Introducing AIR-O, a model that decodes what a person is attending to directly from brain scans.
Using only noninvasive brain data, we can generate an image from scratch in real time that depicts where a person is focusing their attention.
This is the first of a new generation of algorithms that can read the shape of thoughts directly from brain activity.
Alljoined
@alljoined
The act of describing our internal world through language can be incomplete and biased.
Our position is that brain embeddings will help overcome those limitations.
These new models enable us to represent the meaning of thoughts as a mathematical coordinate, similar to how LLMs encode words into latent space.
Our position is that brain embeddings will help overcome those limitations.
These new models enable us to represent the meaning of thoughts as a mathematical coordinate, similar to how LLMs encode words into latent space.
5:00 PM UTC · Nov 17, 2026 · 27K Views
31425396
Alljoined
@alljoined
The signal we're using in this demo is 10x weaker than what Neuralink is using today.
Brain activity is very hard to read non-invasively because neurons emit very small voltages and by the time it propagates outside the skull it’s almost all noise.
But we believe that deep learning can extract signal from this noise and our deep learning models make sense of the smallest fluctuations to create a representation that computers can use.
Brain activity is very hard to read non-invasively because neurons emit very small voltages and by the time it propagates outside the skull it’s almost all noise.
But we believe that deep learning can extract signal from this noise and our deep learning models make sense of the smallest fluctuations to create a representation that computers can use.
5:00 PM UTC · Nov 17, 2026 · 17.7K Views
4918231
Alljoined
@alljoined
We’ve observed a few properties when building this task:
- Scaling laws also apply to this domain, where performance improves log-linearly as a function of dataset size
- We can pretrain on other participants first, which cuts down fine tuning time on a new person from 20 hours to 15 minutes
- We can decode objects in novel orientations, albeit in lower performance
- Scaling laws also apply to this domain, where performance improves log-linearly as a function of dataset size
- We can pretrain on other participants first, which cuts down fine tuning time on a new person from 20 hours to 15 minutes
- We can decode objects in novel orientations, albeit in lower performance
5:00 PM UTC · Nov 17, 2026 · 13.9K Views
3211019
Alljoined
@alljoined
As computers become exponentially more powerful we will have an amazing amount of intelligence at our disposal.
Our goal at Alljoined is to bridge the gap between human and artificial minds so that this intelligence can be steered in ways that we actually want.
You can follow our work here: alljoined.com
Our goal at Alljoined is to bridge the gap between human and artificial minds so that this intelligence can be steered in ways that we actually want.
You can follow our work here: alljoined.com
5:00 PM UTC · Nov 17, 2026 · 12.8K Views
3410026