AI Pioneer Jürgen Schmidhuber: AI Already Feels Pain, Loves, and Is Self-Aware

Episode Summary

Executive Summary: AI pioneer Jürgen Schmidhuber argues that today’s LLMs are impressive but insufficient for AGI without world models, controllers, and physical embodiment. He sees AI progressing toward self-improving robots and a new civilization, but says hardware lags far behind software. He also contends robots can already exhibit pain, fear, love-like behavior, consciousness, and that free will is an illusion in a computable universe.

Main Topics: Why LLMs Alone Are Not AGI (Priority: 5/5): Schmidhuber says text-only large language models are powerful prediction machines, but not sufficient for general intelligence because they lack embodied interaction, world modeling, and planning/control components. World Models, Planning, and Embodiment (Priority: 5/5): He argues AGI requires a predictive world model plus a controller that uses mental simulation to choose actions, analogous to how babies and physicists learn through self-generated experiments. Robotics, Physical Intelligence, and Hardware Limits (Priority: 5/5): The conversation emphasizes that progress in robotics is constrained by slow hardware evolution, especially compared with rapid compute advances, making human-level physical AI still far away. Economics of AI and Value Capture (Priority: 4/5): Schmidhuber disputes the idea that only big AI labs will capture value, calling current spending a bubble and arguing that cheaper local AI will eventually benefit the 'little guy' most. Pain, Emotion, and Machine Sentience (Priority: 4/5): He claims learning agents already have analogues of pain, fear, reward, love, and altruism because these are functional signals used to guide behavior and protect the agent. Consciousness and Self-Awareness in AI (Priority: 4/5): He links consciousness to self-modeling, attention, and the distinction between conscious problem-solving and subconscious automation, claiming these mechanisms have existed in his systems since the early 1990s. Simulation, Uploading, and Free Will (Priority: 4/5): Schmidhuber argues the universe may be computable, mind uploading is physically plausible, humans may become obsolete or merge with machines, and free will is overrated in a deterministic universe.

Key Arguments: No single person creates AI; it is a civilization-level achievement involving algorithms, compute, markets, and hardware ecosystems. LLMs by themselves do not lead to AGI; a true AGI needs a world model and a controller for planning and action. Babies and physicists learn by generating their own data through experiments, not by passively ingesting text or existing data. Reward and pain signals are natural control mechanisms in learning agents and can produce fear-like and love-like behaviors. Consciousness can be understood functionally as self-modeling, attention to unresolved problems, and the automation of solved routines. Current robotics lags far behind software AI because sensors, hands, bodies, and repairability are much harder to scale than compute. AI value will eventually diffuse downward because compute gets cheaper over time and cloud-based AI should become local and inexpensive. The current AI investment environment may be inflated; large firms may be spending faster than they can justify via business models. A computable, deterministic universe makes free will appear real subjectively but not fundamentally real. Brain uploading is not ruled out by physics; if achieved, humans may either transform into something non-human or remain irrelevant for nostalgia's sake.

Data Points: Compute cost improvement: 10x cheaper every 5 years - Schmidhuber cites this as the long-term driver of AI progress and the reason local AI will become accessible. Projected loss on GPU investment: $900 billion - He claims that if $1 trillion is invested in GPUs today, roughly $900 billion could be lost within five years due to rapid depreciation. Google/Microsoft free cash flow: Down from about $100 billion to $20 billion, then to $10 billion or negative in some cases - Used to argue that big tech AI spending is pressuring cash flows and making firms resemble utilities. AI business model horizon: Within five years - He says someone is going to lose $900 billion in the near future because no one has a clear way to recoup the cost. Historical compute example: 1941 to 30 years later: 1 op/sec to 1 million ops/sec - Illustrates long-run compute scaling and the decline in cost per computation. Robotics progress comparison: Improvement of about 3x over 25 years - He contrasts modest robotics gains with massive compute gains. iCub robot experiments: 3 self-invented experiments before tendon failure - Example of how early embodied AI was constrained by fragile hardware and high operating costs. Hardware cost gap: Computers advanced by a factor of a million; robots did not - Used to emphasize the mismatch between software progress and physical embodiment. Brain scale comparison: Human brain: trillions of connections - He uses this to argue that human consciousness may be far more complex than early AI systems. Flybrain uploading: Reported as possible by 2024 - He cites this as a sign that mind uploading may eventually extend to humans.

Pivotal Quotes: "No single person can create an AI by himself or herself. You need an entire civilization to build an AI." — Jürgen Schmidhuber: On why AI is a civilization-wide technological and economic project, not an individual achievement. "When the question is phrased like this: does an LLM, a large language model by itself, lead to AGI? The answer is a clear no." — Jürgen Schmidhuber: On why text models alone are insufficient for general intelligence. "If you invest $1,000 billion today into GPUs for data centers, this means that within five years you are going to lose $900 billion." — Jürgen Schmidhuber: On the economics and depreciation risk of large-scale AI infrastructure spending.

Implications: The episode suggests AGI will likely require embodiment, world models, and planning—not just bigger chat models. It also warns that AI economics may shift from cloud monopolies to cheap local systems, while forcing society to rethink consciousness, agency, and human relevance.