Research
Papers, breakthroughs, and technical deep dives from the frontier of AI research.
DeepMind's Prometheus-7B Achieves Human-Level Scientific Reasoning Without Task-Specific Training
A new architecture trained on structured scientific literature outperforms PhD-level humans on unseen chemistry and physics problems.
Multimodal Models Now Outperform Radiologists on Rare Disease Diagnosis in New Study
A Stanford-MIT collaboration tested 14 frontier models against 200 board-certified radiologists on rare pathology identification.
Why Mixture-of-Experts Is the Architecture That Won 2025
A technical breakdown of why MoE displaced dense transformers as the default architecture at scale, and what comes next.
The Attention Mechanism Turns 10 — How One Paper Changed Everything
A decade after "Attention Is All You Need," we trace the paper's extraordinary impact on compute, culture, and capitalism.
Why Reinforcement Learning from Human Feedback Is Being Replaced
New alignment approaches like DPO and constitutional methods are making RLHF look increasingly like a transitional technique.
Formal Verification Meets LLMs: A New Paradigm for Provably Correct Code Generation
Researchers combine large language models with Lean 4 and Coq proof assistants to generate code that ships with mathematical proofs of correctness.