> Source: easy-learn-ai project, commit 18d79f8
> Date: June 30, 2026
---
1. A Scenario First
You're sitting at your computer, thinking "hotpot for dinner tonight."
Suddenly that sentence appears on screen—not because you typed it, but because the computer "read" your thought.
This isn't the opening of a science fiction novel. This is technology Meta released on June 30, 2026: Brain2Qwerty v2. It can decode the sentences you intend to say from your brainwaves into text.
Accuracy: 61%. The best test participant reached 78%.
What does this mean? Before you even open your mouth, AI already knows what you want to say.
---
2. What Signals Is the Brain Actually Sending?
To understand this, we first need to know: what is your brain doing when you "think" a sentence?
Think of the brain as a giant symphony orchestra. Every neuron is a musician playing with extremely faint electrical signals. When you think "hotpot for dinner tonight," it's not one musician blowing a trumpet—the entire orchestra plays in a specific pattern.
The problem: the performance is too faint and everything is mixed together. It's like standing in a noisy square hearing dozens of people talk at once—how do you pick out one voice?
Meta uses two kinds of "microphones" to capture this performance:
- MEG (magnetoencephalography): measures the weak magnetic fields generated by brain activity. Imagine a super-sensitive compass that can sense the magnetic disturbances caused by neuronal firing. It's highly precise, with millisecond-level temporal resolution—recording exactly "which musician struck the drum at which moment."
- EEG (electroencephalography): electrodes placed on the scalp measure electrical signals directly. Cheaper and more portable than MEG, but the signal is "dirtier"—like recording a symphony on an old tape recorder full of noise.
---
3. From Brainwaves to Text: How Does AI "Understand"?
Even if you record the whole performance, how do you know this passage corresponds to "hotpot tonight" rather than "hiking tomorrow"?
That's where AI comes in. Meta's approach has three steps:
Step 1: Slice brainwaves into "syllables"
Imagine listening to foreign radio you don't understand, but noticing repeated sound patterns—certain sounds always appear at the start of sentences, others at the end. The AI does something similar: it cuts the continuous brainwave signal into small segments and trains a model to identify which phoneme (the smallest unit of speech) each segment corresponds to.
Step 2: Assemble "syllables" into "words"
Once each segment's phoneme is known, the next step is spelling out words—like learning pinyin and combining b-a-o into "bao."
A key point: when the brain thinks of a word, its signal patterns are often not linear. You might first grasp the overall concept of "hotpot," then gradually "unpack" it into the individual characters. The AI must learn this parallel-to-serial conversion.
Step 3: Guess using context
Even after the first two steps, the decoded text still contains many errors. So the AI does something clever: it exploits statistical regularities of language to "guess."
Say it decodes "tonight eat huo-pot" with "huo" unclear. It knows that in Chinese, the thing that follows "eat" and sounds like "huo" is most likely "fire." So it corrects "huo" to "fire."
This is exactly the same principle as your phone keyboard's autocorrect.
---
4. Is 61% Accuracy Actually Good?
You might say: 61%? That's terrible—my typing is far more accurate than that.
But remember: this was decoded purely from brainwaves, with no hand movement and no speech.
It's like someone blindfolded, in a noisy room, hearing your inner monologue and writing it down. Getting more than 60% right is already remarkable.
Moreover, this is sentence-level accuracy—not guessing individual characters, but entire sentences of a dozen-plus characters all correct at once, which is far harder.
Meta's team also noted the best participant reached 78%. This shows results vary widely across individuals—perhaps some people have stronger brain signals, or were more "cooperative" during testing (more focused, fewer stray thoughts).
---
5. Why Does This Matter?
First: opening a door for people who cannot speak
Millions worldwide have lost the ability to speak due to stroke, ALS, spinal cord injuries, and similar conditions. Their brains work fine—they can think and feel—but they can't turn thoughts into speech or text. Existing solutions like eye-tracking typers are slow and exhausting. If brain-computer interfaces mature, they could simply "think," and text appears on screen. This isn't about convenience—it's about restoring personhood.
Second: the ultimate transformation of interaction
Long-term, if brain-computer interfaces become practical, they change interaction for everyone, not just disabled users. All current interaction is fundamentally "translation": a thought is translated into finger movements (typing), tongue movements (speech), or eye movements (clicking), with loss and delay at every step. The vision of BCIs is to skip all translation and let machines receive your raw intent. The road is long—current devices require a huge MEG helmet or a scalp full of EEG electrodes—but technology iterates. Twenty years ago phones were bricks; ten years ago there were no voice assistants.
Third: deepening our understanding of AI—and ourselves
In doing this, Meta's scientists are also answering a more fundamental question: how is language actually encoded in the brain? Neuroscientists haven't fully solved this, and AI is becoming a powerful tool for decoding our own brains.
---
6. Let's Be Sober: What Remains Unsolved?
Challenge 1: The hardware is bulky. MEG requires superconducting materials cooled in liquid helium, with machines the size of a room. EEG is portable but noisy. Everyday use demands major hardware breakthroughs.
Challenge 2: Huge individual differences. Everyone's brain is different. A model trained on one person may perform poorly on another—commercial use would require per-person model training, which is costly.
Challenge 3: The ultimate privacy problem. This is the most unsettling part. If machines can read your brainwaves, can they read thoughts you don't want others to know? Current technology only decodes what you actively intend to say—thoughts already "organized into language" in your brain. But what about the future, when AI might read deeper thoughts you're not even aware of? This isn't alarmism: technology advances faster than we expect, while ethical and legal frameworks are built slower than we hope.
---
7. Closing Thoughts
Meta's Brain2Qwerty v2: 61% sentence-level accuracy, with the best participant reaching 78%. The code will be open-sourced, and the v1 dataset will be released by BCBL.
Behind these numbers and promises is a group of scientists trying to answer an ancient question: can human thought be understood by machines?
The current answer: yes, but only a little.
That little is enough to excite us—and enough to put us on guard.
Exciting, because people trapped by their bodies may soon converse with the world again.
Cautionary, because when machines can truly "read minds," we must ensure they read only the parts we intend to share.
---
*"You focus on moving forward; leave the remembering to me."*