Jacob Effron and Jürgen Schmidhuber distinguish capable screen-based models from broader intelligence operating in the physical world. Jürgen Schmidhuber argues that current language models inherit a strong human bias from web data, while artificial scientists can generate their own training data by acting, predicting consequences and building world models.
Jacob Effron and Jürgen Schmidhuber trace recursive self-improvement from early metalearning and self-modifying neural networks to systems that improve code or weight updates. Jürgen Schmidhuber says practical approaches remain constrained by gradient descent, while formal systems such as the Gödel machine describe a more general but computationally difficult path.
Jacob Effron and Jürgen Schmidhuber emphasize efficiency as a defining property of intelligence. Self-improving systems should treat computation and energy as costs, and Jürgen Schmidhuber expects future architectures to move toward linear or near-linear scaling rather than the quadratic attention cost associated with standard transformers.
Jacob Effron and Jürgen Schmidhuber question whether frontier-model companies can sustain durable technical moats. Jürgen Schmidhuber expects important ideas to spread through academic and open-source communities, predicts falling AI costs, and argues that current infrastructure spending may be misallocated even if the underlying technology continues to advance.
Jacob Effron and Jürgen Schmidhuber also debate AI safety and goal formation. Jürgen Schmidhuber is less concerned about centralized alignment than many researchers and expects autonomous artificial scientists to develop changing goals, while acknowledging that greater independence makes behavior less predictable and requires social feedback and safeguards.
Watch on YouTube