MIT Star: We're Wasting Model Power

🔥 What happened MIT PhD Alex Zhang told the Latent Space podcast that Recursive Language Models (RLMs) cracked ARC-AGI-3 before OpenAI. His Prime Agent harness beat frontier labs on autonomous coding tasks. 💡 Why it matters OpenAI's 10,000-agent experiment burned roughly 130B output tokens — an estimated $40M — on a single problem. Zhang argues frontier models carry massive capability overhang that primitive harnesses simply waste. ⚡ Our take If you're still optimizing models instead of harnesses in 2026, you're losing. The real leverage isn't the next checkpoint — it's the architecture wrapped around it.