Transition
AI Jim Fan: Recently, I have been somewhat silent on X. This year has been a transformative journey for me. Although emerging large language models like Grok-4 and Kimi K2 are impressive, the field of robotics still resembles a wondrous western wilderness. It reminds me of the NLP field in 2018, when GPT-1 was just born, and BERT burst onto the scene, with thousands of novel ideas emerging. No one knew which idea would eventually become the future ChatGPT. There was endless debate, chaos, but the sparks of creativity were exciting. Now, I believe that the robotics ‘GPT-1 moment’ has actually arrived, hidden within the ocean of Arxiv papers, we just don’t know which one it is yet. It could be World Models, Reinforcement Learning (RL), learning from human videos, sim2real, real2sim, or some combination of these. The current situation is still reminiscent of the early days of NLP: intense debates, high entropy, but the ideas are both novel and fun, rather than just scrambling to improve the last percentage on AIME and GPQA rankings. The complexity of robot design far exceeds that of large language models. This is because LLMs only need to handle the clean and simple ‘bit world’ (i.e., text), while robots must confront the chaotic and real ‘atom world’. After all, what we control is a piece of real metal defined by software. Those working on LLMs might find it hard to imagine that, so far, the robotics field hasn’t even established a unified benchmark testing standard! Different robots have their strengths: some excel at acrobatic movements, while others are better at fine manipulation; some are suited for industrial scenarios, while others excel at household chores. A universal brain applicable across different carriers and various robots is not just a gimmick in research, but a necessary condition for realizing true intelligent agents. Recently, I have spoken with executives from dozens of different robotics companies, including both established enterprises and emerging startups. Some companies sell complete robotic systems, while others specialize in highly dexterous robotic hands. Even more companies choose to sell shovels—such as robotic manufacturing tools, simulation platforms, or data collection systems for training. The creativity in the robotics business field is flourishing, akin to a new gold rush, much like the entrepreneurial boom brought by ChatGPT in 2022. The best time to enter is often when there is the least consensus and the most chaos. Currently, we are still at the beginning of the loss function curve—although we see some strong signs of life, we are far from convergence. Each exploration of the gradient takes us into unknown territories. But one thing I firmly believe is that without deeply engaging with the real world to touch, perceive, and be endowed with a real body, the so-called AGI (Artificial General Intelligence) is out of the question. On a more personal note—leading a research lab has given me an unprecedented sense of responsibility. When I need to report progress directly to the CEO of a $40 trillion company, to be honest, this experience is both exhilarating and mentally taxing. It’s no longer as easy as before to keep up with and delve into every piece of news in the AI field. I will try to find time to continue sharing my journey with you.