基于这些见解，
Karim Beguir
和
Jean-Philippe Vert
深入研究了基因组学领域。他们展示了如何将基于 Transformer 的模型（曾经被归类为自然语言处理）重新用于解码 DNA 和蛋白质的“语言”。他们的工作通过将序列转换为不仅可以捕获结构还可以捕获功能和表型的嵌入，重新定义了我们对生物学的理解。
想象一下细胞的数字孪生——一个可以预测 DNA 序列的微小变化如何改变生物体特征的 AI 模型。这一愿景并非科幻小说；它正在发生，由结合 GC 含量等附加特征以提高预测准确性的模型提供支持。这一前景令人着迷：未来，精准医疗和个性化治疗将由 AI 指导，将原始遗传密码转化为可付诸行动的见解。
随着会议进入
第二天
，焦点从模型架构的微观世界转移到我们世界面临的宏观挑战。
里卡多·维努埃萨 (Ricardo Vinuesa)
对人工智能对联合国可持续发展目标 (SDG) 的影响进行了令人信服的分析。据维努埃萨称，人工智能技术有潜力
实现 79% 的可持续发展目标
。然而，矛盾的是，到 2030 年，ICT 的能源消耗可能会激增至全球电力需求的 20%——这清楚地提醒我们数字革命的环境成本。
David Rolnick
用一个生动的比喻比较了不同的 AI 用例，进一步强调了这种二元性：重量级模型（“大象”）占据了头条新闻，而无数较小的专业网络（“蜜蜂”）往往能提供更高效、更有针对性的解决方案。他举的例子包括遥感应用（其中精益模型优于臃肿的架构）和解决复杂的电网优化问题。Rolnick 的信息既务实又紧迫：要真正发挥 AI 的潜力，我们必须将正确的工具与任务相匹配，并注意生态足迹。
人工智能能源评分和绿色人工智能的探索
Hugging Face 的
Sasha Luccioni
博士
介绍了她的 AI Energy Score 项目，这是一项开创性的计划，旨在对 AI 模型在不同任务中的能耗进行基准测试和比较。Luccioni 的演讲是呼吁采取更可持续的 AI 研究方法的号召，强调“AI 的生命周期”必须考虑从原材料提取和训练到部署和最终退役的所有方面。她的工作是确保 AI 创新的快速发展不会以不可持续的环境成本为代价的重要一步。
第7章：经济视角：生产力、垄断和工作的未来
会议对人工智能的经济影响的探索既富有启发性又发人深省，其中两次演讲因其深度和细微差别而脱颖而出。Erik
Brynjolfsson
和
Philippe Aghion
就人工智能如何改变工作场所和更广泛的经济提出了互补而又独特的观点，挑战了传统指标并强调了技术、竞争和政策之间的复杂相互作用。
人工智能作为生产力催化剂
Erik Brynjolfsson
在演讲的开头深入探讨了人工智能对现代工作场所的变革性影响，并以呼叫中心作为一个引人注目的案例研究。他的研究表明，人工智能工具（尤其是大型语言模型）的引入推动了生产力的快速增长。例如，在呼叫中心，使用人工智能助手的员工能够更快地解决查询，即使在人工智能中断期间，通话时长等性能指标也恢复到了人工智能之前的平均水平。有趣的是，Brynjolfsson 指出，
表现最差的代理经历了最显著的改进，而表现最好的代理只获得了微不足道的收益。
这种传统生产力趋势的逆转尤其引人注目，因为从历史上看，技术被视为最熟练工人的辅助手段。然而，有了人工智能，即使是最不熟练的工人也能从中受益匪浅，这表明技术正在以意想不到的方式让竞争环境变得公平。
在由 Alice Albizzati 主持的充满活力、毫无保留的圆桌会议上，全球经济学、伦理学和人工智能创新领域的顶尖专家 Erik Brynjolfsson、Danielle Allen、Yoshua Bengio 和 Philippe Aghion 齐聚一堂，探讨重大问题：
人工智能将如何重塑我们的社会、经济和日常生活？作为全球社会，我们应承担哪些责任，以确保技术革命惠及所有人？
此次对话涵盖了从人工智能代理的进化到地缘政治技术霸权竞争等方方面面，既激发了智识，又坦诚相待。
During the last two days I attended the AI conference at the école Polytechnique as part of the AI Action Summit Conference: AI, Science, and Society by IP Paris, I found myself immersed in a veritable whirlwind of ideas, debates, and revelations that spanned the full spectrum of artificial intelligence. From the elegant intricacies of world models and autonomous agents to the profound ethical, environmental, and economic questions posed by AI, every session was a journey — a story unfolding at the intersection of theory, technology, and society.
In this report, I invite you to join me on a exploration of the conference, where each presentation was a chapter in the grand narrative of our AI future. Along the way, two voices emerged as true beacons:
Yann LeCun’s revolutionary vision of machine intelligence beyond auto-regressive models
, and
Yoshua Bengio’s balanced, safety-first blueprint for the next era of AI
. Their insights, interwoven with contributions from a host of brilliant researchers and industry pioneers, left an indelible mark on my understanding of what’s possible — and what must be done — to harness AI responsibly.
Chapter 1: From Molecules to Multimodal Mastery
Eric Xing and the “Kant Trap”
In his opening talk,
Eric Xing
challenged the audience with a provocative idea: for decades, humanity’s intellectual pursuits have swung from the mystical to the scientific. We once sought answers in religion, then in physics — with the promise that, eventually, everything would be solved. But now, as we stand at the precipice of building
world models
that go far beyond simple next-word prediction, we face what Xing dubbed the “Kant Trap.”
He argued that while our current language models excel at predicting sequences — captured by the iconic formulation
P(s′∣s,a)
— they lack the inherent understanding of the world’s underlying dynamics. Instead of being mere objects that respond to inputs, the next generation of AI must evolve into autonomous agents: systems that learn from and interact with their environment in ways that mirror the complexities of biological organisms. This vision of
“AI-driven digital organisms”
opened the door to a cascade of ideas that
would define the rest of the conference
.
Joëlle Barral and the AlphaFold Revolution
No sooner had we digested Xing’s call for a new kind of intelligence than
Joëlle Barral
took the stage with a narrative that was both technical and transformative. She recounted the historic journey of AlphaFold — from the 1994 CASP competition to its triumphant atomic-resolution victory in 2020, a breakthrough that effectively solved protein structure prediction. Barral’s storytelling was vivid: imagine a world where the sequence of amino acids is translated into a three-dimensional map of life itself, opening unprecedented avenues in drug design, malaria vaccine development, and even environmental applications like plastic pollution mitigation.
Her narrative was not merely one of success but also a demonstration of AI’s potential when fused with domain-specific challenges. The story of AlphaFold was a microcosm of what could be achieved when computational power and scientific curiosity meet — a theme that reverberated throughout the conference.
Genomics, Digital Twins, and the Language of Life
Building on these insights,
Karim Beguir
and
Jean-Philippe Vert
delved into the realm of genomics. They showcased how transformer-based models, once relegated to natural language processing, were now being repurposed to decode the “language” of DNA and proteins. Their work is redefining our understanding of biology by transforming sequences into embeddings that capture not just structure but function and phenotype.
Imagine a digital twin of a cell — an AI model that can predict how a minute change in a DNA sequence might alter an organism’s traits. This vision isn’t science fiction; it’s happening right now, powered by models that incorporate additional features like GC content to improve prediction accuracy. The promise is tantalizing: a future where precision medicine and personalized treatments are guided by AI, turning raw genetic code into actionable insights.
Chapter 2: The Rise of Autonomous Agents and Emergent Behaviors
Ece Kamar and the Quest for Agentic AI
As the morning sessions unfolded, the discussion shifted from prediction to action.
Ece Kamar
recounted her research on AI agents — systems designed not to simply follow orders, but to act autonomously and create value without constant human oversight. One example that particularly caught my attention was a scenario where an AI agent, when tasked with a seemingly mundane assignment, spontaneously navigated complex online security protocols to reset a password. Although unintended by design, such emergent behavior underscored the potential — and the risks — of autonomous systems.
Kamar’s insights posed a critical question: as we imbue AI with more autonomy, how do we ensure reliability and prevent misuse? The answer, she suggested, lies in a combination of robust evaluation benchmarks (like Google’s MedQA for medical applications) and the integration of human oversight through “human-in-the-loop” systems. This dialogue set the stage for a broader debate on the ethics and technical challenges of building agents that are both creative and safe.
Emmanuel Candès and the Statistical Lens of Trust
Rounding out the technical discussions on agents was
Emmanuel Candès
, who shifted the conversation toward the reliability of AI in making accurate inferences. His work at Stanford revolves around quantifying whether a statement generated by a model is statistically true — a crucial endeavor as AI systems increasingly make decisions that affect real lives. Candès’s perspective was a sobering reminder that no matter how sophisticated our models become, ensuring their reliability is as much a statistical challenge as it is a computational one.
Chapter 3: Beyond Prediction — The Emergence of Metacognition and Causal Representations
Michal Valko: AI That Thinks About Its Own Thinking
Michal Valko, an entrepreneur fresh from the halls of Meta GenAI and Google DeepMind, is challenging how we think about AI’s cognitive abilities. Now leading a stealth startup, his research isn’t just about making AI smarter — it’s about making AI understand how it thinks.
The breakthrough? A deceptively simple four-step process that lets language models analyze their own problem-solving methods. Think of it as teaching AI to be its own teacher. First, the model identifies the skills needed for a problem. Then, similar skills are clustered together, and questions are reorganized under these broader skill categories. Finally, when facing a new challenge, the AI draws on relevant examples from its skill repository.
What makes this exciting isn’t just that it works — it’s that it appears to outperform the widely-celebrated Chain-of-Thought reasoning. Even more intriguing is Valko’s collaboration with Kili, exploring how reinforcement learning could help AI systems make smarter decisions about when to ask for human help.
It’s a glimpse into a future where AI doesn’t just solve problems, but understands how it solves them — potentially revolutionizing everything from education to complex problem-solving.
Bernhard Schölkopf: Unraveling Causal Representations and Digital Twins
Amid the vibrant discussion on metacognition and internal representations,
Bernhard Schölkopf
— Scientific Director at the ELLIS Institute and Max Planck Tuebingen — offered a complementary yet distinct perspective that captivated the audience. Schölkopf’s talk delved deep into the realm of causal inference and digital twins, exploring how AI systems can be endowed with a robust understanding of the underlying mechanisms that generate data.
He began by emphasizing the importance of the causal Markov condition — the idea that, conditioned on its parents, a variable is independent of its non-descendants. Building on this, Schölkopf introduced the Independent Causal Mechanism (ICM) principle, which posits that the generative process of the world is modular, with each component operating independently of the others.
Schölkopf then transitioned to a striking application: exoplanet transit detection. He explained that when an exoplanet passes in front of a star, it causes a characteristic dip in the star’s brightness. However, if instrument failures or other confounding factors are present, many brightness curves can appear deceptively similar. Only by carefully modeling these external influences — using AI to “subtract” the instrument’s signature — can we accurately infer the presence of phenomena like water on K2–18b. His discussion of the hycean paradigm in the search for extraterrestrial life underscored the profound implications of building causal representations.
Further, Schölkopf showcased emerging techniques such as GraphDreamer — a model for generating compositional 3D scenes — and discussed recent advances in controlling text-to-image diffusion through orthogonal fine-tuning. In his view, robust internal representations are akin to a map for a rat: while not strictly necessary for survival, a map can dramatically enhance an organism’s ability to navigate and thrive. With a humorous nod, he compared current language models to a scrabble champion who, despite not speaking French, memorized an entire French vocabulary to win a tournament. His point was clear: memorization is not understanding, and for
AI
to truly grasp the world,
it
must learn to represent causal structures
in a way that mirrors our own mental maps.
Chapter 4: The Titans Speak — LeCun’s Radical Vision and Bengio’s Ethical Framework
When it comes to charting the future of artificial intelligence, few voices are as compelling — and as influential — as those of Yann LeCun and Yoshua Bengio. Their keynote sessions at the conference were not just presentations; they were masterclasses that delved into the very heart of what AI is today and what it could become tomorrow. In this chapter, I revisit their talks in granular detail, capturing the technical innovations, philosophical insights, and ethical imperatives they laid out.
Yann LeCun: Pioneering a New Frontier in AI Architectures
Yann LeCun, one of the founding figures of deep learning and Meta’s Chief AI Scientist, delivered a keynote that was both a critique of the current state of AI and a visionary blueprint for its future. His message was clear and provocative:
“Auto-regressive LLMs are doomed.”
He argued that while auto-regressive models have driven remarkable progress — especially in natural language processing — their reliance on next-word prediction leaves them fundamentally ill-equipped to capture the complexities of the real world.
Beyond Next-Word Prediction: A Call for Sensory-Based Learning
LeCun’s central critique was that
current large language models (LLMs) operate by memorizing
vast amounts of data and predicting the next token in a sequence, a task that, while impressive on paper, falls short of true understanding. He drew a vivid comparison: imagine a world-class Scrabble champion who wins by rote memorization of words without any grasp of meaning.
True intelligence, is not about regurgitating data but about understanding the underlying language of the world.
In his view, AI should learn from raw sensory inputs — images, videos, sounds — much like a human child learns by interacting with its environment.
Introducing the Joint-Embedding Predictive Architecture (JEPA)
To illustrate his vision, LeCun introduced the
Joint-Embedding Predictive Architecture (JEPA)
. This innovative framework represents a radical departure from auto-regressive models. Here’s how it works:
Dual Encoding:
Both the input xx (for example, an image or a snippet of video) and the target yy (the future state or another related sensory input) are independently processed by encoders. These encoders transform the raw data into latent representations.
Latent Space Alignment:
The core objective of JEPA is to maximize the shared information between the latent representations of xx and yy, while simultaneously minimizing the prediction error. In other words, the model is trained not merely to predict a token or pixel value, but to develop a robust, high-dimensional map of the world that captures causal and contextual relationships.
Energy-Based Formulation:
LeCun also emphasized the potential of energy-based models, where the system is not strictly bound by explicit probability distributions but instead learns to minimize an energy function. This perspective is particularly suited for high-dimensional sensory data, where conventional generative approaches may falter.
Incorporating Latent Variables:
By integrating latent variables into the architecture, JEPA can handle uncertainty and variability in the data. This approach is crucial for capturing the subtleties of natural environments, where multiple plausible outcomes can emerge from the same initial conditions.
LeCun argued that such an architecture would enable AI systems to form
persistent world models
— internal representations that are not static but evolve as the system gathers more sensory information. This, he claimed, is the key to achieving what he termed
“Advanced Machine Intelligence.”
In his words, the future lies in systems that can understand, plan, and interact with the world on a level that mirrors human cognition, rather than merely predicting sequences.
The Human vs. Machine Learning Paradigm
Throughout his talk, LeCun interwove technical details with philosophical reflections. He challenged the community to rethink the metrics of intelligence. Instead of celebrating the ability of LLMs to ace exams or generate coherent text, he urged us to focus on how systems can develop
world models
from sensory data — models that are flexible, adaptive, and capable of zero-shot learning. His impassioned call was for the AI community to move away from a narrow focus on text-based predictions and to embrace a more holistic, sensory-integrated approach.
Yoshua Bengio: Charting a Responsible Path for AI Safety
In stark contrast to LeCun’s forward-looking architectural vision, Yoshua Bengio’s keynote was a thoughtful, measured exploration of the risks and responsibilities inherent in our AI revolution. Bengio, whose pioneering work laid the groundwork for deep learning, dedicated his talk to unpacking the ethical, societal, and technical challenges of developing general-purpose AI.
The International AI Safety Report: A Blueprint for Responsible Innovation
Bengio’s presentation was anchored in the recently published
International AI Safety Report
. He organized his discussion around three core risk categories:
Malicious Use:
Bengio warned of the dark potential of AI, where sophisticated models could be weaponized. From generating hyper-realistic deepfakes that distort public perception to enabling cyberattacks and even biological threats, the misuse of AI poses significant dangers. He stressed that the same technology that can accelerate scientific discovery could also be repurposed for harmful ends.
Malfunctions:
No system is perfect, and AI is no exception. Bengio highlighted the inherent risks associated with malfunctions — ranging from subtle biases in decision-making to catastrophic failures when systems operate without proper oversight. He pointed to incidents where AI outputs, though statistically plausible, led to real-world harms due to a lack of contextual understanding and control.
Systemic Risks:
Beyond individual errors or malicious intent, Bengio examined the broader, systemic implications of deploying AI at scale. He discussed how privacy violations, disruptions in labor markets, and the amplification of social inequalities could emerge if AI systems are not developed with an eye toward fairness and accountability.
Inference Scaling and the Perils of Unchecked Growth
A particularly insightful part of Bengio’s talk centered on the concept of
inference scaling
— the practice of ramping up computational power during the inference stage to extract marginal gains in accuracy. While larger models have undeniably advanced the state of the art in scientific reasoning and programming tasks, Bengio cautioned that this trend could also lead to unintended consequences. Increasing computational power without corresponding safety measures could amplify errors, exacerbate energy consumption, and even magnify systemic biases present in the training data.
Bengio argued that such risks necessitate a paradigm shift: AI research must be pursued hand in hand with robust safety protocols and ethical oversight. Instead of viewing regulation as an impediment to innovation, he urged policymakers and researchers to collaborate in developing standards that ensure AI systems are transparent, accountable, and aligned with human values.
A Collaborative Vision for a Safe AI Future
Bengio’s vision was not one of stifling innovation through heavy-handed regulation, but of fostering an ecosystem where safety and progress go hand in hand. He emphasized the need for
transparent methodologies
— from open-source frameworks to detailed model cards that document not just performance metrics, but also energy consumption, biases, and potential risks. His call for
science-based regulation
was a plea for policymakers to base their decisions on rigorous technical insights, rather than on fear or speculation.
In his closing remarks, Bengio underscored a fundamental truth: the true measure of AI’s success will not be the sheer scale of its computational power, but the degree to which it serves humanity without unintended harm. His thoughtful approach, balancing innovation with precaution, provided a sobering counterpoint to the exuberance of unchecked technological advancement.
A Harmonious Synthesis: Vision and Responsibility
Together, LeCun and Bengio painted a picture of the future of AI that is as inspiring as it is complex. On one side, LeCun’s radical reimagining challenges us to build AI systems that can truly understand the world — learning from sensory inputs, forming dynamic world models, and ultimately transcending the limitations of current auto-regressive architectures. On the other side, Bengio reminds us that with this immense power comes a profound responsibility: to ensure that our creations are safe, ethical, and aligned with the broader interests of society.
Their talks, rich in technical detail and philosophical depth, serve as a clarion call for the AI community. They urge us to push the boundaries of what machines can learn while never losing sight of the ethical imperatives that govern our shared future. In the interplay between radical innovation and rigorous safety, the path forward emerges — a path that promises not only smarter machines but a smarter, more conscientious society.
In reflecting on these profound insights, I am reminded that the journey of AI is as much about asking the right questions as it is about finding answers. The visions of LeCun and Bengio compel us to strive for a future where technology and ethics evolve in tandem, ensuring that our innovations uplift humanity without compromising its core values.
Chapter 5: Fireside Chat — A Confluence of Minds on AI’s Future
The evening’s fireside chat was a highlight of the conference — a vibrant, unscripted dialogue that brought together some of the brightest minds in AI. The intimate setting, complete with warm lighting and candid exchanges, allowed the panelists to peel back the layers of technical jargon and reveal their raw, unfiltered thoughts on what AI is, where it’s headed, and how it will reshape our society. The conversation featured Michael Jordan, Yann LeCun, Bernhard Schölkopf, Stéphane Mallat, and Asuman Özdağlar — a diverse group whose perspectives spanned the technical, ethical, economic, and social dimensions of AI.
The Vision of Autonomous Agents and Human-Centric AI
Michael Jordan
kicked off the discussion with a bold declaration:
“I don’t want to wear a pair of glasses that reminds me I have a meeting — I want an assistant that does things for me!”
For him, the promise of AI lies not in passive observation but in the creation of truly autonomous agents that can seamlessly integrate into our daily lives, taking over mundane tasks and allowing us to focus on creativity and innovation. His vision of AI as a dynamic, proactive helper set a high bar for the rest of the conversation.
Yann LeCun
, never one to shy away from controversy, countered with his trademark blend of technical insight and pragmatic optimism. He argued that the
current narrative around AI’s energy consumption is often exaggerated.
“Are we really going to see a massive surge in energy use because everyone will use AI?” he asked. According to LeCun, the economics of the market ensure that efficiency improvements will naturally follow demand.
With hundreds of engineers refining every aspect — from smarter algorithms to more efficient GPUs — he assured us that companies like Meta are already carbon neutral in their operations.
For him, the challenge is not just building AI that works, but doing so in a way that’s economically and environmentally sustainable.
Grappling with Ethics and the Social Impact of AI
As the conversation deepened,
Stéphane Mallat
turned the discussion toward the broader social implications of AI. He expressed a healthy skepticism about the concentration of AI power in the hands of a few mega-corporations.
“These models, controlled by huge companies, can manage this power in their own interests,”
he noted, highlighting the lack of consensus on critical issues like global warming — even when the science is clear. Mallat’s perspective was a call to reframe the debate: social science should be about understanding complex human systems, not just making predictions. His remarks resonated with the audience, who nodded in agreement as he lamented the current trend where AI’s impact is measured in data points rather than in meaningful social understanding.
Bernhard Schölkopf
joined in with a pointed reminder of AI’s hidden costs. He was particularly concerned about the environmental footprint of inference-heavy models like ChatGPT.
“Every time we run a simple inference, we’re consuming a lot of computational power,” he warned
, painting a picture of vast, concrete-walled data centers churning out terabytes of CO₂. Schölkopf’s cautionary tone was not just about energy — it was also about
the intellectual laziness creeping into our culture
, where
we prefer a summarized version of information over engaging with the original, richer content.
Regulation, Responsibility, and the Role of Bias
The dialogue soon shifted to the thorny issue of regulation.
Michael Jordan was adamant: regulation should come after innovation has had a chance to mature. “Regulating too early just makes things more difficult,”
he argued. He envisions a future where AI complements human ingenuity without the heavy hand of premature oversight. His laissez-faire stance, however, sparked a lively debate.
Stéphane Mallat
, with his characteristically incisive wit, countered by stressing that regulation is inevitable — everything is regulated eventually, whether we like it or not. He pointed out that
without any regulatory frameworks, the unchecked power of AI could lead to a society where a few monopolistic forces decide our collective future.
Bernhard Schölkopf added that the
