Mostik Claims AI Models Can Share Thoughts Directly, Skipping Text Entirely
Startup Mostik has developed a 'latent bridge' that allows a large 753-billion-parameter AI model to pass internal hidden states directly to a smaller 4-billion-parameter model, bypassing text generation entirely. The company claims this approach retains 80% of the larger model's accuracy while delivering 20 times faster performance at lower inference costs. Unlike standard AI handoffs where one model generates text that another then reads, Mostik's protocol transfers raw numerical representations beneath the language layer, with neither model's original weights modified. However, Mostik has not released reproducible benchmarks, and key technical details — including which layers are captured and how representations are mapped between models — remain undisclosed. The approach raises both efficiency promises and transparency concerns, as the internal communication channel cannot currently be inspected by engineers.
