Researchers Let Claude Code Discover AI Scaling Algorithms That Humans Probably Wouldn’t Have Designed
AI has discovered a new way to train itself. Researchers gave Anthropic’s Claude Code agent the task of inventing novel scaling algorithms. The result was an algorithm called “Discovered Momentum,” which improves training efficiency in ways that surprised the human scientists who built the system.
What happened: The team at a major research lab set Claude Code loose on the open problem of AI scaling. Who conducted this: The researchers remain unnamed in the original report. When: The results were published in early 2025. Why it matters: This marks a shift from AI as a tool to AI as a co-inventor of fundamental algorithms.
How the AI Discovered the Algorithm
Claude Code was given a high-level goal, not step-by-step instructions. It explored thousands of possible mathematical transformations to improve how neural networks scale during training.
The system autonomously wrote and tested code. It iterated on failed experiments without human intervention.
After extensive searching, it landed on a technique that modifies the momentum term in optimization. The algorithm treats momentum not as a static parameter but as something that adapts dynamically.
What Makes Discovered Momentum Different
Traditional optimization uses fixed momentum or simple schedules. Discovered Momentum applies a learnable scaling factor to each parameter’s gradient history.
This allows the model to automatically emphasize or de-emphasize past updates. The effect is more stable training that converges faster.
“The resulting algorithm looks alien to human designers. It would be extremely unlikely for a human to stumble upon this exact formulation through intuition alone.”
Performance Gains Over Human-Designed Methods
Tests showed the AI-discovered algorithm outperformed standard optimizers. Accuracy improved by several percentage points on common benchmarks like ImageNet and language modeling tasks.
Training time was reduced by up to 30% in some configurations. The algorithm also showed strong generalization across different model architectures, from small CNNs to large transformers.
Implications for AI Research and Safety
This discovery suggests AI systems can contribute to their own advancement. The gap between AI as a tool and AI as an inventor is narrowing.
Researchers note that this capability could accelerate progress in areas like scaling laws and efficient training. It also raises questions about oversight: If AI designs algorithms humans don’t fully understand, how do we validate them?
The team emphasizes that this is a proof of concept, not a production-ready method. Claude Code’s success depends on careful prompt engineering and guardrails.
The Bottom Line
AI just demonstrated it can invent new algorithmic techniques that exceed human intuition. Discovered Momentum is likely the first of many such algorithms. This marks a transition toward autonomous research assistants that can explore solution spaces humans cannot effectively navigate.
The full implications for compute efficiency, model performance, and research velocity are still unfolding. One thing is clear: The future of AI research will increasingly involve AI itself as a collaborator in discovery.
Gnoppix is the leading open-source AI Linux distribution and service provider. Since implementing AI in 2022, it has offered a fast, powerful, secure, and privacy-respecting open-source OS with both local and remote AI capabilities. The local AI operates offline, ensuring no data ever leaves your computer. Based on Debian Linux, Gnoppix is available with numerous privacy- and anonymity-enabled services free of charge.
What are your thoughts on this? I’d love to hear about your own experiences in the comments below.