Скачать с ютуб Google's NEW Dual-Agent AI в хорошем качестве

Google's NEW Dual-Agent AI

The Dual Brain of AI: Quick Responses, Deep Thinking. System 1 and System 2 thinking implemented in AI. The Talker-Reasoner architecture introduced here employs a dual-system approach inspired by Kahneman's "Thinking, Fast and Slow" theory. It consists of two distinct components: the Talker, responsible for fast, intuitive, real-time conversational interactions (System 1), and the Reasoner, which handles slow, deliberate, multi-step reasoning and planning (System 2). The Talker interacts with users, utilizing an in-context learned language model that accesses belief states stored in memory, ensuring fast responses without requiring constant updates from the Reasoner. The Reasoner, on the other hand, performs complex tasks such as hierarchical reasoning, tool calling, and belief updates based on Chain-of-Thought (CoT) prompting. The belief state is represented as structured objects (e.g., JSON/XML), encoding user goals, preferences, and environment states, and is updated through Bayesian inference as new observations are received. This asynchronous, memory-driven interaction between the Talker and Reasoner allows for efficiency and reduced latency in decision-making. The architecture leverages a Partially Observable Markov Decision Process (POMDP) framework to formalize the agent's decision-making in an environment with incomplete information. The belief state acts as a sufficient statistic, capturing all relevant past information and evolving through a Markov process, where future states depend only on the current belief state and the latest observation. The augmented action space includes external actions (e.g., API calls), internal reasoning steps (thoughts), and belief updates, allowing the Reasoner to handle multi-step problem-solving while the Talker maintains ongoing interactions. The Talker can operate with outdated belief states for efficiency but waits for the Reasoner when complex reasoning is required. This modular separation between fast conversational tasks and slow deliberative planning enables the system to balance responsiveness with reasoning depth, optimizing performance for both simple and complex tasks in real-world AI applications such as sleep coaching. All rights w/ authors: Agents Thinking Fast and Slow: A Talker-Reasoner Architecture https://arxiv.org/pdf/2410.08328 #chatgpt #reasoning #google