Introduction
Over the past few years, AI for Science has been the赛道 most likely to spark imaginations of "the next Nobel Prize being driven by AI."
The protein structure revolution sparked by AlphaFold, the rise of a batch of AI pharmaceutical companies, and the continuous penetration of large models into scientific research processes have made the outside world believe: AI has the opportunity to reshape life sciences, redefine drug discovery, and even push humanity's understanding of aging, disease, and life itself.
But after truly entering the industry, many people begin to realize that things are far more complex than imagined. Life sciences are not pure software systems, and drug discovery is not a process where a model generates a molecule and automatically succeeds. There are experiments, failures, regulations, clinical trials, real-world feedback, and an extremely slow and heavy industrial structure. Models are important, but the model itself does not equal drug value.
Hua Chenqing was drawn to this problem very early on.
In early 2026, he chose to leave his Yale Ph.D. program to join Aureka as an entrepreneur, serving as the Head of AI Research. What attracted him was not the title, nor another prettier resume label, but something more practical and rarer: large-scale computing power, a real experimental platform, and a flywheel system capable of continuously generating data.
In his view, the competition in AI for Science is no longer just a competition of models. The scarcest thing in the pre-AGI era is not making another single-point model, but the infrastructure that can connect AI to the real world: compute, model, wet lab, data generation, verification, and feedback loop. "Three years from now, if an AI for Science company lacks its own data generation capabilities, experimental closed-loop, and real feedback system, its window of opportunity will be very short." This is a judgment he repeatedly emphasizes.
At Aureka, he drives the construction of systems such as OpenDDE (Open Drug Discovery Engine), biomolecular foundation models, and Codex for Life Sciences. For him, OpenDDE is not just a structure prediction model, nor a replacement under the AlphaFold3 paradigm, but something closer to a drug discovery engine: it attempts to unify structure prediction, functional understanding, molecular design, and experimental feedback into the same system.
And further out, what he truly cares about is no longer just drug design.
As his research continues to deepen, Hua Chenqing becomes increasingly convinced that the microscopic biomolecular world and the macroscopic real world essentially follow the same laws. The biomolecular world model, the scientific world model, and the general world model have no clear boundaries; AI for Biology and AI for Science are just experimental fields on the path to AGI.
He believes that today's large models have already completed most of the foundational capabilities of AGI: language, knowledge, code, tool calling, agentic workflows, data annotation, and long-horizon task execution. But the remaining key parts belong to the not-yet-fully-defined world model and systems capable of continuous self-evolution.
"LLMs are more like the brain, while the world model is more like the internalization of the laws governing how the world operates. AGI is highly likely not something designed all at once, but something that grows out of self-evolving dynamics."
Therefore, while many are still discussing AI for Biology, AI for Science, drug design, or foundation models, this young researcher who just left academia has already set his sights on a larger proposition: how to build the scientific infrastructure for the post-AGI era in advance during the pre-AGI era.
And his latest self-identity is no longer an AI for Science researcher; he is preparing for the post-AGI era.
In this interview, we had the pleasure of inviting Will Hua. Will shared his detailed understanding of AI for Science, AGI, AI capital, and business. Technically, he discussed the scaling laws and reasoning of biomolecular foundation models, how OpenDDE unifies structure prediction and molecular design, and how the biomolecular world model connects to a more general world model.
We have compiled these frontline thoughts and collisions, hoping to give everyone a more intuitive feel. This is a complete reflection by a young researcher at the forefront of AI for Science on the current technology cycle, the life sciences industry, and the future path to AGI. Enjoy~
Z Highlights
Life is a continuous process of finding larger leverage for oneself until one reaches their bandwidth limit. For me, the two biggest changes over the past few years are: transitioning from a thinker to a doer, and shifting from academic worship to truly caring about a person's practical ability, execution density, and capacity to create results.
I believe the current AI for Science industry is essentially producing scientific data, experimental feedback, and real-world validation paths for the future general AI. Three years from now, general AI may have accumulated enough data and capability to begin possessing stronger science-in-AI abilities. By that time, if an AI for Science company lacks its own data generation capabilities, experimental pipeline capabilities, or real feedback closed-loop, it will likely be eliminated.
OpenDDE is the open-source drug discovery engine I am driving at Aureka. We hope it becomes the world model and all-in-one model for AI for Drug Discovery: not only serving structure prediction, but also gradually serving a more complete drug discovery process like pocket identification, binding affinity, complex modeling, and de novo design. We chose to open-source a part of its capabilities because AI for Science should not just be a closed capability of a few companies; it should also serve the broader scientific community and social good.
I believe that in the future, some companies or model formats focusing solely on "drug design models" may gradually disappear. Because we have found a more general approach: turning the large model into an engine, letting it search for drug molecules within the large model's embedding space. In other words, we no longer view drug design merely as a generation problem, but redefine it as a search problem in a high-dimensional intelligent space.
Aureka's most special feature is its data flywheel capability. Aureka's theme is AI infra for TechBio: we simultaneously possess a compute platform, a model platform, and a wet-lab platform. The combined three-in-one capability of compute, model, and wet lab is what provides the opportunity to truly change the R&D methods in the chemistry, biology, and pharmaceutical industries.
AGI is not the end point, but the starting point for civilization to restart; Aureka's goal is to build the life science engine for the post-AGI era in advance during the pre-AGI era.
I believe the development of AI will eventually be summarized into a kind of evolutionary dynamics, just like human development. Human evolutionary dynamics come from human biology; and AI will soon give birth to its own evolutionary dynamics. The realization of AGI is very likely not a one-time model leap, but a gradual occurrence through self-evolving. Inspired by human biology, my team and I are building a framework for a self-evolving system for post-AGI: allowing AI to continuously evolve from data, experiments, feedback, failures, and environmental changes.
01 From Graph Neural Networks to AI for Science: He Believes the World Can Be Modeled
ZP: Welcome Will, it's a great pleasure to speak with you today. Later we will focus on AGI and world models. Let's start with a simple question: Do you think LLMs themselves are the path to AGI?
Hua Chenqing: I think LLMs are a critical path to AGI, but not the complete framework. LLMs have already completed a very important step: they have unified language, knowledge, reasoning, code, tool calling, and agentic workflows into a universal interface. In a sense, LLMs have already accomplished 80% of AGI. Because they give AI very strong data annotation capabilities, task decomposition capabilities, code execution capabilities, and long-horizon workflow organization capabilities. But the remaining 20%, I think, cannot be solved by just making the LLM a bit bigger. The remaining part will definitely move toward world models. Everyone is talking about world models now, but honestly, this field hasn't yet seen a clear architectural paradigm like the Transformer for LLMs or ViT for vision models. We still don't know what the final world model will look like. But I judge that there will definitely emerge a more all-in-one model to learn the true world's state, action, dynamics, feedback, and causality.
LLMs are more like the brain, while the world model is more like the internalization of the world's operating laws. If a system can only talk, write, and call tools, it is not close enough to AGI. True AGI should be able to understand the state of a complex system, propose interventions, predict consequences, absorb feedback, and then update its strategy.
This is also the biggest difference between pre-AGI and post-AGI. In the pre-AGI era, AI is primarily a tool: humans define tasks, humans provide data, humans decide the next step. In the post-AGI era, AI will start to propose its own hypotheses, generate its own data, execute its own experiments, and update its own systems.
So I think the most critical thing in the next one to three years is whether the architectural paradigm of the world model can be defined, whether it can be scaled up, and whether it can be combined with a data flywheel and agentic workflows. If this closed-loop truly works, what we see will not just be a stronger LLM, but a self-evolving AGI system.
ZP: Could you first introduce yourself chronologically? Which key nodes do you think shaped who you are today? You can talk about your background, including when you started getting interested in AI for Biology?
Hua Chenqing: I was actually quite mischievous as a kid, loved to play, and wasn't the traditional hardworking type. What really changed me was when my grandmother got lung cancer during my high school years and later passed away. That event had a huge impact on me. From then on, drug design and drug discovery became etched in my mind.
I grew up in China and only went abroad for college. In 2018, I went to McGill for my undergrad. In the second semester of my freshman year, I took COMP551 Machine Learning, taught by Professor Will Hamilton. Before the class, I didn't know who he was, but later I realized he is a very important figure in the field of graph neural networks.
That course had a big impact on me. Will mentioned a viewpoint: graph can model everything. Because the world is fundamentally relational. There are relationships between people, between molecules and atoms, between residues in a protein structure, and even between events. Looking back now, my journey from graph neural networks to AI for Biology, and then to world models, is actually behind the same question: Can the world be modeled? Can complex systems be represented? Can relationships, causality, dynamics, and feedback be learned by a model?
At that time, I didn't use terms like "post-AGI," but that seed was already planted.
ZP: You did a lot of early research related to graph neural networks. For a freshman, getting into graph neural networks and graph theory isn't easy. Why were you initially drawn to "graphs"?
Hua Chenqing: In a sense, the discipline chose me, not the other way around. What attracted me most about graphs is that they provide a very general way to represent the world. This is actually consistent with how I understand world models today. A molecular system can be viewed as a graph, a protein complex can be viewed as a graph, a cell system can be viewed as a graph. On a more macro level, between two events, two states, or two actions, there can also be some hidden relational structure.
Later, when I did AI for Biology and biomolecular world models, I was essentially still doing the same thing. Except early on I was working on representations of nodes and edges; now I care more about state, action, function, feedback, and how a system evolves.
This is also why I feel graph neural networks were very crucial to me. It didn't just give me a research direction; it gave me a way to view the world.
ZP: It sounds like you've had a desire to figure out how the world works from very early on. How do you think this drive for exploration pushed you all the way from graph learning to AI for Biology, and then to world models?
Hua Chenqing: I recently read a book related to Demis Hassabis, which mentioned a viewpoint from Hinton, roughly saying that human desire for research is greedy. I think that's very accurate. Humans clearly know that AGI has risks, quantum computing has risks, and controlled nuclear fusion has risks, so why must we still do it? Because human desire to explore the unknown is greedy. We know a little, and we want to know more; we understand a system, and we want to understand a larger system. This can also be applied to individuals. I started with graph neural networks because I wanted to understand relationships. Then I moved to AI for Biology because I wanted to understand biological systems. Now I'm doing world models because I want to understand how complex systems operate, how they can be intervened with, and how they evolve. For me, these are not three disconnected directions, but a continuous line moving upward.
Graph learning made me believe the world can be represented; AI for Biology showed me how complex real-world systems are; and world models made me start thinking about whether an intelligent system can not just represent the world, but predict it, intervene in it, and even evolve with it. Everyone has their own bandwidth and their own bottlenecks. At least up to now, I haven't felt like I've reached a bottleneck. I still very much want to keep exploring forward.
ZP: Looking back at this early research phase from 2020 to 2024, which experiences do you think had the biggest impact on who you are today? Was it algorithmic training, engineering implementation, or a renewed understanding of the academic system?
Hua Chenqing: The biggest change for me during this time was turning me from a thinker into a doer. At first, I was mostly doing algorithms themselves, and later I slowly applied algorithms to real applications. For me, that application is drug discovery. I took the hammer of AI and smashed it onto life sciences and drug discovery.
AI made me realize one very important thing: AI is an experimental science. No matter how fancy an idea is, or how beautiful the theory is, if it can't be implemented, can't run on GPUs, can't be trained, and can't work in a real benchmark or real task, it's just an idea. I no longer blindly trust fancy ideas; I care more about whether a system can actually run, scale, and produce results.
Another change was that I gradually shed my academic worship. Between my undergrad graduation in 2022 and starting my master's, I interned in Yoshua Bengio's group. That experience was very important to me. I saw that top scholars are of course very impressive, and Yoshua himself is very hands-on, often showing up in group meetings to seriously discuss questions. But I also saw that a system with a big title can have a lot of chaos, uneven quality, and inefficiency internally.
So I feel this period shaped two of my most important changes: one was moving from thinker to doer; the other was moving from academic worship to truly caring about a person's practical ability. These two changes were crucial for my later decision to join Aureka, do frontier AI research, and build OpenDDE.
ZP: From McGill/Mila, then to Harvard, MIT, and later Yale and joining Aureka, has your work philosophy changed?
Hua Chenqing: I now strongly believe in one viewpoint: Life is a continuous process of applying leverage to yourself until you reach your bandwidth and can no longer apply more. After 2023, I went to Harvard and MIT, and worked with different people on protein design and RNA evolution. The outside world might think that being a visitor or visiting student at these places would make the slope of one's life immediately shoot up exponentially. But my personal feeling is that that period was more about accumulation than a qualitative change. It also made me more clearly realize what I wanted.
So what I care more about now is: Where can I get the biggest leverage? Where can I truly unleash my accumulated abilities? Where can I do more than just write papers, but train truly large models, build real systems, and connect to real experimental feedback? This is also why I ultimately chose to join Aureka. Not because I wanted to jump from one title to another, but because I saw greater leverage there: computing power, a model platform, an experimental platform, a data flywheel, and the possibility of pushing AI for Biology toward a post-AGI scientific system.
In the pre-AGI era, many people are still comparing schools, titles, papers, and citations. But in the post-AGI era, what truly matters will become: Who can build systems, who can generate data, who can connect real-world feedback, and who can turn AI from a tool into a continuously evolving scientific engine.
So my work philosophy change is simple: Don't blindly believe in halos; seek leverage. Don't stop at ideas; build systems. Don't just write pre-AGI era papers; prepare in advance for the scientific engine of the post-AGI era.
02 Leaving Yale for Aureka: What Attracted Him Wasn't a Title, But Compute, an Experimental Platform, and a Data Flywheel
ZP: Let's talk about Aureka. What choices did you go through before joining? Besides the team itself, what was the most attractive factor for you to join?
Hua Chenqing: Before joining Aureka, I actually had other options. Including some overseas AI for Biology startups, some platforms that had already raised a lot of money, and people who wanted me to lead a specific model project. But ultimately I chose Aureka. The core reason wasn't a title, nor a better-looking resume node, but that it gave me a very rare platform: sufficiently large compute, a real experimental system, and the infrastructure to form a data flywheel.
In academia, you're mostly validating ideas; but at Aureka, for the first time, I felt I could bring forward things that normally wouldn't be done for another three to five years and try them today. Like large-scale biomolecular foundation models, large-scale drug discovery engines, and large-scale antibody design systems. People who do AI know that small models and large models, small-scale training and large-scale training, are two completely different species. Many things only show new capabilities after reaching a certain scale. So what attracted me most to Aureka was its willingness to truly invest resources into frontier AI research. Not just making a very narrow application model, but letting me train a more general biomolecular foundation model to do something closer to a post-AGI era scientific engine.
For me, this is a bit like a timeline overtake.
If scientists in the pre-AGI era are still advancing science using papers, benchmarks, and single-point models, then the scientific system of the post-AGI era will definitely be a closed loop where compute, models, experiments, data, and feedback are all connected. Aureka's attraction to me was that it didn't just give me a position, but an environment where I could build this kind of closed loop in advance. So I left Yale not because I didn't believe in academia, but because I felt the frontier of this era had shifted from "who has a better idea" to "who can validate an idea in a sufficiently large system."
In the pre-AGI era, the truly scarce resource isn't smart people, but a platform that allows smart ideas to quickly access compute, experiments, and real feedback.
ZP: Aureka emphasizes functional antibody design, not just binding, but also more complex functions. For a model, is function harder to learn than structure? Why did Aureka choose to enter from this direction?
Hua Chenqing: Function is definitely harder than structure. Structure is more of a physics and geometry problem: what does this molecule look like, how does it fold, how does it bind to a target. But function is a higher-order question: Does it have a biological effect after binding? Does it activate or inhibit? Does it have a phenotype at the cellular level? Does it have therapeutic relevance? These things cannot be fully deduced just from structure.
In the past, the logic of antibody design was: do structure first, then binding, and finally look at function. But what Aureka wants to do is to put structure and function into the same evolutionary closed-loop. It's not a sequential relationship, but a parallel one. The model proposes variants, the experimental platform screens for molecules that truly meet functional goals during protein evolution, and then this data is fed back into the model. This is Aureka's most special feature: it doesn't just have models, it has a data flywheel.
This is also my understanding of post-AGI drug discovery. Future drug discovery won't be a model spitting out a molecule all at once and humans slowly validating it. It will be more like a self-evolving system: the model proposes hypotheses, the experimental system validates them, failure samples are absorbed, strategies are updated, and the next round of design becomes stronger. Function isn't a result to be looked at only at the end, but something that enters the reward, selection, and data flywheel from day one.
ZP: Compared to other antibody design companies, like Chai, is Aureka's data flywheel unique?
Hua Chenqing: I don't really want to talk about Aureka by evaluating others. More accurately, Aureka's difference lies in the fact that we don't understand AI for Biology as a pure model problem, but as an infrastructure problem.
A complete AI for Biology system needs at least three layers of capability: the first layer is compute, the ability to train sufficiently large models; the second is models, the ability to propose candidates, understand structures, and predict functions; the third is the wet lab, the ability to quickly generate real-world feedback. If any of these three layers is missing, the system is not a fully closed loop. Aureka's characteristic is that we have our own high-throughput experimental platform. For a given target or protein, it can generate million-level throughput in a very short cycle. This is crucial for AI because what the model truly needs is not a static database, but an action-feedback trajectory: what mutation I proposed, what the experimental result was, why it succeeded, why it failed, and how the next round should be adjusted.
The AURA012 project is a great example. Before 2023, there were no antibodies for this target. RFdiffusion designs all resulted in 0 hits. In 2023, relying on its foundation model, Aureka designed a weak antibody at 795nM, but it completely failed to meet the criteria for pipeline advancement, so the project was suspended. In 2024, by using the data flywheel to fine-tune the foundation model, an antibody at 127nM was designed. In 2025, using the data flywheel, an antibody at 93pM was designed. We internally observed a scaling law between the structural foundation model and de novo design.
We are not just building a model, nor just an experimental platform, but building a reality engine for biology in the pre-AGI era. Even when general AI becomes stronger, it will still need real-world interfaces, validation systems, and high-quality feedback data. Aureka's data flywheel is the prerequisite infrastructure for this post-AGI scientific system.
ZP: What do you think of the statement that "experimental platforms are one of the infrastructures of AI for Biology"? In life sciences, does compute mean not just GPUs, but also the ability to generate experiments and real data?
Hua Chenqing: I completely agree, and I would say it more directly: in AI for Biology, the wet lab is just another type of compute. The GPU computes in the digital world, and the wet lab computes in the real world. The GPU gives you gradients, and the experimental platform gives you ground truth. The GPU makes the model stronger, and the wet lab lets the model touch reality.
So the compute in AI for Biology cannot be understood merely as GPUs, AIDCs, or training clusters. True compute should include model compute, experimental compute, data compute, and feedback compute. The biggest difference between life sciences and pure software is that you can't just hallucinate an answer in a computer. Whether this antibody has function, whether this molecule has activity, and whether this target has therapeutic value must be validated back in the real world.
That's why I think the next three years will be very brutal for AI for Science companies. If a company only has modeling capabilities, lacks its own data generation capabilities, lacks an experimental pipeline, and has no real feedback closed-loop, then it will likely be compressed away by general AI. Because general AI will become increasingly strong—reading papers, writing code, calling tools, generating candidates, and doing preliminary analysis—these capabilities will become increasingly universal. The barrier of single-point models will decrease. What truly cannot be easily replaced is a system that can continuously produce high-quality scientific data, real experimental feedback, and reusable trajectories.
In other words, AI for Science companies in the pre-AGI era are essentially producing scientific data and real-world interfaces for the post-AGI era. What looks like AI for Science now may become science in AI in the future. By that point, AI won't just assist science, but will internalize scientific experiments, validation, failure, and iteration into its own capabilities. So Aureka's theme is AI infra for TechBio. The "infra" here is not just models, not just GPUs, nor a cloud platform. It includes compute, models, wet lab, data flywheel, verification system, and therapeutic pipeline. All these things put together are the true TechBio infrastructure.
ZP: From the perspective of the entire molecular design pipeline, compute may be gradually approaching saturation, and people will naturally move to open up the wet lab side. What do you think of this trend? If they don't have their own experimental closed-loop, where will pure model AI for Science companies get stuck?
Hua Chenqing: The biggest risk for a pure model company is that in the end, it can only generate predictions, but cannot generate reality.
In life sciences, prediction is not the end point. Predicting a structure, generating a molecule, designing an antibody—these are only the beginning. What's truly difficult is: Can it be expressed? Can it be synthesized? Can it bind? Does it have function? Does it have specificity? Does it have developability? Is it safe? Does it have a chance to enter a real pipeline?
If there is no experimental closed loop, the model will get stuck in two places. The first is data. It will increasingly rely on public data, collaborative data, or synthetic data, but this data will soon be absorbed by general AI, and the barrier will become thinner and thinner. The second place is validation. The model can propose many beautiful-looking candidates, but if they cannot be quickly validated, there is no way to know which directions are truly effective, nor to turn failures into fuel for the next round of evolution. So I think the future competition in AI for Biology will not just be "whose model is bigger," but "who can complete the design-build-test-learn closed loop faster." It is not enough for the model to only handle design; it must be able to build, test, and learn. Otherwise, it will forever remain at the very front end of preclinical discovery, unable to touch the core of true drug value.
ZP: You just mentioned that many of the rules of drug discovery are still in the hands of MNCs (multinational corporations). Is it possible for Aureka to break this structure, or can it only first change the early research stage?
Hua Chenqing: In the short term, what Aureka will change first is definitely the early research stage. Because this is where AI is easiest to enter and can generate the most leverage.
The cycle from drug discovery to clinical trials, and to even later stages, is very long. AI can compress early research a lot, for example, compressing an 18 to 36-month process down to 9 months, or even shorter. But stages like clinical trials, regulation, trading, manufacturing, and commercialization will not disappear immediately just because AI appears. They are determined by long-standing industry systems, risk appetites, and regulatory frameworks. So I wouldn't say Aureka can rewrite the entire pharmaceutical industry tomorrow. A more accurate statement would be: Aureka first turns the very front 20% of drug discovery into machine-speed. In the past, this part relied on human experience, low-throughput experiments, and slow trial-and-error; now we use models, data flywheels, and experimental platforms to turn it into a rapidly iterating self-evolving system.
But in the long run, I think the impact of this will gradually transmit backward. When the speed of early discovery increases by an order of magnitude, the systems further back will definitely be forced to change. Because if the front end can produce more high-quality candidates every year, the traditional preclinical, BD, validation, and decision-making systems will encounter new pressures. They cannot forever digest machine-speed science at a human speed.
In the pre-AGI era, AI can only first embed itself into old scientific processes to help humans accelerate certain steps. In the post-AGI era, AI will begin to restructure the scientific process itself: it won't just provide candidates for the old system, but will form a new closed loop of hypothesis, design, experiment, verification, and decision-making. What Aureka is doing today is establishing this closed loop in the early research phase first.
Aureka is more like building the infrastructure for a post-AGI life science system in advance during the pre-AGI era. Today, it's changing early research; in the future, it might change the very speed at which scientific discovery operates.
03 The Biological World Model: Drug Design is an Experimental Field for AGI
ZP: Let's talk about something more technical. Let's start with OpenDDE. How do you define OpenDDE? Is it a biomolecular foundation model, or a drug discovery engine? Compared to models like AlphaFold3, Chai, and Boltz, what is its most critical technical breakthrough?
Hua Chenqing: I would prefer to define OpenDDE as an open-source drug discovery engine, rather than just a biomolecular foundation model. If it were just a biomolecular foundation model, it would be answering: given a sequence, a structure, and context, I predict a structure. But a drug discovery engine has to answer a bigger question: given a disease context, a target, and a desired function, can I search through a massive space of biomolecular possibilities to find molecules, structures, and functions that truly have value?
So the goal of OpenDDE is not simply to be a replacement for AlphaFold3, Chai, or Boltz. Those models are very important, but they lean more toward being core models of the structure prediction era. And I hope OpenDDE takes a step further: from structure prediction to a unified engine for structure-function-design.
For its most critical breakthroughs, I will mention three points. First, we provided the scaling law for biomolecular foundation models and calculated the slope of this curve. Second, we defined the reasoning of biomolecular tokens. Third, and what I consider the most important breakthrough, is that we unified the training architecture for structure prediction and de novo design. This is also why I say OpenDDE is not a pure prediction model, but a drug discovery engine.
The previous generation of models mainly answered "what does this molecule look like"; OpenDDE wants to answer "how should I search for a new molecule with function, structure, and drug value within a biological system." Looking at this on a larger timeline, it is actually a transition from pre-AGI to post-AGI. In the pre-AGI era, we first turn biology into a modelable, searchable, and verifiable space; in the post-AGI era, AI will directly propose hypotheses, design molecules, execute experiments, absorb feedback, and update strategies within this space. OpenDDE is the early infrastructure for this direction.
ZP: How much computing power is roughly needed to train OpenDDE? Why is it difficult to fully validate a model like this in academia?
Hua Chenqing: A lot. We spent roughly a cluster of a thousand cards and close to half a year of training time. We also calculated that if trained on a single card, the entire GPU time would be about 54 years. This is why it's difficult for a model of this scale to be fully validated in academia. It's not because academia lacks smart people—on the contrary, many of the best ideas come from academia. But a model of this scale is no longer just an algorithm problem; it's an infrastructure problem. You need to build a large-scale GPU cluster from scratch, need long-term stable training windows, need a sufficiently large data system, need an engineering team to continuously maintain the pipeline, and need to constantly perform ablations, evaluations, reranking, and failure analysis. Academia might be able to get some cards, but it's very hard to continuously devote 6-12 months to training a complete biomolecular foundation model from start to finish, and then systematically verify its scaling behavior and downstream capabilities.
ZP: Let's talk about the antibody world model. How do you understand the world model in antibody design? What is its relationship with OpenDDE?
Hua Chenqing: The antibody world model as I understand it is not a standalone antibody generation model, but a closed-loop system. In this system, the state is the current biological context: what the target is, what the epitope is, what the existing parent antibody is, what the historical experimental feedback is, which mutations succeeded, and which failed. The policy is the system deciding what to do next: whether to choose a certain parent, explore a new CDR mutation, optimize affinity, or improve specificity. The action is generating specific new antibody variants. The world model predicts the consequences of this action: what the binding mode is, whether affinity improves, whether the structure is reasonable, whether the function meets expectations, and whether off-target risks are controllable. OpenDDE plays the role of the biological world model and verifier here.
The large model can act as the brain, responsible for proposing strategies, selecting actions, and generating candidates; but candidates cannot be judged solely by logic; they must be put back into the biomolecular world model for verification. What OpenDDE does is exactly this: it tells you how this antibody might bind, whether the structure is reasonable, and whether it has the potential to become a real drug candidate. So we redefined the antibody design problem into a drug search problem within the large model's embedding space.
This is also why I believe that in the future, some independent model formats focusing only on "drug design" might be restructured. Because when LLMs become strong enough, world models become strong enough, and verification systems become strong enough, the design itself will become a search and decision problem, rather than a single-point generation problem. Today's structure might be: LLM proposes, OpenDDE verifies, and the wet lab provides real-world feedback. But in the future, I think it will move toward all-in-one: the next generation of scientific world models will propose, simulate, verify, and update their own strategies. At that point, drug design will no longer be "a model generates a molecule," but a self-evolving system continuously searching, verifying, and evolving in the biological space. This is the shape of drug discovery in the post-AGI era.
ZP: Can this biomolecular world model be placed on an agent architecture? From LLM for Science to scientific LLM, and science in AI, what evolution will happen in between?
Hua Chenqing: Yes. More accurately, I think antibody design will become an agentic world model for drug discovery. In the earliest days, we had standalone LLMs. They could read literature, answer questions, and summarize knowledge. Then came the agent stage, where it could call tools, write code, run pipelines, and do some automated tasks. Further down is agentic LLM: it doesn't just execute a task once, but has loops, memory, evaluators, feedback, and the ability to update its own strategy.
The same evolution will happen in science and biology. The first stage is LLM for Science. The second stage is Scientific LLM. The third stage is Science in AI. At this stage, science is no longer just an external application of AI, but becomes a native capability inside AI. AI can propose its own hypotheses, design its own experiments, call its own wet labs or simulation systems, absorb the results itself, and then update its next round of strategy. This is what I call post-AGI thinking.
In the pre-AGI era, we put AI into scientific workflows to assist human scientists. In the post-AGI era, scientific workflows will be internalized into the AI system itself. AI is not just helping you do science; science happens inside AI. So I think the most important significance of AI for Biology is not just making a better antibody or a better structure prediction model. It is an experimental field for AGI.
Because biology is a real-world system: it has physical constraints, functional constraints, experimental costs, failures, feedback, and multi-scale dynamics. If an AI system can complete the closed loop of hypothesis, design, verification, feedback, and update in biology, it is not just doing drug discovery; it is learning how to intervene in the real world. This is also why I say drug design is an experimental field for AGI. AGI in a chat window is just a manifestation of intelligence; true AGI must be able to enter the real world, understand complex systems, propose interventions, endure failures, absorb feedback, and continuously evolve. Biology is one of the hardest, yet most valuable entry points.
OpenDDE is a piece of infrastructure we are building in the pre-AGI era. Today it serves structure prediction, antibody design, and drug discovery; but in the long run, it is preparing for the scientific intelligence of the post-AGI era.
04 From the 3D Microscopic World to AGI: Why Self-Evolving Might Be the Next Main Theme
ZP: When you work on the biomolecular world model, is there a clear correspondence between it and the general world model? What does this link specifically refer to, and has it spawned a clearer understanding of world models for you?
Hua Chenqing: AI for Biology is actually a very important training ground for general world models. Because biology has both physical constraints and functional constraints; it has both structure and dynamics; it has parts that can be simulated, and parts that must be validated through experimental feedback. It forces your model to not just generate a beautiful structure, but to enter the real-world closed loop: predict, validate, fail, update, and predict again. Philosophically speaking, I feel a world model is essentially searching in a massive space of possibilities. We don't know the right answer from the beginning, but through models, experiments, feedback, and iterations, we constantly eliminate the impossible and approach those states that can truly happen and have true value.
The biomolecular world model does this: in the massive sequence-structure-function space, it eliminates molecules that cannot become drugs and approaches truly effective drug candidates. The general world model also does this: in the massive world-state space, it simulates the consequences of different actions, eliminates infeasible paths, and approaches better decisions and future states. So for me, the biomolecular world model is not just a biology-specific small model. It makes me see more clearly that the first principles of world models are actually unified: understand the state, predict interventions, absorb feedback, update strategy, and then continuously approach better possibilities in a complex world.
ZP: This naturally leads to the topic of AGI and self-evolving. How do you define self-evolving? What is the difference between it and Agents, reinforcement learning, or AutoML?
Hua Chenqing: If you look back at the history of scientific development, many major breakthroughs are actually like an evolution. For example, why was the Transformer born? Before the Transformer, people went through many stages like RNNs, CNNs, attention, and representation learning. The real bottleneck of that era was: sequences were too long, long-distance dependencies were too hard to model, and traditional architectures couldn't scale efficiently. So a large number of researchers continuously tried, failed, and eliminated impossible architectures, and finally, the Transformer was found.
So I would ask a more philosophical question: Was the Transformer created by humans, or did it always exist in some space of possibilities, and was finally discovered by humans? I am now increasingly leaning toward the latter. Many great scientific discoveries and AI architectures fundamentally seem to exist in a massive latent space. Humans continuously eliminate the impossible through experiments, failures, constraints, and feedback, and finally find that possibility. AlphaFold2 is similar. It didn't suddenly drop from the sky, but was a result evolved after a series of conditions matured, like biological data, structure prediction tasks, MSAs, attention, geometry, end-to-end learning, and CASP evaluators. Behind it is not a single-point genius, but the result of the entire field's long-term search, failure, accumulation, and selection.
This is my understanding of evolution: every era has a state, and every state has its own bottleneck. When the bottleneck is solved, the system enters the next state, encounters a new bottleneck, and continues to evolve.
The development of AI is the same. From early representation learning, to Transformers, to GPT-2, to scaling laws, to instruction tuning, RLHF, and reasoning models, to today's agentic LLMs, coding agents, and looped LMs, every step is a new paradigm evolved to solve the bottleneck of the previous stage.
The problem in the past was that this evolution was primarily driven by humans. Humans proposed directions, humans wrote code, humans designed experiments, humans summarized failures, and humans decided what to do next. But this era is very special: for the first time, we simultaneously have a sufficiently strong compute infrastructure, sufficiently large AIDCs, sufficiently strong LLMs as brains, sufficiently strong coding agents, and an increasing number of data flywheels that can automatically generate and validate. This means that AI is no longer just the result of evolution; it is beginning to be able to participate in evolution itself.
I think this is the key to self-evolving leading to AGI. Humanity has its own evolutionary dynamics, coming from human biology, environmental selection, cultural accumulation, and tool use; AI will also give birth to its own evolutionary dynamics. This dynamic may not be the same as humans, but it will continuously advance itself through data, compute, models, agents, experiments, feedback, and self-rewriting. So my understanding of self-evolving is: it is not a smarter agent, nor a bigger model, but a set of evolutionary mechanisms that allows AI to continuously break through its own bottlenecks.
ZP: Why do you think AGI will arrive in this era? Is the core reason the computing infrastructure, the data flywheel, or the maturation of LLMs as brains and coding agents?
Hua Chenqing: I think AGI will arrive in this era not because of a single-point breakthrough, but because several key curves have converged at the same point in time for the first time.
First is the computing infrastructure. Today's computing infrastructure has become strong enough to give us the opportunity to model increasingly complex systems. In the past, many things were not theoretically impossible; rather, compute, data, and engineering systems were not mature enough. Now, GPUs, AIDCs, distributed training, inference infrastructure, and data pipelines are all rapidly maturing, which makes "modeling everything" go from a philosophical vision to an engineering-driven direction for the first time.
Second, LLMs are starting to become a general brain. It is not just a language model, but begins to possess the ability to understand, reason, plan, call tools, write code, and execute tasks. The maturation of coding agents is especially critical, because once a model can write code, debug code, run experiments, and analyze results, it is no longer just passively answering questions; it begins to possess the ability to modify the external world, and even modify its own system.
Third is the data flywheel. In the past, AI's evolution mainly relied on humans providing it with annotated data, designing tasks, writing benchmarks, and giving feedback. But now agentic LLMs can automatically generate data, write code, design experiments, perform annotations, and evaluate results. That is to say, AI is starting to participate in producing its own training data and evolution environment. This is a very fundamental change.
ZP: From the current moment, the era of AGI might be approaching, and self-evolving has reached a feasible stage. What does this moment mean to you personally? Or for Aureka, what significance does this moment give you?
Hua Chenqing: For me, AGI is not an end point, but the starting point for all things.
In the past, we viewed AI as a tool: helping people write code, analyze data, and improve efficiency. But the true significance of AGI arriving is that humanity, for the first time, possesses an intelligent infrastructure that can continuously self-improve and evolve alongside scientific systems. It will gradually turn many goals that seemed distant or like sci-fi in the pre-AGI era into problems that can be advanced through engineering: extending the boundaries of life, understanding life and intelligence itself, exploring more efficient energy, going to Mars, and even moving toward more distant interstellar civilizations.
So I don't think AGI is just a technological milestone. It is more like the beginning of a new stage of civilization. AI will continuously evolve like biology, and this evolution will open up a massive number of new opportunities unimaginable in the pre-AGI era.
For me and Aureka, the significance of this moment is: we are not waiting until AGI arrives to start thinking about the future, but are already preparing for the post-AGI era in the pre-AGI era. What we hope to build is not a short-term tool, but an intelligent system capable of participating in scientific discovery, understanding biological systems, and continuously self-evolving.
AGI arrives, and all things come to life. And we hope to be among the first to bring this capability into the life sciences and real-world creation. AI for Biology is the first experimental field for AGI to enter the real world; what we want to do is transfer this self-evolving capability from life sciences to a larger scientific battlefield.
ZP: If AGI really comes in three years, looking back at today from three years later, what impact do you hope you, your work, Aureka, or the system you described will leave behind? For example, an open-source ecosystem, real drug value, or a path to AGI?
Hua Chenqing: If AGI really comes in three years, I hope that looking back at today from three years later, people will feel we did one thing right in this moment: we didn't just understand AGI as a smarter chatbot, a code assistant, or a productivity tool, but understood it as an infrastructure that can participate in scientific discovery, promote the understanding of biological systems, and continuously self-evolve.
I hope the impact Aureka leaves is not singular. Not just leaving behind an open-source project, nor just a drug pipeline, nor even just a single paper. I hope more that what we leave behind is a new paradigm: letting AI not just predict structures, generate molecules, and screen targets, but be able to continuously propose hypotheses in real scientific problems, execute experiments, absorb feedback, and update strategies, ultimately making scientific progress itself more computable.
Of course, real drug value is very important. Life sciences are not a pure demo; in the end, they must return to patients, diseases, and real-world efficacy. So if looking back at today from three years later, I hope Aureka at least proved one thing: a self-evolving AI system can not only get stronger on benchmarks, but can also produce verifiable value in real drug discovery, helping us reach new targets, new molecules, and new treatment options faster and at a lower cost. At the same time, I also hope we contribute to the open-source ecosystem. Because the era of AGI shouldn't mean intelligence is only held by a few companies, and science shouldn't be advanced only by closed systems. We hope to open up some foundational capabilities, models, tools, and methods so that more researchers, entrepreneurs, and labs can stand on this foundation and continue exploring.
But the most core impact, I hope, is the third point: Aureka can prove that one of the key paths to AGI or post-AGI is science. Because science is fundamentally the process of intelligence continuously understanding the world, intervening in the world, and updating itself from failures. If a system can truly participate in scientific progress, it is not just a model, but approaching a higher-order form of intelligence.
So if AGI really comes in three years, I hope today's Aureka is remembered as: one of the teams that started preparing for post-AGI scientific infrastructure the earliest in the pre-AGI era. What we hope to leave behind is not a short-term product, but a path: letting AI transform from a tool into a participant in scientific evolution, and letting life sciences become the first important entry point for AGI to change the real world.
ZP: For young researchers with a general AI background who want to enter AI for Biology, what advice would you give them?
Hua Chenqing: I would give them three pieces of advice.
First, do not treat biology as just another modality. It is not about treating protein sequences as text, molecules as graphs, and structures as 3D point clouds, and then just applying a model and being done with it. The hardest part of biology is that it is a real-world system: data is noisy, mechanisms are incomplete, experiments have costs, failures are common, and many ground truths need to be generated by yourself. So entering AI for Biology, you can't just be model-first; you need to quickly become problem-first and system-first.
Second, get close to real experiments and real feedback as quickly as possible. In the future of AI for Biology, the scarcest resource will not be yet another larger trained model, but the ability to generate high-quality data, the ability to validate through experiments, and the ability to iterate in closed loops. If a young researcher only stays at the level of benchmarks, papers, and leaderboards, they may be quickly caught up by general AI. But if they understand assays, understand wet labs, understand disease mechanisms, and understand what kind of prediction can ultimately turn into real-world value, they will have a longer-term advantage.
Third, retain the abstract thinking ability that comes from a general AI background. AI for Biology is not just about building biology-specific models. It is actually about studying how an intelligent system can understand complex systems, propose hypotheses, design interventions, execute experiments, absorb failures, and then update its own strategies. This process is highly correlated with AGI, world models, and self-evolving systems. So my advice to young researchers is: don't limit yourself to being "someone who builds protein models" or "someone who builds drug design models." Instead, think: can I build a system that allows AI to continuously learn, validate, and evolve within the life sciences?
So to sum it up in one sentence: the greatest advantage for someone with a general AI background entering AI for Biology is not knowing how to tune models, but the opportunity to redefine biology as a real-world intelligent system problem. But the prerequisite is that you must respect the complexity of biology, stay close to real experiments, build your own data flywheel, and think about closed-loop validation from day one.
05 Let AGI Happen Through Chinese Innovators: Transferring AI for Science Capabilities to a Larger Battlefield
ZP: Finally, regarding AGI, self-evolving, world models, or the inspiration that AI for Biology offers for AGI, is there anything else you'd like to express?
Hua Chenqing: What I want to express last is that the inspiration AI for Biology provides for AGI may be underestimated.
Many people think about AGI from the directions of language, code, mathematics, and robotics. But I believe the life sciences offer another extremely important path. Because life itself is a highly complex, self-organizing, multi-scale, and continuously evolving system. A truly powerful intelligent system should not just be able to answer questions, write code, and do reasoning; it should be able to understand how a complex system changes, how it is intervened upon, how it updates itself from failures, and how it generates new capabilities in the real world.
This is why I have always focused on self-evolving and world models. The core of a world model is not "predicting the next frame," but understanding the dynamics of the world. The core of self-evolving is not "automatic hyperparameter tuning," but whether a system can continuously evolve from its own history, experiments, failures, and feedback. The life sciences happen to be one of the most real and challenging scenarios for testing these capabilities.
So for me, AI for Biology is not a vertical application; it is a training ground for AGI. Drug discovery, protein design, cell systems, disease mechanisms—these all require AI to simultaneously possess the abilities of modeling, reasoning, experimental design, feedback absorption, and long-term strategy updating. If a system can truly work in the life sciences, its capabilities are not just biology-specific; they can be transferred to a larger battlefield: materials, energy, robotics, industrial systems, and even broader scientific discovery.
I also hope that AGI can happen more through Chinese innovators. This is not a narrow expression of identity, but rather my belief that in the next era of intelligence, we should not just use systems defined by others or follow paradigms defined by others. Chinese scientists, engineers, and entrepreneurs should have the opportunity to participate in defining the path to AGI, especially in directions like AI for Science that truly connect intelligence with real-world creativity.
Aureka is currently doing AI for Biology, which looks like it's in the life sciences. But at a more fundamental level, we're trying to build an intelligent system that can understand complex systems, intervene in complex systems, and continuously self-evolve. Today it starts from biology; in the future, it should be transferable to the larger world of science and engineering.
So if I were to sum it up in one sentence: AGI should not just happen in a chat window, nor should it only happen in a few closed laboratories. AGI should happen in science, happen in the real world, and happen in the hands of our generation of Chinese entrepreneurs and scientists. AI for Biology is our starting point, but not our endpoint.
Please note that this interview has been carefully edited and approved by Hua Chenqing, representing only the interviewee's views. We also welcome readers to share their thoughts on this interview by leaving comments.
Z Potentials will continue to provide more interviews with research explorers in fields such as artificial intelligence, global markets, and robotics. We invite you, who are full of vision for the future, to join our community to share, learn, and grow together.