Ivan Nardini

For more than a decade, Neha Patki and I have kept returning to one question: Can a generative model learn a relational database as a connected whole? At Massachusetts Institute of Technology we developed some of the earliest algorithms for training generative models on relational databases. That work eventually became The Synthetic Data Vault. But our work with enterprise customers taught us that the algorithm was only part of the challenge. Real enterprise databases are messy, deeply interconnected and often poorly documented. Important relationships and business rules may be embedded in the data rather than written into the schema. Building useful models at enterprise scale required the software to discover much more of this on its own. Today, we are giving this class of models a name: generative relational models (GRMs). We are also launching SDV 2.0, which brings us much closer to automating how enterprises build them. My conviction remains the same: an enterprise’s data is a strategic asset, and models trained on that data will become strategic assets too. Companies should be able to build and own those models inside their environments, then put them to work across software, analytics and AI. LLMs bring the world’s intelligence to the enterprise. Generative relational models bring the enterprise’s intelligence to AI. Neha and I share more about the journey here: https://lnkd.in/gERNMBHs #GenerativeAI #EnterpriseAI #SyntheticData

42

1 Comment

Wrote a blog explaining "LoRA Fine-tuning Efficiency Under Different Loss Functions." LoRA made fine-tuning cheap, but not always efficient. This post explores how loss functions like Cross-Entropy, Label Smoothing, and Focal Loss change convergence, stability, and generalization in lightweight LLMs. Read here: https://lnkd.in/gh_vm2kG

2

Excited to share that our Research Topic in Frontiers in Artificial Intelligence is now open and welcoming contributions! Key focus areas include:  • Retrieval-Augmented Generation (RAG) for domain-grounded answering and synthesis  • Deep Research systems, agentic AI search, and knowledge work paradigms  • Indexing, representation learning, and scalable retrieval over heterogeneous data (text, tables, images, logs)  • Search relevance modeling and neural reranking for knowledge-intensive queries  • LLM-enhanced recommendation systems (resources, products, next-best-action)  • Query understanding at scale: intent detection, rewriting, clarification  • and more Published in 𝗙𝗿𝗼𝗻𝘁𝗶𝗲𝗿𝘀 𝗶𝗻 𝗔𝗜 (Impact Factor: 4.7 | CiteScore: 7.3). Articles publish as accepted. Manuscript deadline: 𝗔𝘂𝗴𝘂𝘀𝘁 𝟯𝟬, 𝟮𝟬

9

🚀 scikit-learn 1.8 is out 🚀 A big shoutout to the community of contributors who continue to push open-source machine learning forward ❤️ ✨ Key Highlights: ▶️ Expanded Array API support (including PyTorch & CuPy) to run more estimators and metrics on GPUs ▶️ Free-threaded CPython 3.14 support for better multi-threaded performance ▶️ Probability calibration with temperature scaling in CalibratedClassifierCV ▶️ Major efficiency boosts in linear models (Lasso / ElasticNet with gap safe screening) ▶️ Much faster and more robust DecisionTreeRegressor with criterion="absolute_error" ▶️ New manifold.ClassicalMDS implementation for classical multidimensional scaling 🔗 Check the full release highlights: https://lnkd.in/gkXQSbmZ Discover scikit-learn 1.8 and its: 🟢 28 new features 🔵 12 efficiency improvements & 13 enhancements 🟡 9 API changes 🔴 34 fixes 👥 193 contributors (thank you all!) 📖 More details in the release notes: https://lnkd.in/gCrs42se You can upgrade with pip as usual: pip install -U scikit-learn Using conda-forge builds: conda install -c conda-forge scikit-learn #scikitlearn #MachineLearning #opensource #DataScience #Python #ML

1,518

14 Comments

I have developed CiberIA CogState Estimator, a proprietary private model that analyzes the observable behaviour of an AI system to estimate its operational cognitive profile, identify risks, and support its evaluation and monitoring. It is part of CiberIA and is available for client projects and controlled deployments. In this article, I explain how it works and what it brings to AI cognitive security. https://lnkd.in/eUS-kg8b

1

Big day for builders on Microsoft Foundry. We launched Claude Opus 4.6 today. Microsoft is the only platform offering same-day access to the latest frontier models — OpenAI and Anthropic — under one roof. Why this matters for AI-native founders: - Pick the best model for your workload, full flexibility. - Run real A/B testing across frontier & open-source models - Avoid lock-in while scaling with enterprise-grade security & governance built in. - Ship faster with priority capacity and integrated tooling Excited to see what's coming next! Read the announcement: https://msft.it/6040QMYRM #Microsoft #Azure #MicrosoftFoundry #Anthropic #OpenAI #Claude #MultiModel #AINative

83

1 Comment

IBM Research put Helion through its paces by implementing paged attention for vLLM. Interesting early look at PyTorch's beta portable kernel language. Quick takeaways from their experiment: • Code was shorter and easier to debug (Helion handles masking/tiling automatically) • Portable across NVIDIA H100 and AMD MI300X • Strong on decode, still catching up on prefill • Autotuning is powerful but takes hours (vs Triton's heuristics) For anyone running long-context inference (clinical histories, imaging batches), portable kernels like this could open up more hardware options without the rewrite pain. Useful for clinical AI teams running on mixed setups. Full post: https://lnkd.in/gmNZkbxT #PyTorch #vLLM #Helion

4

Open models and datasets are changing how AI gets built, adapted and deployed. See how model architectures, training data, post-training and evaluation are opening new paths across AI, science, healthcare and robotics at #NVIDIAGTC Berlin. View sessions 👉 https://bit.ly/4r0GIeW

1