AI research trends and benchmark releases

AI research announcements, benchmarks and fellowship openings. - OpenAI opened applications for a 6-month paid research fellowship in San Francisco. - Fellowship salary around $220K annually plus relocation support; interviews Jan 2026. - Two 2025 technical breakthroughs cited: Test-Time Scaling → Reasoning Models and MoE → DeepSeek R1. - Prediction: reasoning cost reduction and on-device AI growth in 2026. - NVIDIA and Stanford released NitroGen trained on 1,000+ games and 40K hours of gameplay. - NitroGen dataset and model weights to be publicly released. - NeurIPS 2025 top 50 contributors include many US labs and leading Chinese universities. - Frontier AI no longer requires frontier budgets according to a summary of 2025 trends.
Key insights
OpenAI launches paid 6-month research fellowship
OpenAI announced applications for a full-time, paid 6-month research fellowship in San Francisco with an annualized salary of approximately $220,000 plus relocation support. The program does not require formal ML/AI backgrounds but seeks strong technical foundations; interviews start January 2026.
NitroGen released as generalist game-playing model with open data
NVIDIA and Stanford published NitroGen, a generalist model trained on over 1,000 games and approximately 40,000 hours of gameplay, and announced plans to release the dataset and model weights publicly to support reproducible research.
2025 technical highlights: test-time scaling and MoE advances
Two cited breakthroughs for 2025 include Test-Time Scaling leading to improved reasoning models and MoE architectures contributing to DeepSeek R1, enabling reasoning cost reductions and supporting the view that frontier AI is becoming more cost-accessible.
NeurIPS 2025 contributors show broad global participation
NeurIPS 2025's top 50 paper contributors list includes major U.S. corporate labs (Google DeepMind, Meta, Microsoft) and leading Chinese institutions (Tsinghua, CAS, Peking University), with a smaller set of European universities represented.