MIT report warns of writing tool dependency

🎉 🎉 Thrilled to share that our paper “FedSVD: Adaptive Orthogonalization for Private Federated Learning with LoRA” has been accepted to NeurIPS 2025! Paper: https://lnkd.in/gjm6x6gf Thanks to co-authors: Sangwoo Park, Dong Bok Lee, Dominik Wagner, Haebin Seong, Tobias Bocklet, Juho Lee, Sung Ju Hwang TLDR; We propose a simple SVD-based reparameterization that adapts LoRA while avoiding DP-SGD’s noise amplification—leading to more stable, higher accuracy FL.

our last publication, entitled "Deep Signals: Enhancing Bottom Temperature Predictions in Norway’s Mjøsa Lake Through VMD- and EMD-Boosted Machine Learning Models", with Zeliha Selek, Sertac Oruc, and Mehmet Ali Hınıs, just released. https://lnkd.in/dJ6_w3Eq

How can you equip students with rigorous research skills in an AI-enhanced world? Through a subscription to ProQuest Historical Newspapers: Global, institutions gain access to an unparalleled collection of over 200 premier newspaper titles — providing the continuity, depth and breadth necessary for meaningful primary source research that complements digital tools. Learn more: https://ow.ly/LSJN50WXBQn

NEWS: Anthropic just launched 'Claude Chrome'. This feels like the moment we’ll look back on as when AI integration shifted from “nice to have” to “competitive necessity” across every industry. With an 11% attack success rate, it’s too risky to pilot at the work place … but whoa this is exactly the kind of breakthrough that’s going to reshape how we think about AI in the enterprise. Watching this space closely. https://lnkd.in/erXkhNhU

Scaling RAG to large knowledge bases comes with a challenge: chunks often lose context. A line like: “The company’s revenue grew by 3% over the previous quarter” is meaningless without knowing which company or quarter. Anthropic proposed combining contextual retrieval and reranking to boost retrieval accuracy. Adding chunk-specific context before embedding, combining retrieval with BM25 TF-IDF similarity, and then reranking the top N transforms retrieval: - Contextual embeddings reduced the top-20-chunk retrieval failure rate by 35%. - Combining contextual embeddings with contextual BM25 reduced the top-20-chunk retrieval failure rate by 49%. - Adding reranking post context retrival reduced the top-20-chunk retrieval failure rate by 67%. The result: precise, context-aware retrieval that balances exact matches and semantic understanding. https://lnkd.in/dPqYSa8v

An interactive portal to explore the current state, future potential, and societal impacts of Artificial General Intelligence (AGI). Get an objective perspective on AGI by separating the scientific facts from the hype. Start Exploring 🚀 | View References 📚 https://lnkd.in/dz9zqzMf

Part 3 of my RAG Notes is here: This post explores how knowledge graphs (built from entities, relationships, and evidence paths) push retrieval beyond vectors, making it more transparent and powerful. You can read it here: https://lnkd.in/dw8ahMmK

Next Wednesday, as part of the swyx and Eugene Yan paper club, I will go over the Vision behind GPT-OSS. This paper has many tips for efficiently handling post-training LLMs to achieve competitive performance compared to commercial models. Link to the event: https://lnkd.in/eJYRs6Rn Link to the paper: https://lnkd.in/ebaEGsmN

🚨 Interesting article on AI and democracy by Luciano Floridi & Anna Ascani "Debates on AI governance often focus on regulating risks. This article shifts perspective to examine how AI can augment democratic processes, presenting a critical analysis of the Italian Chamber of Deputies ’pioneering AI initiative" "Using this initiative as a critical case study, we move beyond merely applying the concept of “augmented democracy”. We argue that this real–world implementation reveals key tensions between the goals of efficiency and the preservation of deliberative friction essential to democratic practice." Jasper Kars Elja Daae Paul Suijkerbuijk Monika Milanović Marten Tilstra Maranke Wieringa Jana Mišić Jonne Maas Derkjan Z. Martijn Bekking Medina Gasjimova Jurriaan Parie Ylja Remmits Kars Alfrink 🟥

Interesting initiatives...dare one call them controlled experiments? "The vision of “augmented democracy” offers genuine promise as AI could help parliaments process complexity, engage citizens, and craft better legislation... However, realising this potential requires acknowledging that technology alone cannot solve fundamentally political challenges. AI may augment human capabilities, but democratic activities and developments ultimately depend on citizens ’willingness to engage, representatives ’commitment to serve, and institutions ’capacity to evolve."  Luciano Floridi Anna Ascani