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.
Research
DeepMind's New Architecture Achieves Human-Level Performance on Novel Scientific Reasoning Tasks
A landmark paper from Google DeepMind introduces "Prometheus-7B," a model trained exclusively on structured scientific literature that outperforms PhD-level humans on chemistry and physics problem sets never seen during training.
Latest Updates
Models
OpenAI's GPT-5 Confirms 10M Token Context — Internal Benchmarks Leaked Ahead of Launch
Sources close to the company confirm the new architecture handles book-length inputs natively, with near-zero degradation at the 8M token range.
Robotics
Figure AI's Latest Humanoid Completes Full Factory Shift Unsupervised for First Time
Figure 03 ran an 8-hour manufacturing shift at a BMW partner facility with zero human interventions, marking a milestone for embodied AI reliability.
Policy
EU AI Act Enforcement Begins: What Every AI Company Needs to Know This Month
The first wave of compliance deadlines hit today. High-risk AI system operators face audits, and the fines for non-compliance are not theoretical.
Research
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. Three models came out ahead.
Industry
NVIDIA's Blackwell Ultra Ships 6 Months Early — Hyperscalers Already Retooling Clusters
Supply chain sources confirm early shipments to Microsoft and Google. The GB300 delivers 2.4x the FP8 throughput of its predecessor at comparable power draw.
Models
Meta's LLaMA 4 Released Under Apache 2.0 — The Open Source Moment the Community Was Waiting For
After months of restricted licensing controversy, Meta quietly published LLaMA 4 405B under a fully permissive license. The community is already fine-tuning.
Deep Reads
The Model Evaluation Crisis: Why No One Agrees on How to Measure Intelligence
Benchmarks are broken. Researchers are arguing. Here's what the frontier actually looks like from the inside.
Who Actually Funds Frontier AI? Following the Money in 2026
Sovereign wealth funds, hyperscalers, and defense contractors — the funding landscape has shifted dramatically.