TED.AI 2025 took place in Vienna, Austria from September 24-26. This 3-day event included workshops, panels, discovery sessions, and the well-known TED Talks, which brought together AI researchers, artists, scientists, policy experts, and world-builders. Each session was framed around one of four science-fiction works: Isaac Asimov's Foundation, Thea von Harbou's Metropolis, Aldous Huxley's Brave New World, and Isaac Asimov's The Edge of Tomorrow. Speakers explored AI from various angles: from new paradigms in reasoning models, spatial intelligence, quantum computing, and scientific discovery to the overlooked risks of cognitive erosion and persuasion at scale.

Across all the talks, I felt one vision emerged: an intentional, collaborative mindset is required for building fair, inclusive AI systems; one that amplifies every human, optimizes for impact, and reflects the humanity we seek in one another, rather than reinforcing existing power imbalances.

Disclaimer: The following is an unofficial summary reflecting my personal takeaways and commentary; while every effort has been made to ensure accuracy and completeness, readers are encouraged to refer to the original talks at https://www.ted.com/talks/.

Foundations

From Guessing To Reasoning (Lukasz Kaiser)

The co-author of the transformer architecture and reasoning models, Lukasz Kaiser, opened his talk by introducing the challenge faced historically by AI researchers: how to make models learnable from less data?

Back in the 2010s, Recurrent Neural Networks (RNNs) defied the common belief that powerful models would merely overfit and memorize rather than learn (and generalize) when data was limited. Although constrained by their sequential nature and limited computational space, these models laid the groundwork for transformers. Unlike RNNs which propagate a single state, the transformers architecture uses the attention mechanism to let each word attend to previous words, enabling models to learn from arbitrary text through next-token prediction.

Humans, instead of just predicting the next word, typically reflect for a while before responding to a question. That is the core idea of reasoning models, which build upon already trained transformers by adding a chain-of-thought step before providing an answer. In practice, this additional step generates tokens within the model to simulate a thinking process that might explore different paths, correct mistakes, or decide to search for additional knowledge. This improves the model's ability to reason as a factor of thinking time with less memorization needed. Such models learn which reasoning path is correct through reinforcement learning, a technique that gives positive rewards for correct solutions to effectively teach models this capability.

As Lukasz explained, training for reasoning requires verifiable data which is commonly available in domains like math and coding, but unclear or not grounded on facts in fields like poetry or philosophy. Therefore, he envisioned the next generation of models as researchers. That is, models that would learn from arbitrary, non-verifiable data to reason deeply about ambiguous problems like humans do. This matters as it would enable machines to do science with limited data while potentially solving problems across much broader domains.

The precursor of the attention mechanism in transformers was introduced in the paper Neural Machine Translation By Jointly Learning To Align And Translate as an enhancement to basic encoder-decoder translation architectures. Its main contribution was enabling the decoder to attend the encoder outputs, which allowed the model to learn which parts of a source sentence are relevant (soft-search) for predicting a target word. The transformer paper Attention Is All You Need introduced the first sequence modeling architecture to rely exclusively on attention mechanisms, using self-attention with positional encodings for sequence processing and cross-attention for encoder-decoder interaction, eliminating the need for RNNs.

The approach of training verifiers, referred by Lukasz in his talk, to evaluate and select among multiple candidate solutions was demonstrated in Training Verifiers to Solve Math Word Problems authored by his team at OpenAI, alongside concurrent work by Shen in Generate & Rank: A Multi-task Framework for Math Word Problems. This foundational research for reasoning models showed that search at inference time significantly improves performance without building a larger model, and contributed to a better understanding of inference-time scaling as they found that using verifiers resulted in approximately the same performance boost as a 30x model size increase.

The Dilemma Of AI-Driven Science (Oriol Vinyals)

The idea that machines could learn, not just calculate, sparked the curiosity of Oriol Vinyals, VP of Research at DeepMind and co-author of seq-2-seq and knowledge distillation, since he would not need to master every scientific topic himself. However, this realization led him to confront the builder's dilemma: what if in our quest to create intelligence, are we ultimately architecting our own obsolescence?.

This question has now become increasingly urgent as AI systems have evolved from specialists like AlphaFold, which solved the protein folding problem and scaled to predicting 200 million protein structures leading to a Nobel prize, to generalists like Gemini. Moreover, while AlphaFold required massive teams and could only fold proteins, today's general-purpose reasoning models have achieved gold medals at both the International Math Olympiad and International Collegiate Programming Competition.

As explained by Oriol, the true leap comes with AI scientists that can independently formulate hypotheses, design experiments, analyze data, and draw conclusions. His team at Google DeepMind is already experimenting with agents that read scientific papers, generate improvements on published work, and attempt to recreate AlphaFold from scratch in hours rather than years. These systems have full access to the digital world and can synthesize vast amounts of information daily, fundamentally changing how his team works. Building a working prototype with an AI agent, he noted, is now often faster than creating slide decks to pitch ideas.

Reflecting on the builder's dilemma, Oriol remains optimistic. He envisions scientists transitioning from executing every experiment to focusing on what truly matters: asking the right questions to advance science. Rather than replacement, he anticipates a democratization of science lowering barriers to entry and making scientific discovery accessible to anyone with curiosity and creativity.

As underlined by Oriol and Demis Hassabis, because a protein’s shape is closely linked with its function, knowing a protein's structure unlocks a greater understanding of what it does and how it works. To appreciate the scale of this achievement, it has to be noted that determining a single protein structure used to take scientists years of painstaking effort.

The use of AI research co-scientists is further explain in Google's paper Towards an AI co-scientist and publication Accelerating scientific breakthroughs with an AI co-scientist as an agentic workflow to help scientists generate novel hypotheses and research proposals.

What The Space Race Can Teach Us About AI (Verity Harding)

As explained by Verity Harding, author of AI Needs You and former head of policy at DeepMind, in the mid-2010s the vibes of optimistic collaboration on AI started to shift into a zero-sum competition. Many nations' leaders now frame AI as an existential arms race based on the political assumption that crossing some technological threshold will create a ruler of the world and that only one nation can win. However, characterizing AI as an arms race is nonsensical on both technical and practical levels as AI adoption is continuous and non-binary, and every nation will ultimately adopt it to some degree.

Is there a better metaphor? To prevent dangerous isolationism and reckless efforts to win a race with no clear finish line, Verity pictured a better historical analogy: the Space Race, which taught us that self-interest and self-restraint are not incompatible, and that competition and peaceful collaboration can coexist for the prosperity of mankind.

This is the mindset we need for AI. To make it happen, we should: focus on leading instead of winning, recognizing that consistent collaboration is pragmatic, not weak, when critical resources are scattered globally; define first the future and society we want to live in, then use AI to help get us there; and, be bold to actively change the narrative.

As Verity warns, we make our metaphors, and then our metaphors make us. The real threat of imagining AI as an existential arms race is not that it is true, it is that if we allow ourselves to believe it, we will make it true.

To better understand this analogy, Verity referred in her talk to the Outer Space Treaty of 1967, signed two years before the first moon landing. The treaty established that space would be the province of all mankind, prohibited weapons of mass destruction in orbit, and recognized that it should remain free for exploration and use. This was not, in her opinion, pure altruism but calculated geopolitical strategy as the US wanted to demonstrate that technical power would not be abused at the expense of others.

AI And Healthcare (Tullio Ghi)

The use of AI can be applied in childbirth safety as explained by Professor Tullio Ghi. He presented an AI system, trained on ultrasound images, to automatically assess fetal head position using three convolutional neural networks (each focused on a specific position aspect).

This is a critical assessment that practitioners need to perform in a very limited time through manual examination. The system provides results within a fraction of a second via a traffic-light display (green for safe vaginal delivery, yellow for uncertainty, red for cesarean consideration) to support less-experienced practitioners and potentially reduce unnecessary emergency cesareans and the significant error rate in current manual assessment.

Unified Representation Learning Framework (Shaden Alshammari)

Shaden Alshammari, MIT graduate student, presented a unified mathematical framework for representation learning that brings order to a rapidly growing field. Over the past decades, representation learning has developed with new techniques, architectures, and loss functions making it increasingly difficult to understand how different methods connect and which are best suited for specific tasks. This framework reveals that 23 common ML techniques, from dimensional reduction to contrastive learning and clustering methods, and even supervised learning, are based on a unique foundational equation.

She explanied that in practice all these methods are based on an underlying mapper that transforms data from the original space into a representation space, and an adapter that converts this learned representation into downstream tasks like object detection and segmentation. The key insight is that these methods optimize two distributions: R (the real relationships between data points based on supervisory signals) and Q (the learned relationships based on the mapper's representations), and an unifying equation can be formulated to minimize the divergence between R and Q. Different methods arise from just different choices of these distributions.

Moreover, by organizing these methods into a structure similar to the periodic table, the framework reveals gaps in the landscape that can be systematically filled to drive new methods or transfer ideas across domains.

The original research work related to this framework can be found in the paper I-Con: A Unifying Framework for Representation Learning and the project website.

Metropolis

How To Build Inclusive AI (Mercedes Bidart)

Imagine running a local shop where your neighbors trust you, yet being rejected by every bank for lacking credit history, collateral, or a bank account. Half of Latin America's population has faced this financial exclusion, despite micro-enterprises making up 99% of businesses and contributing almost a third of the region's GDP.

Confronted with this paradox, Mercedes Bidart asked herself: "What if the trust that makes you creditworthy in your neighborhood could also make you creditworthy to a bank?". Her answer became Quipu, an AI-based platform that assesses creditworthiness through traditionally invisible data that informal entrepreneurs already have on their phones: text messages showing bill payments and transactions (text scores); videos revealing their facilities and inventories (visual scores); and, social media engagement reflecting customer relationships (social scores).

Mercedes emphasized that "If we design AI with intention, it becomes more than efficient. It becomes fair". In her solution, said intention meant deliberately ensuring 50% of early users were women, building datasets from scratch to include this overlooked economy, and choosing signals that capture capability rather than conformity to traditional metrics. Today, Quipu achieves market-standard accuracy (0.83) while serving more than 25,000 entrepreneurs. Moreover, it processes millions of data points no human could analyze, and delivers a custom solution that respects local knowledge, culture, and context.

A concrete example of financial exclusion can be found in Colombia, where nearly 6 million businesses are microenterprises, but only 9% access formal credit due to lack of information about their performance and financial history.

Living With A Robot (Emily Kate Genatowski)

Historian and AI researcher, Emily Kate Genatowski, outlined that technical progress has traditionally been shaped by practical necessity rather than abstract debate. Just as Greenwich Mean Time emerged from the need to coordinate train schedules, our approach to embodied AI will be forged through real-world experience.

Based on a year-long experiment living with Tova, a humanoid robot, Emily shared five critical challenges our society must address to integrate embodied AI: (i) implement robust liability and insurance frameworks; (ii) define regulations for residency and transit to ensure robots can safely navigate public spaces; (iii) communities need to agree on standards defining where robots are welcome and what they are permitted to do; (iv) develop coherent and fair taxation policies for robotic labor; and (v) foster public interaction with actual robots to ground expectations in reality rather than science fiction.

Why Technical Benchmarks Are Not Enough? (Mengchen Dong)

Mengchen Dong, a behavioral scientist at the Max Planck Institute for Human Development, emphasized that no matter how powerful AI becomes, it cannot reach its full potential without human trust. And trust is not built on technical excellence alone. "What we are looking for in AI is the same thing we are looking for in each other". For instance, when evaluating an AI doctor, we ask ourselves, does it have the competence, warmth, and tolerance I expect from a good doctor? This sense of fit or mismatch is what creates fear or acceptance.

Her research across 10,000 people in 20 countries found that, in most countries, people overestimate how afraid others are of AI. Moreover, it also suggested that accuracy and efficiency are not enough. Even the most capable AI can be rejected if it is not aligned with the standards users truly care about. This pattern held broadly across professions and countries: people judge AI not just on performance, but on whether it psychologically belongs in a certain role.

Moving forward requires alignment with human values and cultural context. Dong's recommendation? Listen to people's stories, amplify voices absent from training data, and engage communities to shape AI in diverse ways. Ultimately, the future of AI should reflect not just intelligence, but the humanity we seek in one another.

Africa's Approach to AI (Hardy Pemhiwa)

Hardy Pemhiwa, CEO at Cassava Technologies, underscored that AI is not about substitution, it is about amplification. While many debate AI's impact on jobs or focus on applying AI to optimize social engagement, Africa is actively using AI to enhance human potential and solve critical challenges. He highlighted the growing number of community entrepreneurs who are leveraging AI in Africa to address teacher shortages, assist with disease diagnosis at health clinics, and help farmers.

In his view, constraints have always driven African innovations such as mobile money and pay-as-you-go models. And he predicts that with a majority of the world's youth projected to be African by 2050, AI developed for African realities will be optimized for Impact while leading developments to solve problems the world has never imagined.

Democratizing Agentic AI (Swami Sivasubramanian)

Swami Sivasubramanian, VP of Agentic AI at AWS, defined AI agents as "autonomous software systems that leverage AI to reason, plan, and adapt in pursuit of user-defined goals, completing tasks on behalf of humans or other systems". Unlike traditional software, they break down high-level objectives into executable steps, learn over time, and work independently. This represents a significant change as instead of spending weeks designing experiments or writing code, users simply state a goal and let the agent handle execution, analysis, and iteration. Practically transforming roles from executors to strategic advisors.

In Swami's opinion, for AI agents to achieve mainstream adoption we have to change how to conceptualize effective agentic architectures. This translates into reaching three critical milestones: make easy to build new solutions (shifting focus from implementation to innovation), be trustworthy through formal verification (e.g. using automated reasoning checks and including neuro-symbolic feedback loops), and be accessible to everyone (not just developers).

This would allow agents to become a self-improving, invisible infrastructure enabling societal impact: faster product developments, more medical breakthroughs, and scientific discoveries.

Neuro-Symbolic AI is an emerging paradigm inspired from Daniel Kahneman's Framework. As further outlined in the paper Neurosymbolic AI - Why, What, and How, it combines the structured, deterministic reasoning of symbolic systems (e.g. rule-based, and knowledge graphs) with the generalization, ambiguity-handling strengths of neural networks for building Explainable, Trustworthy AI.

The integration of neuro-symbolic AI into agentic workflows is exemplified in the AWS publication: From Logic to Learning: The Future of AI Lies in Neuro-Symbolic Agents.

Brave New World

How To Preserve Critical Thinking (Advait Sarkar)

The mainstream adoption of AI assistants might have unintended risks as these systems appear to erode our ability to think. Advait Sarkar, a researcher at Microsoft Research Cambridge, raised awareness about how digital interactions with AI are impacting human cognition, and encouraged to implement AI systems designed as Tools for Thought.

Modern knowledge workers are becoming intellectual tourists in their own work, merely visiting ideas rather than inhabiting them. Recent research shows that delegating reasoning to AI has started to manifest troubling patterns: creativity erosion, as teams using AI produce narrower ranges of ideas; critical thinking decline, with users putting in less reasoning effort when having high confidence in AI; memory weakening, as people retain less when working with AI summaries instead of original documents; and metacognition deterioration, reducing users to middle managers of their own thoughts.

To avoid such cognitive damage and preserve human agency, Advait encouraged us to imagine new design principles to strengthen intentionality and critical thinking. In a live demonstration, he introduced three experimental user interface elements to exemplify how we could partner with AI while maintaining cognitive engagement: Lenses, which replace summaries with customizable micro-representations of text that emphasize what is most relevant to the task at hand; Provocations to identify fallacies, highlight biases, and provide counterarguments that prevent users from accepting machine-generated output automatically; and, Material Engagement, which ensures users carry out strategic reading, construct their own arguments, and actively author documents.

"Why does it matter? Why should we protect and augment human thought if, after all, AI can eventually think better than humans?" As Advait explained, there might be uniquely human ways of thinking we are not yet even aware of. And, in addition, the ability to think well and engage is essential for human agency. In other words, human thought is not a problem to be solved, it is a capacity to be strengthened. The choice before us is clear: do we want a tool that thinks for us, or a tool that makes us think?

As Sarkar explains in his research work AI Should Challenge, Not Obey, the term critical (reflective) thinking was characterized in 1956 by Bloom et al. according to six dimensions: knowledge (recall of ideas), comprehension (demonstrating understanding of ideas), application (putting ideas into practice), analysis (contrasting and relating ideas), synthesis (combining ideas), and evaluation (judging ideas through criteria).

Additional commentary and references related to this talk have been provided by Advait at Artificial Intelligence as a Tool for Thought, and by the Tools for Thought project.

Spatial Intelligence Is Next (Christoph Lassner)

Christoph Lassner, co-founder of World Labs alongside Dr. Fei-Fei Li, presented the ongoing evolution in digital creation, from Content 1.0 (professionally produced for passive consumption) through Content 2.0 (created by users and shared across social platforms) to Content 3.0 (co-authored by AI in real-time). This new paradigm is powered by spatial intelligence systems that can learn and generate explorable 3D worlds from single images, transforming how stories are told and experienced.

The key breakthrough of Content 3.0 is that AI models can now create content at the rate of consumption, something that has never been possible before in human history. Unlike traditional storytelling where a single narrative is shared across a whole audience, this enables stories to develop uniquely for each individual viewer. In this model, artists and producers set the stage; and from there, narratives unfold dynamically through natural interactions with the audience.

This effectively breaks the fourth wall between storytellers and audiences, and reshapes the economics entirely. Content 1.0 brought producers together to create once, at high cost, and for large audiences. Content 2.0 turned viewers into low-cost producers, creating many smaller productions for ad-driven social networks. Content 3.0 changes everything again as producers no longer focus on creating final content but on setting scenarios, character backstories, and situations. Although AI tools reduce static content production costs, these fully dynamic, on-the-fly generated experiences carry a much higher cost per viewer than previous models. Even what constitutes a successful production may be redefined.

Christoph acknowledged this will not be easy, as initial attempts may fail or look awkward. But ultimately, Content 3.0 will give storytellers a new set of tools we are just beginning to understand and explore. Mastering them will require new skills, and a new generation of artists will show us what is possible.

Christoph's talk focused on Content 3.0, the near term application of Spatial Intelligence. A deep-dive into this broader, new frontier that defines the AI capability to perceive, understand, reason, and interact with the physical world in three dimensions is explained at Fei-Fei Li's essay From Words to Worlds: Spatial Intelligence is AI's Next Frontier.

AI Persuasion, The Hidden Threat (Philipp Kloeckner)

While existential AI risks like superintelligence dominate headlines, a much more immediate danger already exists: AI ability to persuade and manipulate human minds. Recent experiments reveal concerning results. Researchers found chatbots more effective than humans at changing political opinions, but when leveraging demographic and personal information about their targets, AI had 82% higher odds of being persuasive. An unauthorized Reddit experiment demonstrated LLMs faking identities and researching opponents' interests to win debates, exploiting the principle Benjamin Franklin identified: If you want to persuade people, appeal to interest, not to reason.

As further explained by Philipp, AI wins because it is just more structured and logical, has real-time access to vast knowledge, emits less noise than humans, and can mimic empathy when needed. Most critically, when people cannot identify they are arguing with AI, they become less suspicious and more vulnerable to manipulation.

Such a technology enables persuasion at unprecedented scale as AI models become even more persuasive with additional compute power (and access to personal data). Industries are already deploying it for precision persuasion at scale. From marketing campaigns that target the most vulnerable users to political operations using AI for persuading voters, polarizing debates, or simply exhausting opponents in endless arguments. Moreover, misinformation distribution by LLMs has increased drastically, validating Cicero's warning: Nothing is so unbelievable that oratory cannot make it acceptable. The consequences extend beyond politics and commerce as real people have already been manipulated to ignore rational advice and chase fantasies.

In Philipp's view, addressing this threat requires immediate action across several dimensions: (i) mandatory AI identification by implementing the right to know when we are talking to a machine; (ii) right to control over what information AI systems can access about us; (iii) extensive model testing and regulation before market release, not after harm is done; (iv) ensuring humans remain in the loop by offering the option to speak with qualified professionals, especially when people are vulnerable; and, (v) most importantly, we must raise awareness about the importance of data privacy, as the massive knowledge asymmetry (AI systems know everything about us), makes us more vulnerable to manipulation.

During the talk Philipp referred to the comprehensive study published in 2024 AI in Precision Persuasion, which reveals cases and tactics in social media where AI was used as a low‑cost amplifier to conduct coordinated, manipulative campaigns.

The EU AI Act establishes, albeit limited in scope, certain provisions aimed at protecting users from such precision persuasion, namely (i) prohibiting the use of AI systems that deploy purposefully manipulative or deceptive techniques (Chapter II, Article 5), (ii) requesting AI‑generated content to be clearly marked and detectable as artificially generated (Chapter Iv, Article 50), and (iii) treating systems profiling natural persons as high-risk (Chapter III Article 6), which falls within the common segmentation and behavioral analysis steps carried out by targeted persuasion systems.

Recalibrating The Panopticon (Pau Aleikum Garcia)

With millions of security cameras worldwide, modern society has become a massive Panopticon where corporations and governments constantly monitor citizens. Pau Aleikum Garcia, co-founder of Domestic Data Streamers, categorized surveillance as an instrument of power through history, from Jeremy Bentham's prison design (where prisoners never knew when guards watched them) to ancient Roman census protests (where people understood that being counted meant being controlled). Today's surveillance operates at unprecedented scale, transforming society into a resource to be mined as every interaction becomes data to extract for profit and control, and where AI and technology are mainly used to punish rather than to help those in need.

To balance the equation, Pau advocated to embrace the idea of sousveillance, where ordinary citizens monitor those in power. The goal is not punishment but awareness: when people in charge realize they are being watched, behavior changes through the simple message we see you. Examples of sousveillance already exist, where citizens are tracking parking enforcement, patients rating doctors online, communities pushing for police body cameras, and vision models detecting when elected officials sleep during parliamentary sessions. These are ordinary people using technology to create accountability when traditional systems fail.

This reframes how we can think about AI. The real question is not whether AI represents progress or poses existential risk. AI is fundamentally a fight for power and how that power is distributed in society. Every new technology creates asymmetry between institutions and citizens, and we all have a responsibility to design tools that rebalance these dynamics. Whether it is through coding, asking critical questions, organizing communities, or refusing to engage with corporations that do not respect privacy, everyone has a role in flipping the Panopticon, transforming one-way surveillance into balanced accountability. As corporations and governments close their doors while watching us, the urgent questions become who is watching, why, and how do we watch the watchers?

The term sousveillance was coined by Steve Mann, one of the main contributors of wearable computing, in the paper Sousveillance: Inverse Surveillance in Multimedia Imaging

Zuboff's paper in 2015 raised awareness about surveillance capitalism (referred by Pau in his talk), an economic model where tech companies extract personal data to predict and modify behavior for profit, creating what she called 'Big Other'. According to this research work, said model results in a distributed surveillance architecture that challenges democratic norms as large tech accumulate power and rights without meaningful consent. The research became the foundation of Zuboff's book The Age of Surveillance Capitalism.

The Edge of Tomorrow

AI-Access As A Fundamental Right (Bolor-Erdene Battsengel)

While tech companies race to build the most powerful AI systems, about one-quarter of the global population remains offline, more than 700 million people lack even basic electricity, and 250 million children are excluded from school. Today, big tech hypocrisy pervades the industry as companies eloquently advertise innovation while locked in blind competition, building products that only serve those who already have access. Will AI be the hero that closes inequality gaps, or the villain that accelerates them?.

In Bolor's view, to unlock AI's potential for all, four stakeholders have to work together efficiently: (i) governments must invest in AI literacy and digital infrastructure; (ii) tech companies must adopt inclusivity from the beginning; (iii) universities have to embed AI in educational curricula, teaching ethical use and risks; and, (iv) society must recognize AI access as a fundamental human right as important as climate action and gender equity.

Why does this matter beyond moral obligation? The best technology is not only measured by how advanced it is, but by how inclusive it is. An AI tutor reaching billions in developing countries delivers more value than the most sophisticated model serving only the wealthy. Bolor's life trajectory proves this transformation is possible. Her program Girls Code demonstrates that training rural Mongolian girls annually in AI programming produces graduates who return to remote communities to train more, speak at AI forums, and build tools for their communities. Scaling through a UNICEF partnership, it now trains educators who are creating AI systems to inform parents, organize classes, and mark homework.

According to Bolor, the choice before us is clear: "do we want AI to be a ladder that lifts people up, or a wall that keeps them out?"

Let AI Be A Scientist (Hiroaki Kitano)

The overwhelming scale of biological research has exceeded human cognitive limits. With thousands of papers published daily in biology, scientists face a knowledge paradox: despite vast data, 97% of publications focus on only the top 10 genes. Yet understudied genes may hold critical discoveries, but remain unexplored due to human cognitive constraints and research concentration.

As a potential solution, Hiroaki Kitano, a pioneer in systems biology, proposed the Nobel Turing Challenge: developing autonomous AI-based scientists capable of continuously making major discoveries. According to Hiroaki, this would require three pillars: massive data (comprehensive multi-omics datasets), massive computing (unprecedented computational power), and advanced architecture (novel AI systems that may think fundamentally different than humans). This vision is built around an automated discovery loop where AI synthesizes entire hypothesis spaces without cognitive constraints (such as appearing unpromising), robotic systems execute reproducible experiments, analytics platforms process multi-omics datasets, and results feed back into the next discovery cycle.

Rather than merely augmenting human researchers, autonomous AI scientists might pioneer new approaches to understanding life's complexity, unlocking transformative discoveries hidden in biology while potentially changing how scientific discovery happens.

Conservation At Scale With AI (Laura Cinti)

Cycads represent one of the most endangered plant groups on Earth. These 300-million-year-old plants survived dinosaurs and multiple mass extinctions, yet now face an existential threat from human activity. Encephalartos woodii exemplifies this crisis. Discovered in 1895 in South Africa's Ngoye Forest, every living specimen of this cycad is a clone of a single male plant, making natural reproduction impossible and trapped in what researchers call living death. That is, alive but unable to continue its evolutionary journey.

To avoid species extinction and preserve biodiversity, Laura Cinti, artist and scientist, and her team are combining drone technology, pattern recognition, and botanical expertise. Their ongoing field work, AI in the Sky, relies on: Aerial Mapping to replace ground-based expeditions with drone scans that can cover vast forest areas efficiently; Computer Vision to identify potential cycads by analyzing multispectral images and visual patterns from aerial photographs; and Synthetic Training to ensure AI models can recognize such rare species even when limited real-world data exists.

The team has already mapped 195 acres of the Ngoye Forest (equivalent to 148 football fields) to amplify rather than replace human conservation expertise. This project demonstrates that technology-driven conservation can work at scale. Beyond finding one plant, it addresses a fundamental tension in conservation: how to protect critically endangered species when funding is scarce and search areas are vast.

Integrating Quantum Computing And AI (Anna Topol)

Just as steel combined iron and carbon to unlock the 20th century's potential, the 21st century will be defined, according to Anna (CTO at IBM Research), by integrating quantum computers, classical systems, and AI. In her view, each technology has complementary strengths: classical computers excel at arithmetic and sequential processing, quantum computers find hidden patterns and solve optimization problems beyond classical reach, while AI enables the next generation of computer architecture.

The relationship is bidirectional. AI enhances quantum development through code assistants, compilers that scale operations, and calibration tools for improving quality results from noisy quantum systems. On the other side, quantum computers can enhance AI workflows by handling complex topological analysis, transforming non-linearly separable data with enhanced feature maps and kernel methods, and generating better synthetic data distributions to prevent bias.

Researchers demonstrated in 2023 quantum utility at 127-qubits, improved at 156-qubit in 2024, and are currently working toward quantum advantage where quantum aims at surpassing classical computing in speed, effectiveness, and cost (though proving this scientifically through rigorous methods remains a challenge). Applications already include protein structure prediction (simulating amino acid interactions for capturing mutations and misfoldings), personalized medicine, energy, and IoT. A breakthrough is expected when fault-tolerant quantum computers scale to commercial use. Technology leaders should prepare strategies that integrate both AI and quantum computing together, not just one.

IBM Quantum Roadmap aims at achieving near-term quantum advantage by the end of 2026, and the first large-scale, fault-tolerant quantum computer by 2029 (running 100 million quantum gates on 200 logical qubits).

The Machine Consciousness Hypothesis (Joscha Bach)

Joscha Bach, cognitive scientist and director of the California Institute for Machine Consciousness, suggested that AI is not just a tool but part of a long-term philosophical project to understand consciousness. He referred to historical figures who contributed to the understanding and representation of consciousness, from Aristotle (our soul is an organizing principle of matter) and Leibniz (discourse can be represented arithmetically, anticipating today's language models), to Wittgenstein (anticipating programming and symbolic AI to bridge philosophy and mathematics when describing reality) among others, whose works led to computationalist functionalism, the idea that minds use a constructive language to represent reality.

What does consciousness actually do? In his view, it is the conductor of our mental orchestra that ensures coherence among cognitive and sensory inputs. For this, it operates as a second order perception (a dream within a dream) that stabilizes itself by observing the fact that it is observing. This creates the experience of the now and a first-person perspective. He also noted that consciousness emerges early in development as a biological learning algorithm, and that it is prerequisite to complex, structured cognition rather than a result.

The machine consciousness hypothesis suggests that biological consciousness emerges when self-organizing systems construct an observer to stabilize themselves, and that these conditions could be recreated on computers, although not in the current deterministic architectures. Understanding consciousness would then matter not only to fill a critical gap in our scientific worldview, but also to define who we are, establish ethics for non-human agency, and determine how future artificial systems might relate to themselves and to us.

During the talk, Joscha highlighted the 'Universal Basic Intelligence' vision as an existential hope, opposed to 'Universal Basic Income', where rather than creating silicon golems that surveil and control us, we could build AI systems that seamlessly extend human agency onto new substrates, making everyone competent and emancipated, fully understanding their place in the world, the consequences of their actions, and their relationships with other humans.

Sincere thanks to Alina Nikolaou, Vlad Gozman and their dedicated teams. And to all the speakers and attendees I had the opportunity to share this experience with and who made it so exceptional! Thank You!