Recursion Is The Next Scaling Law In AI

The gist

1. Standard LLMs process everything in one forward pass, which makes true iterative reasoning impossible without workarounds that need human-labeled data.
2. A 7 million parameter recursive model hit 87% accuracy on ARC Prize tasks, beating models with over 100 billion parameters.
3. Weight sharing across recursion levels inside a single transformer layer cuts complexity while keeping strong reasoning ability.
4. Future AI designs may pair large models for understanding language with smaller recursive modules that handle the actual reasoning.