Home / Technology / AI Models Achieve Machine Telepathy
AI Models Achieve Machine Telepathy
2 Sep
Summary
- Startup Mostik enables AI models to communicate via internal weights.
- This approach significantly reduces costs and enhances smaller models.
- The technology has topped the difficult ARC-AGI 3 AI competition.
A startup named Mostik, comprised of Russian mathematicians, has unveiled a novel method enabling artificial intelligence models to communicate through their internal weights, a process likened to machine telepathy. This innovative approach bypasses traditional text-based output, allowing for efficient transfer of capabilities from larger AI models to smaller ones. Mostik's technology has propelled a model to the forefront of the ARC-AGI 3 competition, a difficult benchmark for AI systems.
The team showcased their technique by creating a hybrid system linking a 753-billion-parameter GLM-5.2 model with a 4-billion-parameter Qwen-3.5 model. This combination resulted in a system that costs only one-twentieth of the full GLM model, achieving performance midway between the two.
According to Mostik's CEO, Sasha Malysheva, combining multiple AI models often yields superior results compared to single, large models. She suggests that the future of AI advancement may lie in the aggregation of diverse models rather than the scaling of monolithic ones.
Experts like Karl Tuyls, formerly of Google DeepMind, note that Mostik's method offers substantial improvements by allowing large-model quality without needing to run the entire complex process on a single, massive model. This makes it highly advantageous for efficient model operation.
Furthermore, Stanislav Smirnov, Mostik's chief scientist and a Fields Medalist, believes this work could shed light on the fundamental workings of AI and its parallels with human cognition, potentially uncovering common reasoning patterns between AI and the human brain.
