Gregory Renard

Over the last months, I’ve been living a strange paradox: I’m CTO of an AI company for children (PopGamma, with Prof. Albert as AI tutor)… and at the same time I’m having long conversations with parents who are worried about AI for their children. 🤖👨‍👧 Some of the parents in our WhatsApp community asked for something simple: "Can you give us a short, honest guide to talk about AI with our kids?" So we co-wrote (Sreekanth Krishnamoorthy) a brief "AI & Children" charter (in collaboration with AI itself) It covers three things in very plain language: 1️⃣the difference between recommendation algorithms, deepfakes, and generative AI 2️⃣what their limits and risks are for children 3️⃣how to react when kids start using chatbots as friends or therapists The message is simple: communication and transparency work better than fear and restriction. 💬🤝 If you’re a parent, educator or just curious, you can read or reuse the guide here. Feedback and improvements are very welcome – this is a living document. 🙏

20

2 Comments

(𝟕/𝟕) 𝐇𝐞𝐚𝐥𝐭𝐡𝐜𝐚𝐫𝐞 | 𝐌𝐞𝐝𝐢𝐜𝐚𝐥 𝐈𝐦𝐚𝐠𝐢𝐧𝐠 | 𝐌𝐋-𝐌𝐨𝐝𝐞𝐥𝐬 𝐃𝐞𝐠𝐫𝐚𝐝𝐚𝐭𝐢𝐨𝐧 | 𝐃𝐚𝐭𝐚 𝐚𝐧𝐝 𝐂𝐨𝐧𝐜𝐞𝐩𝐭 𝐃𝐫𝐢𝐟𝐭 | 𝐎𝐧-𝐏𝐫𝐞𝐦𝐢𝐬𝐞𝐬 | 𝐈𝐧𝐯𝐢𝐭𝐚𝐭𝐢𝐨𝐧 𝐭𝐨 𝐜𝐨𝐥𝐥𝐚𝐛𝐨𝐫𝐚𝐭𝐞 𝐏𝐚𝐫𝐭 𝟔/𝟕: https://lnkd.in/eFVG4QgA We hereby invite researchers and ML vendors with experience in model degradation caused by data and concept drift in medical imaging to collaborate (On-Prem, without access to new data, with partial or no labeling of new data). 𝐂𝐨𝐥𝐥𝐚𝐛𝐨𝐫𝐚𝐭𝐢𝐨𝐧 𝐠𝐨𝐚𝐥 To conduct pilot research experiments under various conditions. Up to 5. 𝐃𝐚𝐭𝐚 Open-source medical imaging data from your own or the following sources: TCIA: https://lnkd.in/ebYSv3t7  kaggle: https://lnkd.in/ewCaXwAK arXiv: https://lnkd.in/eEJF_VY9  BIMCV: https://lnkd.in/etPKEvdt  BIMCV: https://lnkd.in/e6PWwY_a  OpenNeuro: https://openneuro.org/  NLM: https://lnkd.in/efFjuazR   GitHub: https://lnkd.in/ecQ5W_SS  GitHub: https://lnkd.in/ep2pZ_Th  MONAI: https://lnkd.in/e9qWTR4v  others 𝐘𝐨𝐮𝐫 𝐫𝐨𝐥𝐞 ° Describe the technical details of your real research or business case (number of features, causes of drift, other details) ° Select from open sources the dataset most similar to your case and provide it to us  ° Inform us of the data distributions between the researcher/ML vendor and the clinic that are close to real-world conditions (what is typical for each side, the proportion of contribution from each party) ° Describe the conditions for labeling availability of new data on the clinic side (0%, 5%, or more labeled) ° Specify information about the class of ML architectures used ° Provide information about image pre- and post-processing (processing pipeline) ° Specify the class of loss function and optimizer that will be used in the model ° Confirm the availability of computational resources on the clinic and ML vendor sides (GPU or CPU available / not available) 𝐎𝐮𝐫 𝐫𝐨𝐥𝐞 ° Simulate both the researcher/ML vendor and the clinic (data owner) by splitting your existing medical imaging datasets into multiple parts ° Simulate model degradation due to data drift on the clinic side. Per the simulation setup, the researcher/ML vendor will not have access to the clinic's data ° Handle the federated fine-tuning component ° Present the experimental results to you in a clear, demonstrable manner ° Cover related costs 𝐅𝐢𝐧𝐚𝐥 𝐬𝐭𝐞𝐩 - 𝐏𝐮𝐛𝐥𝐢𝐬𝐡𝐢𝐧𝐠 We will jointly publish the results as co-authors in a paper or use case description. 𝐏𝐫𝐨𝐛𝐥𝐞𝐦 𝐬𝐮𝐦𝐦𝐚𝐫𝐲 𝐢𝐧 𝐭𝐡𝐞 𝟏𝐬𝐭 𝐜𝐨𝐦𝐦𝐞𝐧𝐭

41

4 Comments

Cyber debt: the hidden risk behind AI speed. In AI, moving fast often wins. But AI security is stuck behind. This creates cyber debt: a new attack surface that grows over time. It’s new risks created : model extraction, model inversion, adversarial examples… Data pipelines without integrity checks Models lacking adversarial testing Models deployed without proper protection Each shortcut speeds delivery today but increases exposure tomorrow. 🌊 Don’t let cyber debt drown your innovation. Build resilience before the wave hits.

17

Thrilling to see Anthropic release this report about how they can use Selective Gradient Masking to eliminate dangerous data in a corpus of an #LLM, but I immediately see how this could also be used to eliminate (or at least mitigate) the use of unlicensed content in a model without deprecating it. While this is early, the hope is that this could enable them to remove all reference to certain data that a creator does not want included, while the curated content for their training data can come from licensed, credible work. What do you think? #EthicalAI #commerciallysafeAI https://lnkd.in/gkCwKnpZ

10

3 Comments

Rise of Structured AI: From LLM Wrappers to Agentic AI Data Workflows Technical Implementation: The Table Foundation Model (TFM) Framework The engineering shift from Large Language Models (LLMs) to Agentic Data Workflows requires a departure from standard tokenization. We are implementing a tripartite architectural stack: Iterative Reasoning over Structured Schemas: Unlike standard RAG (Retrieval-Augmented Generation) which treats documents as flat vectors, our agents utilize iterative reasoning loops to navigate relational hierarchies. The agent decomposes a natural language query into a series of multi-step joins and analytical operations, validating intermediate outputs against the database schema before final execution. Semantic Vector Integration for Row-Level Context: By integrating vector databases directly into the tabular engine, we enable semantic search across structured cells. This allows the model to understand that a "Late Payment" in a financial ledger and a "Delayed Shipment" in a logistics table are semantically linked within a broader risk-assessment graph. Latency-Optimized Inference via Symbolic Translation: To ensure real-time utility, the system bypasses the computationally expensive process of re-encoding entire tables. Instead, it employs a symbolic translation layer that converts high-level agentic intent into optimized SQL or Python scripts, executing directly on the data substrate to minimize overhead and maximize throughput. Economic Impact and Radical ROI The move toward "Structured AI" represents a fundamental re-engineering of the corporate cost structure. By deploying autonomous agents that interpret spreadsheets and sensor logs as unified intelligence sources, the ROI is realized through three specific channels: Elimination of Technical Debt: Consolidating thousands of task-specific models into a single Table Foundation Model reduces maintenance costs and infrastructure complexity by an order of magnitude. Autonomous Predictive Synthesis: Data agents do not just summarize; they hypothesize. By cross-referencing CRM tables with external market statistics, agents generate predictive revenue matrices that allow for proactive capital allocation rather than reactive reporting. Granular Operational Transparency: Through Dynamic Skill-Dependency Graphs, leadership can identify the exact "cell" or node in a project lifecycle that is underperforming, allowing for surgical interventions rather than broad, inefficient policy changes. It is found in the reasoning capability applied to the world’s structured data. #AgenticAI #TabularReasoning #GenerativeAI #MachineLearning #DataEngineering #AIArchitecture #EnterpriseAI #LargeLanguageModels #DataAgents #StructuredData #CloudComputing #DeepLearning #ArtificialIntelligence #SystemDesign #FoundationalModels #DigitalTransformation #OpenAI #MicrosoftAzure #GoogleCloud #AmazonWebServices #MetaAI #CTO #LeadArchitect

6

🧱 “SKILL SYNTHESIS” FOR BETTER LLM REASONING A lot of current agent design depends on clever prompting and fragile chains. This paper explores a different direction: compositional skill synthesis. Instead of relying on one long prompt, you build smaller reusable skills and combine them to solve harder problems. If this approach holds up, it changes product strategy: • Skills become assets you can test, version, and reuse • Orchestration and routing become core differentiators • Evaluation becomes cleaner because you can isolate which skill failed This is the kind of work that can turn agentic AI from demo magic into maintainable systems. If your AI product is struggling today, is the problem the model, or the lack of reusable skills? Sources: - https://lnkd.in/eQV2txcd #ArtificialIntelligence #Agents #LLM #AIResearch

1

DeepSeek AI published a technical paper a few days ago, co-authored by co-founder Liang Wenfeng that argues for a new architecture to train foundational models. Essentially moving away from 'muscle head logic'. As the writer of the article I linked below by Larry Dignan states, The DeepSeek paper is a bit wonky. But remember what DeepSeek did to the market a year ago when they released their LLM - it was very performant and low cost. The AI market to date has been driven by muscle head logic. The answer for everything in the last two years has been more muscle. In AI's case that approach means more GPUs, more farm land going away, more energy usage, more water and by all means more data centers, debt and capital expenditures. This approach and consumption is likely to change with deep technical architectural change including at the chip level. Price and performance matter. LLM giants are diversifying their chip bases. Google's TPUs and Amazon Web Services' Trainium are going to be threats to Nvidia as inference becomes the main event. It will be really interesting to see how the industry shapes up here - lots to follow and very interesting! News Article here with link to DeepSeek paper: https://lnkd.in/eN38mSAV

6

3 Comments