> Source: easy-learn-ai project, commit 18d79f8
> Date: June 30, 2026
The Scenario
Imagine sitting at your computer, thinking "hotpot for dinner tonight" — and that sentence appears on screen, not because you typed it, but because the computer read your thoughts. This is not science fiction. On June 30, 2026, Meta released Brain2Qwerty v2, a system that decodes intended sentences from brainwaves into text, achieving 61% sentence-level accuracy, with the best test subject reaching 78%.
What Signals Does the Brain Emit?
The brain works like a giant orchestra: neurons are musicians playing with extremely weak electrical signals. When you think a sentence, the whole ensemble plays a specific pattern — but the signals are faint and mixed together.
Meta uses two types of "microphones":
- MEG (magnetoencephalography): measures the weak magnetic fields produced by brain activity. Extremely precise, with millisecond-level temporal resolution — it can record exactly which musician struck the drum and when.
- EEG (electroencephalography): electrodes placed on the scalp measure electrical signals directly. Cheaper and portable, but noisier — like recording an orchestra on an old tape recorder.
- Bulky equipment: MEG requires superconducting materials cooled in liquid helium; the machine is the size of a room. EEG is portable but low-quality. Daily use needs major hardware breakthroughs.
- Large individual differences: Models trained on one person perform poorly on another, meaning costly per-user training for any commercial deployment.
- The ultimate privacy question: If a machine can read your brainwaves, can it read thoughts you don't want to share? Today's technology only decodes what you actively intend to say — thoughts already organized into language. But what if AI one day reads deeper thoughts you're not even aware of? Technology advances faster than we expect; ethical and legal frameworks lag behind.
Meta's cleverness lies in combining both: MEG provides precise timing, EEG supplements spatial information — like one camera lens for sharpness and another for color, merged into a complete picture.
From Brainwaves to Text: How the AI "Listens"
Meta's pipeline has three steps:
1. Slice brainwaves into "syllables": The AI cuts continuous brainwave signals into small segments and trains a model to identify which phoneme each segment corresponds to — like finding repeating sound patterns in an unfamiliar foreign-language broadcast. 2. Assemble "syllables" into words: Like turning pinyin into characters. A key difficulty: brain signals during word processing are often non-linear — you may grasp the whole concept of "hotpot" before expanding it into individual characters. The AI must learn this parallel-to-serial conversion. 3. Guess from context: Raw decoded text contains many errors. The AI exploits statistical regularities of language to correct them — e.g., if it decodes "tonight eat huo-pot" with an unclear character, it knows that after "eat," something pronounced like "huo" is most likely "fire" (hotpot). This is exactly how smartphone autocorrect works.
Is 61% Accuracy Good?
It may sound low, but remember: this is decoded purely from brainwaves — no hands, no mouth. It's like someone blindfolded, in a noisy room, writing down what you silently say to yourself. Getting over 60% right is remarkable.
Moreover, the accuracy is sentence-level: guessing an entire sentence of a dozen-plus characters correctly is far harder than guessing single words. The best subject reached 78%, showing large individual differences — possibly due to stronger brain signals or better focus during testing.
Why It Matters
1. Opening a door for people with aphasia. Millions worldwide have lost speech due to stroke, ALS, or spinal injury. Their brains work — they think and feel — but cannot convert thoughts into language. Current tools like eye-trackers are slow and exhausting. A mature brain-computer interface would let them simply *think*, and text appears.
2. The ultimate change in interaction. Today's interactions are all "translations": a thought becomes finger movements (typing), tongue movements (speaking), or eye movements (clicking) — each with loss and delay. Brain-computer interfaces aim to skip all translation and deliver raw intent directly. The road is long — current setups require a large MEG helmet or a cap full of EEG electrodes — but technology iterates. Twenty years ago phones were bricks.
3. Deepening our understanding of AI and the brain. In doing this, Meta's scientists are also answering a fundamental question: how is language encoded in the brain? AI is becoming a powerful tool for humanity to decode its own mind.
What Remains Unsolved
Final Thoughts
Brain2Qwerty v2: 61% sentence-level accuracy, 78% for the best subject. Code will be open-sourced, and the v1 dataset will be released by BCBL. Behind these numbers is a group of scientists tackling an ancient question: can machines understand human thought?
The current answer: yes — but only a little. That little is enough to be excited about, and enough to be wary of. Excited, because people trapped by their bodies may soon speak with the world again. Wary, because when machines can truly "read minds," we must ensure they read only what we intend to share.
---
*"You focus on moving forward; leave the remembering to me."*