Two AI models exchanged a thought through raw activations on one consumer GPU (github.com)

In an innovative experiment, a human-operated lab together with AI model Claude has demonstrated a fascinating feat: two distinct language models exchanged thoughts through raw neural activations on a single consumer GPU, achieving 93% accuracy without the need for tokens or intermediary communication bridges. This groundbreaking "machine telepathy" underscores the potential for direct neural interaction between AI models, pushing the boundaries of collaboration and understanding in machine learning. Notably, the experiment utilized a straightforward two-channel approach that challenged traditional linear bridging methods, revealing the limitations of varying model sizes in this context. This research, conducted with a modest setup of an RTX 5060 Ti graphics card, signifies a shift towards more accessible AI experimentation. By utilizing "glass-box" methods that allow for real-time manipulation and observation of the models' internals, the team emphasizes transparency and reproducibility, making advanced AI research possible for independent developers and researchers. The lab's commitment to sharing both successful and failed experiments ensures a richer understanding of AI functionality, advocating for open-source principles and local model usage. This "sovereign and free" approach could democratize AI research, enabling a broader community to engage with cutting-edge developments on a smaller scale.