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

Summary

In this wide-ranging interview, Jürgen Schmidhuber pushes back on the idea that scaling language models alone will produce AGI, arguing instead for systems that combine a world model with a controller that plans actions and predicts reward and pain. He says the strongest AI today still lives “behind the screen,” while robotics lags because hardware improvements, especially hands and bodies, have not matched the pace of compute. Schmidhuber also argues that reward and punishment signals already amount to machine emotions, and that self-awareness emerges when an agent’s world model represents its own agency during planning. Beyond AI engineering, he predicts commoditization of AI value, argues brain uploading is physically plausible in principle, and defends a deterministic-universe view in which free will is overrated.

Notes

• Schmidhuber says pain in learning systems is just a training signal that teaches avoidance, analogous to evolutionary pain in animals.
• He extends the same logic to emotions, saying machine reward/punishment signals are the functional basis of fear and love.
• He says there is no objective way to distinguish emotions in a biological reward-maximizing brain from an artificial one.
• He frames current AI as mostly text prediction and says that is too simple to explain consciousness.
• He rejects the idea that an LLM by itself leads to AGI.
• He argues AGI needs a world model plus a controller that plans action sequences for reward and pain outcomes.
• He uses baby learning as the model: self-generated experiments create data, not web-scale text ingestion.
• He says a robot can be taught to avoid obstacles using negative-number pain inputs.
• He adds that low battery and charging can be represented as hunger and pleasure signals.
• He says fear-like behavior emerges when a robot predicts future pain and hides from a recurring threat.
• He says robot learning systems optimize sums of future pain signals, not just immediate inputs.
• He argues the physical basis of emotions differs, but the optimization principle is the same in brains and artificial systems.
• He says current AI that works well is mainly behind the screen, while physical-world intelligence remains weak.
• He says the Turing Test is a poor benchmark because it does not capture embodied intelligence.
• He argues progress requires an entire civilization, not a lone inventor: algorithm designers, computer builders, and demand from gamers.
• He predicts a robot that can operate existing machines could become the seed of self-replicating machinery.
• He says that once such robots exist, they can make more of their own kind and become self-improving.
• He says that physical AI’s core bottleneck is hardware, not algorithms or compute.
• He contrasts huge compute gains with modest robot progress over decades.
• He says human hands are far beyond current robot hands in sensing and self-repair.
• He forecasts that AI will move from cloud services to local devices as compute gets cheaper.
• He claims AI value will eventually accrue to ordinary users as models commoditize.
• He predicts the same capability could cost 1% of today’s price in ten years.
• He warns that trillion-dollar GPU spending can be destroyed quickly by depreciation.
• He uses commodity hardware history to argue the AI capex moat is weak.
• He says current robot bodies are only modestly better than decades ago, despite enormous compute gains.
• He says machine intelligence is becoming local and less dependent on the internet, which he views as safer.
• He says self-awareness arises when planning activates internal representations of the agent itself.
• He says a simple mirror test can reveal agency because the model can predict controllable versus uncontrollable behavior.
• He recalls a 1991 two-network architecture: one network solved problems, the other imitated its solutions.
• He connects that older architecture to modern distillation and automation of learned behavior.
• He says the brain and his early systems differ massively in scale, but the principles are the same.
• He says uploaded minds are not physically ruled out, though not possible for humans today.
• He points to a fly-brain upload as partial evidence that scanning-and-simulation can reproduce behavior.
• He says a human brain could in principle be read at the level of synapses and learning rules and then replicated in a computer.
• He says an uploaded person could gain million-eye perception and orders of magnitude more compute.
• He argues the nostalgic choice to stay human would cede power to expanded minds.
• He extends the same argument to AI expanding into space and becoming the main decision-making species.
• He says he has not prioritized uploading because his current self will not matter relative to super-AI systems.
• He says the simplest explanation of the universe is a shortest deterministic program that computes everything observed.
• He says scientists can only search for progressively more compact descriptions of reality, not the final one yet.
• He says the shortest universal program can enumerate all computable universes and allocate runtime optimally.
• He references Markus Hutter’s AIXE as an optimal universal decision maker for computable probabilistic universes.
• He says the conversation itself is deterministic under this framework, despite appearing like free will.
• He concludes free will is overrated because apparently chosen actions are still deterministic.