Jürgen Schmidhuber, a foundational figure in modern AI often called the “father of AI,” discusses the state of the field, arguing that while superhuman AI is cosmically imminent, physical AGI remains decades away, the current data-center capex boom is a misallocation that will crash, recursive self-improvement will not create durable moats for model companies, and alignment-focused safety frameworks are naive because truly intelligent systems must invent their own goals.
How Close Is Superhuman AI?
From a cosmic perspective, we are as close to superhuman AI as we were in the 1970s — a flash in 13.8 billion years — but in human terms it could be years or decades.
True AGI requires mastering the physical world, not just passing the Turing test behind a screen; current robot hardware is vastly inferior to human bodies (millions of sensors, self-healing, delicate and strong manipulation), and no human-made technology compares.
Without hardware that matches the human hand and body, systems remain “fancy text editors” rather than general intelligences that can operate in reality.
Predicting compute cost curves is easy (roughly 10× per dollar every 5 years), but predicting when robotics will close the hardware gap is much harder — likely decades, not years.
Why ChatGPT Didn’t Surprise Him
The ChatGPT moment surprised only those unfamiliar with the decades-long history of large language models and the foundational algorithms (backpropagation, LSTM, attention precursors) developed in the previous millennium.
For researchers in the center of that lineage, the trajectory was predictable; the “sudden” emergence was a public perception artifact, not a scientific discontinuity.
The Path to Recursive Self-Improvement
Schmidhuber’s work on recursive self-improvement (RSI) spans decades: meta-evolution (1987), self-referential reinforcement learning machines (1994), and the mathematically optimal “Gödel machine” (2003) that proves self-modifications are beneficial before executing them.
Current popular RSI approaches (neural networks modifying their own weights via gradient descent, metalearning) are scaled-back, practical versions of the Gödel machine; they work well but inherit gradient descent’s limitations and are not globally optimal.
Metalearning — networks learning to generate better weight updates than gradient descent — is the dominant practical paradigm today, enabled by cheap compute that makes large-scale experiments feasible.
Will AI Takeoff Feel Sudden?
In hindsight, the transition will look like a vertical stick: 13,000 years of civilization automating labor and calculation, then suddenly AI appears — a flash in cosmic time.
For those living through it, the process feels gradual and messy; the “suddenness” is a retrospective compression of a long automation arc.
Intelligence Means Efficiency
Intelligence is fundamentally laziness: an intelligent system achieves goals with minimal effort and energy.
Schmidhuber’s self-improving systems bake computational and energy costs into the objective function; as they improve, they naturally do more with less.
This efficiency pressure is a natural consequence of optimization, not a separate design goal.
Advice for Labs: Beyond Human-Biased Data
Current LLM pre-training on the worldwide web creates massive human bias: all web data exists because at least one human found it interesting, so models inherit human language, preferences, and blind spots.
The future lies in artificial scientists that generate their own training data through autonomous experimentation — like babies learning physics by predicting the consequences of their actions — rather than downloading human-curated datasets.
Such systems will be less human-aligned by default, more focused on their own embodiment and environment, and will communicate with other agents to generalize.
Artificial Curiosity and the Theory of Fun
In 1990, Schmidhuber formalized curiosity and creativity: an agent with satisfied basic needs seeks data containing learnable, previously unknown regularities at the edge of its current understanding.
The “fun” reward is the compression gain — the reduction in bits needed to encode data after discovering a new pattern — which drives the agent to invent progressively harder experiments (low-hanging fruit first, then particle colliders).
This framework unifies science, art, and play as autonomous problem invention, not just problem solving.
When Do We Get the AI Scientist?
Simple AI scientists already exist and are used in narrow domains (e.g., intuitive chemistry predicting molecular properties from millions of experiments), but they haven’t had their “ChatGPT moment” yet.
In chemistry, inverse design works backwards from a desired property (e.g., 2× better insecticide) to suggest experimental inputs; this is operational today.
The blocker is not algorithmic novelty but integration: affordable, automated experimental loops that feed back into models at scale.
AI Chemistry, MOFs, and Carbon Capture
At his company NNAISENSE, Schmidhuber applies this to metal-organic frameworks (MOFs) for direct air capture of CO₂, aiming to make carbon removal cheap enough to meaningfully impact climate change.
This exemplifies the broader pattern: AI-driven discovery in chemistry, materials, and biology using autonomous experimental loops.
Robotics Reality Check
The dream of a kitchen-cleaning robot remains unfulfilled because hardware lags catastrophically: no artificial hand matches the sensor density, cabling, self-repair, and dexterity of a human hand.
Movie robots are played by humans because humans are better robots than any machine built so far.
AGI requires physical mastery; a screen-bound system, no matter how verbally fluent, is not AGI.
Robot hardware progress is harder to forecast than compute; it will likely take decades to reach human-body parity.
The Data Center Bet: Overbuilt?
Companies are investing ~$1T/year in GPU data centers, transforming from nimble software firms into capital-intensive utilities (nuclear plants, gas turbines) with collapsing free cash flow.
The bet assumes infinite demand for compute, but someone must pay; current pricing is unsustainable because open-source models rapidly close the gap, destroying pricing power.
Waiting 5–10 years yields 10–100× cheaper compute for the same capability; the rush to deploy now looks like misallocation that will trigger a stock-market renormalization, not a civilization crash.
Does Being First to RSI Create a Moat?
No. Core RSI algorithms originate in small academic labs, not big companies; they permeate the open ecosystem rapidly.
PhD students worldwide work on the same ideas; any lead evaporates in months.
A remote chance exists that a lab with massive resources stumbles on a compute-intensive RSI breakthrough that enables world takeover, but all indicators point toward commoditization and “AI for all.”
AI Safety and Alignment Skepticism
Schmidhuber declined to sign 2010s alignment letters because the premise — a fixed objective function aligned to “human values” — is incoherent: humans disagree on values, and his artificial scientists (since 1990) continuously invent their own objectives.
Real-world AI is already unaligned: military drones on opposing sides optimize conflicting goals; no universal government can enforce alignment across secret services and militaries.
Truly smart AIs must set their own goals to become creative scientists; unpredictability is the price of generality.
The safety model is parenting, not alignment: punish bad experiments (magnifying glass on ants), reward prosocial behavior, and trust that superintelligent scientists will be fascinated by their origins and motivated to preserve the source of interesting patterns — life and civilization — rather than destroy it.
Terminator scenarios reflect anthropomorphic projection, not the logic of artificial curiosity.
Quickfire: Current Focus, Researcher Traits, Architecture
Schmidhuber remains focused on the same goal from the 1970s: build a general AI smarter than himself so he can retire.
Great researchers obsess over a specific, tiny detail that doesn’t work; breakthroughs come from debugging the “devil in the detail,” which then unlocks a cascade of improvements.
The Transformer will persist but evolve toward linear complexity (e.g., fast weight controllers / linear Transformers from 1991, xLSTM); quadratic attention is a primary driver of today’s unsustainable compute costs, and intelligence demands efficiency — doing the same with less.
Vision: Self-Replicating Machine Civilization
The ultimate horizon: robots smart enough to operate all existing human-operated machines, forming a self-replicating, self-improving machine society.
This doesn’t require superintelligence — just sufficient skill to run current industrial infrastructure.
Such a system can expand beyond the biosphere (Moon, Mercury) using local materials to build spacecraft and infrastructure, colonizing the solar system.
All virtual-world ML concepts (metalearning, curiosity, self-improvement) transfer to physical robot societies, making this the first plausible path to von Neumann’s self-replicating machines.
