The world of artificial intelligence (AI) is witnessing a fascinating evolution, and at the forefront of this revolution are a group of brilliant Russian mathematicians and their innovative startup, Mostik. This company has developed a groundbreaking approach that enables AI models to communicate and enhance their capabilities through a unique form of 'machine telepathy'.
The concept revolves around utilizing the mathematical values embedded in the weights of AI models, which are the parameters that determine how a prompt is transformed into an output. By feeding the output of one model into another, Mostik's technique allows for a seamless exchange of information without the need for text generation. This method has proven to be a game-changer, especially in the context of open-weight models.
One of the most impressive demonstrations of this approach is the bridge built between two Chinese open-weight models: GLM-5.2, a massive model with 753 billion parameters, and Qwen-3.5, a more compact version with 4 billion parameters. The resulting hybrid system not only costs a fraction of the full GLM model but also achieves performance that is a harmonious blend of the two models.
Sasha Malysheva, the CEO of Mostik, draws an intriguing parallel between AI and the art of guessing a pig's weight. Just as a group of people can collectively estimate a pig's weight more accurately than a single expert, combining the outputs of multiple AI models can lead to superior results. This concept is not new, but Mostik's approach streamlines the process, making it more efficient and cost-effective.
Malysheva's journey is an inspiring one. Initially doubted by peers who believed the bridge approach was too challenging for a young woman, she persevered and proved them wrong. This determination reflects the spirit of innovation that drives the Mostik team.
The implications of this technique are far-reaching. By enabling the pairing of frontier models with domain-specific models, Mostik's approach could lead to a proliferation of specialized AI models in fields like biology and physics. This could potentially democratize AI capabilities, making them more accessible and tailored to various industries.
Furthermore, the work of Mostik's chief scientist, Stanislav Smirnov, suggests a deeper connection between AI and human cognition. By analyzing the mathematical underpinnings of AI models, Smirnov believes that we might uncover commonalities in how both AI and humans approach complex problems. This could lead to a better understanding of the inner workings of AI and potentially inspire new approaches to machine learning.
In conclusion, Mostik's innovative approach to AI communication and collaboration is a testament to the power of human ingenuity. As AI continues to evolve, it is essential to explore these creative solutions that not only enhance performance but also foster a deeper understanding of the technology's capabilities and limitations.
