Beautiful view on the Eberhardsbrücke bridge in Tübingen, Germany.
I am doing "AI research", as a Tenure-Track Assistant Professor at Westlake University, Hangzhou, China. My group, ENCODE Lab, focuses on Efficient AI, an area trying to build efficient and reliable AI systems that are scalable and self-improving. In layman's terms, we try to make AI models run faster and/or consume less resources (energy, memory, etc.) on hardware. Speed/energy/memory are all about your time/money. So simply put, our job is to save your time/money when you are using AI :-).
Efficient AI is more about computing systems, at the implementation level of AI. If we look at the past history of AI (roughly starting from Turing's great paper in 1950), we might be impressed by the fact that AI is so interdisciplinary a topic that has been studied so broadly by many diverse groups. They try to answer what is human-level intelligence, especially what is the difference between human intelligence and animal intelligence.
We humans approach scientific problems through different levels (by using a kind of very special capability that we still know little about except its name: "abstraction"). In the AI field, if we borrow the wisdom from David Marr, the problem can be split into three levels: the "Computational theory" level (or, I personally prefer to call it the "Modelling Level"), the "Representation and Algorithm" level, and the "Hardware Implementation" level. For simplicity, if we consider the "Representation and Algorithm" level as the "software implementation" level and merged with the "Hardware Implementation" level, then we have two levels: the modelling level and the implementation level.
Both the modelling level and the implementation level are indispensable towards the final goal of solving the problem in the real world. But, in terms of scientific value, nuanced yet important distinction still exists (this distinction usually arises when we try to answer why one paper is more novel or important than another): the modelling level is about what and why; the implementation level is mostly about how.
The computing systems based on the semiconductor are well developed. They are the "body" of AI today. Most works in the efficient AI area are actually about how to improve the efficiency of the "body" (computing systems), of AI. In this sense, efficient AI, in my opinion, is mostly at the implementation level of AI, not the modeling level - I dare to believe that most papers in the efficient AI area prove my opinion. This is why the current efficient AI area works like a post-processing stage in the whole AI pipeline: after we have some super powerful model, we try to compress/accelerate it through various techniques (model compression, token compression, memory optimization, etc.). Another fact to see this is that: many researchers (esp. those with computer system background) may approach the AI efficiency problem by optimizing the memory access, where they do not even care whether the AI model is smart or dumb.
Efficient AI is like a "bridge"; tradeoff is everywhere. What we are doing is deploy AI models (deep neural networks) on hardware, and the work we do acts as a "bridge" between the two. Efficient AI is basically about getting the model training and deployment, and inference more efficient. My group is more focused on model deployment and inference. Some of my students will go on industrial internships for model training.
Efficiency is never a free lunch. Tradeoffs are unavoidable within different aspects: time, space (memory), energy, accuracy, etc. A good man in efficient AI is usually who is familiar with the model (deep neural network architecture), the algorithm (how model is used), and the underlying hardware (e.g., GPU/CPU/NPU/ASIC/FPGA, etc.), namely, a full-stack guy.
Computing systems are hierarchical. For today's AI computing systems, the problem can be split into usually four levels: the algorithm level, the model level, the system level, and the hardware level. We can get efficiency gains from each level. But different levels come with different potential gains and workloads - developing a fast AI inference algorithm (such as INT4 quantization) is easy (in 1 month), but the gain is also relatively limited (e.g., 2-3x speedup). During my Ph.D. period, I focused primarily on the algorithm level; now, my research spans all the spectrum of the computing systems.
Efficient AI is less about the task. AI tasks are unifying in the past several years. Efficient AI is therefore less about specific tasks. We are dealing with deep models, which are typically transformers. The data now is all token flow. As a result, we place no restrictions on the specific tasks we work on; we pursue new topics as they emerge (in Chinese: "追热点"). This is why you may see that my works have covered a wide range of tasks, including image classification, style transfer, super-resolution, text-to-image generation, large video models, robot manipulation, and SCI (snapshot compressed imaging), among others.
If we ask "What is the next big problem in AI now?", it is no surprise to hear many would say: "Obviously, how to achieve strong AI or AGI is the next big problem". The success of LLMs appears so astonishing that many people believe that strong AI or AGI is just around the corner, for the first time in history. The path leading us here is said in many different ways, but the core is the same: "Scaling Law" - scaling the model and feed it with more data, supported with more GPUs. Finger crossed: When the model/data/GPUs are enough, intelligence/magic will "emerge" on its own. This path is phrased in many different ways, such as "data fitting", "data-driven", "trial-and-error", etc. I personally prefer to call it the "empirical path", referring to the terms in the philosophical dispute between empiricism and rationalism.
I believe in the empirical path, i.e., fitting the data can lead to intelligence, but the key question is: to which level? The empirical path can lead us to a certain distance, nonetheless, no sufficient evidence convinces me that this trial-and-error approach can lead us for very far. I thereby firmly believe, a good computational theory is a must to bring us the next level of artificial intelligence in the road to so-called strong AI or AGI. Some may argue, ChatGPT/LLMs are so successful with the black-box deep neural networks, isn't this enough to show mathematics are not indispensable?
Well, a good theory is something arising from empirical data but gets beyond. The value of a good theory may have been well said by Kurt Lewin: "Nothing is more practical than a good theory (in the long run)". If this argument is not strong enough, I wish the following example can say something: Before Claude Shannon formalizes the mathematical definition of "information" in 1948, radio and wireless communication already existed for quite a while (the first radio was invented by Guglielmo Marconi in 1895), just in the dark. Without the light of information theory, we probably can still listen to Lennon's Watching the Wheels via radio; yet Turbo code and LDPC code will very unlikely appear, and we cannot have the Internet as we know it today. Big data will also very unlikely be possible, let alone the current AI boom. So basically, I consider the success of today's LLMs, which might appear so astonishing, is still at a very early stage, like the time when radio was just invented. We are in grave need of the "light".
Where does a good theory come from? If we look at the AI history, we can see many great minds from a variety of fields have been trying to answer the AI modeling problem: philosophy, psychology, neuroscience, cognitive science, in addition to the computer science and engineering. Now, it seems the room is (too) crowded with people of only a few kinds. The computer guys are the dominant force in the AI field; among them, LLM guys are the dominant -- as told by many, "LLM sucks the oxygen out of the room". No wonder, Yann LeCun chanted: "If you are a PhD student, don't work on LLMs" (I may not 100% agree with this statement per se, but 100% agree with the idea behind).
No need to think it hard, we will see that AI problem is clearly not just about computer science and engineering. Opinions from all disciplines should be considered seriously, especially philosophy, psychology, neuroscience, cognitive science, even when we are talking about the technical problems. However, here we will encounter a practically critical question: We involve the ideas from other disciplines, in what capacity, to what level?
I never doubt even a little bit about the formality and seriousness of the researchers in humanities and social sciences, regarding their interest and contributions to the AI problem. Yet, we (I mean those with a science or engineering background) are still self-identified as scientists, vs. those claiming themselves as philosophers or anthropologists, etc. We must set an important boundary to avoid our discussion from slipping into an ambiguous territory that loses our unique identity. The boundary is actually easy to see, if we recall the key characteristic of being scientific, falsifiability: measurable, computable, and replicable.
Philosophers or thinkers can discuss AI in an ambiguous (玄学) fashion - they are entitled with that kind of right, while we, scientist, have to be precise, clearly define what we try to discuss. In some situations, where we have reached the boundary between science and non-science (by non-science, I do not mean they are wrong - just yet to be precisely defined in a scientific way), there is no rigorous tools or concepts at hand, then we are "compromised" to use ambiguous terms. We ought to be careful and realize that is not the ultimate situation it should stay. Respect should be given to those who bravely attempt to make these ambiguous concepts precise, like Claude Shannon did with the measure of information and Alan Turing did with computability. Here, it is worth mentioning the great contributions made by Kant, Wittgenstein, among others, in the philosophy field - before recognizing the world, we need to examine in what capacity the tools (our rationality, language, to name a few) we use are reliable. For the concepts we cannot clearly define now, we should be cautious and refrain from using the terms, yet to be scientific, in a seemingly scientific way - Daemons often hide in these misused details, just like how Wittgenstein has warned: "Most propositions and questions that have been written about philosophical matters are not false, but nonsensical".
Hereby, I believe that the problems should be considered as one.
When I write this part, I try to figure out why I am writing this, along with other questions like what I am living for, etc. Being a researcher is an occupation, a means of livelihood. It is also a vocation, a way of "dasein", which should be examined.
On many occasions, we are accustomed to defining ourselves by a school, a department, a narrow field. I deem this as a necessity brought by industrial division of labor, not the way things ought to be, especially when what we study is AI.
It is not hard to see that AI is a very fundamental topic, one of the "holy grail" problems. No discipline in the past is like this: The creature is to be like the creator. Even if we refrain from a religious perspective such as "we are doing something like playing God", nor can we deny that AI research is probably the most important research of all time, if we take it really seriously.
This means we should not, and cannot, consider it a problem that is purely technical; nor deprive others (including those who do not even code) of the interest in the problem. At least, many AI researchers are electrified by the idea of inventing a super intelligent system that aims to replace common humans laboring in various fields, as many pieces of evidence have projected. But have we ever asked if they want to be replaced? Or, more broadly, why the technology has to keep going? Where will we end up? Throughout the history of science and technology, the ethics problems are ever-lasting and now even more important when it comes to AI.
Now, let’s just look at the builders of AI. As human beings, AI researchers must also answer questions like the meaning of life, and the deeper significance of doing AI. We can explain the technical motivation of a method, but we must also explain the motivation of the people who build this technology — the hidden motivation, I would call it. From my even limited experience, I have found that this hidden motivation determines the external outcomes, and also shapes the technical motivation itself.
If we consider the profound importance of the AI problem and the fact that the main producers of current AI papers are actually a very young group (Ph.D., master, or even undergraduate, high-school students), we might be shocked by this counterintuitive fact: a critical problem that concerns the fate of humanity is entrusted to a group of young people.
Now, the problem is, these students or young researchers (me included), as a group, have been shouldering a burden in a weird way: as scientists, we are supposed to be precise and clear, while the papers are filled with vague and ambiguous terms - what's worse, as time goes by, the newcomers even think it is the norm. In addition, to accomplish the important mission of AI research, we are supposed to input and think carefully; while most young researchers are "forced" to output (publish papers) at an unusually high pace (in Chinese: "卷") - always busy with publishing; no time to think, and hard to stop. The rapidly increasing submissions and low-quality reviews are not a surprise, as a result. Collectively, the community, both the academia and industry, now seems to be rampant with, in Max Weber's word, instrumental rationality. Anxiety is pervasive — something I can clearly sense within my own group — many are burning out. This state itself is non-technical. It hides behind the papers and influences the technical outcomes we see, yet it is not fully discussed.
In my group, I try to help my students to keep up with the pace and meanwhile seeking ways to "jump out". The key is to balance or tradeoff short-term goals (papers, awards, etc.) with long-term goals. The long-term goal hinges on the fundamental motivation of doing research / publishing papers. I personally am greatly inspired by two stories: 1) the address Principles of Research by Albert Einstein in 1918 for Max Planck's 60th birthday; 2) the story how Records of the Grand Historian (Shiji) was written by a father and son: SIMA Tan and SIMA Qian, two grand historians in China in circa 100 BC.
Here we say, motivation matters; perspective matters.
Some may argue these problems belong to philosophy and sociology, psychologists, etc., while we are computer scientists or engineers. They are not our areas. I think this is a valid argument, yet merely in a narrow sense. In a broader sense, we have no other choice but to accept the challenge, once it is realized that we cannot really outsource our own personal questions — the meaning of life, the purpose of existence, what kind of work, research, family, friendship, and relationships we want — to people called "scientists", "philosophers", "psychologists", or "sociologists", simply because they are deemed more "expert" in these areas than we are.
Hereby, I believe that the problems should be considered as one.
If you are interested in joining my lab, please feel free to contact me (refer to the admission.pdf). I also look forward to collaborating interdisciplinarily with researchers from different fields, including philosophy & social sciences, physics, and sustainability, etc.
- March 28th, 2026. (This statement was mostly conceived on a flight back from Tübingen to Shanghai, after CPAL 2026.)
