TLDR AI 2025-05-08

82% of organizations intend to integrate AI agents within 1-3 years.

According to Capgemini research, while only 10% of organizations currently employ AI agents, a large majority (82%) intend to integrate them within one to three years.The time is now for leaders to invest in the right AI platform — one that seamlessly integrates analytics, models, and agents — to maximize value with what's next. Learn more in this C-suite guide to AI agents here.

Claude has been updated with web search support via API, enabling developers to build applications that incorporate real-time, up-to-date information from the internet.

Apple is considering integrating AI search engines like Perplexity and Claude into Safari, reflecting growing momentum behind LLM-powered search.

Amazon's Vulcan is the company's first robot. It has a sense of touch, enhancing handling precision and marking a new phase in physical AI capabilities.

Researcher Sampatt tested O3, an AI agent, against a GeoGuessr expert in the challenging street view geography game. O3 demonstrated strong reasoning and geographic inference but still fell short of consistently outperforming the human master. The experiment highlights both AI's growing spatial understanding and its current limitations in complex real-world tasks.

AI systems like Claude are dramatically impacting coding by automating significant portions of programming tasks. Startups are leading in the adoption of AI coding tools like Claude Code, which is heavily used for front-end development, while enterprises lag behind. As AI capabilities grow, the nature of developer roles may shift towards managing AI systems, potentially accelerating technological advancements.

The Mahmood Lab, using Meta's DINOv2, has developed open-source AI models for pathology, improving disease detection and diagnostics.

Researchers introduce a LoRA-based fine-tuning method for code search that reduces trainable parameters below 2% while improving retrieval accuracy by up to 9.1% for Code2Code tasks.

Google researchers have used LLMs to simplify complex text without losing critical details, improving user understanding while preserving accuracy and nuance.

A new initialization technique, IDInit, ensures stable convergence in deep neural networks by maintaining identity transitions in both main and sub-stem layers.

PyTorch has evolved from a research framework to a foundational platform that powers generative AI. The PyTorch Foundation was expanded to integrate complementary projects and support scalable AI development.

Mastercard's AI program streamlines e-commerce searches, reducing time and friction. Consumers retain purchase control as AI agents can't finalize transactions.

Meta's new AI app introduces a Discover feed that adds a social media twist by allowing users to share their interactions with the AI. This feature aims to demystify AI by enabling comments, likes, and remixing of shared AI posts. The app, which replaces the View app for Meta Ray-Ban glasses, uses a Meta-tuned Llama 4 model and provides advanced voice interaction in select regions.

Meta introduced the AI Defenders Program to help developers detect and prevent misuse of LLaMA models.

OpenAI and the FDA held meetings to explore how AI could expedite drug approvals, signaling a shift toward modernizing regulatory science with machine learning.

Cluely claims its hidden window enables undetectable cheating, sparking competition from startups like Validia and Proctaroo, which say they can identify Cluely users.