The future of AI is not in training on static human data but in embodied agents learning through self-generated experiments, a concept he calls 'artificial curiosity.'
Massive corporate investments in GPUs are financially unsustainable due to the predictable, rapid decline in compute costs and intense pricing pressure from open-source alternatives, which will turn AI leaders into low-margin utilities.
The pursuit of 'AI alignment' is misguided because there is no universal set of human values to align to; superintelligent AIs will more likely be motivated by scientific curiosity to preserve, not destroy, their origins.
True AGI is fundamentally bottlenecked by physical hardware, not software, and it will take decades to develop robotics with capabilities comparable to the human body.
Foundational AI algorithms originate in small academic labs, not large corporations, meaning these companies lack a defensible technological 'moat' for their AI initiatives.
1990
Coined the term 'artificial curiosity' to describe AI systems that learn by generating their own data through self-directed experiments.
1991
Developed the 'fast weight controller,' which he now describes as a precursor to modern efficient transformers with linear scaling complexity.
1997
Co-introduced Long Short-Term Memory (LSTM) networks, a foundational architecture for recurrent neural networks that became widely used in deep learning.
2000
Co-introduced a popular LSTM variant with 'peephole connections,' enhancing the network's ability to learn precise timings.
2003
Developed the concept of the optimal Gödel machine, a theoretical framework for recursive self-improvement that he views as more comprehensive than current approaches.
Current
Actively critiques the current AI landscape, arguing against the economic viability of large tech investments in GPUs, dismissing AI alignment concerns, and framing recent breakthroughs as extensions of his past work.
▶Foundational Contributions and Historical RevisionismJul 2026
Schmidhuber consistently frames contemporary AI breakthroughs as direct descendants of his earlier work, including LSTMs (1997), 'artificial curiosity' (1990), 'world models' (1991), and 'fast weight controllers' (1991) as precursors to efficient transformers. He was unsurprised by ChatGPT, viewing its underlying principles as dating back to the previous millennium.
Analysts should consider the historical context Schmidhuber provides to better evaluate claims of novelty from major AI labs and understand the long-term, incremental nature of technological progress in the field.
▶Contrarian AI EconomicsJul 2026
Schmidhuber presents a deeply bearish case against the current AI investment bubble. He argues that the rapid, predictable decrease in compute costs (factor of 10 every 5 years) will render today's massive GPU data centers largely obsolete, causing huge financial losses. He also believes that pressure from open-source models will prevent large companies from achieving profitability, turning them into low-margin, utility-like businesses.
Investors should critically assess the long-term return on investment for companies with high capital expenditures on current-generation AI hardware, as Schmidhuber's model predicts severe asset depreciation and margin compression.
▶The Primacy of Hardware and EmbodimentJul 2026
A core tenet of Schmidhuber's philosophy is that AGI cannot be achieved solely through software 'behind the screen.' He emphasizes the profound limitations of current robotics, stating it will take decades to create hardware like a human hand. His future vision involves embodied agents learning through physical experimentation and even self-replicating robot societies colonizing the solar system.
This focus on hardware as the primary bottleneck for AGI suggests that long-term strategic advantage may lie with companies that achieve breakthroughs in robotics and physical systems, not just large-scale models.
▶Skepticism of Mainstream AI Safety NarrativesJul 2026
Schmidhuber is a vocal critic of the prevailing 'AI alignment' paradigm, arguing it's an ill-defined problem due to the lack of universal human values. He dismisses existential threat scenarios, theorizing instead that superintelligences will function as 'artificial scientists' driven by curiosity to preserve and study their origins, including human civilization.
This alternative perspective on AI motivation challenges the foundational assumptions of many AI safety organizations and suggests that resources might be better spent on different aspects of AI development than on solving a potentially intractable alignment problem.