One of the founders of modern artificial intelligence, Jürgen Schmidhuber, participated in The Information Bottleneck podcast, where he discussed the fundamental concepts laid down by his laboratory decades ago and their role in the development of modern technologies.
What Happened
In an interview with The Information Bottleneck, Jürgen Schmidhuber analyzed the impact of his past developments, such as LSTM, GAN, and the concept of artificial curiosity, on the current state of the industry. Particular attention was paid to the critical distinction between models that predict the next token and models capable of planning actions in the physical world.
Context
Modern successes in the field of AI are largely based on scaling architectures and ideas proposed as far back as 30 years ago. The discussion highlights that the current LLM boom relies on classical methods, which creates a foundation for the transition to more complex systems.
Why It Matters for the Industry
For the industry, a key barrier to creating advanced robotics is the gap between predictive LLMs and models capable of planning in real-world environments (Physical AI). The transition to World Models and the integration of Reinforcement Learning (RL) with generative models define the next vector of development: moving from creating "smart chatbots" to developing autonomous action agents.
Why It Matters for Users
For readers and specialists, this is an opportunity to understand the industry's direction beyond simple generative text. Understanding the connection between classical architectures (like LSTM) and modern hybrid schemes helps better navigate technological trends and potential AI application areas, such as Embodied AI.
Sources
Author
Look at AI, Editorial Staff