Oasis Capital Dialogue with Researcher Zhou Hao: Full of Curiosity

Aim for the top, and you'll likely land somewhere in the middle; dare to imagine what others won't, and you just might produce research you're truly satisfied with.

We sat down with Dr. Hao Zhou, Associate Researcher at the Institute for AI Industry Research (AIR) at Tsinghua University, to discuss the new paradigm of AI-powered scientific exploration. Enjoy.

Oasis Capital: Could you briefly introduce your research focus?

Dr. Zhou: During my PhD, I worked on syntactic parsing — essentially analyzing the grammatical functions of words in a sentence, or the relationships between them. This was crucial for traditional NLP because these word relationships represent the latent features that sentences expose to downstream applications. In 2016, I started working on machine translation, and in 2017 I joined ByteDance to focus on text generation. GPT models are all autoregressive: given the previous word, generate the next one. By then, academia had already largely figured out their architecture; everything else was engineering. So around 2019–2020, I began researching parallel text generation models, aiming to build the next generation of text generators — non-autoregressive models that generate all words in a sentence simultaneously. This approach has several advantages:

First, speed — it can fully saturate GPU compute.

Second, bidirectional modeling. Left-to-right is always unidirectional, which creates problems with bidirectional dependencies; future context can't be modeled well. "Generating everything at once" requires bidirectional modeling.

Third, no sequential bias. Defaulting to left-to-right assumes the generation target is sequential. Text is obviously sequential, so autoregressive generation works fine. But you can't know where a small molecule begins, and proteins have severe bidirectional dependencies. This sparked my research interest, and I developed a highly general-purpose model. Initially, non-autoregressive models achieved only 18% accuracy, compared to 29% for autoregressive ones. By 2022, many of our parallel generation models were outperforming autoregressive approaches.

I value the latter two characteristics more. After 2022, I focused on protein generation and design, as well as molecular generation for pharmaceutical and new materials applications. I see far broader research space here than in NLP. The future of AI for Science and AI as Science — AI-assisted science and AI as a new science itself — is vast. It can help traditional natural sciences like biology, chemistry, environmental science, and materials science achieve real breakthroughs.

Generating text, proteins, molecules, or genes — all involve complex structures composed of discrete symbols. The former goes from words to sentences; the latter from atoms to molecules, or individual amino acids to protein sequences. My past work in text generation, applying discrete-symbol generative models to new data modalities, is my strength and what I want to pursue.

Oasis Capital: You were already quite accomplished in text generation at ByteDance. Could you share what your main work was there?

Dr. Zhou: At ByteDance, I primarily worked on generation and pretraining, reaching GPT-2 scale. I developed AdsGPT, specialized for ad generation, which worked very well. My research there focused on general-purpose controllable generation — controllable via discrete or continuous methods — developing a series of algorithms layered atop large models to generate exactly what you want with 100% controllability.

For parallel generation, I wanted to build a toolkit to jump from the autoregressive paradigm to parallel generation. The generative models worked out, but large-scale parallel pretraining remained blank due to prohibitive costs.

Oasis Capital: Is personalized medicine the main application scenario for your research?

Dr. Zhou: Not just drugs — materials are also molecules: fluorescent small molecules, or new energy solar thin-film materials, all involve molecular design. Even enzymes, usable in bio-manufacturing to transform industry. My research relates broadly to biology, chemistry, and environmental materials. If it can be commercialized within ten years, it would bring enormous changes to human life.

Oasis Capital: Does your research direction align with where AlphaFold is heading?

Dr. Zhou: AlphaFold advances "protein structure prediction." My direction is "generation" — whether through pretraining or other methods, the goal is designing personalized products. For example, rapidly designing robust drug molecules given a disease, or designing proteins given functional requirements. This parallels machine translation: given English, generate Chinese.

Oasis Capital: Has the LLM breakthrough directly helped or advanced work on complex problems like proteins?

Dr. Zhou: That's an open question. Direct help certainly exists — Facebook's ESM, for instance, essentially applied BERT from text to proteins. But whether this is the ultimate path remains debatable. If you directly apply LLMs, you might reach 40–50% effectiveness; how do you get to 80–90%? Should you try a different path that could directly yield 80–90% results? Even if it's harder going now, once you break through, you might find clear sailing ahead.

Oasis Capital: In AI-assisted drug design, are there core problems the entire academic community is trying to solve? Have you encountered bottlenecks or difficulties in your research?

Dr. Zhou: Everyone is working on various generative models — given a target, generate molecules, and so on — all still exploratory. During modeling, you sometimes need to borrow tools from CV or NLP, but in practice you may not find suitable ones, or what academia considers currently suitable may lack future potential.

Modeling approaches for this problem keep evolving. Data is also a bottleneck — how to break through? Obtaining data is harder than getting text or drugs, and generated content is also difficult to interpret. Data can't be uniformly distributed across the entire data space; there's inevitably bias, requiring solutions to model robustness. While we haven't reached convergence yet, once consensus emerges, it means certain problems in the field have been solved. From my research perspective, one bottleneck is scientific exploration — what constitutes truly good modeling? Another is timing: only when consensus is reached and confidence established will people be willing to "brute-force" scale. Of course, consensus might only come after brute-forcing — a chicken-and-egg problem. We also face insufficient machine resources; GPUs are all going to text research, market prices are skyrocketing, cloud services are raising prices, procurement costs have doubled, and we're dealing with hoarding issues. The pressure is quite substantial.

Oasis Capital: Why did you choose to pursue AI for Science?

Dr. Zhou: In 2019, Weiying Ma was head of our AI lab, and he left a deep impression on me. He needed to track academic progress weekly across NLP, CV, robotics, and I organized study groups for this. I expressed that "generation" had great potential, and many generation-focused groups abroad were already moving into AI for Science — molecular and protein design. Ma and the company found this direction attractive. Though still in early stages, its prospects, social impact, and value exceeded text generation. So our work received company encouragement. Especially when COVID-19 hit in 2020, AIGC moved from virtual to physical. What is the "C" in AIGC? It's things you see in everyday life — a major turning point. AIGC currently generates well in "virtual" domains; if it can also break through in "physical" domains, it would undoubtedly transform human life.

Oasis Capital: In this wave of AI enthusiasm, where do you think AI for Science, particularly in biomedicine, proteins, and materials, will make its impact?

Dr. Zhou: ChatGPT falls right in my research area. I've witnessed its milestones one after another, and I deeply relate to its developmental journey. Even as an industry insider familiar with generative pretraining methods, I never anticipated such massive success in scale and effectiveness. Just ten years ago, text generation was almost unimaginable; now AI has sparked global AGI discourse. So I'm confident in generative AI's prospects.

Directly transplanting traditional NLP to proteins would be counterproductive. Text pretraining's philosophical principle is distributional semantics — similar contexts share similar meanings. Design mathematical principles from this philosophy, then do autoregressive generation — that's the training objective. The combination of both, including current practices, I believe puts NLP on a reasonable path. With proteins, things get complicated. The philosophical principle remains unknown; we need to explore its mathematical principles to solve it. For computer scientists, given a philosophical principle, you can start trial and error. Currently, our focus is determining proteins' philosophical principle, designing reasonable mathematical principles, and pairing principle-driven training objectives for extensive experimentation. We're conducting generative pretraining on proteins, hoping for breakthroughs.

Oasis Capital: For AI for Science, especially domain-specific disciplines, are domain experts also important?

Dr. Zhou: Absolutely. Here's an example: if AI development lacks domain experts, their understanding of data will be insufficient, potentially leading to obviously biased models that ultimately malfunction. Second, data often has redundancy. Take molecules: you could treat all atoms as coordinates in Euclidean space, essentially geometric point clouds. But this completely denies molecules' intrinsic structure — they can't be arbitrary point clouds in arbitrary Euclidean space, but rather points on specific manifolds within Euclidean space. Defining good modeling approaches that capture intrinsic structure and effectively reduce data modeling redundancy is crucial.

Collaborating with domain experts pays dividends. For example, 10 million data points might not suffice, but with expert collaboration, 10,000 points could train a fairly ideal model.

Oasis Capital: Currently, many domestic disciplines still lag behind foreign counterparts in domain understanding. Regarding gaps in domain-specific knowledge or research itself, do you think AI can compensate?

Dr. Zhou: It's possible. Precisely because certain disciplines have gaps is where leapfrogging becomes valuable. For example, materials design has centuries of accumulated experience, passed down through generations. Breaking existing limitations through human effort alone is nearly impossible. But if AI can extract this knowledge from data, you only need domain experts to assist analysis, and success becomes possible. Another example: internal combustion versus electric vehicles. In the past, ICE technology had huge gaps — years of German and Japanese accumulation gave them advanced technologies and processes delivering high performance, which we struggled to match and might never surpass. But with EVs, as a layman, I feel the technology resembles smartphone tech, suddenly narrowing the gap with foreign technology.

Chinese experts have very rich knowledge reserves in specific domains. With AI assistance, it's like switching tracks — gaining more convenient and efficient tools, potentially enabling leapfrogging.

Oasis Capital: Can AGI's mathematical reasoning capabilities, as predicted by academia, be applied to reasoning-assisted protein design or materials exploration?

Dr. Zhou: That's another branch of AI for Science. There may be commonalities, but they're not necessarily identical. Generating proteins or designing molecules may not involve mathematical capabilities. Just as brilliant mathematicians with strong creative abilities may not be able to develop drugs. Protein design capability comes from seeing vast amounts of data, then transforming previous physical simulations of proteins into direct predictions, parameterizing first principles. How machines align with such knowledge principles is hard to say; from my perspective, it may not be the same capability.

Oasis Capital: AI for Science aims to solve many problems. For instance, the computational path — will future quantum computing advances affect this?

Dr. Zhou: If quantum computing truly works, certainly. If quantum chemistry can really calculate, it's like having something where, given a target and a molecule, you can model their every movement every second — no need for further prediction.

Oasis Capital: If researchers vigorously advance AI for Science, will scientific development accelerate dramatically in ten to twenty years? Might cross-disciplinary influences proliferate rapidly?

Dr. Zhou: I think it's quite possible, and we hope to work in this direction. Both the Ministry of Science and Technology and the National Natural Science Foundation attach great importance to this.

Oasis Capital: Even domestically, are there different schools of thought in AI for Science approaches?

Dr. Zhou: Definitely. Infinitely subdivided, different people have different ideas and want to do things differently. Even among computer scientists, some start from NLP, from generative AI, from graphs, from computation, from domains. Everyone is climbing Mount Everest from different paths — all path-dependent. After convergence, the starting point becomes clear.

Oasis Capital: Do you think practical attention to AI for Science has increased?

Dr. Zhou: I think it's natural growth — overall AI attention has increased, and people are discovering many things are genuinely feasible. If NLP didn't have so much left to do, people would look for other directions to explore.

Oasis Capital: Globally, where are the leading research institutions in your field distributed?

Dr. Zhou: It's still unclear — a hundred boats racing. MIT, UW are quite good; at these schools, many domain experts are getting into computing while computer scientists are moving into applications. Who will do well may only become apparent in a few years; it's still early.

Oasis Capital: You returned to academia from ByteDance AI Lab. What motivated this decision?

Dr. Zhou: Several reasons. I spent five years at ByteDance, fortunate to grow with the company's rapid expansion. I learned how to do business, how to collaborate with business lines from a central platform — negotiating, allocating, managing — while doing much research around business needs. Five years is a milestone; I felt ready to change perspective or environment. Also, at the company I had short-term collaborations with many students. What I wanted to do was AI for Science, but from any company's perspective, this direction is too far from commercialization, too risky. Academia has higher tolerance for research risk.

Oasis Capital: How do you understand AGI?

Dr. Zhou: I think AGI is a concept. There are 100 AGIs in 100 people's minds. Some think if you can have voice conversations indistinguishable from humans, that's AGI; some think AGI is superintelligence... By common understanding, given current LLM capabilities, many consider it already a semi-finished AGI. What we should focus more on is: what's the purpose of achieving AGI? To obtain an AGI itself, or to solve specific needs in life and applications?

Celebrating Vitality

What do you think is technological vitality?

Technological vitality comes from researchers' curiosity about the world.

—Dr. Hao Zhou

Institute for AI Industry Research (AIR), Tsinghua University

Oasis Capital is a new-generation Chinese venture capital firm dedicated to discovering the most vital entrepreneurs of the next decade in China, growing alongside them to create long-term value. "Celebrating Vitality" is Oasis's vision and mission. This vitality is both the direction of structural transformation in our era and the resilient, evolutionary force of entrepreneurs.

Oasis Capital focuses on early and growth-stage investments, with individual ticket sizes of $3 million to $30 million USD, concentrating on robotics, artificial intelligence, technology services, and other fields, empowering China's technology-driven new service upgrades.