Forget the hype around superintelligence; a new wave of AI is making recursive self-improvement accessible to everyone, not just giant tech labs. This means you might soon be able to build your own AI assistants that get smarter over time.
The buzz around AI lately has been all about frontier labs creating models that can train themselves, potentially leading to superintelligence. But what if this technology could be used for more practical, everyday tasks? Experiments show that it's not only possible but surprisingly effective, offering a vision of AI development that's less centralized and more democratic.
Kicking things off, a simple experiment involved using a tool called AutoResearch, developed by AI superstar Andrej Karpathy. By feeding instructions to a powerful AI like Claude and providing the necessary computing power, the system autonomously refined a smaller language model. While early results were a bit rough – think endless repetition – the model showed clear signs of improvement and more coherent output over time.
The journey continued with a more complex, real-world application: automating parts of a newsletter. Using a platform from Prime Intellect, a startup focused on democratizing AI training, a custom model called Frontier_Paper_Curator was built. This AI was tasked with finding and summarizing relevant research papers, essentially learning and improving its own performance through a feedback loop.
Vincent Weisser, CEO of Prime Intellect, emphasizes that the goal is to empower every company with advanced AI training capabilities. "We don't want one centralized, almost godlike intelligence, we want a billion intelligences that go into all the niches that create beautiful things," he stated. This approach contrasts with the current model where a few major players control the most advanced AI, and it highlights the risks of relying too heavily on a single, powerful AI like Anthropic's recent Fable 5 model.
While the ultimate goal of recursive self-improvement is for AI to generate novel insights, the tools available today are already impressive. After just a short period, a surprisingly capable model for curating research was created. Although it still needs some fine-tuning, this development signifies a promising step towards AI that can continuously learn and assist us, potentially freeing us from tedious tasks and unlocking broader creativity.
