Envisioning

🐘

Imagine a group of blind people encountering an elephant.

Each person reaches a different part of the animal and describe what they are feeling. The one holding the leg senses something like a tree. The one grabbing the trunk feels like something like a snake, and so on.

Depending on who you ask, you will receive a different response about what they are witnessing. And despite their distinct witnesses, they are each correct in their individual assessment of the situation.

How do you know who to listen to?

This metaphor helps illustrate what it is like trying to wrap your head around AI.

Developing at the intersection of numerous fields, AI is the culmination of computer science, mathematics and statistics, but also linguistics, ethics, psychology and many more areas of research. Countless experts are collaborating on furthering AI, and we can safely claim that nobody understands the whole thing.

I had the privilege of joining a conference which brought together people from as many such fields as possible. Over 48 hours in Vienna I experienced an information download like nothing else at the TEDAI Vienna conference. I absorbed myriad conversations, presentations and workshops and immersed myself between people making it happen. Scientists, developers, policy-makers, entrepreneurs, artists, enthusiasts and just about everyone who cares about & works with the the state of AI and who could afford this three day deep dive.

While it is impossible to share the full extend of the experience, what stands out is simply how much is happening. Some might question how far the advances in AI might take us, and there are many ways of invalidating the hype. Yet there is no question of how this moment represents a shift like nothing we have experienced. Research and investment in AI could stop tomorrow, and there would still be decades of groundbreaking work to be released. The self-reinforcing loop will only keep tightening and bring us closer to a technological singularity.

I used to have my doubts. Despite being deeply into science-fiction visions of the future, and having followed emerging technology professionally nearly two decades, I had my questions. The promise of AI seemed too good to be true, and the many AI winters had me believing that this would all blow over. That is not going to happen.

In the words of Stewart Brand; we are as gods and might as well get good at it.

Instead of our typical newsletter format, this week I have compiled my notes from the talks and selected previous videos from the speakers, to help you get a sense for what went on.

We got to:

🧠 Witness a computer read brainwaves in real time. A student of Prof. Chin-Teng Lin used LLMs trained on biomarkers to achieve EEG-to-text, or in other words: silently vocalizing and thinking about a sequence of words from a set list of sentences and have the computer identify which phrase was thought.
🤝 Learn the the importance of radical inclusivity when gathering data from your users and how different technologies are either controlled by us, or controls us by Selena Deckelmann, CPO at the Wikimedia foundation.
🕹️ How World and Human Action Models (WHAM) can be generated using transformer models in order to simulate gameplay and create new modes of human-machine collaboration (Dr. Katja Hoffman, Microsoft senior AI researcher).
📜 How two PhD students and digital archeologists managed to unwrap and read papyrus scrolls which had been previously lost to the eruption of Vesuvius nearly 2.000 years ago. They employed segmentation and optical flow techniques to non-destructively reconstruct ink and writing from a CT-scan of the carbonized scrolls. Incredible. Youssef Nader & Julian Schilliger.
🧑‍🤝‍🧑 Why our grandchildren might be the last generation to read and write, if generative immersive media forms keep developing through AI. Victor Riparbelli from Synthesia.ai made a compelling argument for how our fragmented attention span and surge in large language & diffusion models might lead to a near future where we all interface primarily with avatars.
🇪🇺 How the EU could get ahead of the rest of the world in AI development by embracing techno-optimism from GitHub CEO Thomas Dohmke. We need a revolution in mindset to become unstuck from our collective economic & intellectual rut. Teaching code, facilitating startups and investing in infrastructure should be our top priority. Thomas asked us to keep an eye on the GitHub Universe conference next week for insights into GitHub’s upcoming centaur tools.
🌐 How to be more inclusive of low-resource languages and alphabets from outlier cultures using a novel approach for tokenization by Aleph Alpha CEO Jonas Andrulis. Their solution (apparently called TFree) solves out-of distribution behaviors using trigrams and multi-class activation. Make of that what you will! This should lead to more energy efficient models and better language adoption of AI.
🤗 HuggingFace CEO Thomas Wolf on how to transcend ourselves by embracing the possibility of AI “just working” for us. What if it all goes right? Who should intermediate our extended intelligences, and what happens if these solutions are not resilient? If AGI brings us god-like powers, how do we reduce the risk of relying on deprecated tools?
🎤 An actual rap battle between comedian Chris Turner and ChatGPT with surprising and remarkable results.
🪦 The risk of misalignment between humanity and our robotic digital successors. How do we instill human values in our AI, and how do we decide which traits to pass on? Philosopher Andrea Lavazza from Pegaso University recommends an interdisciplinary dialogue to discuss the ethics of new technologies to avoid the end of the world.
🧑‍⚖️ The challenge of governing AI development from EU AI Act co-creator Gabriele Mazzini. Regulation should not be opposed to innovation, and in order to unlock our personal flourishing with AI might come with significant risks if we don’t design our technology carefully.
🔍 The terrible experience of having deepfakes made of yourself. Irish Social Democratic and Labour Party politician Cara Hunter shared a heartbreaking story asking for responsibility and regulation in how AI is deployed.
💧 Computer scientist Shaiolei Ren shared research on the water consumption of AI training and inference, highlighting that 500ml of water is wasted for every 10-50 LLM queries. The unevenly distributed effects of this water footprint can affect vulnerable populations disproportionately.
🛰️ AI training in space - Vít Růžička from University of Oxford shared how ML is being used on satellites to reduce the time spent between identifying methane leaks and notifying the responsible. Reduced methane emissions are crucial for containing the climate emergency and AI can help us avoid paralysis.
👩‍💻 We need a code of conduct for code. NVIDIA VP or Enterprise and Automation Rama Akkiraju shared her experience with deploying AI in the enterprise in order to solve real-world problems. Rama proposed a framework for building your own (FACTS):
- Freshness (RAG & vectorization)
- Architecture (building flexible systems)
- Cost Economics (determining ROI and being strategic about model use)
- Testing (how to evaluate)
- Security (compliance & security)
🕵️‍♀️ Investigative journalist Hilke Schellmann evaluated the use of screening tools in hiring processes, and how deeply biased some of the AI-assisted HR hiring processes can be. I especially liked the idea of affective computing, or developing systems that are cognizant of the users emotions, especially when it comes to life altering moments like work and unemployment.
✨ Poetic AI: artist Ferdi Alici and violinist Phoebe Violet brought together data and art in a beautiful and unique performance by translating billions of data points into an immersive real-time data sculpture. Everything you want, you already are.
🌀 Google DeepMind VP of Research Raia Hadsell went deep into how AI is leading to countless scientific breakthroughs and how we are barely scratching the surface of what will become possible at the current rate of change. From longer and more accurate hurricane and weather predictions to modeling the entire earth as a digital twin ecosystem in order to better understand the effects we are causing the planet.
🤥 How everyone lies, and whether AI can be used to detect it. Psychology and computer science researcher Riccardo Loconte showed how AI can be trained to help us spot fake news and other forms of deception. Understanding how our models make decisions (explainability) might even help us understand ourselves better.
🐙 Liquid Neural Networks (LLNs) as an alternative approach for making our AI models more efficient and performant by emulating biological brain structures. There might be countless potential architectures for evolving AI and Ramin Hasani of Liquid AI made a compelling case for how their research at MIT CSAIL has led to novel techniques like liquid time-constant networks. Give it a try on playground.liquid.ai
🦾 The science of intelligence. Professor Oliver Brock helped us understand their research at TU Berlin for identifying the principles of intelligence which could lead to AGI. For example, instead of pre-programming every potential failure state, what if we could build robots which interact with the world by solving problems like biological intelligences? Think 7 parameters instead of 7 billion. How do simple rules lead to complex behaviors? And would that solve alignment?
🧑‍🚀 We have seen nothing yet. Scientific Director at Swiss AI Lab and personal hero of mine Jürgen Schmidhuber is sometimess called the “father of AI” and shared a unique and deeply personal vision of the future. Schmidhuber has been a pioneer in the field since the 1990s and helped lay the foundations for GPTs decades before anyone believed such things would be possible, and co-developed the most cited AI research paper of the 20th century on Long Short Term Memory (LSTMs). According to him we will see autonomous AIs transitioning from the virtual world to the real world through meta-learning systems employing agency and self-improving hardware. He predicts we will see AIs in the physical world starting around 2029 and that it will inevitably evolve towards expanding into outer space, long beyond our planet. Captivating and memorable.

Meetup Today

We are hosting the first Artificial Insights online meetup later today. Meet other readers and learn about the community is working on. We’ll do a round of introductions and share around two topics: your views of AGI and personal projects you have been exploring around AI. Join us!