About the Event
Join us for the next Claude Community Budapest Meetup, an evening about what happens beneath the prompt.
You might already build with Claude Code every day, run agentic workflows, or just wonder how these models actually think. Either way, this edition takes a step back from tools and tutorials and looks at the questions shaping where AI goes next: How do frontier models reason and revise their own hypotheses? Can AI learn to teach people, and to coordinate with other agents? And is scaling really the only way forward, or do we need to rethink the building blocks themselves?
Two researchers working at the edge of these questions, from Google DeepMind and Pázmány Péter Catholic University, will share their work. Then we keep the conversation going over drinks.
Who Should Attend?
The talks go deep, but the room is open to everyone who's curious:
- Engineers & ML Practitioners: Get a look under the hood at how reasoning models think, and why robustness matters as much as raw capability.
- Founders & Product Builders: Understand the reliability, cost, and agent-coordination challenges behind the products you're building on.
- Researchers & Students: Connect with people working on AI cognition, efficient compute, and alternative neural architectures.
- The AI-Curious: Skip the hype and hear directly from people doing the research. Bring your questions; the break and the networking are made for them.
Agenda
- 05:30 PM: Doors open. Registration, Networking & Drinks
- 06:00 PM: Welcome Remarks
- 06:10 PM: Botos Csaba, Senior Research Scientist at Google DeepMind (https://www.linkedin.com/in/botos-csaba/?isSelfProfile=false). Title: TBA
- 06:50 PM: Technical break
- 07:00 PM: Horvath Andras, Associate Professor at Pázmány Péter Catholic University, Faculty of Information Technology and Bionics (https://www.linkedin.com/in/andr%C3%A1s-horv%C3%A1th-35906a8b/). Title: Beyond Scaling: Neural Complexity, Adversarial Robustness, and the Future of AI
Speakers
Botos Csaba: The thread tying my work together is teaching, which I think is the overlooked problem in theory of mind: you cannot teach a mind without first modelling how it thinks. So I study how frontier reasoning models form, test, and revise hypotheses, and how their internal representations line up with human cognition, using controlled behavioural and neural experiments. I then ask how that understanding lets machines teach humans, by tracking what a learner knows and choosing the tasks that move them forward, and lets artificial agents teach and coordinate with one another. Bio: https://botcs.github.io/academic/
Talk abstract: The Grand Challenge of Theory of Mind: Teaching (https://botcs.github.io/academic/grand-challenge-of-tom/). Also presented as Reason to Play @ NeurIPS'26 Sydney (https://www.youtube.com/watch?v=8yEItbOE_7A).
Horvath Andras: My research focuses on making artificial intelligence more efficient by exploring the interplay between algorithms and the computing systems that implement them. I work on the development of efficient AI algorithms, particularly in machine learning and computer vision, while investigating how emerging technologies such as memristors and unconventional computing architectures can offer new approaches to information processing. Through international research projects, including DARPA-UPSIDE and EU HORIZON project PHASTRAC, I have contributed to exploring alternative computing paradigms that bring algorithm design closer to the underlying hardware. Alongside my research and academic work at Pázmány Péter Catholic University, I have served as chair of an IEEE CASS community focused on memristor array computing, contributing to the development of this emerging research field.
Talk abstract: Artificial intelligence has made remarkable progress, largely driven by increasingly large models, more data, and greater computational resources. However, important challenges remain, including adversarial vulnerability, reliability, and the growing complexity and cost of neural networks. These limitations raise a fundamental question: can we improve the capabilities and robustness of neural networks not only by scaling their size, but also by rethinking their fundamental building blocks?
In this talk, I will discuss the current state of AI and some of its key limitations, with a particular focus on adversarial robustness. I will then present research exploring alternative approaches to neural network design, in which increased computational and functional complexity is introduced at the level of individual neurons rather than solely through larger network architectures.
Important Information
- Registration: Claude Code Community reserves the right to approve or reject any registration. Please ensure your information is accurate, as incomplete or fake profiles will be declined.
- Capacity: To ensure everyone's safety, we reserve the right to deny entry at the door if the venue reaches its maximum capacity.
Thank you and feel free to contact me, my DMs are always open! Tamas
