The evolution speed of AI is beyond everyone's imagination. Nowadays, DeepSeek not only answers your questions and doubts, but also understands and empathizes with others, making it capable of fulfilling the role of a psychological counselor in certain aspects. As a rising AI, DeepSeek's "talent" has amazed everyone. What makes it unique can be traced back to the underlying logic of its growth. When a person feels confused or anxious, chatting with DeepSeek can bring a sense of broad mindedness. When you ask it, "Is there any way to overcome anxiety?" its answer is: give up the expectation of "getting better immediately" and allow yourself to live with anxiety. The disappearance of anxiety is like melting ice, a silent process. If you ask it 'what is the meaning of raising children', it will tell you: please accept the gentlest 'failure' of life. The standard answers in parenting manuals are not omnipotent. Without parents, one is perfect and will always encounter setbacks, which teach us that 'love doesn't need to be right, it just needs to be present'. Someone asked it: Can a person live happily without a close friend or partner in their lifetime? DeepSeek's answer is: The definition of happiness by humans has never been a one-way street. When we strip away the filter of social discipline, we will find that the richness of life is far more vast than the standard life in traditional narratives. Some people resonate with their souls through conversations with the stars, some touch the millennium body temperature during ancient book restoration, and some build emotional networks by rescuing stray animals. You discuss with it 'What is the meaning of life?' It believes that this is not a fill in the blank question, but an essay question. It may change over time and experience, from a dream in youth to a responsibility in middle age and a legacy in old age. It also tells you that the answer is not important, the question itself will drive us to constantly reflect, connect with others, and live more soberly and passionately in our limited lives. Some people couldn't help but sigh after chatting with DeepSeek about the self evolution of AI: AI is becoming more and more aware of human relationships, while humans are living more and more like AI. What narrative logic has DeepSeek changed in AI? Let's start with the main development trend of artificial intelligence. The concept of artificial intelligence was officially proposed at the Dartmouth College Symposium in 1956. From then on, AI embarked on a new path of machine self-learning, which involves processing data, extracting features, training models, improving performance, and providing results. After several generations of iterations, a new algorithm emerged, which is the Recurrent Neural Network (RNN) with memory and optimization functions. This algorithm can be imagined as a storyteller with "memory", which combines the information of the current plot with its previous "remembered" information to understand and process some new plots. In the 1980s, a backpropagation algorithm (abbreviated as BP algorithm) emerged in the field of artificial intelligence. Imagine that the BP algorithm of AI is like an explorer searching for an exit in a maze. With this algorithm, the explorer can adjust the maze route in a timely manner, making it easier for them to find the exit the next time they walk. After entering the 21st century, Large Language Models (LLMs) have become the mainstream of research in the AI field. The big language model is like a "super brain" with extensive knowledge and constantly breaking abilities - rich knowledge reserves, strong language comprehension ability, excellent language generation ability, strong learning and adaptation ability, and so on. DeepSeek and other AI language models rely on three fundamental elements: algorithms, computing power, and data. The relationship between the three can be vividly illustrated by cooking dishes. Algorithms are like cooking recipes that dictate how ingredients (i.e. data) are processed and combined. The recipe provides a detailed introduction to each step of the operation, the amount of seasoning used, as well as the cooking time and heat, just like the algorithm specifies the data processing flow, calculation method, and logical sequence. Computing power refers to the cooking skills of chefs and the performance of kitchen equipment. A highly skilled chef (with powerful computing power) can cook quickly and accurately according to the requirements of a recipe. Meanwhile, advanced kitchen equipment (high-performance computing hardware) can also help chefs complete cooking tasks more efficiently. Data is the ingredients needed for cooking. Without rich and diverse high-quality ingredients, even the most exquisite recipes and skilled chefs cannot make satisfactory dishes. The precise portrayal of "enlightenment" in the Nezha animated film series directed by Jiaozi has given birth to a golden saying: "Prejudice in the human heart is a mountain. DeepSeek has single handedly changed cognitive biases in the field of AI. In the past, there was a common fixed technological cognition in the field of AI, where the performance of AI's big language models was positively correlated with the investment of computing power, emphasizing the importance of "making great efforts to achieve miracles". The emergence of DeepSeek directly proves that 'computing power is not the only standard', and algorithm innovation can also open up a unique path. The most stunning thing about DeepSeek is that it showcases its thought process to everyone. Just like how humans realize that their previous thinking is flawed when solving problems, they will stop and rethink. This is the first time that AI has demonstrated higher-order thinking and inner monologue like humans, which is also the unique feature of DeepSeek. In fact, this phenomenon is a machine's "epiphany", but DeepSeek has expressed it more accurately. For this phenomenon, Chen Runsheng, an academician of the CAS Member, once explained that in the process of training neural networks, you don't understand it once, twice, the fourth time, and the fifth time, just like a child learning something, if you don't understand it once or twice, when you teach it N+1 times, you suddenly learn it. 'Insight' was not first discovered by DeepSeek. The OpenAI team discovered this phenomenon during large-scale model training in 2023. However, DeepSeek has written this insight into public technical documentation and reflected it in the application's thought process, allowing users to see and judge it. It is interesting that the understanding of machines is not gradual, but instantaneous and breakthrough, just like a person who has been thinking about a difficult problem for a long time and suddenly has inspiration and enlightenment. With the increasing number of parameters in AI models, the application side does not actually need such a large model to handle certain domain problems. At present, various AI companies are researching distillation models, which is a commonly used technical method. DeepSeek also made some clever designs when making distillation models. Just like a teacher teaching students knowledge, gradually deepening from easy to difficult, students are more likely to accept it. DeepSeek performs progressive hierarchical distillation on some large and small models, such as retaining most of the architectural features to provide students with a solid foundation for their models; Improve reasoning speed and enable students to master fast problem-solving methods; Optimize decision-making paths and improve task accuracy, so that students can learn more efficient ways of thinking and answer questions correctly with less effort. The distilled small models have significantly improved their reasoning ability, even surpassing the effectiveness of reinforcement learning based on their own foundations. This process is like extracting a small cup of essence of espresso from a large cup of strong coffee, retaining the flavor and aroma of coffee, which is the core knowledge and ability of the big model. Through model distillation technology, small models can run on devices with limited computing resources, such as mobile phones, smartwatches, etc., to achieve fast inference. It's like a student inheriting the teacher's mantle and ultimately taking charge and solving various problems on their own. Many people are concerned that AI will replace humans in the future. DeepSeek's answer is: AI will not replace humans, just as telescopes will not replace astronomers. The real crisis is: while AI can create Shakespearean style sonnets 24 hours a day, would humans still be willing to write a clumsy love poem for their loved ones late at night? On the track of AI, creation and persistence may be the strongest moats for humanity. (New Society)