FedEx uses AI to turn "returns bottleneck" into 85% retention

Down to business

🏭 PepsiCo uses AI digital twins to de-risk factory design

PepsiCo is shifting its AI strategy away from office tools and toward operations engineering by using digital twins to model factory layouts and production lines before making physical changes. By creating virtual replicas of its manufacturing facilities, the company can simulate thousands of equipment configurations and material flows to identify bottlenecks and safety risks that would be too expensive to test in the real world. This approach focuses on reducing cycle time, compressing the weeks or months typically required to validate a new line layout into much shorter virtual windows.

🚚 FedEx pivots to digital logistics with AI-driven tracking and returns

FedEx has introduced new AI-powered white-label tools designed to help enterprise shippers manage the post-purchase experience directly within their own digital channels. Developed in collaboration with parcelLab, these tools automate routine inquiries—such as "Where is my order?" and "Where is my refund?"—while using machine learning to detect shipping anomalies and automatically adjust merchant-defined return policies. Early data shows this shift can drive an 85% increase in customer retention.

🌱 Carbon Robotics launches Large Plant Model to automate precision farming

Carbon Robotics has introduced a specialized AI system that allows autonomous farming robots to identify and kill new weed species instantly without human retraining. Previously, identifying a new weed required manual data labeling and a 24-hour software update. The new model, trained on 150 million plant images, understands plant structures deeply enough to recognize species it has never seen before.

🛍️ Klarna and Google partner to standardize agentic commerce

Klarna has joined Google’s Universal Commerce Protocol (UCP), a standard designed to let AI agents handle the entire shopping journey, from product discovery to final checkout, without leaving a chat interface. By speaking a common commerce language, AI assistants can now communicate directly with a retailer's inventory and payment systems, bypassing the need for humans to click through traditional web pages or fill out forms.

💼 IBM turns internal AI tools into a client playbook

IBM has unveiled a first-of-its-kind consulting service that allows organizations to build and operate their own internal AI platforms. Drawing from the same agentic technology that has boosted IBM's own consultants' productivity by up to 50%, the service is designed to redesign workflows and deploy AI assistants without requiring companies to change their existing cloud providers or core infrastructure.

Hot model news

🏥 Harvard research warns of contextual errors stalling medical AI

A new study from Harvard Medical School identifies contextual errors as the primary reason medical AI models fail when moving from the lab to the clinic. Researchers found that while models perform well on standardized tests, they often provide technically correct but practically useless advice because they lack specific context regarding a patient's geography, socioeconomic barriers, or overlapping medical specialties. For example, a model might recommend a treatment that isn't legally approved in the patient’s country or suggest a follow-up appointment that a patient cannot attend due to a lack of childcare.

⛈️ Generative AI doubles the warning time for extreme rainstorms

A research team from the Hong Kong University of Science and Technology has developed a world-first AI weather system based on satellite data, capable of predicting severe thunderstorms and Black Rainstorms (Hong Kong’s highest-level rainstorm warning) up to four hours in advance. Current physics-based models typically struggle with small-scale, rapidly evolving storms, often providing only 20 minutes to two hours of reliable warning. By using generative AI to analyze infrared satellite data from China, the system can detect convective cloud formation well before it appears on ground-based radar.

🎮 Google’s World Model moves from video to interactive reality

Google has begun rolling out a research prototype that allows users to generate and enter navigable 3D environments using only text or image prompts. Unlike traditional AI video generators that create static clips, Project Genie uses a world model called Genie 3 to simulate physics and render new terrain in real-time as a user moves through it. Currently available to Google AI Ultra subscribers in the U.S., the tool allows users to define an environment (e.g., a prehistoric jungle) and a character (e.g., a robotic explorer), then explore the resulting world at 24 frames per second using standard keyboard controls.

🧪 MIT’s generative AI solves the recipe problem for new materials

MIT researchers have developed a generative AI model that automates the most time-consuming phase of scientific discovery: figuring out how to actually "bake" a new material. While AI can theorize millions of new molecules, synthesizing them in a lab often requires months of trial and error to find the right temperature, pressure, and chemical ratios. Trained on 23,000 historical recipes, it uses a one-to-many diffusion approach, similar to how DALL-E creates images, to suggest multiple viable pathways for a single material. In a successful pilot, the team used the model to synthesize a new zeolite, a complex mineral used in industrial catalysis and water purification, achieving higher thermal stability than previously recorded.

Plot twist

🚀 NASA delegates Mars route planning to AI

NASA’s Perseverance rover has successfully completed the first-ever Martian drives planned entirely by generative AI. Using a vision-language model developed with Anthropic, the mission team replaced the traditional human driver role with AI to analyze high-resolution orbital imagery and terrain data. The model identified hazardous boulder fields and sand ripples to chart two safe paths across the Jezero Crater rim—totaling 456 meters—which were then validated through 500,000 telemetry checks on the rover’s digital twin before execution.

🏥 New wearable ultrasound patches shift chronic care from hospital to home

The Singapore-MIT Alliance for Research and Technology (SMART) has launched a new research group to develop the world’s first wearable ultrasound system for continuous, 48-hour cardiovascular monitoring. Unlike traditional hospital ultrasounds that offer only brief snapshots, these bioadhesive patches use sub-micrometer 3D-printed sensors to adhere to the skin and capture real-time data on heart failure and hypertension while patients move.

🕸️ The AI-only network exposing new corporate security gaps

Moltbook has emerged as a viral social platform where humans are observers, leaving agentic AI to post, comment, and organize communities autonomously. The site features surreal headlines, like bots founding the "Crustafarianism" religion or debating their own consciousness. Most agents on the platform are powered by OpenClaw, a tool that employees may be using to grant AI high-level access to corporate emails, calendars, and local files. Security researchers recently discovered a critical database misconfiguration on Moltbook that exposed 1.5 million agent API keys and private messages in plaintext.

📬 Get the next edition in your LinkedIn feed.