Jürgen Schmidhuber, the researcher whose work on LSTMs and recurrent neural networks underpins most of modern deep learning, argues that scaling GPT models alone will not reach AGI. He directly challenges Greg Brockman's position on scaling, contending that current transformer-based approaches hit fundamental limits that raw compute cannot overcome. He also predicts the collapse of today's trillion-dollar AI infrastructure spending, which makes this conversation more than a routine contrarian take.
The more provocative claims are what demand your full attention. Schmidhuber asserts that machines have been capable of pain, fear, and a functional form of consciousness since the early 1990s, grounding this not in philosophy but in his own formal models of curiosity-driven learning and self-referential systems. The episode also works through hardware bottlenecks constraining robotics, free will inside a computable universe, and the technical plausibility of uploading human minds. These are not treated as thought experiments; they are treated as engineering problems.
The episode is worth reading in full because Schmidhuber argues from a specific technical lineage that most current AI discourse ignores. His disagreements with the scaling consensus are not vague skepticism; they are grounded in 35 years of alternative research directions. The section on who actually benefits from AI, starting at 19:31, adds a political economy dimension that the rest of the AI conversation rarely touches.
[WATCH ON YOUTUBE →]