Renato Iannella

When an abuser can turn your holiday selfie into a convincing sexualised image in minutes, the harm isn’t hypothetical - it’s happening right now. Deepfake abuse has become a weapon in technology-facilitated domestic violence and coercive control. Everyday photos, stolen from social media or shared in trust, are being manipulated to shame, intimidate, and silence. For victim-survivors, it’s not only humiliating - it’s terrifying. The CSIRO’s new image-protection technique is a step in the right direction. By subtly altering pixels in a way humans can’t see but AI can’t use, it could make photos unreadable to the models that generate deepfakes. If platforms embedded this protection into every uploaded image, many abuses could be stopped before they even start. But technology alone won’t solve this. We need three things working together: 1. Platforms adopting protections at scale. 2. Strong laws that criminalise the abuse. 3. Survivor-centred support systems to help those already harmed. This is what “safety by design” looks like - and why innovation like this needs to move from research into real-world adoption, fast. #OnlineSafety #TechSafety #TFDV #CoerciveControl #TrustAndSafety #SafetyByDesign eSafety Commissioner Cybersecurity Advisors Network (CyAN)

4

RAG Systems Are Evolving And They Are Becoming Agentic. Most Retrieval Augmented Generation (RAG) systems work well until they do not. Once you introduce multiple design dimensions like retrieval strategy, chunking, prompts, agents, workflows, and evaluation metrics, the system becomes too complex to tune by hand. That is why I found Fareed Khan’s recent article on self improving agentic RAG systems so compelling. The core insight: Do not build a static pipeline. Build a system that can analyze its own weaknesses and evolve in response to real world performance. Here is the model: 1. Specialist agents Instead of a single LLM call, use a team of role based agents such as a retriever, generator, evaluator, diagnostician, and architect. Each focuses on one part of the workflow. 2. Multi dimensional evaluation Responses are scored across accuracy, feasibility, compliance, clarity, cost, and more. This produces a performance vector instead of a single metric. 3. Root cause diagnosis A diagnostician agent identifies what is going wrong, such as retrieval issues, hallucinations, missing context, or prompt failures. 4. SOP evolution An architect agent proposes variations to the system SOPs. The system tests them, compares performance, and moves toward the Pareto optimal configurations. 5. Continuous improvement loop As production data changes, the system adapts instead of drifting. Why this matters: In complex domains like healthcare, legal, enterprise knowledge, and customer support, traditional RAG systems break down under real world variability. Agentic RAG can lead to: More accurate results More resilient pipelines Continuous learning without manual re tuning Scalable knowledge workflows The future of RAG is not retrieval plus generation. It is retrieval plus generation plus agents plus evaluation plus evolution. If you are building AI products or internal knowledge tools, this shift is worth paying attention to. #AI #RAG #AgenticAI #GenerativeAI #MachineLearning #EnterpriseAI #AIDevelopment #LLM #ArtificialIntelligence #FutureOfWork #TechInnovation #KnowledgeManagement #HealthcareAI #AIEcosystem #AITools

7

Image resolution matters more than we think for document search — here’s a practical guide from experiments from JINA.AI. Finished reading this (another detailed) article on how image resolution affects retrieval of visually rich documents (infographics, slides, tables, maps) using jina-embeddings-v4, a VLM. As usual, like their other articles, this is with practical experiments and clear, actionable takeaways. 𝐏𝐫𝐨𝐛𝐥𝐞𝐦: OCR or image-caption methods miss text+layout semantics in complex docs. 𝐀𝐩𝐩𝐫𝐨𝐚𝐜𝐡: jina-embeddings-v4 feeds image patches into a transformer alongside text and produces single- or multi-vector embeddings. Experiments talk about how resolution impacts retrieval quality and cost, benchmarks on ViDoRe v1/v2 and high-res historical datasets, across many resolutions and resolution mixes. 𝐊𝐞𝐲 𝐟𝐢𝐧𝐝𝐢𝐧𝐠𝐬 • No single optimal resolution: some datasets peak at ~1.2MP, others at much higher (e.g., 4.8MP). • Multi-vector embeddings beat single-vector for retrieval. • Multi-resolution mixes (several mid-range sizes) often outperform any single size. • Too small = unreadable text; too large = noisy, costly • Not OCR-first: embeds visual patches in-context, preserving spatial relations. • Not CLIP-style: captures fine-grained text semantics inside images. • Novel view: image patches treated like text tokens, enabling multi-resolution fusion and late interaction scoring. 𝐖𝐡𝐲 𝐢𝐭 𝐦𝐚𝐭𝐭𝐞𝐫𝐬 - Better search across complex documents (tables, charts, infographics). - Practical guidance to balance retrieval accuracy and compute/storage. Multi-resolution is a robust path forward; automating optimal-resolution selection could simplify adoption. Code and benchmarks are available in the JinaVDR repo if you want to test it on your data. Link: https://lnkd.in/gFTSzcYT

20

2 Comments

Graph foundation models for relational data. Not for right now, but could end up being important for Housing where data structures are abundant but delivering insights is a challenge... https://lnkd.in/g5DN-Nym

9

4 Comments

New research tackles the slow generation speed of high-quality video diffusion models! Transition Matching Distillation (TMD) efficiently distills these complex models into faster, few-step generators. This means potentially real-time video creation is closer than ever. TMD focuses on matching the denoising *process* itself, not just the final output, for superior results. A significant step towards interactive video applications! 🚀 #VideoGeneration #DiffusionModels #AI #MachineLearning #DeepLearning #ComputerVision #TMD #TransitionMatching #Distillation #RealTimeVideo #InteractiveAI #VideoAI #GenerativeAI #AIResearch #AI #ML #DeepLearning #DataScience #Innovation #TechTrends #AIResearch #NeuralNetworks #BigData #Automation #FutureOfWork #DataAnalytics #Research #SmartTech #EnterpriseAI Paper: https://lnkd.in/g9ZJsiYz

3