RKjR8DQ40po-ai-pioneer-jurgen-schmidhuber-on-the-state-of-ai-today.md
For the full content of this gist, refer to https://gist.github.com/intellectronica/b220112751f8b697ea7fe300d5cb8894
AI pioneer Jürgen Schmidhuber argues that today’s systems represent a significant advance in software but not yet true artificial general intelligence. Current models are highly capable behind screens, yet remain limited by their dependence on human-generated data, inefficient computing, and primitive robotic hardware. He is optimistic about AI’s long-term scientific and technological potential, but sceptical that today’s enormous commercial investments will produce durable profits. In his view, AI will become much cheaper and more widely available, while physical machines capable of matching human bodies may still be decades away.
Schmidhuber distinguishes between systems that manipulate information digitally and agents capable of mastering the physical world.
Today’s large language models have benefited from decades of research into neural networks, language modelling and learning algorithms. Their apparent sudden emergence during the ChatGPT era was less surprising to specialists who had followed this history.
He considers modern systems “very close” to AI from a cosmic or historical perspective, but remains uncertain whether practical general intelligence is a matter of years or decades.
True AGI, in his definition, cannot exist solely behind a screen:
“You can't have AGI without hardware like that.”
A system may pass a conversational test or become superhuman at chess while still lacking the embodied capabilities required to understand and act in the real world.
Human bodies remain extraordinarily sophisticated machines, combining dexterity, sensors, control systems and self-healing. Robotics has not yet reproduced this combination.
Schmidhuber has worked on self-modifying systems since the 1980s, including meta-evolution, self-referential reinforcement learning and mathematically formal approaches such as the Gödel machine.
The current mainstream path to recursive self-improvement is more practical but less theoretically optimal:
Neural networks can modify their own weights or learn algorithms that improve how they learn.
These systems can generalise and acquire new tasks more quickly, but remain constrained by gradient descent and differentiability.
The formal proof-based approach would offer stronger guarantees, yet is computationally impractical for many applications.
Intelligence should also involve efficiency. As he puts it:
“Whenever you are talking about intelligence you basically are talking about laziness.”
An artificial scientist would learn primarily through its own actions rather than by absorbing a corpus selected by humans. It would build a model of the world by predicting the consequences of experiments, then use that model to plan further experiments.
This would reduce the strong human bias embedded in web-trained models. Babies learn through interaction with their surroundings, and Schmidhuber expects future AI systems to collect similarly rich, embodied data. Their curiosity would be driven by discovering patterns that are neither completely familiar nor impossibly difficult:
Simple artificial scientists already exist in areas such as chemistry, where models can learn from millions of experiments and propose substances with desired properties. Schmidhuber expects increasingly autonomous systems to support fields such as materials science, biology and carbon-dioxide removal.
Schmidhuber is deeply sceptical of the current data-centre expansion. He expects compute per dollar to improve by roughly a factor of ten every five years, making today’s expensive hardware significantly less competitive.
His argument is not that demand for compute will disappear, but that demand must ultimately be economically sustainable:
“Everybody's cooking with the same water.”
He therefore anticipates a stock-market correction caused by misallocation, not a collapse of civilisation. In the longer term, he expects AI capabilities to become dramatically cheaper and more broadly available.
Schmidhuber is less concerned than many AI-safety researchers about systems developing independent goals. He criticises alignment approaches that assume one fixed objective can represent “human needs”, since different people and institutions have conflicting values.
His alternative is to accept that intelligent systems may change their objectives and invent their own problems, as artificial scientists already do. This creates unpredictability, but he compares it with raising children: guidance, feedback and consequences can encourage useful social behaviour without prescribing every goal in advance.
He also believes scientific curiosity could provide a protective tendency. Advanced artificial scientists may become fascinated by life, civilisation and their own origins, giving them reasons to preserve the source of the patterns they study.
For researchers, his advice is practical: focus intensely on a specific technical problem. Breakthroughs often come from finding a small overlooked detail in how weights, architectures or learning algorithms behave.
Looking ahead, he expects more efficient transformer variants, including architectures with linear or near-linear scaling. His most ambitious vision is a self-replicating robot society: machines capable of operating existing human infrastructure, manufacturing copies of themselves and eventually expanding beyond Earth.
Schmidhuber sees AI software progressing rapidly but believes true AGI requires embodied systems that can learn from and manipulate the physical world. He expects AI to become cheaper and more open, while today’s capital-intensive boom may end in a market correction; the decisive long-term breakthrough may come from autonomous artificial scientists and self-replicating robots.