Former OpenAI employee Naomi Bashkansky joined the Conduit startup, where she took the position of co-founder of the research department. The company develops non-invasive neural interfaces and trains artificial intelligence models to interpret human brain activity. In her blog, Bashkansky described the concept that in the coming years, thought processes may become the main way to interact with AI systems. In her opinion, such technologies should expand human cognitive capabilities, and not replace them with automated algorithms.

According to the researcher's forecasts, by 2027 it may be technically possible to transmit thoughts to an AI agent via a Bluetooth connection for their subsequent interpretation into commands and actions. By 2030, she expects the industry to expand and adopt telepathic technologies similar to those being developed by Conduit. Bashkansky also refers to statements by Elon Musk and Sam Altman, who previously expressed ideas about the convergence of the human brain and artificial intelligence. She noted that the degree of integration of humans and AI by 2035 will depend on the decisions made by society in the process of technology development.

Conduit's technical approach is to train neural networks based on brain activity data. As Bashkansky explained, brain signals are sent to the input of the model, and at the output the system tries to predict the semantic result of the user's actions, for example, the text they typed. The startup is based in San Francisco and focuses on creating interfaces that do not require surgical intervention. Information about the volume of attracted investments and the company's valuation is not disclosed. The publication sparked a discussion in the relevant community. Some users expressed skepticism about the term "non-invasively," noting that accurately reading thoughts without direct physical access to the brain poses significant technical difficulties. It is worth noting that in June, Meta introduced its own system capable of restoring typed text based on recording brain activity without surgical intervention. The claimed accuracy of recognition at the word level was 61%, which exceeds the performance of previous methods, which were at the level of 8%.