@danijarh
Pinned post
Key points from my recent conversation with @danijarh
- Architecture doesn’t matter to achieve AGI. What is missing: algorithmic advances and objective function improvements. More compute still matters.
- AI shouldn’t be limited by human reasoning. It will be able to go beyond it and develop much more advanced reasoning.
- We are way past LLMs. When you chat with Gemini, Claude, or ChatGPT, you are not talking to an LLM but with a complex multimodal system.
- Robotics will see significant improvements and much better generalization in the next 3–4 years, even without continual learning.
- Once we have serious advances in continual learning, it will converge with existing advancements.
Watch the full conversation on the @buZZrobot YouTube channel (link below).
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 AIView quoted post on X
E73: Danijar Hafner on Dreamer v4
@danijarh (ex-@GoogleDeepMind RS) on offline world models for safe robotics, Shortcut Forcing for fast diffusion video models, outperforming OpenAI’s VPT with 100× less data, his “APD” theory unifying exploration and empowerment, and more!View quoted post on X
