Tech news in 3 minutes

The AI world is getting ‘loopy’

28 d ago

At Meta’s @Scale conference on Friday, Claude Code creator Boris Cherny addressed a question about whether AI “loops” are a real trend or just hype. He affirmed they are “for real,” comparing their significance to the shift from hand-written source code to AI agents. Cherny described a transition where agents now prompt other agents to write code, making loops as important as earlier steps. He shared examples from his own work: one agent continuously seeks to improve code architecture, another finds duplicated abstractions, and both submit pull requests endlessly as code evolves. This concept expands agentic AI by authorizing swarms of agents to work continuously in the background, requiring significant trust in AI as models improve. Cherny noted that recursive loops are not entirely new—they are foundational in computer science—but agentic loops follow non-deterministic logic, where a subagent decides when to stop. Simple implementations like the “Ralph Loop” (named after Ralph Wiggum) ask the model to review its work and check if goals are achieved, preventing drift. These loops align with the push for more test-time compute. As OpenAI researcher Noam Brown stated, contemporary models can solve nearly any problem with enough compute. For hill-climbing tasks like code improvement, loops can keep making incremental advances until a threshold is reached—or indefinitely if compute is available. However, AI loops consume far more tokens than simple chatbots, with no cost ceiling. While this benefits token-selling companies like Anthropic, it can be expensive for others. Yet, with proper oversight of token spend and drift, the potential benefits may outweigh the costs, making loops a promising next step for practical AI work.