AIs that set their own goals - learning general purpose world models for efficient planning & acting

AIs that set their own goals - learning general purpose world models for efficient planning & acting

by Jürgen Schmidhuber

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQDArtificial general intelligence (AGI)
🎙 Jürgen Schmidhuber 👥 3K 📅 April 19, 2026 ⏱ 40 min 👁 538 📄 expert opinion 🧭 2026-08-16
Available in: English (current) Français

Keywords

world modelsartificial curiosityreinforcement learninghistory of AIJürgen Schmidhuber

Summary

In this keynote, Jürgen Schmidhuber presents his vision for AI systems that set their own goals, based on learning general-purpose world models for efficient planning and acting. He traces the origins of modern AI back to the 1990s, highlighting his own contributions such as LSTM, artificial curiosity, and fast weight networks. He argues that current large language models are limited because they passively download data, whereas true intelligence requires active exploration and self-invented experiments, as seen in babies. He describes a framework with a controller and a world model, where the controller generates actions and the world model predicts consequences, enabling planning and learning. He emphasizes the importance of intrinsic rewards, such as curiosity, to drive exploration. Schmidhuber also discusses the historical development of neural networks, from the least squares method to backpropagation and deep learning, and notes the exponential increase in computational power. He contrasts the success of AI in virtual environments with the challenges of real-world robotics, where hardware reliability is a major issue. He concludes by outlining his current work on building robot babies that learn through self-invented experiments, aiming to create machines that can operate in the physical world and eventually lead to self-replicating machine civilizations.

200 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides valuable insights into the principles of goal-setting and world models in AI, drawing on decades of research. Schmidhuber’s argumentation is coherent and well-supported by his own work and historical examples. He effectively explains complex concepts like artificial curiosity and fast weight networks in an accessible manner. However, some claims, such as the attribution of backpropagation, are contested, and the talk is more of an expert opinion than a systematic review.

Scientific Rigor, Source Quality, Title Accuracy

The talk demonstrates high scientific rigor in its technical explanations, but the lack of explicit citations in the video is a limitation. The historical claims are generally accurate, though some are debated. The title accurately reflects the content, focusing on goal-setting and world models. The description provides a link to the conference website, but no direct references to papers are given.

149 words

Title / Content Match

The title accurately reflects the content, which focuses on AI systems that set their own goals through world models and intrinsic motivation.

Quality & Reliability

8/10

The talk is given by a leading AI researcher with a strong track record. It presents historical claims and technical concepts that are generally accurate, though some historical attributions are debated. The content is well-structured and grounded in the speaker's own research, but lacks detailed citations in the talk itself.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

Concurring Sources

Dissenting Sources

  • Wikipedia: Backpropagation — The talk attributes backpropagation to Linnainmaa (1970), but some sources credit other researchers like Rumelhart, Hinton, and Williams (1986). This is a debated point in the history of AI.

Contribution & Novelties

The talk provides a comprehensive overview of Schmidhuber’s long-standing research on goal-setting AI and world models, connecting historical foundations to current challenges. It emphasizes the importance of intrinsic motivation and self-invented experiments, contrasting with passive learning in LLMs. The talk also highlights the gap between virtual and physical AI, and the need for robust hardware.

Pour aller plus loin :

  • Artificial Curiosity — Wikipedia article on the concept, directly relevant to the talk’s core idea.
  • World Model — Wikipedia article on world models, central to the talk’s framework.
  • Jürgen Schmidhuber’s publications — List of his papers, including those on LSTM and artificial curiosity.

103 words

Radar Profile

The radar profile shows high scores in quantity of information and technical level, reflecting the depth and breadth of the talk. The quality of information is also high, but the reliability is slightly lower due to the lack of explicit citations and some contested historical claims.

Reliability 8/10

💬 Sur les 0 commentaires analysés, aucune tendance n'est disponible.