Oasis Capital Dialogue with Professor Fang Fei: AI for Good

While pursuing commercial opportunities in the AI wave, how to leverage AI to address social problems and improve public welfare deserves equal attention and exploration. Today, we are joined by Professor Fei Fang, Assistant Professor at the Software Institute of Carnegie Mellon University's School of Computer Science and Sloan Research Fellow, to delve into this field and identify the challenges. Enjoy.

Oasis Capital: We see that your research focuses on the intersection of game theory and AI. Could you introduce your latest work? Has this wave of AI revolution brought any inspiration or momentum to your field?

Professor Fang: My current research centers on multi-agent interaction, with three main branches:

First, computational game theory. We study strategic interactions among multiple agents, such as computing equilibrium strategies; learning agents' reasoning during gameplay, and inferring unknown elements among agents, such as each agent's utility function.

Second, multi-agent learning. This includes multi-agent reinforcement learning, as well as work on improving the efficiency of multi-agent reinforcement learning and enhancing policy interpretability.

Third, AI for Social Good. This primarily addresses collaborative decision-making among multiple stakeholders. Examples include wildlife conservation, food safety, and transportation-related work.

All three directions fall under AI, including the intersection of game theory and AI. Many people may not consider game theory part of AI, but the entire multi-agent system field is indeed AI — it's just that computer vision and natural language processing are more widely known. Top conferences for computational game theory include AAAI, IJCAI, and AAMAS, with some work also submitted to EC.

The integration of computational game theory with what is commonly perceived as AI — such as CV or NLP — is actually quite common. Text-based games that emerged with LLMs are inherently games of strategy. For example, a paper published two years ago on the game Diplomacy (https://www.science.org/doi/10.1126/science.ade9097) combined reinforcement learning and natural language processing, using linguistic strategy to form alliances between nations. With extensive natural language interaction and strategic gameplay among multiple agents and decision-makers, this represents new work influenced by large language models.

Oasis Capital: The recent Stanford Smallville paper — if we assume 25 individuals represent 25 nations, would their negotiations require game theory to function and evolve?

Professor Fang: Strategic interactions would naturally emerge. But in terms of actual solution methods, we don't necessarily need to directly compute Nash equilibrium. In a small town, you simply simulate each agent. In simple cases, you just make each agent a learning agent — because their individual learning processes and utility functions differ, strategic interactions naturally arise. For instance, if I design a learning agent whose observation space includes other agents' positions, identities, current states, and so on, strategic behavior emerges when conflicts of interest occur, without explicitly modeling the scenario as a game and computing equilibrium.

Algorithm design focuses on how each agent learns and updates its strategy. But when we analyze the system, we ask: if everyone learns this way, will it converge to equilibrium? That's the kind of work we commonly do in multi-agent learning research.

Oasis Capital: Regarding multi-agent systems, where can reinforcement learning actually be deployed? Gaming?

Professor Fang: Take Dou Dizhu (a Chinese card game), for example — you might use AI autoplay. The issue is that with multi-agent reinforcement learning, the AI becomes extremely strong, but not all games require a dominant autoplay AI or overpowered NPC. Sometimes we aim to make agents more "human-like" rather than unbeatable. We're working on this aspect too, which could be considered one path to deployment.

The most likely direct real-world applications I'm aware of would be mechanism design for online advertising auctions, and some transportation scenarios like ride-sharing platform dispatch — DiDi has work using multi-agent learning for dispatch optimization.

Oasis Capital: So when multi-agent systems need to find equilibrium, that's where your research comes in. Does this conflict with Transformer-based models or Diffusion text-to-image models?

Professor Fang: There are many Transformer-based models in our field applied to game theory, mechanism design, or multi-agent learning. But the usage is quite different. Xiaotie Deng's group has done Transformer-based mechanism design — for instance, when many people bid for an item, who gets it and how much do they pay?

You may have heard of Decision Transformer, which uses Transformer models for offline reinforcement learning.

Also, the Diplomacy paper mentioned earlier incorporates LLMs to embed textual data. The agents negotiate strategies using natural language, so LLMs are needed to model this component.

Oasis Capital: What does "offline" mean here?

Professor Fang: It refers to using only previously collected data from agent-environment interactions, without generating new interaction data.

Oasis Capital: Where do you think the next breakthrough in multi-agent learning research will come from?

Professor Fang: The last major breakthrough I would point to is the Diplomacy strategy paper. The next one is hard to predict — different researchers focus on different directions. For example, we focus on interpretability. We've found that interpretability is essential in practical applications. When collaborating with NGOs on AI for Social Good, our partners lack computer science backgrounds, yet need to implement our solutions. We must thoroughly explain what our work addresses, how it works, and what impact it will have.

We want to use multi-agent reinforcement learning to solve real-world problems. For instance, traffic safety requires interpretability. Our group hasn't used large language models yet, but I know other groups are using them for interpretability-related work.

Oasis Capital: What motivated you to work on social good, and how has implementation gone?

Professor Fang: That's an interesting question. When I started my PhD, I worked on security game theory, using game theory to design patrol routes for ferries. After publishing this work, I attended conferences and attracted interest from other fields. Through discussions, we realized that ferry patrol route design resembles forest patrols — both involve designing patrol routes that require human execution, with the goal of maximizing ferry protection or, in forests, minimizing poaching and illegal logging.

Our initial thought was simple: since the problems are similar, just deploy the old algorithm to the new domain. We discovered that challenges vary dramatically across domains and cannot be directly transferred. We then began collaborating with wildlife conservation organizations, continuously refining the algorithm for deployment. Currently, several protected areas and national parks in multiple countries use our algorithm. That was how I entered the social good space.

Through this process, I found that building connections and trust is remarkably difficult. It took about two to three years before the other party was willing to try our algorithm. Unlike companies, which have strong incentives to push forward if the algorithm generates revenue, conservation organizations have many priorities that rank above news collection and organization, so our collaboration kept getting postponed. Slow progress, reduced communication frequency, plus staff turnover — all of this extends the磨合 period. Spending two to three years from initial NGO contact to deployment is quite normal.

Oasis Capital: Are there applications more relevant to ordinary people?

Professor Fang: We focus on serving underserved communities. Ferry safety actually affects the general public, though people don't perceive it directly. I believe future applications touching more people's daily lives will likely be in transportation and online advertising auctions, both requiring multi-agent learning and game theory work.

We published a paper on bidding strategy for ride-sharing platforms. We modeled what pricing strategy would make all drivers willing to accept platform dispatch — accepting dispatch constitutes Nash equilibrium. We proposed spatio-temporal pricing: pricing based on time and location, not just supply-demand gaps.

Our pricing mechanism is elegant and can achieve equilibrium, but actual deployment differs. From a research perspective, we care about making everyone willing to follow the pricing strategy and accept dispatches. We've been invited to Lyft and Uber to explain the underlying principles. Companies don't first solve for equilibrium before deploying pricing — other considerations exist, and the final outcome is that our research gets integrated into their own methods rather than being used directly.

We also want to develop multi-agent learning applications for transportation that can be directly used, which is quite challenging. Traffic deployment involves high requirements for compliance, privacy, and security. We're working to advance both the methodology and deployment partnerships simultaneously.

Oasis Capital: If Robotaxi becomes widespread in the future, could your transportation research be directly applied to dispatching Robotaxi?

Professor Fang: A key consideration is whether one company controls all Robotaxis or multiple companies operate separately. If it's one company with a control center, this is originally a single-agent problem, but because optimization is difficult, it may need to be decomposed into multi-agent approaches for tractability. In this case, game theory isn't heavily needed — it's more about multi-agent learning.

If multiple companies' vehicles are operating, with competition between companies, then you must consider the game theory among several major players. Should a company deploy more vehicles? Improve algorithms? Enhance vehicle-traffic light coordination to sell more vehicles and provide greater convenience to owners? These are all game-theoretic questions.

Oasis Capital: Your collaboration results with Uber and Lyft also reflect a gap between research and commercial deployment — even perfect research doesn't guarantee perfect commercial solutions, right?

Professor Fang: Yes. In our AI for Social Good deployment work, we've deeply experienced various complexities. Wherever there's deployment, there will inevitably be gaps. For example, AI conferences need to assign paper reviewers, and our algorithm has been used for reviewer assignment. We only considered two factors: overall conference paper quality and randomness. We focused on "random assignment" — adding randomness reduces the possibility of reviewers exploiting system loopholes to get their papers accepted. But in actual deployment, organizers considered far more factors: reviewer geographic diversity, balanced seniority backgrounds for the same paper's reviewers. The final result was that organizers integrated our algorithm with their constraints to produce an overall solution.

Oasis Capital: How do you identify pain points and problems in social good?

Professor Fang: Wildlife conservation emerged from the ferry route planning work. Food rescue came from chance encounters with the CEO of a Pittsburgh platform, plus volunteering a few times myself and discovering room for improvement. For new directions, we find problems by observing social issues — for example, with homelessness, we investigate whether NGOs or even government agencies are working on it, then explore whether AI can integrate with and enhance existing efforts.

Oasis Capital: Is there any connection between AI for Social Good and AI value alignment work?

Professor Fang: There is relevance. For example, food rescue involves platforms and volunteers — when people are involved, coordination is needed. Everyone has different preferences: too many notifications annoys some people, too few makes others feel undervalued. What we can do now is treat everything we can find as features and feed them into machine learning: how many tasks this volunteer has completed, how long they've been on the platform. More personalized design is something we haven't started yet.

Oasis Capital: Could you specifically introduce how your wildlife conservation work operates?

Professor Fang: One aspect is designing patrol routes for rangers. Situations vary by location. For example, in China's northeastern provinces, forestry bureau staff lead teams on reserve patrols. In such cases, WWF collaborates with the forestry bureau, providing AI-generated poaching risk prediction maps and suggested patrol routes for the bureau's reference and implementation. Another aspect is news collection and organization for managers in wildlife and environmental conservation. Why do this? For instance, if news searches reveal that a dam will be constructed somewhere, potentially causing significant ecological impact, WWF staff use our information to assess whether they need to lobby local governments or construction units, persuading them to conduct proper environmental assessments and more comprehensive environmental protection.

Oasis Capital: Besides WWF China, do you currently collaborate with other domestic organizations?

Professor Fang: Not yet. If today's conversation with Oasis Capital can help us find more collaborators, we would be very interested in exploring what work we could initiate.

Oasis Capital: Why does building trust with NGOs and charitable organizations take even two to three years?

Professor Fang: Because both sides have different priorities. Just signing data sharing agreements can take half a year. Unlike companies, which have strong motivation to push progress if my algorithm helps them make money, conservation organizations — do they want news information? Yes. But they have too many priorities that rank above news collection and organization, so our collaboration keeps getting postponed. Slow progress, reduced communication frequency, plus staff turnover — all of this extends the磨合 period. Spending two to three years from initial NGO contact to deployment is quite normal.

Oasis Capital: Without commercial interest driving it and with low priority demand, what sustains your long-term commitment to this path?

Professor Fang: Currently, the food rescue platform in Pittsburgh uses our algorithm in its backend, coordinating between the platform and volunteers, directly affecting thousands or tens of thousands of volunteers, and indirectly affecting food rescue distribution. Our papers contain specific data on how many volunteers and communities the platform serves, and how many tons of food are transported annually — all powered by our algorithm. The ability to deploy this work and make an impact is the source of our sense of achievement and motivation.

Oasis Capital: Indeed, your research is somewhat more specialized compared to NLP or CV.

Professor Fang: Just look at how many fewer papers we publish annually compared to those two fields (laughs). Actually, the multi-agent learning subfield produces quite a lot of papers.

Oasis Capital: Just as Transformer opened people's minds, LLM emergence has enabled certain breakthroughs. So in the short term, paper output has also increased?

Professor Fang: Yes. We originally entered this direction because deep reinforcement learning became popular in 2015. Everyone started thinking about whether deep reinforcement learning could be applied to multi-agent systems. Multi-agent learning papers were driven by single-agent deep reinforcement learning algorithms — for example, PPO (Proximal Policy Optimization) is used everywhere. Yi Wu's group extended PPO to multi-agent MAPPO (Multi-Agent Proximal Policy Optimization), which works very well and solves many games that previously required complex game theory.

Oasis Capital: Compared to NLP, has your research direction been relatively less impacted by LLMs and industry?

Professor Fang: Indeed. Game theory and AI for Social Good haven't been impacted much — we can actually use LLMs to enable more social good applications, so there's more we can do.

Multi-agent learning is somewhat in competition with industry, and industry has more resources, so academia is relatively disadvantaged.

Interpretability research fares better, because what's needed isn't resources but human explanation — how to make things clear and acceptable to people.

Oasis Capital: In China, mention reinforcement learning and embodied intelligence inevitably comes up. Having talked with you this long without mentioning it — is it because multi-agent learning has low relevance to robotics? Or other reasons?

Professor Fang: There is relevance. But our group hasn't worked on complex embodied robot interaction yet, because it's very difficult and there aren't particularly compelling applications. For example, warehouse management involves multi-agent problems, but may only address robot route conflicts — essentially a traffic problem, not embodied intelligence. Single-agent human interaction is already very difficult; adding coordination makes it even harder. And making every agent in a multi-agent system a fully mobile robot — the challenges there are undoubtedly enormous.

Incidentally, solving multi-robot route conflict problems falls under "multi-agent path finding" — to date, the reinforcement learning approaches haven't yet reached the level of traditional non-reinforcement learning methods.

Oasis Capital: So could one say traditional methods are actually superior for industrial deployment? If so, why continue working on reinforcement learning?

Professor Fang: Traditional methods work very well for small-scale scenarios, but may fail in large-scale scenarios, requiring new methods that can break through limitations. Meanwhile, new learning methods may achieve breakthroughs that surpass traditional methods. From a research perspective, we need to keep exploring the boundaries.

Oasis Capital: Academia in China widely struggles with GPU procurement. Does your research require substantial computing power?

Professor Fang: Pure computational game theory requires computation but not that many A100s. Our work focuses more on designing algorithms to improve computational efficiency rather than relying on multiple GPUs. Some of our work studies how to learn each agent's utility function — again emphasizing methodology over computation. Our group has only 5 GPUs, sufficient for computational game theory work, but tight for multi-agent learning. Different research directions require different resources. AI for Social Good applications increasingly involve text data and computation, where LLMs can be used.

Many of our deployed applications involve NGO collaborations — food safety, food rescue platforms — which can't possibly have many GPUs, and we can't demand resources from them. So if there are computational requirements, we handle them ourselves and then deploy lightweight solutions.

Future research will certainly need GPUs. We want to work on combining NLP with multi-agent learning, and with game theory, which obviously requires more resources. We're exploring how to better secure computational resources.

Vitality

What do you think is technological vitality?

In my view, technological vitality takes many forms. A new technology that immediately impacts society is one form. Inspiring and driving the next wave of technological innovation is another. Enabling humanity to achieve deeper understanding of the world and ourselves is yet another form.

——Professor Fei Fang

Carnegie Mellon University

Software Institute, School of Computer Science

Oasis Capital is a new-generation venture capital firm in China, dedicated to discovering the most vital entrepreneurs of the next decade and growing with them to create long-term value. "Vitality" is both the direction of structural transformation in our era and the resilience and evolutionary power of entrepreneurs.

Oasis Capital focuses on early and growth-stage investments, with individual investments ranging from $3 million to $30 million, concentrating on robotics, artificial intelligence, and technology services, empowering China's technology-driven new service upgrade.