Peter Norvig

Nobel prizes are returning to the Central Coast, thanks to breakthroughs in quantum mechanics that paved the way for today's breakthrough computing labs. Michel H. Devoret of Yale and UC Santa Barbara and John M. Martinis of UCSB joined John Clarke of UC Berkeley as the 2025 laureates in physics. The award was given for discoveries on "quantum mechanical tunneling and energy quantisation" during experiments in the 1980s. It's nice to see the Nobel prizes awarded to UCSB after a gap of over a decade. And to see UCSB's excellence in quantum mechanics and quantum computing get overdue recognition.

26

3 Comments

The Emerging World of Multi-Turn Interactions in Large Language Models, NeurIPS 2025: https://lnkd.in/gnQig_rJ Artificial intelligence is entering a new stage—one where conversations with machines unfold not in single questions and answers, but in long chains of exchanges. These extended interactions, known as multi-turn conversations, are becoming central to how we use Large Language Models (LLMs) today. What Are Multi-Turn Interactions? A “multi-turn” interaction simply means a conversation or task that happens over several steps, each one building on the previous. Instead of a one-off query, the user and AI move through a sequence—much like solving an equation piece by piece, where each line depends on the one before it. This matters because real-world tasks are rarely completed in a single command. Planning a research project, guiding a piece of writing, troubleshooting software, or coordinating actions across tools all require continuity and memory. Why They Matter Now Today’s LLMs are far more capable of holding these extended exchanges than earlier systems. They can: keep track of previous steps, adapt to evolving goals, make decisions in complex environments, and manage long, dynamic conversations. As AI takes on roles that resemble digital assistants, collaborators, or even agents acting inside software interfaces, these multi-step capabilities become essential. Key Challenges Ahead Researchers are now focusing on several big questions: 1. Teaching AI to act over many steps How can we train models to operate in multi-step environments—like navigating a computer interface or using external tools—when feedback is only given occasionally? 2. Staying aligned with human goals During long conversations, an AI may drift or misinterpret intentions. Researchers want models that stay connected to what users truly need across many turns. 3. Building healthier human–AI relationships As interactions stretch over hours or days, how can AI adapt without becoming unsafe, biased, or overly confident? 4. Measuring long-term performance We need better ways to evaluate whether an AI can stay consistent, think ahead, and carry out a sequence of actions reliably. A Field Taking Shape Workshops focused on multi-turn interactions—such as the one linked below—are helping define this research direction. They provide spaces for scientists, engineers, and designers to discuss methods, share experiments, and plan the next steps for building AI systems that can truly collaborate across time.

5

1 Comment