AI is changing everything rapidly. As a PhD student whose work sits close to AI, I try to keep up with developments in models such as LLMs and agents. I am always a little afraid of falling behind in how effectively I work.
Recently, I have begun thinking seriously about the main focus I should choose for the rest of my PhD journey. Computational photography or imaging? They no longer seem like the safest choices for finding a good job. The markets are relatively small and, I imagine, concentrated in a handful of mobile-device companies. Even within computational photography and imaging, generative techniques such as diffusion models and flow matching are appearing everywhere.
Another major development is automated research. More and more startups are trying to build AI systems that can conduct research, both to advance AI itself and to accelerate other sciences. This raises an interesting and challenging question: what kind of research should humans focus on? If you choose to work on something that AI may master within five years, is it really worth pursuing? Would it be like deciding to become a professional Go player just before AlphaGo defeated the world's strongest players? Even a Fields Medal-winning mathematician has chosen to work on AI safety rather than continue to focus exclusively on mathematics. So, what are the truly important—and relatively “safe”—problems for humans to study if we want to build brilliant careers?
My personal, tentative answer is:
1. Help push the boundaries of (artificial) intelligence
If we cannot beat AI in our own domains, why not join it? Becoming a researcher or engineer who is deeply familiar with cutting-edge AI technologies and models seems like a relatively safe choice. Imagine working at a frontier lab such as OpenAI, Anthropic, or Google DeepMind: you would have a clearer sense of the boundaries of the most advanced models and would work alongside some of the most experienced people in AI. So, in 2026, how can we contribute to pushing the boundaries of AI? I have a very personal list:
- Core LLM methods (“old,” but still important): pretraining, post-training, data, and evaluation.
- Long-context learning: enabling models to learn across a lifetime, if that is possible.
- Memory for AI: closely related to long-context learning; perhaps strong memory systems will help us build models that can use longer contexts more effectively.
- Machine-learning systems: more efficient infrastructure and more capable systems for both researchers and users.
- Multimodal systems, including world models: AI cannot learn only from text. I doubt that we can achieve true intelligence solely through textual input, no matter how sophisticated the models become.
- AI safety and alignment: ensuring that increasingly capable AI does not harm society. This will become only more important as AI models continue to advance.
2. Keep doing work that genuinely interests me
Everyone needs personal interests. As AI models become more capable, they are dramatically reducing the time required to write code and plan experiments. Just as we all have hobbies in daily life, researchers can have their own “research hobbies”—topics they pursue because they genuinely enjoy them. Mine are computational photography, imaging, and graphics. Deep down, I believe that passion is all we need—and that it will always matter. Frankly, although LLMs are incredible and have become indispensable in my daily life, they have never given me quite the same sense of wonder as the “black magic” of computational photography. One example is The Moon Camera: even if multiple parallel universes existed, I doubt there would be a single one in which I could come up with that idea! These magical projects may not push the boundaries of “intelligence,” but they inspire me enormously. At heart, this is the kind of research I love: making the invisible visible and showing people something they have never imagined.
For me, graphics is all about art. I love high-end visual artistry such as rendering, where every pixel is worth designing. Another aspect that fascinates me is that graphics gives us a God’s-eye view from which we can create the visual content we see. We simulate the underlying principles of the world and show how things change. Pretty cool, isn’t it?