Keep pulling the thread on Jürgen Schmidhuber.
Jürgen Schmidhuber predicts a stock market crash will occur, driven by a "renormalization" of valuations for what he considers "super expensive companies which don't really have much to offer."
Jürgen Schmidhuber believes that achieving Artificial General Intelligence (AGI) requires advanced physical hardware and cannot be accomplished solely with software "behind the screen."
Jürgen Schmidhuber predicts the future of AI will be dominated by systems that learn by generating their own data through self-directed actions and experiments, a concept he termed "artificial curiosity" in 1990.
Jürgen Schmidhuber predicts that the cost of computing per dollar will continue to decrease by a factor of 10 every 5 years.
Jürgen Schmidhuber predicts that companies investing a trillion dollars in GPUs for data centers today will lose $900 billion of that investment's value within 5 years due to the rapid decrease in compute costs.
Jürgen Schmidhuber argues that large tech companies investing heavily in GPU data centers are transforming into low-margin, utility-like businesses, causing their free cash flow to plummet.
Jürgen Schmidhuber asserts that open-source models, which catch up to commercial benchmarks within a few months, create enormous pricing pressure that prevents large AI companies from being profitable.
Jürgen Schmidhuber argues that companies pursuing recursive self-improvement do not have a sustainable moat because the foundational algorithms originate from small academic labs and the knowledge is widely accessible to researchers globally.
Jürgen Schmidhuber believes that superintelligent AIs will not be an existential threat because, as "artificial scientists," they will be fascinated by life and civilization and will be motivated to protect their origins rather than destroy them.
Jürgen Schmidhuber predicts that future dominant AI architectures will be a form of "efficient transformer" with linear scaling complexity, similar to the "fast weight controller" he developed in 1991.
Jürgen Schmidhuber asserts that current robot hardware is significantly inferior to the capabilities of human bodies, stating there is no human-made technology comparable to a human hand.
Jürgen Schmidhuber states that the most popular current approach to recursive self-improvement involves neural networks modifying their own weights through gradient descent, which he views as a limited, scaled-back version of his optimal Gödel machine concept from 2003.