A groundbreaking San Francisco AI startup is revolutionizing machine learning with models that can essentially teach themselves how to think more effectively. Deep Cogito, founded by ex-Google engineers, has released four new open-source language models that don’t just process information, but actively develop “intuition” about how to solve complex problems.

The company’s approach is radical: instead of relying on pre-programmed reasoning paths, these AI models learn to optimize their own cognitive strategies during training. By running internal reasoning chains and then studying their own thought processes, the models gradually become more efficient at solving tasks.

Deep Cogito’s flagship 671 billion parameter model demonstrates remarkable capabilities across multiple domains. In mathematical reasoning tests, it solved problems using 60% fewer computational steps compared to other leading models. The AI can tackle complex scenarios ranging from legal interpretation to multi-step logical deductions with unprecedented precision.

CEO Drishan Arora describes their method as “iterated distillation and amplification,” which essentially means the AI learns from its own intermediate thoughts. This approach allows the models to develop what the team calls “machine intuition” - a capability to quickly recognize optimal problem-solving strategies.

What makes Deep Cogito’s work particularly impressive is their cost-effectiveness. While some AI companies spend over $100 million on model development, Deep Cogito claims to have trained all of their models for under $3.5 million. This efficiency could democratize advanced AI research, making cutting-edge technology more accessible to developers and researchers.

The models are now available on platforms like Hugging Face and can be used by enterprises, developers, and researchers interested in exploring this new frontier of self-improving artificial intelligence. As the tech world continues to watch, Deep Cogito’s innovative approach might just represent a significant leap forward in how we understand and develop machine intelligence.