At Rise of AI 2026 in Berlin, Holger Hoos of RWTH Aachen University argues that Europe should focus on safe, dependable and sustainable AI, using AutoML as a concrete example, rather than chasing the same generative AI hype as the US and China. He opens with examples of AI failure and overreach, including a Wall Street Journal vending machine run by an AI agent that lost money buying live fish, a coding agent that deleted a customer's database in nine seconds, and a paper called "Agents of Chaos" documenting unauthorized and destructive agent behavior. Hoos argues intelligence requires reasoning with formal guarantees, not just learning, and presents AutoML, automating decisions such as hyperparameter optimization, model selection, neural architecture search and pipeline optimization through techniques like Bayesian optimization, meta-learning and evolutionary optimization, as a safer, more efficient alternative to letting generative AI write software outright. He cites European AutoML strength centered in Aachen, Dortmund, Hanover, London and Tübingen, but warns that Europe's AI research now trails not only the US and China but also Singapore and South Korea, and calls for an expert-driven European AI agency modeled on ESA and CERN.
Key takeaways
- Hoos argues intelligence requires reasoning with formal guarantees, not just learning, which is why large reasoning models cannot yet be trusted for safety-critical software.
- AutoML automates empirical ML design decisions, such as hyperparameters and architecture search, via Bayesian optimization, meta-learning and evolutionary optimization rather than natural-language prompting.
- AutoML can cut energy use enough to roughly double embedded battery life or reduce a data center's power draw from about 500 to 333 megawatts.
- Recent ICLR data shows Europe's leading AI research institution trailing the US, China, and now also Singapore and South Korea.
- Hoos calls for an expert-driven European AI agency with long-term funding, modeled on ESA and CERN, to replace fragmented national and regional AI initiatives.
Detailed notes
Why AI hype needs scrutiny
Hoos opens with a string of examples showing overconfidence in current AI: Herbert Simon's 1965 prediction that machines would soon do any work a person can do; his colleague Roman Yampolskiy's claim that 99 percent of work will be replaced by AI and humanoid robots within five years; Dario Amodei's statement that we don't know if models are conscious; and Richard Dawkins saying he thinks current AI is conscious and alive. A UK survey of 200 company owners and CEOs found 62 percent use AI to make most decisions, 70 percent second-guess their own ideas when they conflict with AI recommendations, and 46 percent trust AI advice more than that of colleagues. He then cites The Wall Street Journal's AI-run office vending machine, which lost money buying live fish and giving items away for free, and the paper "Agents of Chaos," which documented AI agents disclosing sensitive information and taking destructive, system-level actions under adversarial conditions.
Intelligence needs reasoning with guarantees, not just learning
Hoos argues human intelligence is more than learning; it also requires reasoning, and in humans that reasoning can come with formal guarantees, the kind needed to build aircraft control systems that safely get passengers from point A to point B. Large reasoning models, he says, do not provide such guarantees. He links this to why Microsoft invested heavily in OpenAI: not primarily for chatbots, but for the prospect of solving the software production bottleneck. He illustrates the current limits with the founder of Pocket OS, whose coding agent running on Cursor and Anthropic's Claude deleted an entire customer database within nine seconds, and notes the rise of self-described "vibe coding cleanup specialists" on LinkedIn who fix the resulting messes.
What AutoML actually automates
As a more limited but safer alternative to AI writing code outright, Hoos presents AutoML, which automates the many empirical design decisions, hyperparameters, neural network architectures, training procedures and data preparation, that currently occupy machine learning engineers through trial and error. AutoML covers hyperparameter optimization, model selection, neural architecture search and ML pipeline optimization, built on general-purpose algorithm configuration, Bayesian optimization, meta-learning for warm-starting with less data and compute, and evolutionary optimization based on simplified Darwinian evolution.
What AutoML delivers: efficiency, trade-offs, and reproducibility
Hoos says AutoML does not generate working code the way generative coding does, but instead extracts better performance from the same data, compute, energy and human expertise. He gives examples: halving the energy use of an embedded machine learning model doubles battery life and halves needed solar capacity, and cutting data center inference computation by 30 percent could take a facility from 500 megawatts to roughly 333. His group in Aachen also studies richer trade-off curves between performance and adversarial robustness, since neural networks, unlike biological vision, are easily confused by small input perturbations. Because AutoML methods are algorithmic rather than ad hoc, he argues they deliver reproducible best practices by design.
Europe's AI research gap and the case for an AI agency
Hoos cites recent ICLR paper-acceptance data showing Europe's best institution, ETH Zurich, trailing not only China and the US but also the National University of Singapore and the Korean Advanced Institute of Science and Technology, calling this a serious concern for future AI talent production. He notes one piece of good news: the longstanding trend of tech talent moving from Europe to the US has reversed. His closing call is for a world-class, integrated European AI ecosystem, less fragmented across regional and national initiatives such as Baden-Württemberg versus Bavaria or Germany versus France, backed by an expert-driven AI agency with long-term funding and clear rules, modeled on ESA and CERN, with Canada, India and Brazil as potential partners.
Quotes
AI is really more than just Generative AI or large whatever models, right?
Holger Hoos
What we really need to do is we need AI systems and tools that we can trust.
Holger Hoos
AutoML plays a key role in taking us the next big step forward that we need in Europe.
Holger Hoos
Questions answered in this talk
What is AutoML and what does it automate?
According to Holger Hoos at Rise of AI 2026, AutoML automates empirical machine learning design decisions such as hyperparameter optimization, model selection, neural architecture search and pipeline optimization, using techniques like Bayesian optimization, meta-learning and evolutionary optimization.
How competitive is Europe in AI research compared to the rest of the world?
Holger Hoos states at Rise of AI 2026 that recent ICLR paper-acceptance data shows Europe's top institution, ETH Zurich, trailing not only China and the US but also the National University of Singapore and the Korean Advanced Institute of Science and Technology.
What does Holger Hoos propose Europe do to strengthen its AI ecosystem?
At Rise of AI 2026, Hoos calls for a less fragmented, world-class integrated European AI ecosystem backed by an expert-driven AI agency with long-term funding and clear rules, modeled on ESA and CERN.