Redpoint reposted this
In our latest Unsupervised Learning, Patrick Chase and I talked with Percy Liang, Stanford professor and co-founder of Together AI. Percy is driving some of the most critical advances in AI research. We explored o1 and the evolution of AI to autonomous agents capable of long-term problem solving, why interpretability is hard, and the future of AI safety in complex, real-world systems. Here are some of my favorite parts of our chat: 🎮 Building "The Sims" of AI Percy discussed creating a digital twin of our world (with two of his students) where AI agents interact with each other autonomously, mimicking social dynamics. “It’s like building a digital society—agents can announce they’re running for mayor or spread information through conversations, much like real-world social interactions.” This allows for fascinating possibilities, studying everything from social behavior to policy experiments. For example, researchers could implement a Covid-style masking policy in order to learn how the virtual society and its “residents” would adapt. ⚖️ Regulate AI -- Downstream Percy says we need a balanced, proactive approach to AI regulation, emphasizing transparency and disclosure over heavy-handed restrictions. “We need regulation that helps us understand the risks and benefits, but focusing too much on upstream control could be ineffective or overly blunt. Instead, we should empower downstream decision-makers with clearer insights, like providing 'nutrition labels' for AI models.” This kind of regulation would allow for more informed and adaptable oversight. 📊 How to Evolve Evals Percy thinks traditional AI benchmarks are becoming insufficient. “As models claim they can do anything, we need evaluations that reflect their real-world capabilities, including how they handle diverse, open-ended tasks.” He pointed to innovative methods, such as using language models themselves to generate new test scenarios, as key to tracking progress and ensuring AI models are both accurate and reliable in increasingly complex environments. 🔍 Interperability Needs a Solve As AI models grow more complex, interperability becomes increasingly difficult yet critical, especially in regulated industries. “We’re moving further away from understanding why a model makes certain decisions,” Percy says, “especially without access to the model’s weights or training data.” It was a ton of fun hearing from one of AI’s leading thinkers and I hope you'll enjoy listening to our conversation: YouTube: https://lnkd.in/gANgXC8K Spotify: https://spoti.fi/3zyvSa1 Apple: https://apple.co/3zMboKS