Google DeepMind: World Models and Scalable Objectives for AI Improvement скачать в хорошем качестве

Google DeepMind: World Models and Scalable Objectives for AI Improvement

Can LLMs lead us to AGI? Discussing the most promising path to AI progress and AGI with a Research Scientist at Google DeepMind. We talk about reasoning, data, compute, architecture, and other factors to better understand the current trends in AI research. This time BuzzRobot spoke with Danijar Hafner, Staff Research Scientist at Google DeepMind, about his research and findings in AI World Models, AI Temporal Abstraction, and AI Scalable Objectives. Predictive models paper: https://danijar.com/project/dreamer4/ Breaking long-term task into subgoals paper: https://danijar.com/project/director/ Designing objectives for AI to self-improve beyond human input paper: https://danijar.com/project/apd/ Find more works by Danijar Hafner here: https://danijar.com/ Join BuzzRobot: Newsletter: https://buzzrobot.substack.com/ X: https://x.com/sopharicks Slack: https://join.slack.com/t/buzzrobot/sh... Support us: ko-fi.com/sophiaaryan Timestamps: 00:00 Intro 01:03 Best AGI strategies 07:37 Continual learning 11:23 Dreamer 13:04 Nested learning 19:02 Objective functions 22:38 Embodiment in training 24:55 Scaling world models 29:27 Transferring knowledge to robotics 31:56 Pre-training vs. reinforcement learning 36:57 Future direction for AI