AI & ML Research 3 Days

Last updated: July 20, 2026, 5:30 PM ET

AI Safety and Bias

OpenAI shared lessons from deploying long-running AI models, detailing new safety risks and improved safeguards. Researchers have found that AI than humans to exhibit bias during the hiring process, potentially impacting résumé screening. This bias is not inherent but learned from the data AI models are trained on.

LLM Engineering and Deployment

Towards Data Science explored several practical applications and engineering techniques for large language models. One article demonstrated how to run Claude code agents for over 24 hours, aiming to boost engineer productivity. Another discussed "Loop Engineering" with adaptive parsing for handling flat tables via Azure and figures using vision LLMs as a last defense. Further "Loop Engineering" insights covered parsing PDF documents cheaply by starting with simpler methods and escalating to heavier parsers only when necessary for complex pages. A separate piece on "Loop Engineering" focused on RAG question parsing, describing a small loop that precedes retrieval and involves reading a document and formulating a question to guide the process. In enterprise settings, building an AI-native data platform with data agents, AI-powered QA, and governance remains a challenge for many companies.

Model Training and Decision Making

An explanation for beginners broke down the concept of backpropagation, detailing how neural networks learn step by step. In the realm of distributed systems, the concept of Byzantine Fault Tolerance was examined, posing the question of how to make decisions in untrusted environments. For financial technology firms, a practical guide suggested combining pre-churn scoring with uplift modeling to enhance customer retention. However, even when AI agents perform well in evaluations, their cost-effectiveness can still be a barrier, with one case showing a successful agent being "killed" by a CFO because its resolutions human replacements.

Global AI Landscape and Data Handling

The competitive landscape of AI development was highlighted, with China's AI models reportedly causing internal conflict within the US AI community. In data management, a method was presented for automatically assigning categories to uncategorized rows within Power Query and DAX, crucial for effective reporting and aggregation.