AI & ML Research 24 Hours

Last updated: September 11, 2026, 6:04 AM ET

Tool Use & Data Agents

Google's ToolGrad introduces an efficient method for generating tool-use datasets using textual "gradients," streamlining how agents learn to call external functions. OpenAI, meanwhile, rolled out a Data agent in Chat GPT Work, letting teams connect company data, surface insights, and build interactive dashboards through natural language rather than SQL or BI tooling.

Scientific & Communication Workflows

In antimicrobial research, César de la Fuente's lab is using Codex and ChatGPT to mine living and extinct genomes for drug candidates against resistant infections. On the practitioner side, a Towards Data Science guide explains how to 5x communication effectiveness with Claude Code, arguing that clearer intent signals produce better agent output.

Cost & Explainability

A new piece proposes adaptive model routing — moving from static model assignment to intelligent, task-level LLM selection to cut inference costs in multi-agent systems. Separately, one analysis argues that SHAP can't explain agentic AI fraud, exposing a fresh explainability gap as autonomous agents drive fraud detection decisions. Another essay asks when to retest our assumptions, warning that a model is only as reliable as the beliefs beneath it.

Infrastructure & Climate Engineering

MIT Technology Review reports that powering AI is fundamentally an architecture problem: a July 22, 2026 transmission fault in Ashburn, Virginia knocked over 3 gigawatts of load off the grid in seconds, underscoring medium-voltage redesign needs. A San Francisco nonprofit published a road map of experiments and infrastructure needed to make informed solar geoengineering decisions. The Download also covers a God-driven crypto collapse that wiped out investors, alongside the solar engineering roadmap.