AI & ML Research 3 Days

Last updated: July 25, 2026, 2:30 AM ET

AI & ML Research

Organ Preservation and Medical AI

Researchers are making significant strides in organ preservation, aiming to extend the viability of organs outside the body to address donor shortages. Supercooled kidneys have been successfully transplanted into pigs in a "landmark achievement", offering hope for improved transplant outcomes. In parallel, AI is being integrated into healthcare, with Chat GPT launching a "Health" feature that allows users to securely connect medical records for personalized insights. Google AI is also developing Symptom AI, a conversational agent for everyday symptom assessment, and AI is accelerating drug discovery by aiding scientists in designing new medicines.

Advancements in LLM Architectures and RAG Systems

New developments are enhancing the capabilities of Large Language Models (LLMs) and their application in Retrieval Augmented Generation (RAG) systems. Tabular foundation models are now outperforming traditional gradient-boosted trees on benchmarks for predicting missing spreadsheet data. For RAG, a "Loop Engineering" approach is being explored to create LLM cascades, moving from cheaper local models to more powerful hosted ones, incorporating validation loops. Addressing RAG errors, a distinction is made between extraction errors and hallucinations, proposing a "typed generation contract" to improve accuracy. Further work on RAG involves iterating on top-k retrieved candidates and using a sufficiency signal to select the best one.

LLM Memory, Agents, and Runtime Development

Efforts are underway to improve LLM memory and agent performance, alongside tools for building custom inference runtimes. One researcher has developed a "usage-reinforced decay engine" for AI agent memory, inspired by the Ebbinghaus forgetting curve, to prioritize important information over recent data. Challenges in scaling LLM agents are being highlighted, with one instance showing that adding more agents actually slowed down the system due to hidden costs in asynchronous tasks. For those looking to build their own LLM infrastructure, a guide is available on creating an LLM inference runtime from scratch on an H100 GPU. Additionally, a walkthrough demonstrates how to build an LLM agent capable of writing and executing code using the OpenAI Agents SDK and Docker. Security for AI agents is also a focus, with "Skill Spector" being developed for detecting vulnerabilities in agent skills through static analysis.

AI Infrastructure and Scientific Discovery

Major commitments are being made to advance AI infrastructure and accelerate scientific discovery. Google is investing $40 million in AI tokens and credits for the Genesis Mission to push the frontiers of scientific discovery commits $40M. OpenAI is announcing "Project Camellia" in Effingham County, Georgia, focusing on responsible energy, community investment, jobs, and access to Codex. Furthermore, OpenAI is outlining its commitment to advancing American science by collaborating with the U.S. Department of Energy and national labs to leverage frontier AI for accelerating discovery.

Broader AI Applications and Lessons Learned

Beyond core research, AI is being applied in diverse fields, with lessons being drawn from years of machine learning experience. News organizations are adopting AI tools to enhance reporting, grow audiences, and optimize operations using AI. A reflection on 8.5 years of ML experience emphasizes the importance of patience, optimism, discipline, project focus, and effective teamwork. The complexities of data science are also explored, with an anecdote illustrating the significant disparity between an $8 million resolution and a $5,000 settlement in a flight overbooking scenario.