Science

Self-improving AI? 🤖 

That's the claim by Sakana AI in their recent paper, "The Darwin Gödel Machine" - "AI that improves itself by rewriting its own code".

I just finished reading the paper and it's not as scary as it sounds! They built an AI *Agent* that can edit its own code, but the underlying LLM (mostly Claude 3.5 Sonnet or o3-mini) stays fixed.

Still a long way from AI actually inventing new ML architectures and training its own upgrade.

We're safe - for now 😅

They do see big improvements in the overall system's performance as the agent improves itself, but not beyond existing state-of-the-art

As I see it, the use case could be for AI companies building agentic products - like Cursor or ChatGPT - to get better results faster, with less manual crafting. 

paper: https://arxiv.org/abs/2505.22954
blog post:

Founder & Creative Director @ https://flowzed.com • AI researcher, blockchain architect & computational artist • MSc Nanotechnology @ Oxford

They do see big improvements in the overall system's performance as the agent improves itself, but not beyond existing state-of-the-art

As I see it, the use case could be for AI companies building agentic products - like Cursor or ChatGPT - to get better results faster, with less manual crafting. 

paper: https://arxiv.org/abs/2505.22954
blog post: https://sakana.ai/dgm/