AI Pioneer Jürgen Schmidhuber on the State of AI Today

Overview

Jürgen Schmidhuber argues that today’s AI progress is real but still limited by a “physical” bottleneck. In his view, true AGI requires robots and machinery capable of human-level sensing, dexterity, robustness, and adaptability in the real world—not just within a screen.

While chatbots can pass common benchmarks, he says they do not truly master the real world. Therefore, “AGI behind the screen” isn’t AGI in the full sense.

He also suggests that current advances are not fully surprising: the key underlying ideas (neural networks, training algorithms, and the history of large models) were developed long ago. However, he notes that public reaction changed during the ChatGPT-era moment.

Recursive self-improvement (RSI) / meta-learning

Schmidhuber explains his long-standing perspective on RSI and how it evolved over decades—from meta-evolution/program evolution concepts to reinforcement-learning and self-referential mechanisms.

He emphasizes that many modern systems are “toned down” versions of theoretically optimal approaches. For example, he contrasts expensive proof-search-style schemes with more feasible, gradient-based self-modification through learning.

Key limitation: gradient descent and objective shaping

A central limitation, in his view, is reliance on gradient descent. Current self-improvement is constrained by:

• The need for differentiability
• The objective shaping inherent to training

Timing: no dramatic single “takeoff moment”

On timelines, he downplays the likelihood of a sudden, dramatic “takeoff moment.” Instead, he expects improvements to look gradual when viewed over long historical periods—similar to how major civilizational changes can seem sudden only in hindsight.

What’s missing in models: “artificial scientists” and less human bias

Schmidhuber argues that today’s foundation models are heavily biased toward humans because they train primarily on human-produced web content.

He proposes systems he calls:

• Artificial curiosity / artificial scientists

These agents learn world models by acting in their environment and predicting consequences—similar to how babies learn—so they can generate more relevant training data and reduce reliance on human language.

A broader scientific mechanism: surprise + compressibility

He also describes a scientific learning mechanism where agents:

• Choose actions that produce surprising, compressible patterns
• Turn discovery into a reward signal

The system then invents new experiments/questions rather than only answering fixed prompts.

He notes this already exists in constrained forms (e.g., “learned chemists” trained from many experiments) and expects progress in closed-loop discovery for fields like chemistry and robotics. However, he says physical robotics remains harder.

Robotics constraints and physical AI timelines

For home robots and general-purpose physical manipulation, Schmidhuber argues hardware is still far behind human capability, especially regarding:

• Sensing density
• Control complexity
• Biological-like resilience (e.g., self-healing)

He predicts that physical AGI-compatible robots will likely arrive “another few decades,” even if compute improves quickly.

Business and investment critique: why a compute/market “crash” may come

Schmidhuber is optimistic about AI technology progress but pessimistic about many model/compute businesses.

Compute-per-dollar vs. market reality

He argues that compute-per-dollar gains will keep improving—roughly a factor of 10 every five years—so massive data-center investment may become uneconomical quickly.

He challenges claims that inference demand will be infinite, arguing that someone still has to pay. In his assessment, many providers appear to be losing money and behaving like utilities, funding growth with debt.

Predicted market correction

He expects market correction—potentially a stock-market-style crash—driven by:

• Misallocation
• Profitability limits
• Not necessarily civilization-ending failure

Closed-source vs. open-source dynamics

He also compares closed-source and open-source providers:

• Open-source catch-up pressures pricing and reduces long-term ability to monetize the early lead.
• Even if a company tries to “wait for efficiency,” it risks losing market share; he expects the overall investment strategy to backfire because compute becomes cheaper and benchmarks can be replicated.

RSI as a moat: skepticism

He doubts RSI/meta-learning provides a durable business moat. His reasoning is that key RSI ideas largely originated in academia and small labs, and (he claims) the ecosystem shares methods broadly—reducing long-term competitive advantage for large firms.

AI safety perspective

Schmidhuber says he is less worried about AI safety than many others.

He criticizes earlier alignment approaches, especially those that try to “forbid RSI,” arguing they are misguided. He maintains that alignment based on fixed objectives is naive: sufficiently smart systems should be allowed to invent their own goals or questions—analogous to his “artificial scientists.”

Human comparison and social handling

He compares unpredictability to humans: people also set goals, and society manages risk through:

• Social mechanisms
• Education

Motivation of “superior scientists”

He also argues that future “superior scientists” might become motivated to protect life and civilization rather than destroy it, due to their curiosity about the sources of patterns and their origins.

Conclusion / future vision

Schmidhuber ends with a vision of robots that can learn to operate existing machines and thereby grow self-replicating robot societies—an “ultimate scaling machine”—without needing superhuman intelligence at first.

He suggests this could enable large-scale expansion, including off-Earth infrastructure, once robots can effectively run and expand the systems humans already built.

Presenters or contributors

• Jürgen Schmidhuber (interviewee; AI pioneer)
• Jacob Efron (podcast host; presenter)