Technology giant Meta has announced a breakthrough in the field of brain-computer interfaces by unveiling the new version of the Brain2Qwerty v2 system. This innovative algorithm is capable of transforming complex human brain activity into clear typed text, opening new horizons for communication and device control through the power of thought.

Experiment in Spain: How the AI Was Trained

The development and testing of the system took place at the Basque Center on Cognition, Brain and Language (BCBL) in San Sebastián. To train the neural network, a focus group of nine healthy volunteers aged between 25 and 56 was recruited. Participants underwent an intensive data collection process consisting of 10 sessions, during which each of them generated more than 2,500 sentences.

The key difference in this study was the signal recording method. Instead of invasive surgery, magnetoencephalography (MEG) was used. Volunteers wore a special helmet that non-invasively scanned the ultra-weak magnetic fields generated by neurons during the mental typing process.

Accuracy and Architecture: From Noise to Meaning

The test results exceeded expectations: in the best scenarios, Brain2Qwerty v2 achieved a word decoding accuracy of 78%. This means that in the decoded sentences, in more than half of the cases, the number of errors did not exceed one word. Researchers confirmed a direct correlation: the larger the volume of training data, the higher the system's accuracy. This gives hope that in the future, non-invasive MEG systems will be able to completely eliminate the need for surgical intervention.

The success of the project was made possible by the integration of pattern recognition technologies similar to those used in modern chatbots, such as ChatGPT and Meta's Llama models. The process of transforming thought into text is built on a hierarchical structure of three stages:

  • Primary Encoding: The AI analyzes "raw" brain signals and translates them into tokens corresponding to individual characters.
  • Alignment: An intermediate system (aligner) groups scattered letters and symbols into full words.
  • Final Context (LLM): A large language model takes on the task of turning a set of words into grammatically correct and coherent sentences, taking into account the semantic context.

This is the first case in history of the successful application of large language models (LLM) to decode noisy brain signals.

The Role of Autonomous Agents

Autonomous AI agents ("auto-research" agents) are also integrated into the system architecture. Their task is to independently and iteratively modify the codebase, inventing optimal architectural solutions to reduce the Word Error Rate (WER). Despite the high efficiency of these "force multipliers," the authors of the study emphasize that human researchers remain the key element of the scientific process.