvement of the code capability of the model means that it can automatically optimize its training script, data cleaning process, and evaluation methods. This is equivalent to a scientist starting to improve his experimental equipment, enabling him to achieve higher accuracy in the next round of experiments. 3. Self enhancement at the knowledge level (generation → verification → regeneration) The system can automatically propose hypotheses, verify them, and iterate. This shifts the growth of knowledge from sequential advancement to circular advancement. Of course, Opus4.5 did not truly enhance its modeling capabilities; At present, all mainstream models - Opus, O3, GPT-5 family, Gemini 3, Qwen - cannot directly modify their own weights or architectures; Opus 4.5 doesn't work either. Its self-improvement is mostly at the level of reasoning. But there is no technically insurmountable gap between self-improvement at the inference level and self-improvement at the model level. Self improvement at the inference level can be seen as a testing ground at the model level. 3、 Regional RSI is actually a multi domain one The value of domain specific RSI is not only reflected in its modeling capabilities, but also in its ability to directly accelerate the scientific research chain It enables materials science to generate candidate structures faster, biopharmaceuticals to screen targets faster, engineering designs to converge faster, production processes to grow automatically, and software output to increase exponentially. Every field can form its own 'local self enhancement cycle', and each self enhancement cycle can bring about accelerated productivity and further acceleration Conclusion: Not only is there no foam, but it is also underestimated The emergence of Cloud Opus 4.5 and Gemini3 both point to the same fact - intelligent systems are gaining conditions to improve themselves, and we are getting closer to AGI/ASI and technological singularities. This change did not bring about market frenzy, It is because these advances have occurred too quietly and deeply, to the point where most people have not even realized that the thin white line on the horizon is actually a tsunami three thousand feet high in the distance. Most organisms in the market that only possess linear thinking abilities see the daily fluctuations of candlesticks; This is why the market is not only free of foam, but also seriously underestimated!