LLMs for Ground Delay Programs, AI Pilot Training

Hi, here's my take on todays news - which represents the second slice of the last week of AI in Aviation happenings around the world:

Last issue I talked about the FAA's SMART program and the three prime contractors fighting for it. What's quietly interesting is that academia is building the same muscle in parallel. Max Li's group at the University of Michigan is using LLMs to help FAA traffic managers draft ground delay programs - that's the same strategic flow-management problem SMART is going after, but from a research lane rather than a $32.5B procurement.

Pilot training is shifting too - the U.S. Army has narrowed its fixed-wing training outsource to two finalists, both leaning heavily on AI/VR debrief and adaptive syllabi. Expect the Air Force to follow. And on the broader AI front, OpenAI launched GPT-Rosalind, a biology-tuned frontier model deployed under a "trusted access" model - US institutions only, application-gated. This could be the same way in which a regulated aviation domain-specialist frontier model reaches market, as hinted at in EASA latest draft regulations.

Enjoy.

University of Michigan Builds LLM Tooling for FAA Ground Delay Programs

Aerospace America profiles Max Li's research group at the University of Michigan building LLM tooling to help FAA traffic managers draft and refine ground delay programs. The work targets the same strategic flow-management problem as FAA's SMART initiative but from an academic research angle, suggesting parallel commercial and research lanes toward AI-assisted air traffic management.

Pratt & Whitney Engine Installed on Northrop's YFQ-48A Talon Blue CCA — Flight Test Imminent

Pratt & Whitney confirmed its adapted PW500-family engine has been installed on Northrop Grumman's YFQ-48A Talon Blue, the second Collaborative Combat Aircraft increment-1 contender, with flight testing imminent. Combined with the USAF Anduril Fury demonstration last week, this sharpens the three-way CCA race ahead of the summer 2026 production decision.

Baykar's K2 Kamikaze UAV Completes AI-Coordinated Swarm Patrol Test

Five Baykar K2 kamikaze UAVs completed an AI-coordinated formation, patrol, and mission-coherence test in Türkiye. The 800kg platform carries a 200kg warhead over 2,000km of range, signaling Türkiye's push into attritable, AI-native strike mass for contested airspace - a natural extension of the TB2 playbook.

U.S. Army Picks Two Finalists for Fixed-Wing Pilot Training Outsource

Defense One reports the U.S. Army has narrowed its fixed-wing pilot training outsource contract to two finalists, both proposing heavy AI/VR debrief and adaptive-syllabus systems. The contract would shift a significant portion of initial pilot training to commercial providers and accelerate adoption of AI-driven debrief tooling across military aviation.

OpenAI Launches GPT-Rosalind — a Gated, Biology-Tuned Frontier Model

OpenAI launched GPT-Rosalind, a biology-tuned frontier model trained on 50 common biological workflows with direct access to major public biology databases, deployed under a "trusted access" structure limited to approved US institutions. It sets the precedent for how a regulated, dual-use domain frontier model reaches market.

Sam here again - hope you have enjoyed reading this edition. The through-line across this week's second slice - from the UMich ATC research to OpenAI's gated biology model - is the same hard question: how do you deploy AI responsibly inside safety-critical, regulated domains? That is the work we do at Deepsky - consulting on AI adoption and automating compliance and audit checks across your exposition so your documentation moves as fast as the technology. If that sounds useful, contact me at admin@deepskyai.com or visit deepskyai.com.