Ilya’s List, Part 14: Pointer Networks
I know for a fact that I studied this one early on, I thought I wrote a blog post about it but I guess not. Pointer networks are a pretty cool idea. You put in a string of words. The output is just the position of one of the words. For example: Is Ridley a cat or a dog? If we put this into a pointer network, it would return 3, because in position 3 of our sentence above is the word Cat, and Ridley is a cat. This is kind of a cool way of doing word processing, because the results can ONLY be words from the input. While there is still a chance of hallucination here, that hallucination is limited because we can only use words for the output that we used on the input. For example, if we ask whether Ridley is a cat or a dog, the model can’t respond with “Star Wars.” Here is kind a how I see this in my head: I also think of it kinda like a Ouija board, because you can only select the limited words that are already there. The AI model is like the spirit of a dead ancestor, guiding the selector to the right word. Not a whole heck of a lot to see here. On to the next! Ilya’s List, Part 14: Pointer Networks was originally published in Ilya Sutskever’s 30 Foundational Papers of AI, Explained
Ilya’s 30 Foundational Papers, Part 5: GPipe
Hello all. Apparently I missed number 5 and 14 in my quest to understand all that is Ilya Sutskever’s AI knowledge foundation. So anyways, let’s talk about the GPipe paper. I think the best way to explain the paper is using animation. Normally, how you train an AI is, you get your neural network, you load it onto a GPU, and you run some data through it. If the results are bad, then you backpropagate and change the weights and biases so that the results might be better next time. But AI models are getting bigger and bigger. Kimi K3 just came out and it is 2.8 trillion parameters, each one in 16 bits of precision. That is like… a lot of terabytes. And as of right now, even the best GPUs in the world can only hold around 100 gigabytes. So we are going to have to split our huge model up, so that the first few layers are on one GPU, the next few layers are on another GPU, etc. until we have the whole thing loaded onto GPUs. Why do we do this? Why are we loading the model onto GPUs? Because a GPU can train an AI model 60 times faster than a regular old CPU. Instead of taking hours, training can take minutes. Instead of taking 2 months, training can take 1 day. However, today’s models are
Ilya’s List, Number 20: Scaling Laws for Neural Language Models
This paper starts out not very surprising. More data makes language models better. More training makes language models better. More parameters make language models better. Absolutely shocking. Please remain seated. The paper is Scaling Laws for Neural Language Models , and it is one of the papers from Ilya Sutskever’s famous reading list. The big idea is that language model performance does not improve randomly. It follows predictable mathematical patterns. That is the important part. If model performance is predictable, then training a giant AI model becomes less like wizardry and more like engineering. Still expensive engineering, but engineering. Loss, or: How Embarrassed Is the Model? The paper talks about cross entropy loss. Very roughly, this measures how surprised the model is when it sees the correct next token. If the sentence is: The cat sat on the ___ And the correct next token is “mat,” a good model assigns “mat” a high probability. A bad model confidently guesses “spreadsheet,” “mitochondria,” or “Belgium.” The lower the loss, the less embarrassed the model should be. For language models, loss is related to perplexity: Perplexity = e^loss So a loss of 5 means the model
Ilya’s 30u30, Part 19: DeepSpeech2
Hello, all you Sutskevites! We are back with Part 19 of my “Ilya’s 30u30” series, which means I only have one more article to go before I hit my goal for the year. This is huge, because it proves that if you set a realistic goal, work consistently, and publicly embarrass yourself, you too can almost accomplish something by August. Today’s paper is DeepSpeech2 , from 2015. And I have to say, this one is either pretty simple, or I am missing something obvious and one of you is going to politely destroy me in the comments. The basic idea is this: Back then, speech recognition still used a lot of hand-made stuff. People would take audio, study it, and write programs that looked for little clues in the sound. Like, “This pattern probably means this person is making a B sound,” or “This curve probably means an O sound.” That is very smart, but it is also a lot of work. You are basically trying to teach the computer how to listen by writing a giant rulebook. DeepSpeech2 said: what if we do not write the giant rulebook? What if we just let the AI learn from the sound directly? Sound Can Become a Picture One thing I have always liked about speech recognition is that sound can be turned into
Ilya’s 30u30, part 18: The Return of Complexodynamics
Ok now this is an interesting one. We talked earlier about how I think Complexodynamics is complete bullhockey. The authors give an example of how if you have a glass half full of coffee, and you pour some milk into it, that is not super interesting at first because half of the glass is coffee and half is milk. Nothing interesting there. But after a few seconds the coffee and milk start to mix and these beautiful swirling patterns form. The authors claim that this is more interesting to watch than just regular coffee. They also claim that after like an hour or so the coffee and milk mix into a uniform lighter brown liquid, which is less interesting. So according to the authors, systems sometimes start out not very interesting, then become more interesting, then become boring again. They say the universe is like this: Not interesting before the Big Bang, now it’s interesting, and it will be less interesting when the universe suffers heat death or collapses in on itself. The authors separate Complexodynamics from Entropy. Entropy states that the natural order of the universe is from organization to chaos. I think about this like my house: it takes zero effort to make it messy, clothe
Ilya’s List, Part 17: Relational Recurrent Neural Networks: What If an RNN Had More Than One…
Ilya’s List, Part 17: Relational Recurrent Neural Networks: What If an RNN Had More Than One Thought at a Time? A normal recurrent neural network is pretty easy to describe. It takes in two things: 1. A token vector 2. A state vector The token vector represents the current word, or sometimes the previous word the model predicted. The state vector represents what the model “knows so far.” Then the RNN produces two things: 1. A prediction for the next token 2. An updated state vector Then we feed the next token and the updated state vector back into the RNN again. That is why it is called recurrent . The model keeps looping its own memory back into itself, like a snake eating its own tail, but with more matrix multiplication and fewer mythology majors. At a high level, it looks like this: current token + old state -> next token prediction + new state The state vector is doing a lot of work. It is basically saying: “This is what we are saying, what we have said before, what I think matters, and everything I currently know about this sequence.” That is impressive. It is also a little insane. Because all of that information is mashed into one single vector. The Problem With One Big S
Ilya’s 30u30, Issue 16: Variational Lossy Autoencoders and Art Forgery
Imagine two art forgers working in a dusty loft. The first forger has one job: look at the original painting and write a description. The second forger has one job: take that description and paint a copy. One day, the first forger walks over to a famous painting, studies it for about four seconds, writes two words on a slip of paper, and hands it to the second forger. The note says: Mona Lisa. Then the first forger goes to lunch. The second forger stares at the note. Then he stares at the empty canvas. Then he stares at the door where the first forger has already disappeared, probably to buy a sandwich with money he did not earn. But the second forger is good. Annoyingly good. He has studied thousands of paintings. He knows how faces are shaded. He knows how brush strokes flow. He knows how backgrounds fade into distance. He knows how eyes, hands, fabric, shadows, and little color transitions usually work. So he grumbles, picks up his brush, and starts painting. And somehow, he finishes a pretty convincing Mona Lisa. When the first forger comes back from lunch, the second forger points at the finished painting and says: “This is ridiculous. You cannot just write ‘Mona Lisa’ and lea
Ilya’s Papers, Part 13: Resnets, AGAIN?
A few posts back I wrote about ResNets, or Residual Neural Networks. The short summary is, for each layer of a neural network, you have your input and your output. ResNets do something clever: they send the input through that layer to calculate the output. And then, they add the input to the output. So it’s like this: input (5) → layer → output (20) input (5) → layer → output (20) + input (5) input (5) → layer → output (20) + input (5) → 25 Ok now why on earth would we do this? We just transformed the input into the output, why would we be adding it back in again? Well, because it really helps to stabilize training. For a long time, overfitting was a huge problem (see the internet for more information on overfitting) and we couldn’t train neural networks with thousands of layers. That was a big limitation, because more layers means more fun. Residual networks made it so that we finally can train networksi with thousands of layers. So the paper in question, “Identity Mappings in Deep Residual Networks” talks about why this works so well. In the original ResNet paper, researchers would add the input of a layer back to the output of the layer, and then run that sum through an activati
Ilya’s List, Episode XV: Comparing Things
A couple of blog posts ago, I complained that Graph Neural Networks are poorly named. Then, with the confidence of a man walking directly into a glass door, I announced that I would call them Relationship Neural Networks instead. I regret that now. Because I recently discovered that there is already a real thing called Relation Networks . AI terminology, everybody. A beautiful field where every name is either too vague, already taken, or both. Anyway, today’s topic is Relation Networks, and they are one of those AI ideas that feel almost suspiciously simple. Like, “Wait, that’s it?” simple. But also surprisingly profound. At their core, Relation Networks help AI systems do something humans do constantly: Compare things. And more specifically, compare things in relation to a question . The Problem: Comparing Things Is Harder Than It Looks Imagine we have three sentences: Jack is 2 meters tall. Brad is 1.5 meters tall. Betsy is 1.75 meters tall. Now we ask: Who is taller, Brad or Betsy? For a modern language model, this is easy. GPT-5 is not exactly sweating over “1.75 is bigger than 1.5.” This is not where the Nobel Prize gets handed out. But in the dark ages of 2017, models were mu
Ilya’s Papers, Part 13: The Dawn of Attention
Ilya’s Papers, Part 12: The Dawn of Attention It was a quieter time, a simpler time. Dinosaurs roamed the earth. “Uptown Funk” by Bruno Mars was at the top of the charts, and no one had heard of ChatGPT. I am referring, of course, to the ancient year of 2015, which is basically prehistory in AI. Transformers did not exist yet. “Attention Is All You Need” was still in the future. But even before transformers, researchers had already started exploring an important idea: attention. I have written about RNNs before, and this post will make a lot more sense if you already know the basics. Still, I’ll try to explain this one as painlessly as possible, because RNN papers have a special talent for making people sad. The basic problem Imagine you want an AI system to translate a sentence from one language into another. Back then, a common approach was this: • One RNN reads the whole sentence in the source language • It compresses everything it learned into one final hidden state vector • A second RNN uses that vector to generate the translated sentence In simple terms, the first RNN reads the sentence and hands the second RNN a single bundle of notes. That sounds fine… until the sentence ge
Many Tongues, One Intelligence: Building the AI Commons for Language Diversity | AI House Davos 2026
How can we ensure that the next generation of AI systems reflects the world's linguistic and cultural diversity rather than narrowing it? What infrastructure is needed to preserve and amplify language diversity in the age of AI? This interactive evening session explores how shared AI infrastructures, from open data to equitable compute access, can empower linguistic diversity and prevent AI systems from homogenizing global languages. Through short expert inputs and open networking, participants exchange ideas on inclusive AI innovation for the common good, examining how multilingual AI, inclusive datasets, and accessible compute can create an AI Commons that serves all languages and cultures, not just dominant ones. Speakers Annette Oxenius (ETH Zurich, Vice President Research and Member of Executive Board) Stéphanie Lacour (EPFL, Vice President Support to Strategic Initiatives) Antoine Bosselut (Swiss AI, Co-Lead of Apertus Project) Martin Tisné (Current AI, Founder and Chair) Alexandra Baumann (Swiss Federal Department for Foreign Affairs, Ambassador, Head of Division for Prosperity & Sustainability) Moderator Nina Frey (ICAIN, ETH Zurich, Executive Director) (c) AI House Davos 2
The Long Horizon: A Swiss Morning on the Future of Intelligence | AI House Davos 2026
After a week of conversation at the AI House under the theme A Human Intelligence Shift, Friday turns our attention to The Long Horizon. The Swiss Morning sets the stage for looking beyond immediate breakthroughs to ask how intelligence may evolve, whether it will converge across cultures, fragment into competing systems, or diversify in new and surprising ways. Building on Switzerland's neutrality and strong innovation heritage, the day connects global debates to local realities, exploring how a small, future-focused nation can translate worldwide conversations into meaningful action at home while offering a steady voice to the international community. Speakers Bernhard Kratzwald (EthonAI, Co-Founder and CTO) Davide Scaramuzza (University of Zurich, Professor of Robotics and Perception) Harald Kröll (Chipmind AG, Founder and CEO) Nadja Christoffel (Prättigau–Davos, Head of Regional Development) Philipp Hölzenbein (Raven, Co-Founder) Rolf Pfister (Lab42, Co-Founder & Director of Research) Pascal Kaufmann (Mindfire, Founder and CEO) Tilman Eberle (Davos Tech Summit, GM) Moderators Florian Herzog (FHGR, Director of Studies AI in Software Engineering) Rebecca Brauchli (Zurich Universi
From Co-Pilots to Agentic Autonomous Economies | AI House Davos 2026
The emergence of agent economies is inevitable. The rapid rise of autonomous AI agents is leading to a new economic layer of agent-agent interactions beyond human pace. These agents are beginning to act not just as tools, but as autonomous participants, transacting, negotiating, and coordinating in digital environments at scale. This shift brings immense potential for efficiency and innovation, but also serious risks and questions around fairness, control, and societal impact. This panel convenes AI researchers, economic game designers, policymakers, and ethicists to explore how we can proactively shape these emerging economies and build the critical infrastructure, identity, and oversight that will foster trust and coordination in these high-speed ecosystems. Speakers David Evan Harris (UC Berkeley, Chancellor's Public Scholar) Gordon Liao (Circle Internet Group, Chief Economist and Head of Research) James Landay (Stanford HAI, Co-Director) Rod Beckstrom (Investor, Author, Tech CEO, Geopolitician) Moderator Kenneth Cukier (The Economist, Deputy Executive Editor) (c) AI House Davos 2026 Founders & Strategic Partners: ETH AI Center, Merantix, G42, Hewlett Packard Enterprise, EPFL AI
Sustainable AI: From Climate Science to Efficient Computing | AI House Davos 2026
As climate pressures intensify, AI is emerging as a critical tool for understanding risks, strengthening adaptation, and guiding smarter decisions. This session explores how new frugal AI models are improving climate efficiency, how advanced optimization tools can enhance energy systems and industrial processes, and how next-generation forecasting is reshaping resilience planning. The discussion highlights where AI delivers real environmental impact today and what breakthroughs are needed to scale sustainable solutions for tomorrow. Speakers Eric Enselme (Independent Board Advisor) Varun Sivaram (Emerald AI, Founder and CEO) Himanshu Gupta (ClimateAi, Co-Founder and CEO) Moderator Ghjulia Sialelli (ETH AI Center, PhD Fellow) (c) AI House Davos 2026 Founders & Strategic Partners: ETH AI Center, Merantix, G42, Hewlett Packard Enterprise, EPFL AI Center, The University of Tokyo Presenting Partner: KPMG
Unprecedented Scale – Building Startups in the Age of AI | AI House Davos 2026
AI is reshaping the startup playbook, with vibe coding and instant access to powerful models lowering the barrier to build. This panel brings founders, investors, and tech leaders together to explore how speed, data, and automation now drive advantage for AI-native companies and how unicorn-style growth is transforming markets. The discussion examines shifting economic dynamics across industries, venture, government, and ecosystem partners, how investors identify real potential, and what it takes to stand out when anyone can ship fast. Panelists share strategies for scaling responsibly, competing globally, and building impactful, sustainable AI-first organizations in this new era of company creation. Speakers Alex Ilic (ETH AI Center, Co-Founder and Executive Director) Andrew Ng (DeepLearning.AI, Founder) Andy Hock (Cerebras Systems, Chief Strategy Officer) Laura Modiano (OpenAI, Head of Startups EMEA) Moderator Nicole Büttner (Merantix Momentum, Founder and CEO) (c) AI House Davos 2026 Founders & Strategic Partners: ETH AI Center, Merantix, G42, Hewlett Packard Enterprise, EPFL AI Center, The University of Tokyo Presenting Partner: KPMG
The Fabric of Society: AI in Critical Infrastructure | AI House Davos 2026
AI is reshaping critical infrastructure, from data centers and power grids to satellite and communication networks. Rising compute demand, tighter reliability requirements, and deeper automation are forcing rapid changes in how these systems are designed and managed. With atomic energy remaining a vital source of stable baseload power, its role grows even more important as AI-driven systems increase overall electricity demand and dependence on reliable generation. This panel examines how AI can improve efficiency and operational resilience, while also creating new vulnerabilities, cyber risks, and capacity challenges. Speakers Francesco Sciortino (Proxima Fusion, Co-Founder and CEO) Frederic Werner (ITU, Chief Strategic Engagement Division, AI for Good) Rafael Mariano Grossi (International Atomic Energy Agency, Director General) Moderator Vijay V. Vaitheeswaran (The Economist, Global Energy and Climate Innovation Editor) (c) AI House Davos 2026 Founders & Strategic Partners: ETH AI Center, Merantix, G42, Hewlett Packard Enterprise, EPFL AI Center, The University of Tokyo Presenting Partner: KPMG
Shadow AI: The Hidden Layer of Intelligence | AI House Davos 2026
This panel explores "Shadow AI" models, tools, and autonomous agents operating outside formal oversight. As AI proliferates beyond institutional control, risks around security, governance, and accountability intensify. The discussion examines how organizations and policymakers can detect, audit, and manage hidden AI systems, and how to balance open innovation with the systemic risks posed by unmonitored or unauthorized intelligent agents. Speakers Menna El-Assady (ETH Zurich and Faculty, ETH AI Center, Professor) Navrina Singh (Credo AI, Founder and CEO) Petar Tsankov (LatticeFlow AI, Co-Founder and CEO) Moderator Akhilesh Tuteja (KPMG, Global Leader, Cyber Security) (c) AI House Davos 2026 Founders & Strategic Partners: ETH AI Center, Merantix, G42, Hewlett Packard Enterprise, EPFL AI Center, The University of Tokyo Presenting Partners: KPMG
Sovereignty vs. Collaboration: Who Owns the Future of Intelligence? | AI House Davos 2026
In a world where data is currency and AI drives progress, sovereignty is no longer just a buzzword. It's a battleground. Who truly controls the future of intelligence: nations, corporations, or the networks that span them? This session challenges the assumption that independence and global collaboration are mutually exclusive. Is the quest for sovereign AI a path to resilience, or a recipe for isolation and stagnation? We confront the hard questions: Can secure, compliant, air-gapped data coexist with the open exchange that fuels innovation? Does sovereignty demand walls, or can it thrive on bridges? This panel debates whether sovereignty is a shield or a cell and explores how bold, pragmatic strategies can reconcile autonomy with the undeniable benefits of global collaboration. Speakers Denise Wong (IMDA, Assistant Chief Executive) Juan Ramón Troncoso-Pastoriza (Tune Insight, Co-Founder and CEO) Talal Al Kaissi (G42 & Core42, Group Chief Global Affairs Officer & Interim CEO) Thierry Pienaar (HPE, Chief Tech Officer, HPC & AI WW) Moderator Irena Bednarich (Vice President, International Government Relations and Strategic Partnerhips, HPE) (c) AI House Davos 2026 Founders & Strategic
Inclusive Intelligence: Empowering Societies to Shape AI | AI House Davos 2026
How can societies actively shape the future of AI through inclusion, literacy, and broad civic participation? What mechanisms enable informed communities to influence AI governance? This panel explores how societies can move from passive recipients to active shapers of AI futures. The discussion examines strategies to strengthen public engagement in AI policymaking, close knowledge gaps through accessible education programs, and build tools that enable transparent, participatory governance. Speakers share insights on how informed communities can influence the development, deployment, and oversight of AI systems that impact daily life and democratic decision-making—from university-led education initiatives to civic assemblies and cross-border research collaboration. Speakers Baroness Joanna Shields (Responsible AI Future Foundation, Executive Chair) Basma AlBuhairan (C4IR Saudi Arabia, Managing Director) Joël Mesot (ETH Zürich, President) Thomas Hofmann (Technical University of Munich, President) Moderator Spriha Srivastava (CNBC, VP & International Executive Editor of Digital) (c) AI House Davos 2026 Founders & Strategic Partners: ETH AI Center, Merantix, G42, Hewlett Packard Enter
AGI Night | AI House Davos 2026
An evening to rethink what AGI truly means. Rather than seeing intelligence as stored knowledge or fixed patterns, we explore it as a dynamic, evolving process, one that has shaped human cognition across millions of years. How close are we to creating machines that don't just perform tasks but continuously reshape their own models of the world? What risks and possibilities arise when intelligence becomes augmented, accelerated, or even decoupled from biology as we know it now? We examine the promises and perils of transformative AI, and widen the discussion into cultural, philosophical, and spiritual dimensions. What does it mean for humanity when a new kind of mind emerges alongside us? And how might this shift redefine creativity, purpose, and the future of human evolution? Speakers Gary Marcus (NYU, Professor Emeritus) Max Tegmark (MIT & Future of Life Institute, Professor & President) Richard Socher (you.com, Co-Founder and CEO) Moderator Jack Symes (Durham University, Philosopher) (c) AI House Davos 2026 Founders & Strategic Partners: ETH AI Center, Merantix, G42, Hewlett Packard Enterprise, EPFL AI Center, The University of Tokyo Presenting Partners: KPMG
AI Generation: Rethinking Primary and Secondary Education | AI House Davos 2026
In recent years, the use of AI, including generative AI, has been expanding in primary and secondary education. This panel discusses the use of AI in education from a wide range of perspectives, including educational benefits and concerns, impacts on children's development, challenges in implementation and operation, national and local policies, and global trends. Speakers Masami Hagiya (Institute for AI and Beyond, Director) Don Passey (Lancaster University, Professor, School of Social Sciences) Mary Webb (King's College London, Professor of AI in Education) Moderator Toshinori Saito (Seisa University, Professor, Graduate School of Education) (c) AI House Davos 2026 Founders & Strategic Partners: ETH AI Center, Merantix, G42, Hewlett Packard Enterprise, EPFL AI Center, The University of Tokyo Presenting Partners: KPMG
The Next Wave of Deep Tech and Future Use | AI House Davos 2026
As generative and agentic AI move from early adoption into true mainstream use, a new set of frontiers is rapidly coming into view. This session explores the next wave of AI innovation beyond today's paradigms, examining where the technology is heading, which breakthroughs still stand between us and the next leap forward, and how these advances could reshape entire industries throughout the 2030s. From multimodal intelligence and autonomous experimentation to large-scale simulation and the emergence of new human-machine interfaces, the panel maps out trajectories that could redefine how AI is built, deployed, and experienced. Speakers Adrian Locher (Merantix Capital, Co-Founder and General Partner) Andrey Khusid (Miro, Founder and CEO) Deepak Pathak (Skild AI, Co-Founder and CEO) Nal Kalchbrenner (Project Prometheus, Founding Member) Moderator Jamie Heller (Business Insider, Editor in Chief) (c) AI House Davos 2026 Founders & Strategic Partners: ETH AI Center, Merantix, G42, Hewlett Packard Enterprise, EPFL AI Center, The University of Tokyo Presenting Partners: KPMG
Quantum Computing as the Next Frontier in AI Capability | AI House Davos 2026
Quantum computing is emerging as the next frontier in AI, promising new forms of capability far beyond today's systems. This panel explores what breakthroughs might become possible when quantum techniques are applied to machine learning, and how close we truly are to achieving useful quantum advantage for AI. The discussion also looks at the broader implications for security, accelerating human discovery and global governance as quantum power grows. Speakers Grégoire Ribordy (IonQ, VP Science & Technology, Founder, ID Quantique) Marina Marinkovic (ETH Zurich, Professor Computational Physics) Nathan Baker (Microsoft, Lead Architect of Quantum) Steve Suarez (HorizonX, Founder and CEO) Moderator Anu Unnikrishnan (ETH Zurich, Executive Director Quantum Center) (c) AI House Davos 2026 Founders & Strategic Partners: ETH AI Center, Merantix, G42, Hewlett Packard Enterprise, EPFL AI Center, The University of Tokyo Presenting Partners & Guest Country: KPMG, The University of Tokyo, Poland
From Orbit to Insight: AI’s Role in Shaping a Smarter Planet | AI House Davos 2026
Space exploration isn't just about reaching new frontiers - it's about unlocking breakthroughs that transform life on Earth. This panel explores how AI in space is accelerating innovation and driving real-world impact across smart cities, precision agriculture, disaster response, climate modeling, and global sustainability. From analysing satellite imagery to training AI models aboard the International Space Station, and even the future development of orbital data centres, space-powered intelligence is transforming global challenges into opportunities for a smarter and more resilient world. Speakers Samantha Cristoforetti (European Space Agency, Astronaut) Bernhard von Weyhe (European Space Agency,Senior Media Relations Officer) Mark Mozena (Planet,Vice President of Government Affairs) Moderator Kirk Bresniker (HPE Labs, HPE Fellow, Vice President & Chief Architect) (c) AI House Davos 2026 Founders & Strategic Partners: ETH AI Center, Merantix, G42, Hewlett Packard Enterprise, EPFL AI Center, The University of Tokyo Presenting Partners: KPMG
What is a Job? NextGen Organizational Strategy, Structure, and Scale | AI House Davos 2026
AI is redefining how organizations create, coordinate, and capture value. As enterprises become AI-native, traditional organizational models are being dismantled and the boundaries of what a 'job' is are being reshaped entirely. Collaboration between humans and AI agents is transforming organizations from rigid structures into dynamic systems of intelligence. Acting as co-workers and autonomous economic actors, they reshape work, markets, and accountability. This panel focuses on how human–agent organizations are transforming industries and redefining business models and what this means for strategy, leadership, and the future of work. Speakers May Habib (WRITER, Co-Founder and CEO) Alessandro de Luca (Merck, Head of Digital Enterprise Solutions and Group CIO) Judith Wiese (Siemens, Chief People and Sustainability Officer and Member of the Managing Board) Moderator Ashish Madan (KPMG Germany, Managing Partner, CTO) (c) AI House Davos 2026 Founders & Strategic Partners: ETH AI Center, Merantix, G42, Hewlett Packard Enterprise, EPFL AI Center, The University of Tokyo Presenting Partner: KPMG
NYT Op-ed: Silicon Valley is at an inflection point
2025. ‘Terrified’ federal workers are clamming up , 2025. The foundations of America’s prosperity are being dismantled , 2025. Microsoft’s hypocrisy on AI , 2024. AI is taking water from the desert , 2024. Inside the chaos at OpenAI , 2023 The new AI panic , 2023. Cleaning up ChatGPT takes heavy toll on human workers , 2023 The U.S. is turning away from its biggest scientific partner , 2023 Artificial intelligence is creating a new world order , 2022. U.S.-China tensions fuel outflow of Chinese scientists from U.S. universities , 2022. How Facebook got addicted to spreading misinformation , 20…
What is ChatGPT? What to know about the AI chatbot
2023. China seeks a quantum leap in computing , 2022. AI has cracked a key mathematical puzzle for understanding our world , 2020. This is how AI bias really happens—and why it’s so hard to fix , 2018.
The hidden workforce that helped filter violence and abuse out of ChatGPT
podcast, 2023 Can you make AI fairer than a judge? Play our courtroom algorithm game , interactive, 2019. We analyzed 16,625 papers to figure out where AI is headed next , data viz, 2019. What is AI? We drew you a flowchart , graphic, 2018. What is machine learning? We drew you another flowchart , graphic, 2018.