文本
语音
视频
AI原生应用
即刻接入AI能力
与所有人共创智能MiniMax CEO Keynote at WAIC 2025: Everyone's AI
The 2025 World Artificial Intelligence Conference (WAIC) kicked off this week in Shanghai, bringing together global leaders, innovators, and trailblazers under the theme “Intelligent Era, Shared Future.” A standout moment from the opening day was a bold and forward-looking keynote by MiniMax Founder & CEO Junjie Yan, titled “Everyone’s AI.”
And it wasn’t just a talk—it was a call to action.
Full Keynote Transcript:
From Deep Learning Pioneer to AI Visionary
“Everyone’s AI” is more than a slogan—it reflects the core belief that has guided my journey over the past 15 years. From being among the first PhD students in China focusing on deep learning when Dr. Hinton began designing AlexNet, to witnessing the turning point of AlphaGo that brought AI into the public spotlight, to founding MiniMax—one of the first large language model companies in China—just before the rise of ChatGPT, I’ve consistently explored a single, evolving question:
What exactly is AI? And more importantly, what role should AI play in our lives—and how can it serve everyone?
Unlocking Creativity at Scale
As our models continue to improve, we're seeing AI gradually become a true driver of productivity. Tasks like writing data analysis tools, which once required manual coding, are now efficiently handled by AI-generated software.
At one point, I considered building an app to track developments across the AI landscape. But it soon became clear that an AI agent could perform this task more effectively and with far greater efficiency.
AI is not only a more powerful form of productivity, but also becomes a stronger source of creativity. Take the example of “Haibao,” the beloved mascot from the 2010 Shanghai World Expo. Now, if we want to create new iterations of Haibao that reflect local culture and current trends, AI can do it even better. From Xuhui Academy to Wukang Mansion, AI can instantly reimagine Haibao in real time, embedding it into the city's evolving identity and generating a wide range of creative designs with just one click.
Take the recent viral example of Labubu. In the past, creating a high-quality creative video featuring Labubu could take up to two months and cost tens of thousands, or even millions. With Hailuo, our advanced AI video generation model, high-quality animations that once took weeks and significant production budgets can now be created within a day at a fraction of the cost. In just the past six months, users have generated over 300 million videos using Hailuo, demonstrating its growing global impact and making creativity more accessible than ever.
Beyond Expectations: AI as a Capability Multiplier
AI is also being applied in ways we never anticipated: decoding ancient scripts, simulating a flight, or designing a space telescope. As models become more powerful, these once-unimaginable ideas are becoming increasingly feasible. With just a bit of collaboration, anyone’s ideas can turn into reality.
At MiniMax, we see AI not just as a tool, but as a foundational capability—a multiplier that enhances human potential across domains.
We anchor our work in two core beliefs:
- AI is a capability—a way to amplify human skills.
- AI is sustainable—its power continues to evolve and improve.
AI has the unique ability to continuously evolve. As humans, we face natural limits in how fast we can learn or work. But AI does not. As our models improve, they help us build even better models in return.
At MiniMax, over 70% of our code and 90% of our data analysis is now supported or generated by AI. The gains in efficiency are significant, but more importantly, they signal a shift in how we work and create.
From Labeling to Reasoning: Teaching AI to Think
A year ago, data labeling was essential to model training. Today, with smarter models, we can automate basic annotation tasks and empower human labelers to focus on high-level supervision and reasoning. The task has shifted from providing answers to teaching models how to think.
Another key direction is teaching AI through dynamic environments. Whether it's integrated development environments (IDEs), agent systems, or sandbox games, we're seeing strong results from reinforcement learning. With clear goals and rich interaction, AI can achieve sustainable, scalable progress.
All signs point to one conclusion: AI is improving rapidly, and the ceiling is nowhere in sight.
A Pluralistic Future, Not a Monopolized One
With such rapid development, it’s natural to ask: will AI be monopolized?
We believe the answer is no. The future of AI will be pluralistic, driven by three key forces:
- Diverse model alignment objectives.
Models are being trained for different goals—reliability, emotional intelligence, imagination, and creativity. These differences reflect the distinct values of the organizations behind them, leading to diverse model behaviors and long-term coexistence. - Rise of multi-agent systems.
Advanced AI applications are increasingly powered by networks of specialized models rather than a single foundation model, making it harder for any individual model to hold a monopoly advantage. - Acceleration of open-source innovation.
Many of the most creative AI systems now emerge from the open-source community. While top-tier performance may still come from closed models, open ecosystems are catching up rapidly.
Together, these trends point to a future where AI is not concentrated in the hands of one player, but shared and advanced by many.
For AI to truly be "Everyone's AI", it must be cost-effective and available to all.
At the same time, we believe that AI will become increasingly accessible, and the cost of using it will become more manageable.
Over the past year and a half, the size of AI models hasn’t grown significantly, even though the computing power available to us has increased. Why is that? For any practical model, inference speed is a critical factor. If a model runs too slowly, it discourages users from adopting it. That’s why everyone is focused on finding the right balance between the number of model parameters and the level of intelligence.
In the past, model size generally increased in proportion to advances in chip technology. We know that chip performance has historically doubled about every 18 months, and model growth followed a similar pace. But now, despite the increase in available compute, model sizes haven’t grown much. So where is all that extra computing power going?
First, let’s talk about training. Over the past six months, the growth in scale has started to slow down, but the cost of training a single model hasn’t significantly increased. Much of the compute is now being used for research and exploration. And as we know, successful research depends not only on compute, but also on effective experiment design, strong R&D teams, and brilliant ideas. As a result, the gap between companies with massive compute and those with less is narrowing. Even companies with relatively limited compute can still drive efficient innovation by continuously improving their experiment design, problem-solving strategies, and organizational structures.
Now, let’s look at inference. In the past year alone, the inference cost for leading models has dropped by an order of magnitude. Thanks to major advances in compute networks and optimization algorithms, we believe it could drop by another order of magnitude in the next year or two. In short, we don’t expect the cost of training individual models to rise significantly.
We believe that with continued innovation, AI development won’t necessarily remain a high-cost industry, though compute usage will keep growing. Even though the cost per token will drop, the number of tokens used is increasing dramatically. A year ago, a ChatBot conversation might have used just a few thousand tokens; today, a single Agent interaction can consume millions. As AI is applied to more complex and practical problems, more people will use it — and use it more often.
A Shared Vision for AGI
Ensuring that AI is affordable and accessible to all has always guided our view of its development. Intelligence with Everyone isn’t just our slogan — it’s the founding vision that drives everything we do. We believe AGI will become a reality—and when it does, it must serve the many, not the few.
If AGI is achieved, it won’t come from one company alone—it will be built together with users, and it should belong to all of us.
We are proud to be part of this global effort, and we invite everyone to help shape the future of “Everyone’s AI.”
