Guest Post: The Coding Personalities of Leading LLMs*
GPT-4o, Claude, Llama tested on 4,400+ coding tasks: each LLM has a distinct risk profile. Newer models aren't always safer — data from Sonar's code quality analysis.
Modular manifolds treat neural network layers as geometric modules for stable, scalable optimization. A deep dive into Thinking Machines Lab's approach.
Thinking tokens let AI models reason before answering at massive compute cost. How hidden reasoning reshapes AI economics, and whether the bubble holds
The History of Reinforcement Learning: From Thorndike to GRPO
From the early trial-and-error concepts to today’s breakthroughs with RLHF, PPO, and GRPO, and where to go next, according to Andrej Karpathy and Richard Sutton.