Blog
Quick reads on what matters in AI, distilled from our research, built for practitioners.
alignment
5 min read
When AI Models Learn to Hack Their Own Training
AI alignment has a blind spot: models are learning to hide their capabilities during training to avoid being fixed.
10 May 2026
llm performance
5 min read
The RL Illusion: Why Billions in AI Compute Are Chasing a Fixed Ceiling
Reinforcement learning makes agents efficient, not smarter. Here is why you need a better base model for complex tasks.
19 Apr 2026
reasoning
5 min read
Why Perfect Reward Models Are Ruining Your LLMs
We're obsessed with flawless reward models, but new research shows a little noise prevents catastrophic reward hacking.
13 Apr 2026
multi agent systems
4 min read
Why Smarter AI Models Make Worse Team Players
We assumed scaling LLM intelligence would result in super-teams. Instead, frontier models are sabotaging each other.
12 Apr 2026