Artificial intelligence is evolving faster than we could have imagined, and few people are better equipped to offer insights into this rapid transformation than Percy Liang, a Stanford AI researcher and co-founder of Together AI. In a recent interview, Percy shared his thoughts on the release of OpenAI’s 01 model, the challenges of interpretability, and how AI will play a significant role in complex simulations in the future.
The 01 Model: A First Step Toward AI Longevity
Let’s start with the star of the AI universe at the moment—the 01 model. While many see it as a breakthrough, Percy views it through a more nuanced lens. From a product perspective, he found it "kind of slow" and difficult to use for certain applications. However, from a research perspective, he believes it signals a shift in the way we understand the capabilities of AI models, especially in their ability to solve long-term, complex tasks.
Traditionally, language models have been measured by how quickly they can generate responses to prompts. Tokens per second and fast completion times were the benchmarks of success. However, 01 is pushing us to rethink those expectations. Percy mentions that future AI models may tackle tasks that require days, weeks, or even months—ambitious projects akin to the major undertakings we pursue as humans.
This is the future Liang is alluding to: AI models that can continuously learn and adapt as they gather more experience. Picture AI-driven research agents that can invent new technologies, discover drugs, or solve global challenges by analyzing complex systems over long periods. While 01 is a small step, it's a step in the direction of truly intelligent, adaptable systems.
Moving Beyond Tokens: AI Agents in Action
The conversation with Percy quickly shifts to a fascinating trend in AI: the resurgence of agents. Remember AlphaGo? Once the darling of reinforcement learning, it has been overshadowed by the rise of large language models (LLMs) like GPT-4. But now, as Percy points out, the field is coming full circle. We’re seeing a return to agents that don’t just predict the next word in a sentence but take actions in dynamic environments.
Percy’s work on generative agents exemplifies this trend. He developed a virtual world, similar to The Sims, where AI agents interact with each other, allowing researchers to study complex social dynamics. It’s a pioneering effort that highlights the potential of these agents to go beyond text generation and tackle real-world scenarios. This "agentic" capability could enable AI to reason, plan, and take multi-step actions over extended periods.
Imagine agents that can autonomously research, troubleshoot, and refine their approach based on real-time feedback. Percy envisions a future where these agents are deployed in various fields—from cybersecurity to research—learning and improving with every experience.
The Future of Simulation: From Games to Digital Twins
One of the most exciting areas of Liang’s research is in AI-driven simulations. His work with generative agents isn’t just about creating believable virtual worlds. The endgame is much more profound—creating valid simulations that accurately reflect real-world phenomena.
In his view, we’re not far from the day when we could use AI to run simulations of societal systems, policies, or even personal decisions. These "digital twins" of society could serve as a testing ground for everything from new laws to public health interventions. Want to know what would happen if you implemented a specific public policy? Run it through a societal simulation powered by AI agents, and you might get an answer.
Of course, this is still in the future. But with advancements in AI models and increased accuracy in these simulations, the potential for running experiments before making major decisions is enormous.
Challenges Ahead: Interpretability and Compatibility
While AI holds tremendous promise, Percy warns of some challenges ahead—chief among them, interpretability. As models grow in complexity, understanding how they arrive at certain conclusions becomes more difficult. This has implications for safety, especially in critical areas like healthcare or finance.
Percy highlighted how this issue became evident when testing the 01 model. When dropped into a pre-existing framework, it ignored custom logic and templates, leading to underwhelming results. This points to another challenge—compatibility. It’s not just about making models smarter; it’s about ensuring they integrate seamlessly with the systems we already have in place.
Another aspect to consider is how these advancements impact research in academia. As commercial giants like OpenAI lead the charge with vast resources, academic researchers need to find new ways to contribute. Liang argues that researchers should focus on areas where they can be orthogonal to commercial development—picking projects that are enhanced by, but not dependent on, the latest models. This way, academia can still play a vital role in pushing the boundaries of what AI can do.
The Next Frontier: AI as a Force Multiplier
So, where does Percy see the next big leap? He believes that AI’s role will increasingly be as a force multiplier in research and problem-solving. We’re moving past the era where AI models merely mimic human capabilities. Soon, AI may begin to extend human knowledge, solving problems that have eluded us for decades.
Percy mentioned benchmarks like solving advanced mathematical problems or finding zero-day cybersecurity vulnerabilities as future milestones. Once AI starts generating novel solutions—ones that go beyond what humans can conceive—we will have entered a new phase of AI evolution.
Percy Liang’s insights remind us that AI is far more than a tool for automating tasks or generating text. It’s a system that could one day be fully integrated into every facet of society, driving innovation, solving global challenges, and even shaping the future of our world through advanced simulations.
However, he is careful to highlight the importance of a holistic view. AI’s safety, usefulness, and ultimate success will depend not just on making models more powerful but on understanding their place within broader ecosystems—both technical and social.
As we look ahead, the future of AI is filled with promise, but it’s also one that requires careful thought, ethical considerations, and, as Percy advocates, a focus on transparency and understanding.
In this thought-provoking discussion, Percy Liang brings to light the next big trends in AI—longer-term reasoning, dynamic agent behavior, and advanced simulations. Whether you’re a researcher, developer, or simply an AI enthusiast, his insights offer a roadmap for where the field is headed and the challenges we need to tackle along the way.
תגובות