Marcin Detyniecki

At AXA, our teams have always been driven by a shared curiosity to tackle the most complex challenges. I am proud of that same drive in my Research team, who developed an AI-powered tool designed to analyze geopolitical risk. This prototype helps our experts navigate geopolitical risk to better assess its long-term impact across different regions. The results speak for themselves: weeks of analysis and countless data points are condensed into just 2 hours — freeing our analysts to focus on the strategic thinking and nuanced judgment that only a human expert can provide. This research paves the way for advanced prevention services, helping us move beyond protection and toward proactive partnership. It’s a steppingstone to create more value for our clients in a world of constant uncertainty. Here's a short clip from our recent General Assembly, where Samuel Pouyt explains how we're empowering our talent with next-generation tools — unlocking a whole new frontier of expertise. Congrats Samuel, Marcin and the entire team AXA Group Operations Matthieu Caillat Marcin Detyniecki Samuel Pouyt Yann Bry Alexandre Delidais Antimo Di Donato Vincent De Ponthaud Arnaud Balthazar Julie Bonneau Olivier BOILEAU Virginie Sésia 

190

6 Comments

New paper alert! 🚨 How does the perfect dataset look? One option: a dataset where you know everything about it. This is exactly what we tried to do in our paper: “Beyond the Loss Curve: Scaling Laws, Active Learning, and the Limits of Learning from Exact Posteriors” 🧵👇 It all started on Twitter. A few months ago, I posted an old idea: using normalizing flows as oracles to get exact posteriors, which could be really cool to try with modern tools. Arian Khorasani saw the post, reached out, and this collaboration was born. The core idea is to train class-conditional normalizing flows on real images, freeze them, and now you have a synthetic world where you know the exact true posterior p(y|x) for every single image. This means you can finally answer: how close is my model to the best it could do? Why does this matter? Every time you see a loss curve flatten, you can’t tell if the model learned everything learnable, or if it’s stuck. Our oracle decomposes total loss into: Aleatoric (irreducible) - the noise in the data itself Epistemic (reducible) - what the model should still learn So after we worked hard to train these models (why you need so many GPUs 🥲 ), what did we find? Finding 1: Hidden scaling laws 📉 Epistemic error follows a clean power law (KL ∝ N⁻ᵅ) that keeps shrinking even when total loss looks completely flat. Your model is still learning - standard metrics can’t see it because the aleatoric floor dominates. Finding 2: Rethink robustness - Since we have exact likelihoods, we can compute the exact KL divergence for any distribution shift. We find that class imbalance (changing which classes are more common) barely scratches accuracy, while Gaussian noise at collapses it. Why? Because the imbalance only shifts p(y), the features stay intact. Noise corrupts p(x|y) directly, destroying what classifiers actually rely on. Knowing how much the distribution shifted tells you almost nothing. Finding 3: Active learning with ground-truth uncertainty. In active learning, people typically use entropy to choose which samples to label next. But predictive entropy mixes aleatoric and epistemic signals - a high-entropy sample might just be inherently ambiguous rather than informative. With our oracle, you can decompose these and select genuinely informative samples. We found that this gives a big improvement in sample efficiency. Finding 4: Soft labels from exact posteriors beat hard labels. If you have p(y|x), you can use it! The posteriors encode inter-class similarities and the uncertainty structure. Training with soft posteriors instead of one-hot labels outperforms hard-label training and achieves near-perfect calibration. Huge thanks to Arian Khorasani, who led the project brilliantly, co-authors Nathaniel Chen, Yug D Oswal, Akshat Santhana Gopalan, and Egemen Kolemen, and to Alex Alemi for valuable discussions. Questions and feedback are very welcome! 🙏 Paper: https://lnkd.in/eTh982Nc

131

6 Comments

🧠 AI Feels Alive — But That’s Not the Same as Being Alive Lately, I keep seeing the same idea repeated in different forms: AI already feels conscious. It talks well, reflects, advises, even challenges us. Surely it’s only a matter of time. But there’s a quiet mistake hidden in that intuition — not a technical one, a human one. A useful way to think about it came from neuroscientist Erik Hoel (https://lnkd.in/eW5dbzwb). He argues that consciousness, if we want to treat it seriously, has nothing to do with how impressive the output looks. It has to do with something much simpler: does experience leave a mark? Here’s an everyday analogy. Imagine a colleague who always says the right things. Emails are perfect. Meetings go smoothly. Advice sounds thoughtful. But there’s a catch: every morning, they wake up without memory. Yesterday’s mistake doesn’t make them wiser. A tough conversation doesn’t change how they show up tomorrow. Success doesn’t shape judgment. Failure doesn’t create caution. We wouldn’t call that a person in the full sense. We’d call it a very polished tool. That’s roughly where today’s large language models are. They don’t grow from interaction. Conversations don’t reshape them. Experience doesn’t scar them. Once deployed, they are frozen. What looks like learning is actually replay — not transformation. Biologist and philosopher Michael Levin (https://lnkd.in/efMKHVTj) puts it beautifully in another context: memory isn’t an archive. Memory is how a system rebuilds itself over time while preserving meaning. Humans don’t stay the same because they remember everything accurately. We stay ourselves because the past quietly rewires who we become next. That’s the difference. 💡 A conscious system isn’t defined by how convincing it sounds, but by whether life changes it. Whether mistakes hurt. Whether experience reshapes priorities. Whether time leaves fingerprints. This distinction matters far beyond AI debates. It matters for how we design organizations. Many companies optimize for consistency, predictability, compliance. Processes become fixed. Roles become rigid. People are discouraged from changing how they think — as long as outputs look right. Over time, organizations can start to resemble large models: extremely capable, highly articulate, and strangely ... static. Efficient, yes. Alive? Not really. ❓ The uncomfortable question isn’t whether AI will become conscious. It’s whether, in chasing efficiency, we’re building systems — and cultures — that slowly eliminate the very thing that makes humans human: the ability to be changed by experience. #AI #PhilosophyOfAI #Consciousness #Leadership #FutureOfWork #OrganizationalCulture #HumanCenteredAI #LearningOrganizations

1

I'm thrilled to announce that I'll be speaking again about Semantic Croissant and the latest CODATA Cross-Domain Interoperability Framework (CDIF) developments at the FORUM-MD Workshop, “AI in a Metrological Context: AI for Metrology and Metrology for AI”, taking place at the Parc de Saint-Cloud in Sèvres, France on 11th of March. My talk in the session “Data Ready for AI: How Do Data Quality and Semantics Interact with AI Applications?” will explore how data and structured metadata, prepared for machine-actionable formats, can significantly enhance AI applications. By leveraging the semantics of data, we can help AI systems better understand the meaning behind datasets and bridge knowledge gaps in areas such as interoperability. At the same time, AI tools have strong potential to support experts in the digitalization of diverse data sources, helping to automatically harvest metadata and semantic information. See you there tomorrow if you're around! https://lnkd.in/ehkgaHiQ #AI #DataSemantics #Metadata #Interoperability #Metrology #SemanticWeb #Croissant #CDIF

69

4 Comments

🤖 Following the interview I shared yesterday (link in the first comment), a number of messages and calls turned into interesting conversations. A few points really stood out: • We are seeing rapid advances and impressive demos across AI, autonomy, and deep tech. • But many of these capabilities are not reaching real operational use at scale. • That gap is starting to create real frustration. • The same pattern appears in AI powerful agents and tools exist, but scaling them inside real organizations is a different story. • Especially in high-responsibility environments where validation, reliability, and trust are critical, and where regulatory requirements must be met.  • In defense deep-tech, the challenge is even sharper: breakthrough capabilities require long processes of integration, testing, and adoption before they reach the field. • Not because of resistance to innovation but because the system has to operate differently when lives and security are at stake. 👉 Which raises a few key questions - curious how others here are thinking about this: • How do we better bridge the gap between innovation and deployment? • What needs to change to allow deep tech to scale into real operational systems? • And how do we do this while maintaining responsibility and trust? This feels like one of the defining challenges right now.

14

3 Comments

🌽🌾 A Multi-Scale Feature Fusion #DeepLearning Network for the Extraction of #Cropland Based on #Landsat Data ✍️ Huiling Chen et al. 🔗 https://brnw.ch/21wZadu

8

Well done, Arlequin AI! Arlequin AI was a key sponsor for the Boston TAG Party this year. They are technical leaders and enthusiastic supporters of academic mathematics communities. In a world where AI is changing constantly and the discourse centers on a few key players and architectures, I am personally thrilled to see investors leaning into the value, promise, and power of new, mathematically informed neural design!

9

💬 Yann LeCun: “LLMs aren’t a bubble — our expectations are.” According to Meta’s chief AI scientist, the current wave of large language models is not an investment bubble. These systems already have strong practical value and will continue to deliver real-world impact. The real “bubble,” he argues, lies in the belief that scaling LLMs alone will lead to human-level intelligence (AGI). True progress will require new conceptual breakthroughs, not just more data and bigger clusters. > “We’re missing something fundamental.” — Yann LeCun #AI #DeepLearning #LLM #YannLeCun #AGI #MachineLearning #AIResearch #MetaAI

2