How to fully automate AI research

With every passing month, AI models get better at most tasks that an AI researcher does in their work.¹ Yet for all these gains, today's models only assist human researchers, falling far short of automating research completely. What will it take to build AI systems that can fully replace human researchers, and why aren't we there yet?

The Current Paradox

Current AI systems present something of a paradox. Their performance on narrow research tasks already exceeds that of most human researchers. However, any researcher who has worked with them quickly notices the need to keep AI agents on a very short leash. Despite good benchmark scores and impressive demos, there are clearly core capabilities that human researchers have that our current systems are missing.

Missing Capabilities

We've previously highlighted some of these shortcomings: lack of reliability in complex reasoning, poor long-term planning capabilities, and overly narrow focus on specific domains, among others.² But why are these capabilities missing in AI systems to begin with? We train them on more compute and data than humans have access to in their entire careers, and we can run millions of parallel experiments, and yet it's still not enough.

The Efficiency Gap

On some level, the answer has to be that our learning algorithms have been and remain much less efficient than the human brain. Deep learning researchers often point to this and say that it's a sign the field needs new paradigms. But we think there's a more specific and actionable explanation.

Beyond Information Processing

The key insight is that human researchers don't just process information—they actively construct and test hypotheses through experimentation, collaborate with others, and build upon decades of accumulated knowledge in sophisticated ways.³ Current AI systems, despite their impressive capabilities, lack the integrated reasoning frameworks that allow humans to navigate uncertainty and make breakthrough discoveries.

Our Integrated Approach

At Genesis AI Labs, we're developing research environments that capture these missing elements. Our approach combines geometric deep learning for understanding complex relationships, reinforcement learning for long-term planning, and generative modeling for creative hypothesis generation. These aren't separate tools but integrated components of a unified research framework.

The Full Research Lifecycle

The path forward requires building AI systems that can engage in the full research lifecycle: identifying important problems, designing experiments, interpreting results, and iterating based on findings.⁴ This means creating environments where AI agents can learn not just from data, but from the process of discovery itself.

Looking Ahead

We believe this approach will unlock the next phase of AI development, where systems don't just assist researchers but become capable research partners, accelerating the pace of scientific discovery and bringing us closer to artificial general intelligence.