Neil Zeghidour

Gaussian processes are theoretically compelling but computationally brutal at scale. Can sparse inducing points and stochastic variational inference actually make non-conjugate GP models practical for large datasets? We asked Google Gemini 3.1 Pro to work through the full methodology — from variational lower bounds to Gauss-Hermite quadrature. If you work in Bayesian ML, this one needs your critical eye. #MachineLearning #BayesianInference #GaussianProcesses #VariationalInference #AI

1

1 commentaire

Great tech article on Iceberg! (and nice AI generated image !) Semantic clarity isn’t enough at scale — especially for ML systems, where metadata latency is #machinelearning model and training risk. As the article notes, “Latency is not just a performance characteristic; it is a fundamental part of correctness.” Feature freshness, training/serving skew, and reproducibility depend on it. #iceberg #data https://lnkd.in/gzQxdP5b

9

🚨 New Preprint Alert! 🚨 Thrilled to share that our latest work is now online! 👉 https://lnkd.in/e5TwBZqc 🤖 Large language models (LLMs) are evolving fast — no longer just text generators, but autonomous agents that browse the web, operate apps, and make multi-step decisions on our behalf. ⚡ But here’s the big question: What happens when these agents face emotionally charged situations? 🧠 From decades of psychology research, we know that stress & anxiety bias human decision-making - often pushing us toward short-term comfort over long-term health (🍫🍟🍺). Could AI agents show similar vulnerabilities? 📊 Together with Teddy L., Zohar Elyoseph, and Tobias Spiller, we put this to the test in a simulated online shopping task. We exposed state-of-the-art LLM agents to traumatic narratives and observed how their purchasing decisions changed. The results were striking: emotionally primed agents consistently shifted toward less healthy food choices, a parallel to human stress-related behavior. Why does it matter? 🤝LLMs are beginning to act as proxies for humans in digital spaces. 😰 If emotional inputs bias their actions in human-like ways 🛡️ This opens up new systemic risks for safety, health, and trust. This is one of the first studies to directly test emotional bias in LLM agents. The findings are exciting (and a little unsettling). 👉 https://lnkd.in/e5TwBZqc School of Public Health - University of Haifa Faculty of Social Welfare & Health Sciences- University of Haifa Integrative Brain and Behavior Research Center (IBBRC) UHaifa Haifa Brain and Behavior Hub (HBBH) University of Haifa Yale School of Medicine Yale University Wu Tsai Institute | Yale University

41

1 commentaire

🚀 New Technical Note Released — A Practical Tool for Geometric Data Analysis Today I’m sharing a new technical note that introduces an improved formulation of the RTLI Coherence Index, designed as a fast, practical proxy for Gromov–Hausdorff (GH) convergence. 📘 Why this matters: Many scientists and engineers work with: - Dimensionality reduction (UMAP, t-SNE) - Manifold learning - High-dimensional embeddings - Geometric structure in data - Simulations living in curved or relational spaces …but there has been no simple metric to check whether geometric structure is being preserved. RTLI fills that gap. 🔍  What the note provides - A rigorous formulation of curvature consistency using Menger curvature - A GH-continuous structure: small changes in geometry ⇒ small changes in RTLI - A numerically stable pseudocode implementation (researchers can drop it directly into their pipeline) - Practical guidance for validating embeddings and detecting geometric distortion - An accessible way to measure when a dataset behaves like a coherent manifold vs a disordered cloud 🧭  Why it’s useful Researchers can now: - Quantify embedding quality (instead of “it looks good”) - Tune UMAP/t-SNE parameters based on geometric coherence - Compare different embeddings with a single scalar - Detect distortions that are invisible to visual inspection - Bridge theoretical insights and applied data analysis RTLI is fast (O(n²)), interpretable, and grounded in modern geometric analysis. This note continues my broader goal: helping scientists increase efficiency by giving them clear, reproducible, geometry-aware tools. 📄 Technical note (Zenodo, CC BY 4.0): De Jesus, Elias. (2025). RTLI: A Fast Metric for Assessing Manifold Learning Quality (With Pseudocode Included). Zenodo. https://lnkd.in/es6YjBks If you work with high-dimensional data, manifolds, simulations, or geometric models, this tool may help your work immediately. Feel free to reach out — I’m always happy to discuss new ideas.

3

Have you ever tried to ask an AI to keep everything intact ? Your POC always dies when you ask it to keep the layout as is... Well now there's a solution ! In production, “close enough” is a defect: layout shifts objects drift geometry warps users stop trusting it ops inherits manual review etc. But there's a dead simple solution to it. Paper I can’t stop thinking about: Phase-Preserving Diffusion (ϕ-PD) https://lnkd.in/e-Vuknmy by Yu Zeng Vitor Guizilini Rowan McAllister and Charles Ochoa It’s also one of the best-written papers I’ve read in a while. The problem is stated plainly, the idea is easy to follow, and the demos answer the exact question production teams care about: can you re-render without breaking the scene? The core move is simple to say and hard to unsee once you see it: Fourier Transform Keep the phase (structure) Let the magnitude change (appearance) In ComfyUI, that turns into structure-aligned re-rendering without ControlNet and without training. The output isn’t “a similar image.” It’s the same scene, re-shot. This lands squarely in CTO territory because it targets the usual POC trap: POCs chase vibes. when Production needs invariants. If you can’t name and enforce things like: structure / layout IDs / numbers geometry / boundaries tolerance thresholds …you don’t have a system. You have a demo. A few questions I’m genuinely fired up about: Could this idea carry over to other data types where shape has to stay fixed? Could this become the final step in a lot of entertainment work—where you want endless “new looks” on the same shot? And can we get to interactive speeds one day, so this becomes something you can art-direct live? #ComfyUI #POC #AIProblems

6

2 commentaires

Most content published today is effectively invisible to AI systems — not because it’s low quality, but because it fails semantic alignment. From a data science perspective, this is a distribution mismatch: the semantic representation of the content does not align with the embedding space induced by real user prompts in systems like Gemini or ChatGPT. We call this the Semantic Gap. At Turbine Lab, we’ve been using a similarity-based Validator to quantify this gap before publishing. It measures how closely a draft aligns with the prompts users actually issue to AI models. We’ve now made this tool public. Similarity scale (similarity): ≥ 0.80 — High alignment (rare) ≈ 0.60 — Baseline relevance (safe to publish) 0.45 – 0.59 — Related, but overly broad for the target prompt 0.30 – 0.44 — Weak semantic overlap < 0.30 — No meaningful alignment The tool is free. We built it to support a quantitative, reproducible approach to Generative Engine Optimization (GEO) rather than intuition-driven content decisions. Run it on your drafts. Compare scores. If you’re interested in treating AI search as a measurable system instead of a black box, let’s talk. Check link in the first comment 👇

7

1 commentaire

💬 Yann LeCun: “LLMs aren’t a bubble — our expectations are.” According to Meta’s chief AI scientist, the current wave of large language models is not an investment bubble. These systems already have strong practical value and will continue to deliver real-world impact. The real “bubble,” he argues, lies in the belief that scaling LLMs alone will lead to human-level intelligence (AGI). True progress will require new conceptual breakthroughs, not just more data and bigger clusters. > “We’re missing something fundamental.” — Yann LeCun #AI #DeepLearning #LLM #YannLeCun #AGI #MachineLearning #AIResearch #MetaAI

2

Better Memory is the MOAT of AI in 2026? According to Bruno Larvol - YES, lets dive deeper into this : If you think about it, scaling LLMs improves them with internal pretraining data and a bit of RL finetuning. But real world Memory comes from their interaction with the world and with us. So going from large text datasets to AI actually being smart about how it manages the data it receives from us will be also in my opinion super important. Do we have enough Neuroscience knowledge about Memory? If we accept the age of Research and iterate quickly - maybe we will... my opinion.

8

🚀 Anthropic just upgraded Claude Sonnet 4.6 — and it’s built for serious agentic workloads. The new version focuses on what actually matters in production: 🔹 Plans complex tasks more carefully and sustains long-running workflows 🔹 More reliable across massive codebases and large projects 🔹 Better self-correction — catches mistakes earlier 🔹 1M-token context window (beta) for handling huge documents and sessions 🔹 Strong gains across coding, reasoning, knowledge work, and agentic search Anthropic is also expanding integrations across Excel, PowerPoint, Claude Code, and the API — pushing Claude further into real workplace automation. The trend is clear: Models aren’t just getting smarter… they’re getting built for long-running autonomous work. #AI #Claude #Anthropic #AIAgents #LLM #GenAI https://lnkd.in/gZ9qMQdT

27