TLDR AI 2025-03-28

The early days of search were heavily keyword-based. Later, ML and NLP enabled search engines to interpret the semantics behind queries, allowing for more accurate and contextually relevant results.

Today search is changing again with the rise of AI and especially genAI — and companies that sleep on these changes risk falling behind.

This white paper by Algolia explores how AI has transformed both the front-end and back-end landscapes of search technology, including:

• The evolution of search
• The current state of AI-driven solutions
• Challenges and opportunities in this space
• The specific impact of AI on front-end and back-end systems

Download the white paper →

OpenAI has outlined its ongoing efforts to support AI-native cybersecurity research. They now include microgrants and expanded areas like model privacy and agentic security.

Two new studies from Anthropic trace internal pathways in Claude models, uncovering planning behaviors, shared conceptual spaces across languages, and instances of fabricated reasoning.

Anthropic's latest report revealed how Claude 3.7 Sonnet is used across coding, education, and healthcare tasks. It offers new insights into how AI augments or automates different occupations.

A great paper that talks through some of the challenges and unsolved problems facing the field when trying to understand how knowledge and skills are represented internally in a language model.

Mixture-of-Mamba introduces modality-aware sparsity to state-space models (SSMs), enabling efficient multi-modal pretraining. It achieves similar performance to Transformers with significantly lower computational cost across text, image, and speech modalities.

This paper introduces two methods that use diffusion models to generate counterfactual explanations in image regression tasks, highlighting trade-offs in sparsity and quality between pixel and latent space approaches.

Over 20K attendees roamed downtown San Jose for 2025 GTC last week; AI developers learned from Bria.ai How to build predictable, controllable, on-brand visual content at scale while respecting artists' rights.

See a few of the key topics they covered:

• The Future of a Sustainable Economy Around Visual Generative AI is Built on Licensed Data, Not Scraped (a.k.a stolen) Data.
• Visual AI Cookbooks (GitHub projects) For Visual Generative AI Model Customization
• Explore All Bria APIs for Content Creation

Get your free API token to play with Bria today, with 1000 credits on your new account.

PET-MAD is a universal interatomic potential model trained on the MAD dataset that is capable of predicting energies and forces for complex materials modeling across the periodic table.

A new metric that uses attention maps to quantify internal bias in vision models. It is capable of identifying confounding features and going beyond group-based accuracy disparities.

Guidance-Free Training (GFT) removes the need for Classifier-Free Guidance (CFG) in visual generative models, reducing computational cost without sacrificing performance. Unlike distillation-based approaches, GFT trains from scratch and requires minimal modifications to existing codebases.

Inpainting support for the new powerful diffusion model Flux.

This post walks through how to fine-tune high-performing reranker models using Sentence Transformers 4.0, outperforming many large public rerankers with smaller, specialized versions.

Aardvark Weather, developed by the University of Cambridge, offers rapid AI-driven weather forecasting on desktop computers while needing less data and computing power than traditional models. It surpasses conventional systems like the U.S. GFS in accuracy across multiple metrics and provides customizable, location-specific predictions. The system's accessibility democratizes forecasting, making it viable even in regions with limited infrastructure.

Follow this step-by-step tutorial to install Ollama locally, run it in a container, connect to Kibana, and run local RAG and inference with DeepSeek R1. Read the blog

Cornerstone is transforming enterprise training with XR and the Llama 3.1 model, enabling the fast creation of personalized simulations that feature intelligent virtual mentors and real-time multilingual capabilities.

Browser Use, a Y Combinator startup, enhances AI agent navigation on websites by converting site elements into a text-like format.

Raj Aggarwal is leaving AWS, where he served as GM of generative AI, to start a new company.