HOLY SHIT thought-to-action is almost here meta just published two...
vittorio@IterIntellectus
82 views
Feb 17, 2025
~2 min read
Advertisement
1
HOLY SHIT
thought-to-action is almost here
meta just published two papers showing they can decode thoughts into text with 80% accuracy, in real time, using non-invasive brain-computer interfaces.
pardon my french, but WHAT THE FUCK?
thought-to-action is almost here
meta just published two papers showing they can decode thoughts into text with 80% accuracy, in real time, using non-invasive brain-computer interfaces.
pardon my french, but WHAT THE FUCK?
0:17
2
they used magnetoencephalography (MEG) and EEG to record brain activity while 35 typists typed sentences
then, they trained a deep learning model (brain2qwerty) to predict the typed text from brain signals.
I knew we were getting here but not this fast. absolutely insane.
then, they trained a deep learning model (brain2qwerty) to predict the typed text from brain signals.
I knew we were getting here but not this fast. absolutely insane.
3
with MEG, the model got to 32% character error rate on average, meaning it got the right letter 7-8 times out of 10.
for the best subjects, it decoded full sentences perfectly, outside of training data
this isn't sci-fi. this is now.
for the best subjects, it decoded full sentences perfectly, outside of training data
this isn't sci-fi. this is now.
4
up until now, the best brain-to-text BCIs required invasive brain implants (think elon’s neuralink)
non-invasive EEG-based systems sucked, maxing out at 43% accuracy on basic tasks
non-invasive EEG-based systems sucked, maxing out at 43% accuracy on basic tasks
5
their deep learning model + MEG setup does two crucial things:
- it uses motor signals from typing to boost accuracy
- it uses transformers + language models to predict text more efficiently
they basically built a brain-powered autocorrect.
- it uses motor signals from typing to boost accuracy
- it uses transformers + language models to predict text more efficiently
they basically built a brain-powered autocorrect.
6
thought-to-text BCIs are coming faster than anyone expected.
right now, it still requires an MEG scanner (big, not portable). but once they adapt this to wearable MEG sensors, it’s so over
silent communication, direct thought-to-text, instant AI commands.
right now, it still requires an MEG scanner (big, not portable). but once they adapt this to wearable MEG sensors, it’s so over
silent communication, direct thought-to-text, instant AI commands.
7
it means:
- direct brain control over digital interfaces
- speech restoration for people who can't talk
- silent communication for soldiers, operators, or anyone needing stealth
- and soon full mind-machine integration
- direct brain control over digital interfaces
- speech restoration for people who can't talk
- silent communication for soldiers, operators, or anyone needing stealth
- and soon full mind-machine integration
8
we are approaching a world where your thoughts can be read, transmitted, and executed as actions
the interface between mind and machine is dissolving
welcome to the future.
the interface between mind and machine is dissolving
welcome to the future.
9
still, right now we cannot “write” inside the brain yet, we'd need to stimulate neurons directly.
four ways to do it:
- ultrasound (focused brain activation)
- TMS (transcranial magnetic stimulation, magnetic pulses to shape neural firing)
- brain implants (microstimulation of cortex, neuralink-style)
- optogenetics (future, precise light control over neurons)
when this gets solved, the interface between mind and machine is gone
four ways to do it:
- ultrasound (focused brain activation)
- TMS (transcranial magnetic stimulation, magnetic pulses to shape neural firing)
- brain implants (microstimulation of cortex, neuralink-style)
- optogenetics (future, precise light control over neurons)
when this gets solved, the interface between mind and machine is gone
10
if you could “draw” letters inside the brain, you could write thoughts into existence.
focused ultrasound might be the first real shot at this
beams of ultrasound can activate specific neurons without surgery, it's already used to stimulate motor areas & treat disorders, and in theory, it could “paint” words into your language cortex.
right now, it’s imprecise. but once it scales, silent thought-to-text becomes real
focused ultrasound might be the first real shot at this
beams of ultrasound can activate specific neurons without surgery, it's already used to stimulate motor areas & treat disorders, and in theory, it could “paint” words into your language cortex.
right now, it’s imprecise. but once it scales, silent thought-to-text becomes real
11
brain-to-text is coming. brain-to-brain or AI-to-brain is next.
once we solve writing into the brain, you could:
- “download” a new language overnight.
- control machines with pure thought.
- receive silent messages directly into your mind
this tech is moving faster than people think
once we solve writing into the brain, you could:
- “download” a new language overnight.
- control machines with pure thought.
- receive silent messages directly into your mind
this tech is moving faster than people think
