Su Yang

Matthieu Jonglez from Progress talking about how to use AI with #semantic #agentic #RAG on structured and unstructure data from a companies’s private systems “The greater the context, the greater the insights” #progressdatasummit

56

2 commentaires

Excited to share that the GenSec CTF we ran at DEF CON 33 with Airbus to let the community explore how human-AI collaboration can speed up cybersecurity was a success. Overall: • Nearly 500 participants completed initial challenges • 85% found it useful for learning AI security workflows • 23% were using AI for cybersecurity for the very first time More details: https://lnkd.in/g9ekMFQQ #Cybersecurity #AI #DEFCON #education Royal Hansen Evan Kotsovinos Amanda Walker Pankaj Rohatgi Nicolas L. Carlos de Segovia Brian Roddy Luca Invernizzi Jennifer Pullman Yanick Fratantonio Marianna Tishchenko Lenin Simicich Naomi Black Dominik Świerad Kajal Rahimian Giovana Bordieri Ferraz Anne Wang Diana K.

94

5 commentaires

Google and LLMs don’t just crawl your website — they construct an Information Graph from it. What is an Information Graph? In NLP, every sentence becomes an Entity–Attribute–Value (EAV) triple. Example: ChatGPT → is → the best AI chatbot ChatGPT → is → one of the best AI chatbots Search engines and LLMs then use fuzzy logic to decide what’s consistent, what’s opinion, and what creates conflict. Even small differences like “best” vs “one of the best” change the weight of consensus around an entity. What about Open Information Extraction? It means crawlers can extract facts from across the web — without a fixed schema. The more sources agree on a value range for an entity–attribute pair, the stronger the “plausibility” of that fact. This is how web-wide consensus is built. For SEO and LLMs, consensus matters at two levels: Internal site consensus → Your own site must present a consistent Information Graph. Too many conflicting statements reduce trust. Web-wide consensus → Becoming a Topical Authority means being the source that defines the “truth range” in your domain. Over the past two years, we built a GPT Crawler that scans websites for information conflicts, suggests corrections, and even powers chatbots that only speak with the site’s own consensus. This was one of the Agents we introduced in our “Orchestra of Agents” lecture (Topical Authority Course 2.0). We’ve also published a case study on expanding topical maps using EAV triples: 👉 https://lnkd.in/dQ8FDGsd 💡 Why does this matter? Information Graphs explain information density, quality, and accuracy. OpenIE systems use them to separate facts from perspectives. Relevance Configuration and our Algorithmic Authorship Rules (shared 3 years ago) help shape how your site communicates with these systems. The takeaway: Search engines and LLMs always rely on Information Graphs. If your site is inconsistent, you weaken its authority. If your site sets consensus, you define the truth range for your industry. We will continue protecting the legacy of Bill Slawski by focusing on research, not hype. 👉 Join our community or course: https://lnkd.in/dp8mGy8E Stay with research and reason. #SEO #TopicalAuthority #ChatGPT

142

26 commentaires

🧠"Less is More: Recursive Reasoning with Tiny Networks" — a new take on AI reasoning from Samsung SAIL Montreal. This paper challenges the classic idea that bigger is always better in LLMs. Instead, it shows how smart recursion with tiny networks outperforms huge models (like Deepseek, Gemini, Claude) on hard tasks — Sudoku, Maze, and ARC-AGI — using a fraction of the data and parameters: • Core Problem: LLMs often stumble on tough reasoning puzzles. Hierarchical Reasoning Model (HRM) was an inspired solution using two small networks, biological analogy, and deep supervision. But — is complexity required? • New Approach: Tiny Recursive Model (TRM) simplifies the design radically — using just one tiny (2-layer, 7M parameter) network and pure recursion. No biological hierarchy, no fixed-point math — just step-by-step improvement. • Results: TRM beats HRM at generalization, pushes state-of-the-art on tough benchmarks (Sudoku-Extreme test accuracy from 55%→87%, ARC-AGI-1 from 40%→45%, ARC-AGI-2 from 5%→8%). Models like Gemini 2.5 Pro and o3-mini fall short at higher cost. • Implications: Recursion, not scale, unlocks smarter reasoning on sparse data. Deep supervision and minimal architectures may hold the key for future LLM and AGI advances. • Real-world Design: TRM removes unnecessary complexity — no extra networks, math, or biological analogies. Results fuel new ways to build efficient, reliable AI for reasoning-intensive tasks. What's next for recursive models? Can smaller, adaptive architectures redefine how we build LLMs for real-world intelligence? 📄 Attached is the original research PDF for full details. Curious how this approach might influence future model architectures?

39

🌍 BREAKING: The UN just made history on AI governance. For the first time, the United Nations General Assembly has adopted a resolution creating two global mechanisms to guide how humanity navigates artificial intelligence: 🔹 Independent International Scientific Panel on AI – 40 experts worldwide, tasked with producing evidence-based assessments on AI’s risks, opportunities, and impacts. 🔹 Global Dialogue on AI Governance – an annual forum for governments, industry, civil society, and academia to build shared approaches and bridge divides. 💡 Why it matters: AI is advancing faster than policy. Without coordination, we risk fragmented rules, growing inequalities, and weakened trust. These mechanisms aim to: • Promote safe, secure, and trustworthy AI • Support capacity-building in developing countries • Ensure transparency, accountability, and human oversight • Advance open data, open models, and interoperability ⚖️ For now, this resolution is non-binding — but it lays the foundation for international cooperation. Much like other global compacts, its true power lies in shaping norms, aligning priorities, and building trust across borders. This is a historic step: AI governance is moving from national debates to the global stage. 👉 What do you think — will global cooperation help ensure AI serves humanity, not the other way around? Link to the full resolution in comments 👇 #AIforGood #AIGovernance #GlobalGoals #UN

162

58 commentaires

GenAI impact assessment, 3 years later : Igor Carron CEO, portrait of LightOn started in 2016: from photonics for AI to start to building LLMs since 2021 now building sovereign foundational models. Simple, generic and repetitive tasks can easily be replaced by GenAI but people are expected to provide their expertise and will have more time to in the future. GenAI today portrait by MANGEANT Fabien Chief AI & Data Officer Air Liquide : not to replace people, but to augment people, and to improve operations. Yves Nicolas AI Group Program Director at Sopra Steria : GenAI is not magic, brings value in some use cases such as customer relationships or regulation gap analysis, but not everywhere. In European Trustworthy AI Association summit

38