AI & ML Research 3 Days

Last updated: July 24, 2026, 11:30 AM ET

AI Agents and RAG Systems

Researchers are exploring methods to improve Retrieval Augmented Generation (RAG) systems, addressing issues like hallucination and memory decay. One approach, "Loop Engineering for RAG Generation," details a cascade from cheap local models up to flagship hosted models, focusing on cost and a validation loop. Another paper on RAG as extraction errors rather than true fabrications, proposing seven patterns for a typed generation contract to mitigate this. To combat context window limitations and memory issues in AI agents, a "Usage-Reinforced Decay Engine" was built to prioritize important information over simply the newest. However, scaling these agent systems can lead to unexpected performance drops; one investigation actually slowed down their system due to tiny CPU tasks becoming a bottleneck. Further work on RAG context engineering to prevent hallucinations, arguing that these systems often faithfully answer the wrong context. For those looking to build their own AI infrastructure, a guide from scratch, detailing the process of packing weights and capturing CUDA graphs on an H100. Additionally, a method for that can write and run code is presented through a hands-on walkthrough using the OpenAI Agents SDK and Docker. Security within these agent systems is also a concern, with a tool called "SkillSpector" in agent skills, though it highlights the gap between static analysis and real-world security judgment.

AI in Science and Healthcare

AI is significantly impacting scientific discovery and healthcare advancements. Google is committing "$40 million in AI tokens and credits" for the Genesis Mission to accelerate scientific frontiers. A new development in organ donation showcases "supercooled kidneys" in a landmark achievement, offering potential breakthroughs in organ preservation. AI is also being leveraged to "design the next generation of medicines" more efficiently, a process that is traditionally expensive and failure-prone. In healthcare, Chat GPT is launching a new feature called "Health in Chat GPT" that lets U.S. users securely connect medical records for personalized insights. Google AI is also developing "SymptomAI," for everyday symptom assessment. Looking towards future computing, researchers are working on "a quantum computer that learns from its errors" to improve accuracy. OpenAI is also deepening its commitment to national science initiatives, outlining its plan to "advance American science" by working with the U.S. Department of Energy and national labs to utilize frontier AI for accelerated discovery.

Enterprise AI and Infrastructure

Businesses are increasingly adopting AI to automate tasks and improve operations. OpenAI is introducing "OpenAI Presence," a proven enterprise AI agent platform designed for deploying voice and chat agents in customer and internal workflows. News organizations are finding ways to "use AI to advance their vital missions" by strengthening reporting and operations, with OpenAI tools supporting journalists. NTT DATA Group is already seeing benefits, having "cut incident analysis to 30 minutes", enabling secure AI adoption for its 9,000 employees. To support smaller businesses, OpenAI has launched the "Chat GPT for Small Businesses program," AI skills and grow. In terms of infrastructure development, OpenAI is also announcing "Project Camellia in Effingham County, Georgia," focusing on responsible energy and community investment for building AI infrastructure. Meanwhile, related to energy infrastructure, a "power line that could reshape New York’s grid is hitting snags" as it faces development challenges.

Machine Learning Engineering and Lessons Learned

Experienced professionals in machine learning are sharing their insights and practical advice. After "8.5 years of ML," include patience, optimism, discipline, focusing on projects, and effective teamwork. The development of more efficient AI systems is also a focus, with efforts to create "homegrown Chinese chips" aiming to reduce reliance on foreign technology. While exploring new technologies, the complexities of data science can sometimes lead to unexpected outcomes, as illustrated by a story of "an overbooked flight". For engineers looking to understand the underlying technology, a guide offers insights into "how to build your own LLM runtime from scratch" on an H100 GPU.