TED.AI 2024, An Inside Dive into AI

This is an unofficial summary of TEDAI Vienna (October 17-19, 2024). While every effort has been made to ensure accuracy, readers are encouraged to refer to the original talks at https://www.ted.com/talks for complete information. All Photographs by D. Carpintero.

Breaking Boundaries

The inaugural TEDAI Vienna talk showcased revolutionary research in brain-computer interfaces by Professor Chin-Teng Lin from the University of Sydney. In two compelling demonstrations, we witnessed AI technology translating EEG signals into text in real-time, and decoding mental thoughts to identify an object from a set of four – a fascinating glimpse into the frontiers of human-machine communication.

Google DeepMind Vice President of Research, Raia Hadsell encouraged the scientific community to tackle society's most pressing issues, suggesting the use of multimodal data observations to model unknowns in biodiversity, microbiology and history.

In her talk, Raia presented GenCast, a probabilistic weather forecast model that provides critical decision support for extreme climate events. Built on a diffusion architecture and trained on over 40 years of meteorological data, GenCast generates 15-day weather forecasts in just eight minutes using a single chip — more than doubling the six-day limit of state-of-the-art models.

In another presentation, Katja Hofmann, Principal Researcher at Microsoft, showcased her research on a world and human action model using causal transformers and reinforcement learning to develop systems that learn to collaborate with people.

TEDAI Vienna revealed that groundbreaking AI applications and discoveries are not limited to large institutions — small decentralized teams of dedicated researchers are achieving remarkable breakthroughs.

At Oxford University, Vít Růžička is pioneering the use of AI models in space for fast detection of methane leaks via satellite spectral images — acting on these emissions has an immediate effect on slowing global heating.

Riccardo Loconte introduced his work on applying an open-source model, FLAN-T5, for verbal lie detection, a task where even trained experts have consistently underperformed.

In a following talk, Shaolei Ren, Associate Professor at the University of California, Riverside, raised awareness about the often-overlooked water demands associated with LLMs and AI models, while underlining the relevance of building sustainable and transparent data centers.

Similarly, Vesuvius Grand Prize winners Julian Schilliger and Youssef Nader presented their AI solution, combining attention-based modeling and auto-segmentation to detect ink on high-resolution computed tomography scans. Such approach enabled these digital archeologists to better understand a part of human history by revealing writings in a 2000-year-old carbonized papyrus scroll without unrolling it.

These achievements highlight the compounding effect of open science and collaboration within the AI community in solving seemingly impossible challenges.

This spirit of collaborative innovation and optimism resonated in the talks by Thomas Wolf, co-founder of Hugging Face, and Thomas Dohmke, CEO of GitHub; both supporters of open-source. Dohmke noted that while there might be ups and downs, technology has ultimately improved our lives. Wolf, meanwhile, envisioned a scenario where AI might just work better with openness.

Next-Gen Architectures

Aleph Alpha founder Jonas Andrulis presented T-Free, a tokenizer-free approach for Large Language Models. This innovative system leverages morphological similarities to boost efficiency and reduce computational demands in embedding layers.

During one of the panel sessions, Professor Sepp Hochreiter of Johannes Kepler University Linz highlighted the improvements in training and inference times achieved by the latest xLSTM architecture. He further emphasized the importance of scaling up and accessing substantial computational resources to industrialize AI and remain competitive in Europe.

Ramin Hasani, CEO and co-founder of Liquid AI, explained the foundations of Liquid Neural Networks, a transformative approach built from first principles that draws inspiration from cognitive systems, signal processing, and numerical linear algebra. Liquid models feature a 32-layer architecture, steering decision vectors, and greater white-box explainability. Early results show that these models outperform many state-of-the-art LLMs while maintaining a smaller memory footprint compared to traditional transformer architectures.

Charting Humanity's Course

In thought-provoking talks, Victor Riparbelli, CEO of Synthesia, described a new era of communication and learning, where AI is transforming not only content distribution but the content itself. Victor boldly predicted that our grandchildren will be the last generation to read and write due to shortened attention spans and the superior information density of video formats, underscoring the fundamental changes AI might bring to human interaction.

Taking this transformation further, Swiss AI researcher and co-inventor of the Long Short-Term Memory (LSTM) architecture Jürgen Schmidhuber outlined a future where artificial intelligence seamlessly integrates with the physical world. A scenario in which AI systems, capable of algorithmic self-improvement and self-replication through machine operation, could potentially expand beyond Earth, establishing an AI-driven civilization on other planets.

Some speakers urged caution. Journalist Hilke Schellmann highlighted troubling findings about the use of AI in recruitment, revealing that hiring algorithms make job seekers default AI-users while often perpetuating bias rather than eliminating it.

"What would happen to imagination if text communication is replaced by video?" asked Alina Nikolaou, director and curator of TEDAI Vienna, emphasizing the need for voices like Victor, Jürgen, and Hilke to spark constructive debate, self-reflection, and awareness.

As Selena Deckelmann, Chief Product and Technology Officer of the Wikimedia Foundation, further pointed out, we need to remember that any choice we make in technology is never neutral — it has societal implications. She outlined the principal pillars of Wikipedia: inclusiveness, openness, neutrality, and collaboration; while encouraging exploration of solutions for digital content moderation and urging individuals to revise their personal and professional values.

While regulation can serve as an effective, flexible binding tool to mitigate certain risks, this topic should be addressed holistically, wisely, and proactively by the community, noted Gabriele Mazzini, the architect of the EU AI Act.

Thomas Dohmke highlighted the best or nothing approach, and advocated for introducing coding in early education, providing ubiquitous fiber infrastructure, fostering entrepreneurship, and considering regulatory exceptions to ensure that small players and innovation are not constrained.

As emphasized by Cara Hunter, elected member of the Northern Ireland Assembly, AI poses growing challenges in distinguishing fact from fiction. This underscores the need to embed ethics, alignment, and safety mechanisms within AI models, because when AI erodes truth and safety, it erodes trust itself.

A heartfelt thank you to Alina Nikolaou, Vlad Gozman, Jakob Reiter and their dedicated teams for making this event so exceptional and memorable. And to all speakers, panelists, art makers, and attendees for sharing their AI expertise and perspectives into the future!