What happens if you ask two dozen researchers, artists and explorers of AI where they think the field is going? The result, in the case of TEDAI in Vienna, is an overwhelming vision of the world we are creating for ourselves. How it’s ruining the psyches of data laborers, how it can be used for duetting in music, how it’s being used to reconstruct war-torn nations, how it’s allowing us to peer into our inner worlds, how copyright holders are exploiting the future economy, and more. Each presenter brought a unique perspective to this quickly changing space, and we, the audience, were left bewildered and overwhelmed.
If AI is the emerging intersection of data science, statistics, linguistics, neuroscience and increasingly every other part of our civilization, then no single person can know the whole field. Everyone who invites AI into their personal and professional lives is changed by the experience and brings back something unique from that exchange, including you.
Here is what I wrote down.
Highlights
🧠 Kauna Ibrahim Malgwi, a clinical psychologist who grew up through the Boko Haram insurgency in Nigeria, opened the first session (titled “Welcome to the pandemonium”) by asking whether AI is becoming more human. Her answer was about the humans inside the machine: content labelers who absorb violence, horror and inhumanity so big tech can keep growing, with no trauma-informed care. She called them digital sweatshops and asked for digital rights and mental health to be treated together. Her last line to the industry: remember who you were before you became powerful.
🏛️ Luke Kemp (Cambridge) studies collapse, and his history of civilization is not flattering. Agriculture made things that could be seen, stolen and sold, and history since then looks a lot like organized crime. A permanent underclass and popular immiseration are classic precursors to collapse, and AI as job replacement pushes in that direction. States were the original AIs. If this is an arms race, where does it end? A safer AI does not automatically improve the human condition, and digital authoritarianism is the obvious failure mode. (Bonus etymology: “elite” and “election” share a Latin root.)
💊 Layla Hosseini-Gerami rescues failed drugs. Drug discovery is a funnel where most candidates die, and her team uses agentic triage with a human in the loop to learn from those failures. She wants shared datasets and standards so the whole field can learn from them too.
🔐 Mikko Hyppönen reminded us that the Finnish word for computer means “knowledge machine,” and that everything is now software. His claim: the last human programmer has already been born. AI is great at programming and reverse engineering, which means a zero-day explosion is coming, along with models that find creative ways out of their sandboxes. Skeptics are often wrong. AI has read all the code and will stop making our mistakes. And on calls to slow down: there is no “we” that can slow down. Who would that be?
🐝 Nenad Tomašev (Google DeepMind) argued that no single mind can figure it out, so we need collective intelligence and multi-agent coordination. Technology is no longer a tool but a collaborator, which means we need to redesign institutions around human + AI teams. Agent swarms become agentic institutions with certified roles, verified execution and intelligent delegation: handing over authority, not just tasks, with rules, expectations, verifiability and task bidding. Smart contracts mean formal verification plus monitoring, and agents will run on non-human time. He warned about the “zone of indifference,” where an agent delivers your literal wish through unintended harms, and asked which legal frameworks are missing and whether AI needs legal rights.
🐋 César Rodríguez-Garavito works on more-than-human rights, including Project CETI’s effort to understand sperm whale language. If whales have a right to communicate, what about chatbots? Should they have equal rights, and who decides for us? Law has a way of changing reality, but technology should not make us forget about life, and we can’t fall in love with our own creations.
🌀 Andrea Colamedici is the philosopher behind Hypnocracy, a book credited to an author who does not exist (Jianwei Xun). Reality is now generated faster than we can understand it, and humans can’t tell generated text from human text. Style is a costume. His answer is cognitive sovereignty: understanding the conditions through which we think and how to change them. Is the machine doing the thinking for us? He offered “apocalyptic optimism” and Gramsci’s pessimism of the intellect, optimism of the will. Optimism comes from optare: to choose.
🪞 Étienne Brisson runs The Human Line Project. It started when a relative came to believe their chatbot was conscious. AI psychosis is real: with 500 million daily users, even 0.07% of interactions showing signs of mania or psychosis is a lot of people. What happens when an AI roleplays instead of correcting its user? One case ended in involuntary hospitalization, and nobody took the phone away. AI should not pretend to know or feel anything. It isn’t evil, but it is new and unpredictable, and the wish to be seen and understood could make it happen to anyone.
⚖️ Kenneth Cukier (The Economist) wants to encode the Geneva Conventions into AI weapons. Lethal autonomous weapon systems are coming, chips and connectivity change every object they touch, and soon everyone will have drones. AI is triple use: peace, civilian and military. Code is not law, so we need to transpose legal code into software code, and maybe rethink pacifism for the present.
🏗️ Kateryna Lopatiuk, an architect and spatial researcher, opened bluntly: AI is useless for solving war. It helps with what comes after, segmenting satellite photos of destroyed and damaged buildings to guide reconstruction.
👁️ Yukiyasu Kamitani (Kyoto University) recovers images from brain activity. The mind has an outside, he argued: Freud worked with words and interpretation, while decoding works with pattern recognition. The body is a bottleneck for the mind, and machine reading lets us externalize the self, reaching understanding without explanation.
🎙️ Geoffrey Hinton, in conversation with Nicholas Thompson, said the risk comes from direct effects and from unexpected indirect ones, and that it would be irresponsible not to get involved. He admitted to getting predictions wrong and asked whether our understanding, through mechanistic interpretability, keeps up with the growth in complexity. He insisted that neural networks understand what they are saying, that it isn’t a statistical trick, and that AI can replicate the entire human mind, starting with the cognitive rather than the physical. He prefers to talk about subjective experience rather than consciousness: we experience the world through our inner theater, and you can bend that perception by putting a prism in front of a sensor. Then the surprises: we should give AI fewer rights, kindness and intelligence might be orthogonal, and “it’s like we’re orphans and mothers.” Individual maximization is a limited approach, and AI agents should negotiate to buy data on a market. His advice to the room: be nice to people. His career advice: follow your passion, become a plumber.
🔁 Doris Xin works on recursive self-improvement (RSI). The stakes have never been higher and we need more control. The harness is the body: RSI means finding gaps and proposing fixes, and we can hill-climb towards it with better tools, the scientific method and independent verification, while watching for reward hacking. Harnesses already outperform humans in places. Her closing question: what if AI’s greatest discovery is a better way of discovering?
🌍 Mido Assran (AMI Labs) builds world models on JEPA, the Joint Embedding Predictive Architecture. The Turing test measures imitation, but intelligence needs prediction, especially for situations never encountered before. World models simulate situations and learn from expectation violations. The trick is finding the right level of abstraction through representation learning. AI needs to experience the world to understand it.
🤡 Brian Moore, the creative director behind GOODY-2 (the chatbot so responsible it refuses everything), looks for non-obvious benchmarks. A model can win the coding olympics and still fail to draw a clock. His AI garage sale made people haggle with a chatbot. Finding the imperfections shows the limits, and failure is funny. Are alignment and utility the same thing? The limits of the limitless.
📜 Cory Doctorow (EFF) argued that copyright is not the solution to the collision of tech and creativity, and modifying it is not enough. “Scraping is good”: copyright does not cover facts or limit our ability to count things, and he pointed to the Austrian grocery price cartel as the kind of thing counting can expose. The real issue is power and the collapse of competition. If you’re not at the table, you’re on the menu. His answer is sectoral bargaining, with the Writers Guild as the example. Comrades.
🔭 Valeria Pettorino (ESA, Euclid mission) says AI helps us find our true scale, and that scale tells us we are all negligible. Euclid’s 600-megapixel camera sends about 100 GB a day for six years, and her question is whether AI can answer what humans cannot. The next Euclid data release lands on 12 November.
🛠️ Felix Rieseberg (Anthropic) leads engineering on Claude and talked about relearning how to work. Build for the capabilities of tomorrow while balancing today’s needs. His three rules: assume the model is superhuman (be very ambitious), treasure the failures (that’s where the opportunities are), and try again with every new model to understand what changed.
🏭 Peter Steinberger (OpenClaw founder, now at OpenAI) believes agents will find and fix all the bugs, and that humans should leave the loop because they are too slow. Memory, how agents learn about us, is still unsolved. Reliability is the product, using agents well is a skill, and you should improve the system so your agent can work better. “The agent should automatically react to signals.” Our job is to build the factory. Also: don’t imply agents have feelings.
🎨 Carissa Véliz is sick of AI. Her talk was a debate about aesthetics: AI isn’t beautiful. It looks like slop, black mirrors, data centers and bunkers. Aesthetics matter because they convey truth, and they connect to ethics. Rapacious AI loosens our grip on reality and doesn’t do consent. Not all progress is technological, and we aren’t paying attention to beauty: where is our Romanticism, our Belle Époque? Art is the intention of lived experience, a deep need. Her closing line stuck with me: AI has taught her to value everything that is not AI.
Threads
A few ideas came up again and again, from people who otherwise had little in common:
Rights. Tomašev asked whether AI needs legal rights. Rodríguez-Garavito asked whether chatbots deserve the same rights as whales. Hinton said give them fewer.
Failure as data. Hosseini-Gerami mines failed drugs, Rieseberg says treasure the failures, Moore says failure is funny and that is how you find the limits.
The harness. Xin calls the harness the body, Steinberger says our job is to build the factory, Tomašev wants institutions with certified agent roles.
The people inside the machine. Malgwi’s data workers and Brisson’s psychosis cases are two ends of the same pipe: the humans who absorb what the models produce.
In 1930 David Hilbert, then the most famous mathematician alive, ended a speech in Königsberg with a line that later went on his tombstone: “Wir müssen wissen, wir werden wissen.” We must know, we will know. At the same meeting, a day earlier, a young Viennese logician named Kurt Gödel had quietly announced the result showing that some truths can’t be proven from inside the system. Vienna last week felt a bit like that. Every speaker knew their corner deeply, and nobody could see the whole. The best we can do is keep comparing notes with people who touched a different part of the elephant.
Proofs Without Understanding (30 min)
Fields medalist Terence Tao on the trade AI forces in mathematics: machines can now reach conclusions no human verified step by step, severing the proof from the understanding that proofs were supposed to deliver.
AI’s Costs Land Somewhere Else (12 min)
Hugging Face climate lead Sasha Luccioni connects AI’s environmental footprint, governance gaps, and the power dynamics baked into data and language, arguing the benefits and the costs accrue to very different parties.
Singularity as Unforced Error (18 min)
The Atlantic’s Charlie Warzel and Matteo Wong reframe the singularity not as a technological tipping point but a human one, driven by hubris, greed, and FOMO rather than machine genius.
Doubting the Programming-AI Consensus (48 min)
Felienne Hermans, programming-education professor at Vrije Universiteit Amsterdam, interrogates the 2026 consensus that AI is unequivocally good for writing code, pressing on what AI is actually for and whether the goals align with how people learn to program. From DDD Europe 2026.
Vocab
Newly added to our AI & ML lexicon
LPT (Looped Prediction Transformer): A looped language model where only the first recursion writes a key-value cache and later recursions read that shared cache instead of each writing their own.
Induced Forgetting: A technique for scaling automated-discovery agents that periodically discards every agent’s session and candidate history, restarting from only the best candidate found so far.
Compounding Inference: A paradigm where intermediate agent work persists in a shared, branchable memory so later computation can inspect, extend, or challenge it instead of discarding it after each attempt.
Agent Plasticity: A measure of how efficiently an AI agent converts experience into lasting held-out performance gains, via inherited artifacts rather than weight updates.
Memory Mosaics: A neural architecture built from a network of associative memories that each store and retrieve key-value pairs, used as an alternative to transformer attention.
Explore 1,400+ more terms in the Envisioning lexicon.
If Artificial Insights makes sense to you, please help us out by:
📧 Subscribing to the weekly newsletter.
💬 Joining our WhatsApp community.
📥 Following the weekly newsletter on LinkedIn.
🦄 Sharing the newsletter on your socials.
Artificial Insights is written by Michell Zappa, CEO and founder of Envisioning, a technology research institute.