Kilian Bron’s dreams lead him to the summit of Huayna Potosí, at more than 6,000 meters above sea level, for a never-before-seen alpine mountain biking project. A committed and risky project, one that made him doubt himself and sometimes feel afraid, but also pushed him to go beyond his own limits. Through his bike, Kilian expresses himself, travels the world, meets incredible people, and discovers landscapes that leave a deep mark on him. This talk tells the story of how far a dream can take us, when we dare to find out what we are truly capable of.

Seas the Day spent 165 days at sea, setting off from Peru, South America, in May 2025. Rowing up to 15 hours a day, they travelled west across 8,213 nautical miles to Cairns, Australia. On Saturday, October 18th, 2025, Miriam Payne and Jess Rowe became the first all-female team to row the Pacific Ocean mainland to mainland, non-stop and unsupported.

Embracing the challenges of the high seas, the duo used this opportunity to empower especially young women, break several world records, and raise nearly £120,000 for The Outward Bound Trust, a UK charity that helps children develop lifelong skills through adventure.

The tale of their Great Escapade is one of two halves. It highlights the challenges faced when their rudder broke just seven days and 350 miles into their first attempt, forcing them to return to land and plan their re-start. On their second attempt, it acknowledges the sheer grit and determination given to keep spirits high when battling driving rain and wind, 30ft waves, sleep deprivation and losing an oar to the ocean. And, it details some of the more technical issues with their power system and emergency hand pump watermaker’s failure and focuses on how resilience and a positive mental mindset to manage these issues.

Seas the Day’s story is a testament to the power of a positive mindset and the sheer drive to continue going forward. The duo will inspire audiences to embrace their own personal challenges and discover the limitless possibilities that lie beyond their own comfort zones.

As AI systems get better and better, we come across unexpected bottlenecks: The technology seems to be ready, but often, the people who should use it aren‘t. In this talk, Gregor Schmalzried looks at the most intriguing current and upcoming developments in generative AI, particularly with agentic systems, and takes them one step further: In a world of abundant technology, which products and services become valuable? And what can we do to reduce fears, inspire minds, and bring everybody along for the ride of a lifetime?

AI safety is often framed as an ethical question, but for businesses shipping real products, it is also a matter of security, trust, and legal integrity. Unsafe model behavior can expose secrets, amplify bias, enable abuse, and create serious reputational risk. This talk gives developers and product teams a practical introduction to AI alignment and AI safety through the lens of how systems actually fail in production.

From there, Áron Erdélyi, László Balázsik, and Benjamin Balogh examine how the AI red teaming field is moving from manual probing to automated adversarial evaluation at scale, including static test sets, agentic jailbreaking, and optimizer-driven attack discovery. They close with a practical threat model, concrete mitigations, and an overview of the current AI security landscape, giving developers a grounded framework for building safer, more secure AI systems.

Understanding cosmology - the fundamental physical laws that govern our Universe - requires big telescope data and a look into features that have so far escaped human-only analyses. To this end, Artificial Intelligence can be used in multiple ways. With suitably designed architectures, it can be a tool for accurate calibration of data and control of systematic errors. Replacing costly calculations by emulation, it can enable analyses that otherwise would be computationally infeasible. Trained on numerical simulations, it can serve as a model for complex astrophysical phenomena that exceeds the abilities of human theorists. And perhaps LLMs are about to not just augment but replace the human scientist as the origin of ideas and interpretation. In this talk Prof. Dr. Daniel Grün will give an overview of these developments and challenges.

In 2025, the performance of LLMs in verifiable domains like math, coding tasks, and reasoning increased rapidly.
Tools like Claude code or Codex became a standard tool in the repertoire of software developers as a result. What was the mechanism behind this rise in capability?

To answer this question, Dibya Chakravorty, Bernhard Altaner, and Debsankha Manik invite you to take part in their journey of competing in ARC-AGI-2, one of the hardest benchmarks targeting abstract reasoning capabilities at a time before agentic harnesses were commonplace.

When they started their work, frontier LLMs had near-noise level performance in that benchmark. Nine months later, in January 2026, they had devised an agentic harness whose performance surpassed the human baseline. They argue that dynamically using as much compute as necessary for the complexity level of the task at hand, makes LLMs perform significantly better at reasoning tasks.
More precisely, in 2025, LLMs finally allowed for efficient "test-time compute scaling" during inference by leveraging two trends: reasoning within the chain of thought, and using tool calls (in their case, a python interpreter). By combining these two principles, i.e., interleaving the reasoning process with tool calls, we realized that even smaller models could significantly improve their capabilities and exhibit features of fluid intelligence, at least in verifiable domains. While this paradigm is now generally accepted in the field, their journey provides a behind-the-scenes view of a rapidly developing field from the perspective of a small AI Lab - and some practical lessons.

Large Language Models are becoming more capable, but also more expensive to run, deploy, and customize. This talk explores how REAP and related compression techniques can make LLMs practical for edge and resource-constrained deployment.

Join this talk for an inside look at TNG's AI Research Team as Dr. Andreas Rabenstein and Dr. Fabian Klemm share their journey setting up an 8×NVIDIA B200 GPU cluster for distributed training and inference. They will walk you through the process – from hardware setup and network configuration to software stack decisions and performance tuning.

This talk covers the technical challenges they faced, some of the experiments they ran, and lessons learned. They will discuss container orchestration, distributed training frameworks, and inference optimization, sharing benchmarks and results from their test workloads.

Whether you are planning your own GPU cluster, curious about the B200 architecture, or interested in AI infrastructure, this session provides practical insights and honest reflections on what worked and what didn't.

Reinforcement learning (RL) has emerged as the primary paradigm for scaling base models into autonomous agents across software engineering, research, and beyond. Yet scaling RL poses fundamentally different challenges than pre-training. Its asynchronous and online nature introduces new problems, from curriculum design and synthetic task generation to off-policy training dynamics and large-scale sandboxed rollout execution.

In this talk, Konstantin Dunas traces how RL scaling has evolved from early approaches and why this shift creates new infrastructural bottlenecks. He will examine challenges such as training stability, training–inference policy mismatch, and off-policy learning, and how prime-rl, Prime Intellect’s open-source training framework, enables others to apply these approaches to train their own agents. The talk will also outline what comes next for large-scale LLM-RL.

Artificial Intelligence is advancing rapidly, but teaching machines to understand the visual world still often depends on huge amounts of carefully labeled data. This talk introduces self-supervised learning: an approach that allows AI systems to learn from images and videos without needing every example to be manually annotated. Using videos as a natural source of visual experience, we can help AI learn from motion, change, and continuity -much like humans learn by observing the world around them. Prof. Dr. Yuki Asano will give an accessible overview of why this matters for the next generation of vision foundation models, what makes videos such a powerful training signal, and how these ideas could support future applications in areas such as robotics, medical imaging, and general visual understanding.

For years, processing confidential data with AI required on-premise infrastructure. TNG has followed this path, but the reality is that the hardware costs and operational complexity of maintaining private clusters are often impractical for rapid scaling. Since using untrusted cloud GPU providers bears the risk of loosing confidentiality, it is necessary to investigate how to build trust through hardware-enforced security while retaining the flexibility of cloud resources.

In this talk, Dr. Benjamin Merkel, Dr. Marcel Lippmann, and Dr. Fabian Roll demonstrate how Intel Trust Domain Extensions (TDX) and NVIDIA Confidential Compute enable cryptographically verifiable execution environments on untrusted hosts. Leveraging remote attestation, they maintain workload integrity for sensitive operations like LLM inference. They share their learnings about these technologies, trust boundaries, and operational challenges.

Artificial Intelligence is no longer just for engineers, yet its inner workings often feel like complex magic hidden behind high-dimensional math. How do we make sense of modern AI without getting lost in the technical jargon? The answer lies within ourselves.

In this session, Daniel Klingmann and Max Dietsch will bridge the gap between biological and Artificial Intelligence using intuitive, everyday analogies. They will take a high-level journey through the "Digital Mind," exploring how AI mirrors human psychology and the way a child learns. They will discover how machines are being built to think, learn, and evolve just like we do. No PhD required - just a healthy dose of curiosity.

The internet never forgets. We've all heard that before. But what does that really mean? In this talk, Dr. Dennis Schulz and Gregor Ziegltrum show the chaos, the biggest insights and the weirdest details that come back when you invoke your European right to your own digital history.

They asked hundreds of companies for their personal data: Online shops, social networks, advertisers, health insurers, dating platforms, and so many more. They received everything from hand-redacted PDFs to gigabytes of CSV and JSON files, and tried to reconstruct their online and offline lives. They turned to AI to handle the heaps of data, and before they knew it, it had assembled an almost complete biography of their lives.

In this talk, Dennis and Gregor will unpack all the datasets they got: Small online shops, medium-sized tech companies, and the Silicon Valley giants. Using this data, they try to analyze their own behavior: Have they truly never ghosted anyone while dating online? Can they tell where they went on vacation ten years ago? And do they even want to know what their search history reveals? In true “what could go wrong?” spirit, they will run a live demo on our their data. Let's face the most embarrassing things companies already know about us!

How does AI change the way that we approach legacy code migrations? In this talk, Dr. Miguel Obando Leitón, Dr. Moritz Maus, and Lukas Hauser will share their insights from migrating COBOL to idiomatic Java.

Modernizing a legacy application landscape, how hard can it be? What if it's older than most of your employees?Forgotten knowledge, unstructured architectures, and cost pressure are challenges that are commonly faced during such code migrations. Even worse: When the systems are still in use today, a simple migration turns into open-heart surgery. Good thing, there is AI assistance.

In this talk, Miguel, Moritz, and Lukas will demonstrate how they leverage modern agentic coding tools to enable new migration paths, which would have been nearly impossible without AI.

Due to the acute shortage of dental assistants, the documentation workload is increasingly shifting to dentists – with consequences: Compromises in treatment quality, longer waiting times, economic losses, and legal risks from incomplete records. Martha Büttner, Sebastian Zett and Dr. Julian Bollig present Dampsoft's innovative AI-powered speech documentation system: A seamlessly integrated solution for practice management systems that transcribes treatment conversations and generates corresponding documentation. They provide insights into the software architecture – from audio preprocessing and live transcription to AI-generated summaries – and discuss implementation challenges. They also introduce an evaluation framework serving as a foundation for quality assurance and development. Currently in beta testing, the system demonstrates how AI not only optimizes processes but ultimately enhances patient care.

Operators face enormous logistical challenges when expanding fibre optic networks. The trench network alone extends over a length that is almost equivalent to the circumference of the earth. Precise billing for these massive construction projects requires innovative solutions. Current methods for verifying route lengths rely on on-site personnel who must walk the constructed sections with a measuring wheel. This process is not only time-consuming and labour-intensive, but also hardly scalable given the dimensions involved.

In their talk, Thorsten Nieser, Dr. Sebastian Stamminger and Dr. Tetyana Kyrey will present an approach that uses AI image recognition for the automated analysis of road surfaces and pavements. Georeferenced camera images and LiDAR scans serve as the data basis. A deep learning model retrained by them evaluates the data and identifies relevant surface structures. The solution presented enables a significant improvement in quality through full inspection instead of random sampling, while also significantly speeding up the billing process. In addition, machine evaluation ensures greater objectivity and traceability of results.

The presentation focuses on practical implementation and the ‘surprises of real life’ - and on providing a clear answer as to where AI helps and where it does not.

Before something becomes big, it often looks ridiculous first. AI roleplay is treated like a strange fringe space. That is convenient. Millions of people are writing with language models into characters, worlds, and lives they cannot find anywhere else. The quietest cultural form of this decade, and one of the largest. Whoever shapes it shapes how a generation learns intimacy and storytelling in the digital age.

Born from fanfiction, pen-and-paper, and the possibility of being immortal. No one planned this. Neither did Dana Sophia Hylla. She built part of the infrastructure for it on her own. On the largest platform in this field, her name was not even allowed to be spoken for a time. Then they called.

LoreBary did not begin with a vision, but with a mistake. A single Colab notebook, meant to break through barriers, and with it the question of what emerges when you stop asking for permission. Created out of a refusal to lock people into rigid systems, it makes room for creative tools that do not belong only to those who can afford them.

How do you build something that does not keep turning the same wheel? What do you owe the people who write on it - and who have already seen everything?

Markus Knall offers insight into a newsroom in the midst of transformation. Ippen.Media, with 70 websites, 27 million monthly readers, and 4,000 articles per day, is rethinking publishing from the ground up. The change is profound. Since the introduction of ChatGPT in November 2022, AI-generated content has surpassed human-written content in quantity. At the same time, with Google’s decline in importance, a business model that had worked for years is collapsing.
Publishers are being squeezed from two sides: by AI systems that deliver direct answers, and by the creator economy competing for the same attention. Ippen.Media has already deeply integrated AI into its editorial workflow - from interactive data‑journalism projects (for example, real‑time election maps for Bavarian and Hessian local elections) to AI‑supported topic series and automated quality checks.
The talk demonstrates how a fully AI‑driven editorial pipeline works in practice and shows that AI is seen not as a replacement for journalistic expertise but as a productivity multiplier. The next major paradigm is not the mobile web or social media, but the “agentic web,” in which AI agents decide which content users will even get to see.

LLM progress now depends heavily on one practical issue: training stability at scale. Sparse Mixture-of-Experts (MoE) models are especially sensitive, since routing drift can overload experts, collapse utilization, and stall learning.

In this talk, Lucas Atkins will share an "anti-loss-spike" playbook from a recent open-weight run: a 400B-parameter MoE with 13B active parameters per token, trained for 17T tokens with an unsmoothed loss curve and zero loss spikes. Lucas Atkins will start with the failure pattern we saw, router drift, overload, MaxVio divergence, and plateau, then cover the fixes that restored steady convergence: bounded and momentum expert-bias updates (SMEBU), z-loss for logit stabilization, a precision fallback from MXFP8 to BF16, better balancing objectives, and data/packing choices that reduced step-to-step variance.

Many companies want to harness the great potential of AI to increase efficiency, speed, and quality in a targeted manner. However, ready-made solutions like ChatGPT or Claude are often problematic from a data protection perspective and often too generic to be truly integrated effectively and fruitfully into everyday work. Creating an alternative offering that provides similar basic functions but also targeted integrations for specific use cases is challenging.

In this talk, Christopher Köhldorfer, Dr. Jannis Zeller and Johannes Esslinger will show how they set up a modular architecture of the AI platform at Bank Frick to enable the long-term vision of scalable automation of knowledge-intensive tasks along the entire value chain.

Large Language Models have moved from research curiosity to production reality at a pace that has left security teams scrambling. Organizations are racing to integrate LLMs into agentic browsers, AI-integrated development environments, customer service chatbots, and internal knowledge systems – often without fully understanding the novel attack surfaces they're introducing.

This talk examines security challenges at the frontier of LLM adoption: Prompt injection attacks turning helpful assistants into unwitting accomplices, data exfiltration risks hiding within conversational interfaces, and the dangers of giving AI agents the ability to take real-world actions on behalf of users. Drawing from academic research, real-world implementations, and documented incidents, this talk will discuss what's working, what's failing, and what companies need to understand as these systems become embedded in critical workflows.

For tackling engineering tasks with advanced AI capabilities, one can use multi-agent systems that orchestrate specialized tools. These tools encapsulate large parts of the underlying engineering knowledge. Due to the generally complex nature of engineering tasks, the respective multi-agent systems oftentimes require additional care regarding their robustness. In this talk, Dr. Marc Eric Vogt, Heinrich Schmidt, and Dr. Thomas Hugle show how they achieve this necessary robustness by using a state-based multi-agent system and deterministic logic in key locations. While they focus on the engineering case, the general lessons learned should apply to all multi-agent systems that tackle complex problems with a large number of underlying variables.

Lenny Miao and David Ji will share a commercial review and the story of how they built and scaled MiniMax to a high-performance AI platform. Alongside this, they will talk about how teams use their models in practice, identifying the best real-world use cases.

Ever-increasing computing power, rapid advances in deep learning, and growing amounts of medical data are transforming medical imaging. In image-guided interventions, however, leveraging these developments requires careful integration of multiple technologies and robust system architectures. ImFusion combines classical image processing, real-time visual computing, and modern machine learning to develop advanced software for cutting-edge medical devices.

Dr. Wolfgang Wein will present examples from projects where multiple real-time imaging streams such as ultrasound, video, and X-ray are processed jointly; outline the architectural principles behind their software stack; and show how ImFusion evolved towards hybrid solutions spanning medical devices, browser-based applications, and cloud backends. The talk concludes with perspectives on trends and challenges in bringing innovative software into the regulated medical device ecosystem.

How do you turn “something with AI” into a productivity tool for legal research that is actually used in practice? Using beck-chat as an example, Daniel Petró and Maximilian Huber showcase the journey from the first RAG prototype through a beta integration into beck-online to the standalone platform chat.beck.de.

They explain how the project evolved across three dimensions: technically, in terms of UX, and professionally within the legal domain. In particular, they discuss the technical concepts required to achieve this goal: semantic search, RAG, agentic AI, and the evaluation of AI products. The talk does not require expert knowledge and is designed to be accessible to a general audience.

The presentation provides concrete insights from real-world evaluation and usage, as well as the steps that were necessary to transform an experiment into a reliable legal research tool.

ScanCode is the open-source industry standard for scanning software for licenses and copyrights – a substantial real-world system with more than 60,000 lines of Python behind it. Conventional wisdom says rewriting something like that would take a team of engineers and a seven-figure budget. Dr. Maxim Stykow and Dr. Adrian Braemer wanted to find out how far $20,000, three weeks, and a swarm of AI agents could really get them.

This is their war story. They first attempted the rewrite in February 2025, when coding agents were still new. It was promising, but too supervision-heavy to be practical, so they shelved it. Then in early 2026, the picture changed: Better models, better harnesses, and better workflows made it possible to push agents much further on long-running engineering tasks.

Maxim and Adrian will share the gritty details: the mix of third-party frontier models and open-weight models running on TNG’s own Skainet infrastructure. The dramatic quality gap between models – where stronger ones solved in minutes what weaker ones failed at for hours. The point where an agent introduced thousands of tests and slowed iteration to a crawl, forcing them to teach it to optimize its own feedback loop before continuing. Progress lost to accidental git resets. Orchestrator agents forgetting about their own subagents and fighting themselves. The “one more prompt” addiction. And the surprising power of AGENTS.md, living documentation, and tight feedback loops with custom linters and tests.

Whether you care about large software rewrites, autonomous coding, or simply getting more leverage as a developer, this talk is a practical field report from the current frontier of agentic engineering – what worked, what didn’t, and what’s coming next.

AI chips remain the foundational building blocks of modern datacenter infrastructure. This talk provides a deep dive into the critical specifications of AI hardware and the current competitive landscape. Gilles Backhus will explore the technical differentiators across architectures and how they impact the KPIs that define success: model size, context length, cost-efficiency, and performance metrics such as tokens per kWh, per-user speed and rack-space density. Finally, he will showcase Tensordyne’s latest innovations, demonstrating how quantitative hardware leaps unlock next-generation applications – such as real-time 4K AI video generation.

With every release of large language models (LLMs), people often dive straight into their performance on relevant benchmarks like GPQA or SWE-bench Verified. Differences of a few percentage points compared to the competition are quickly interpreted as progress, falling behind, or a breakthrough. Critics say, however, that many benchmarks have little or no significance and are disconnected from reality.

In this talk, Florian Brand addresses the challenges of LLM evaluations: From the differences in the implementation of benchmarks and the effects of different parameters, to the necessary infrastructure. In addition, problems in creating benchmarks, especially for agentic systems, will be discussed, as these pose new challenges for the design of the evaluations and the infrastructure.

This talk is a field guide to MCP, A2A, generative UI, and related standards. MCP has rapidly become the de facto standard for tool calling in AI systems – not because it was perfectly designed, but because it was the first open protocol to gain real traction. What began as a pragmatic interface for connecting models to tools has now become the foundation of a much broader ecosystem.

In this talk, Mathias Burger and Áron Erdelyi trace how MCP evolved from an early standard into widely adopted infrastructure across major LLM providers and thousands of public servers, and examine the challenges that have come with that growth: Fragmented implementations, unresolved security risks such as prompt injection and tool poisoning, and increasing context bloat as servers expose ever-larger tool surfaces.

But tool calling is no longer the whole story. As AI systems become more capable, they also require richer interfaces, longer-running workflows, asynchronous communication, and ways to share artifacts across systems. That pressure has driven the emergence of adjacent ideas and protocols: MCP Apps and A2UI for generative user interfaces, and A2A for long-running, multi-agent interactions.

Mathias Burger and Áron Erdelyi cut through the noise to clarify what these agentic protocols actually specify, what they do not, and how they fit into the larger AI landscape. The result is a practical map of where these protocols stand today and how these efforts are shaping the next generation of agentic systems.

For over a decade, Sonic Pi has been used in classrooms around the world to teach programming through music. Students explore sequencing, loops, data structures, and algorithms by coding beats, basslines, and shimmering synths. At the same time, professional musicians have taken Sonic Pi to stages and concert halls, using programming itself as an instrument of musical expression.

Now the story continues with Tau5, a new live coding system focused on collaboration. In music, concurrency, distribution, and fault tolerance aren’t just safety nets - they are uncompromising demands. When embraced, they become powerful creative tools for human expression and connection.

In this talk, Dr. Sam Aaron reflects on a year of building Tau5 using agentic AI, specifically Claude Code, as a primary collaborator in prototyping, architecture, and implementation. While AI is often seen as a tool for implementing already “solved” systems, this experience suggests a different possibility - one that challenges many of our assumptions about how software is made.

The audience will explore how AI can support deep systems programming, and why doing so requires rethinking our approaches to engineering, creativity, and education. Ultimately, this is a story about learning to work with AI - not asking what it can replace, but how it can improvise alongside us.

Generative AI inferencing has quickly become as indispensable as browser search, scaling usage across market sectors and audiences. But what do you do when online is not practical, desired, or secure for the data and content involved? Can these new AI PC’s power AI models offline and be functional alternatives to costly online services?

Jannick Jones from Intel, and Markus Schüttler from TNG will answer this question while demonstrating AI Playground, an open-source project from Intel in partnership and collaboration with TNG. This easy to install and operate tool looks and acts like your favorite online AI tool, but leverages over 60 open-source models to perform a variety of AI features like chat, vision, document search, tool calling, image generation, image editing, and video generation, all from a single prompt experience while offline.

The Scrum Guide is strict on the product owner being one person, which works well for a single team – but when you start adding multiple teams to the same project, that person quickly becomes a bottleneck. What if your AI coding tool could help you stay up to date with everything that's going on? In this talk, Jonathan Rohland will demonstrate how he combines deterministic scripts with AI processing to solve the following use cases:

- Staying up-to-date with all the work happening in a 40-person project
- Updating timelines in an ERP-integration project spanning 200 stories
- Turning MS Teams transcripts into human-reviewed meeting notes that stand out from the slop

The audience will learn how he turned Cursor into a workflow platform, what tooling he uses, and what it takes to build something similar yourself.

The uncomfortable part of coding agents is not that they write code. It is that they write plausible code. At Kilo Code, agents now open real pull requests, including OpenClaw instances running on KiloClaw while nobody watches. The summary sounds right. The tests pass. Still, a reviewer has to ask: Did the agent understand the task? In this talk, Marius Wichtner follows those PRs from creation to production: How to read beyond the summary, test what matters, and decide whether a plausible change deserves to ship.

It happens very often that developers are tasked with creating programs that can do more than one thing at a time. In recent years, "async/await" has become the main set of primitives provided by most mainstream programming languages to make this possible.

In this talk, Loris Cro will show how Zig recently re‑implemented its entire I/O stack from scratch in order to provide developers with an I/O implementation that decouples asynchrony from concurrency – giving Zig developers the ability to reuse code across different concurrency models in a way that has never been seen before.

Zig is a new low-level programming language that some consider the spiritual successor of C for its focus on simplicity and its toolchain that integrates seamlessly with C/C++ projects. No prior knowledge of Zig is necessary to follow this talk which will also feature code snippets from other mainstream programming languages. Anyone who uses concurrency in their day job and is curious about how these systems operate behind the scenes will find something of interest in this talk.

In this talk, Stephan Ilaender defines digital sovereignty not just as a security measure, but as a critical competitive advantage. By strictly avoiding vendor lock-in through open-source standards (e.g., Kubernetes, openDesk, openCode), STACKIT ensures that European companies remain agile and independent of non-European providers. This independence creates a unique market position: It protects against geopolitical risks - such as the U.S. Cloud Act or digital tariffs - while ensuring operational continuity and data supremacy. For customers, this means the freedom to migrate workloads and scale without being tied to proprietary ecosystems.

Within the last six months, one could observe the ease with which coding agents can create computer games. In this talk, Christoph Bergemann and Davin Schneider present their platform "Playspawn", which allows users to easily create browser games without any technical knowledge. They want to enable everybody to create their own games.

In Playspawn, casual game creators can generate their computer games using a simple chatbot. The authors will then feed their creation into a stream of games similar to other social media platforms. Christoph and Davin show how they built the complete platform without ever writing a single line of code. They highlight the game creation process, which is a dynamically started agentic system, featuring capabilities like image, audio, and 3D model creation. Then they show how these integrate into a fully functional social media platform. They also discuss issues such as deployment to the cloud or handling errors during the generation process.

The audience will also learn about the implications of unattended AI generation in coding on other aspects of the software development process, especially operations, quality assurance, and security.

Since reigning World Champion Garry Kasparov lost to the chess computer Deep Blue in 1997, computers have been regarded as stronger than humans in chess. Twenty years later, the world of chess engines was revolutionized once again by the first AI-based engine, AlphaZero, and today all leading chess programs make use of neural networks. How does technological progress affect the human understanding of chess? Has the playing style of top players noticeably changed since 2017? In this talk, Matthias Blübaum examines both the opportunities and risks of AI’s influence on chess and attempts to forecast how the game might continue to evolve in the years ahead.

The Rennteam Uni Stuttgart e.V. is among the most successful student motorsport teams worldwide, having achieved a total of eleven world championship titles. Following a concise overview of the team structure and its history, the presentation focuses on the hands-on development of an electric Formula Student race car. The technical deep dive covers system architecture, vehicle design, and the interdisciplinary integration of mechanics, electronics, and software. A key emphasis is placed on the driverless system, which combines advanced sensor technology, precise localization, and autonomous trajectory planning. The presentation highlights the successful collaboration between students and industry partners in the development of future-oriented technologies and their validation under real competition conditions.

Quantum computing becomes relevant where physical breakthroughs give rise to concrete technological opportunities. Using neutral atoms as an example, this talk explores how modern quantum hardware is evolving from a promising field of research into a platform of growing industrial relevance. It will focus on the technological foundations of this approach, advances in scalability, control, and error correction, and the question of which real-world problem classes may be addressed in the future. The talk will highlight opportunities in simulation, optimization, and materials research, as well as the prerequisites for meaningfully positioning quantum computing within existing innovation and technology landscapes. The result is a perspective that is both technical and application-oriented, showing how quantum hardware and quantum algorithms are gradually being translated into industrial applications.

Quantum computing promises substantial speedups for a narrow but important class of problems. Yet today’s discourse often oscillates between exaggerated expectations and unnecessary fear. In this talk Prof. Dr. Patrick Glauner provides a realistic roadmap of what quantum computers will - and will not - achieve over the next decade, highlighting practical constraints such as decoherence, error correction overhead, low levels of programming abstraction and various other often-overlooked challenges. Attendees will leave with a clear understanding of where quantum computing creates genuine value, where classical systems remain superior, and how organizations can prepare without falling for the hype.

Gravitational waves - tiny ripples in spacetime - open a new window into the universe. They are produced by violent cosmic events such as merging black holes and may even carry information from the earliest moments after the Big Bang. Following the discoveries by the ground-based detectors LIGO and Virgo, the next step is observation from space. The ESA mission LISA, planned for 2035, will detect gravitational waves in the millihertz range, inaccessible on Earth. Three spacecraft will form a laser interferometer with arms 2.5 million km long. Achieving this requires atomic-scale precision, perfect free fall, and advanced digital techniques such as time-delay interferometry. At the same time, thousands of overlapping signals must be separated and interpreted reliably. LISA combines precision engineering, data analysis, and fundamental physics to create a new way of exploring the cosmos.

While a single two-input NAND gate suffices for all of Boolean logic in digital hardware, no comparable primitive has been previously known for symbolic mathematics: Computing elementary functions such as sin, cos, sqrt, and log has always required multiple distinct operations. In this talk Dr. hab. Andrzej Odrzywołek will discuss the consequences of the discovery that a single binary operator, eml(x,y)=exp(x)-ln(y), together with the constant 1, generates the standard repertoire of a scientific calculator. In EML (Exp-Minus-Log) form, every such expression becomes a binary tree of identical nodes, yielding a grammar as simple as S -> 1 | eml(S,S). This uniform structure also enables gradient-based symbolic regression, using EML trees as trainable sub-circuits. He will also discuss related exhaustive search methods, the rationale behind them, and common misconceptions that surfaced after the result went viral on social media.

What do we know today about the origin of COVID-19? Was SARS-CoV-2 the result of a zoonotic spillover, the largest accident in human history – or even an intentional release?

This talk examines the current evidence regarding the emergence of the virus, with a particular focus on unusual genomic features and whether they align with research proposals that have since become public. From a bioengineer’s perspective, key data points will be placed in context, evidence will be distinguished from uncertainty, and open questions for further scientific and societal scrutiny will be discussed.

The current hype surrounding humanoid robots is enormous, but a look at reality reveals a different picture: most companies simply aren't ready for them yet. Yes, humanoid robots will eventually make their way into our production facilities, but this shift won't happen as quickly as many currently believe. That makes it all the more critical to separate vision from practice now, and lay the right technological foundations today for this next stage of automation.

How can you make robot learning accessible to everyone? In this talk, Steven Palma introduces LeRobot, a Hugging Face initiative designed to bring Artificial Intelligence into the physical world by making AI for robotics more accessible through end-to-end learning.

In this talk, he will take the audience on a comprehensive tour of the LeRobot platform, exploring its core pillars for open-source robotics:

- Robots & simulation: Utilizing affordable, real-world hardware like the SO-101 arm, alongside robust simulation environments

- Control & teleoperation: Streamlining data collection through unified robot teleoperation and control systems

- Datasets: Navigating the explosive growth of community-driven datasets stored on the Hugging Face Hub - a surge approaching 30,000 datasets that may signal the "ChatGPT moment of robotics"

- Models: Training with ready-to-use, state-of-the-art (SOTA) Vision-Language-Action (VLA) models in PyTorch, including Action Chunking with Transformers (ACT) and SmolVLA.

- Inference & evaluation: Testing and evaluating policies across various layouts and goals using environments like LIBERO.

Finally, Steven Palma will highlight the endless possibilities unlocked by the global community, such as multi-task co-training for zero-shot and few-shot motion transfer. He will also cover the latest advancements in the LeRobot v0.5.0 release, which supercharges open-source robot learning in every dimension.

Has this ever happened to you: You just spent €50,000 on a humanoid robot, but it's not very fun at parties, so you have to train it to dance using reinforcement learning in a simulation environment on your €100,000 GPU cluster? What do you mean "NO"?

In this talk, Thomas Endres and David Winderl will teach you how you can train your humanoid robot to dance using Reinforcement Learning (RL) in simulation environments. They're gonna kick things off with an introduction into RL, covering the basics of policies and rewards. After that, they’ll take their first steps in simulation environments with the introduction of NVIDIA Isaac Sim and the Unitree RL Lab. Our speakers will run through the steps of shaping a walking policy, one hilarious fall at a time. Finally, they're going to shake things up with a dancing policy: You will learn how you can retarget human motion from videos to the Unitree G1 robot and train an imitation policy that can dance just as poorly as you.

For the grande finale, our humanoid robot will close the session with a dancing performance, not missing out on a single step.

At last year's Big Techday, it was boldly announced on stage: “Next time, we’ll bring a fully DIY humanoid robot.”
Turns out … building one is way harder than ordering it on the internet.

In this talk, Dr. Florian Gather, Tobias Weyer and Fabian Lübbe rewind to that moment of bold overconfidence and fast-forward to the reality of 3D-printed gears, tangled cables, lost nerves, and finally a humanoid robot that (mostly) does what it’s told. Instead of spending 50k€ on a polished platform, they took the scenic route: Printing, assembling, breaking, fixing, and eventually training their own humanoid robot based on the Berkeley Humanoid Lite.

They’ll walk you through the full journey: From CAD explosions and 3D part mass production to first movements, many failures, and the glorious moment of a successful first biceps curl. Along the way, they dive into reinforcement learning for locomotion, explore how ROS helps to keep the chaos under control, and discuss what really makes humanoid robotics so challenging.

Of course, they won’t just talk about it - they’ll show it live on stage. Best case: It walks. Worst case: It becomes a very embarrassing modern art performance.

Do you wish you had an equal that can play Counter-Strike just as godlike as you? Thanks to recent development in Robot Foundation Models (RFMs), robots can learn how to play video games just like you, imitating your best and worst behavior.

In this talk, Thomas Endres and Jonas Mayer show you how you can use imitation learning to teach an RFM to play arbitrary games. They draw first blood with an introduction into Robot Foundation Models. You'll learn how they work, what kind of data they are designed to process and how that makes NVIDIA's GR00T perfect for playing video games. After that they quick scope the concept of imitation learning, what kind of data they need to train their GR00T and how they can collect it from games. Finally, they'll tell the story of leveling up their model from total noob to global elite - one failure at a time - and share all the things that they learned along the way.

Want to see for yourself just how well GR00T can play? Thomas and Jonas challenge you to beat their best models in the GR00T Gaming League in "Kesselhaus"!

Cognitive scientist Steven Pinker discusses the intellectual journey that led him from the study of language and visual cognition, through a synthesis of evolution and computation to explain the human mind, to his defense of the concept of human nature, culminating in his advocacy of the reality and future prospect of human progress.

Following his afternoon talk, chess grandmaster Matthias Blübaum will take on an exciting challenge in the evening: In a simultaneous chess exhibition, he will face 20 opponents at the same time.

Will the ambitious amateur players pose a threat to Matthias Blübaum? Or will the grandmaster follow in the footsteps of Vincent Keymer, who won all 26 of his games in a simultaneous exhibition at Big Techday 23?