Modern AI models can generate millions of new materials in minutes. The problem is that most of them turn out to be chemically unstable and unsuitable for use in real products – from chips to rockets. This forces engineers to spend enormous computational resources sifting through unsuitable options, leaving only a small fraction. Researchers from MIT have developed the CrysVCD system, which embeds chemical stability rules at the very early stage of generation. Instead of filtering the results post-factum, CrysVCD ensures that each new structure adheres to the fundamental laws of chemistry concerning atomic electrons. This has allowed for a high degree of stability (up to 70% in rigorous tests) while creating materials with desired properties, such as high thermal conductivity or dielectric permittivity, important for electronics. According to the developers, CrysVCD acts as a universal 'DVD player' for any material generation models, both existing and future. The system reduces the process of creating stable materials by approximately 1000 times, cutting computational costs from weeks or months to a few steps. This opens doors for more efficient research and development, especially for small companies and scientific laboratories with limited resources that previously could not afford the expensive selection process. The new approach has already demonstrated the ability to generate materials suitable for the semiconductor industry and data center cooling systems. This is a significant step from theoretical research to the real application of AI in materials science, where stability and functionality have always been key but hard-to-achieve parameters.