Pattern Recognition Lab

A critical assessment of the EU Cloud and AI Development Act (CADA)

Earlier this week I argued that the EU AI Act should be abandoned in favour of an AI sovereignty agenda built around compute, capital, data, deployment and speed. On 3 June 2026, less than three weeks before that argument, the European Commission published its proposed Cloud and AI Development Act (CADA). On paper, CADA looks like the start of exactly that agenda. It promises AI factories. It promises EuroCloud federation. It promises sovereignty assessment. It promises that European public procurement will favour services with verified Union assurance levels.

CADA Is The Right Idea Without The Engine: A Critical Assessment Of The EU Cloud And AI Development Act Earlier this week I argued that the EU AI Act should be abandoned in favour of an AI sovereignty agenda built around compute, capital, data, deployment and speed.

Tying Down the Wild: Charting the Universal Map of Interpretable AI

The modern digital landscape is ruled by colossal Artificial Intelligence models—systems whose predictive power is staggering. These Deep Neural Networks (DNNs) can classify images, translate languages, and drive complex decisions with superhuman accuracy. Yet, this very power comes with a profound, often terrifying, side effect: opacity. We can observe the input and the output, but the vast, labyrinthine computations occurring in the middle remain a black box. For any critical application—from medical diagnosis to autonomous driving—this lack of transparency is not just an academic nuisance; it is a safety risk, a regulatory hurdle, and an epistemic barrier. We need to know *why* the AI made a decision, not just *what* the decision was.

Tying Down the Wild: Charting the Universal Map of Interpretable AI The Elusive Quest for Understanding in the Age of Giant Models

Seeing Through the Veil: How AI is Minimizing Radiation Doses in Life-Saving Angiography

For doctors tackling critical vascular emergencies—such as aneurysms, blockages, or malformations in the brain—Digital Subtraction Angiography (DSA) remains the gold standard. DSA is a sophisticated imaging technique that uses X-rays combined with a contrast agent flowing through the body's vessels. By taking an initial "mask" scan before the contrast is injected and then subtracting it from a subsequent "fill" scan, clinicians can isolate and visualize the blood flow pathways, effectively making the vessels glow against the background tissue. This detailed view is crucial for guiding complex procedures like coiling an aneurysm or removing a clot.

Seeing Through the Veil: How AI is Minimizing Radiation Doses in Life-Saving Angiography Illuminating the Urgent Need for Safer Medical Imaging

Multi-temporal calving front segmentation

Across the planet’s icy edges, marine-terminating glaciers are acting as one of the most sensitive thermometers of our changing climate. These massive rivers of ice spill into the sea, and the precise location where they meet the ocean—the calving front—is not just a geographical feature; it is a critical indicator of the glacier's health and its contribution to global sea-level rise. The dynamic shift of this front, driven by physical forces like ice flow and submarine melting, holds the key to accurately quantifying ice mass loss. For scientists aiming to monitor these shifts across vast, remote regions, satellite imagery is the preferred eye in the sky.

Seeing Through the Fog: How Temporal Sight is Charting the Retreat of Glaciers The Urgent Call of the Frozen Frontiers

AI’s progress is fast and we need to accelerate to keep its pace. Yesterday I had the chance to present my talk „What‘s next in medical AI?“ at DLDxHealth. It was energising to dive into the latest breakthroughs, discuss the hurdles we face in translating research into real-world health solutions, and explore how we can collectively push the field forward.

I was thrilled by the thoughtful questions from the audience – everyone is clearly as passionate about responsible, impactful medical AI as I am. The conversation touched on data quality, regulatory pathways, and the importance of interdisciplinary collaboration. Those moments reminded me why sharing ideas in venues like DLDxHealth is so valuable.

If you missed the event, the slide deck is available on GitHub, so you can catch up on the key points and visualisations I shared. I hope the material sparks new ideas and encourages more dialogue around accelerating medical AI responsibly.

Thanks to the organizers and all who joined the discussion. I’m looking forward to the next steps and to keeping the momentum going together.

I've just published an op-ed on the Anthropic Fable 5 / Mythos 5 ban — and what it should tell Europe.

The U.S. government just ordered Anthropic to suspend access to its Fable 5 and Mythos 5 models for any foreign national — including foreign-national Anthropic employees. Anthropic is complying while disagreeing publicly with the technical basis.

The strategic lesson is brutal: frontier AI is no longer just software. It is industrial capacity, scientific acceleration, medical capability, economic sovereignty. A capability that can be withdrawn by directive is a capability you do not control.

In the piece, I argue the EU AI Act should not be patched or delayed — it should be abandoned and replaced by an AI sovereignty agenda built around compute, capital, data, and speed. Europe needs AI factories, not paper factories.

https://open.substack.com/pub/akmaier/p/the-fable-5-ban-shows-why-europe-765?r=56pbmv&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Glad to be part of the informal kick-off of the Bavarian Foundation Model initiative at DLDxHealth. It was inspiring to see the momentum behind this effort, especially with the support of Minister of Science of Bavaria Dr. Markus Blume, BAIOSPHERE, and a host of Bavarian universities – TUM, LMU, FAU, Uni Würzburg and many more.

Seeing the 56 Mio EUR project finally taking off feels like a real milestone for the region’s AI research landscape. The energy in the room was palpable: researchers, industry partners, and policymakers all sharing a common vision of building large-scale foundation models that can benefit health, science and society.

I’m especially excited about the collaborative spirit that’s already emerging. When you bring together top-tier universities, a forward-thinking initiative like BAIOSPHERE, and strong political backing, the possibilities for breakthrough research multiply. It feels like we’re laying the groundwork for a new era of AI-driven health solutions that are both robust and responsibly developed.

A big thank you to everyone who helped make the kick-off possible – the organizers at DLDxHealth, the ministerial team, and all the university colleagues who contributed their expertise. I’m looking forward to the next steps, the research partnerships that will form, and the real-world impact these models can have.

Stay tuned for more updates as the project progresses. It’s an exciting time to be part of Bavaria’s AI ecosystem!

Data Extrapolation From Learned Prior Images for Truncation Correction in Computed Tomography

Computed Tomography, or CT, scans are foundational tools in modern medicine, allowing doctors to create detailed cross-sectional images of the inside of the human body. When these scans are performed using Cone-Beam CT (CBCT) systems—often employed in dynamic interventional procedures like surgery or angiography—they face a critical physical limitation: the detector cannot cover the entire patient. This is known as data truncation.

Closing the Gaps: How AI is Learning to See Beyond the Edges of a CT Scan The Unseen Challenge of Medical Imaging

Germany Needs to Treat Foreign AI Talent Better

This week, one of my group members — an Indian citizen, soon-to-be junior research group leader at Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU), and someone holding a valid German visa, a valid work permit, and a freshly signed five-year contract — was refused boarding on a Lufthansa flight back to Germany. As I write this, she is stranded in China, in the middle of what had been a highly successful business trip building partnerships with top medical-AI research institutes across India and China. The rest of our delegation is back in Erlangen. She is still there, on her own, while my colleagues and I try to untangle a paperwork problem that should never have existed in the first place.

Germany Needs to Treat Foreign AI Talent Better This week, one of my group members — an Indian citizen, soon-to-be junior research group leader at Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU), and someone holding a valid German visa, a valid work permit, and a freshly signed five-year contract — was refused boarding on a Lufthans...

Quantum's New Lower Bound: Two March 2026 Papers Compressed the Timeline for Breaking Elliptic-Curve Cryptography

For decades, the bedrock of modern digital life—from securing online banking to validating blockchain transactions—has rested upon the strength of Elliptic-Curve Cryptography (ECC). This mathematics, elegantly complex, provides a nearly unbreakable digital lock, making today’s encrypted conversations and stored assets safe from brute-force attacks. However, the looming development of sufficiently powerful quantum computers poses an existential threat to this foundation. The question of *when* that threat materializes—the point often termed Q-day—has become a subject of intense, sometimes speculative, debate. Now, a flurry of significant scientific papers released in March and June of 2026 have acted like a sudden, decisive tightening of the screws, materially shrinking the public estimate for when cryptographically relevant quantum attacks on ECC could become a reality.

The Countdown Clock Shrinks: How New Quantum Papers Compressed the Threat to Digital Security For decades, the bedrock of modern digital life—from securing online banking to validating blockchain transactions—has rested upon the strength of Elliptic-Curve Cryptography (ECC).