It's worth noting that this isn't new technology. This paper is specifically about how their new technique provides a small but statistically significant improvement on existing techniques.
The fact that they provide code and dataset is really praiseworthy.
Reminds me of the scene in series Incorporated where a megacorp uses this sort of tech to interrogate an employee from a competitor mega-corp to get at a trade secret.
It's a little spooky. Oh and that series was a dystopian series because ofc it was
So attended an interesting talk a couple years ago:
- fMRI and/or brain implants are the best to figure out brain waves
- but they are expensive or invasive
- EEG is a lot cheaper and easier but not as precise
- BUT what if you used LLMs to analyze EEG data taken at the same time as brain implants etc
The answer seemed to be that "yes, you can get better than traditional EEG data using EEG + LLMs". Curious to see where this ends up and hopefully not that like that Black Mirror episode with the brain scanning leading to murders.
If one were to go about translating brain waves from dogs to meaning, we'd run into a big problem immediately: vocabulary resolution.
What I mean by that is we'll have a very limited number of words to which a dog's brainwaves can be translated to since we aren't able to understand them beyond their basic instincts of food, survival, fear, affection towards their owner etc.
There is just no way to go past what we have already observed by their behavior since dogs can't talk or write.
I do wonder how animals think. Perhaps this resolution would also be the theoretical maximum?
>There is just no way to go past what we have already observed by their behavior since dogs can't talk or write.
There are many dogs that have been trained to press buttons corresponding to words, in the extreme case tens/hundreds of buttons/words, and they can even construct rudimentary sentences. It doesn't seem insane to me that we could perhaps do a very rudimentary version of this for dogs, given a large enough training set.
The ultrasound technique there is more like MRI, static imaging, not measuring dynamic electrical signals. Also, regardless of static vs moving, all of this hinges on massive relatively expensive devices (ultrasound is never going to be dirt cheap or miniaturized compared to a smartwatch or even a VR headset) and for the subject to remain perfectly still and probably go through regular calibration sequences. All for <75% accuracy on simple classification tasks. The information mixes in the propagation out of the skull, erasing information. It's analogous to trying to do a row hammer attack on a CPU from outside the computer case.
Also, in the Meta result here, "while actively typing" is actually quite different than passive mind reading because the motor cortex sits nicely near the top surface of the skull, and the muscle memory from past typing makes for a nice well formed signal to measure and classify. It's the same trick over and over since the Brown BrainGate days where you can have people perform or imagine movements and get a decent but not good classification result, and it never gets much better after that trick is exploited. Project dies, VCs and grant writers forget or never appreciated the effect, time goes by, a grad student or corporate research lackey rediscovers it, media puts out an article claiming mind reading is here, and the cycle repeats...
This wouldn’t work. Apart from it being very difficult to get people to sleep deeply enough for them to dream in machines like this, the brain state is very different whilst asleep compared with awake. Also the data generated from typing would be very different than thought since it’s likely picking up on broad electrical activity in the primary motor cortex.
Source: spouse works in a sleep lab studying dreams with MRI
They are already paying scarce labor like medical doctors and lawyers hundreds of dollars an hour to create training data. The RoI for training data is high because it can be used to train many models.
Interesting -- really excited for the future of human-brain interfaces and just in general more interface exploration enabled by large transformers. I'm already very excited by voice, although wish I could get something akin to the subvoc common in scifi novels. Seems like it would be an easier path than human-brain and would allow me to use voice models in public.
As an aside, disappointed by the very low quality of comments on this article here.
The fact that they provide code and dataset is really praiseworthy.
It's a little spooky. Oh and that series was a dystopian series because ofc it was
- fMRI and/or brain implants are the best to figure out brain waves
- but they are expensive or invasive
- EEG is a lot cheaper and easier but not as precise
- BUT what if you used LLMs to analyze EEG data taken at the same time as brain implants etc
The answer seemed to be that "yes, you can get better than traditional EEG data using EEG + LLMs". Curious to see where this ends up and hopefully not that like that Black Mirror episode with the brain scanning leading to murders.
- Let's go see what's on the other side of this door, friend, maybe there's food !!
Ok, friend, here you go. - opens door.
- Wow super cool, now let's go see what's on the other side of this door, friend, maybe there's food !!
What I mean by that is we'll have a very limited number of words to which a dog's brainwaves can be translated to since we aren't able to understand them beyond their basic instincts of food, survival, fear, affection towards their owner etc.
There is just no way to go past what we have already observed by their behavior since dogs can't talk or write.
I do wonder how animals think. Perhaps this resolution would also be the theoretical maximum?
There are many dogs that have been trained to press buttons corresponding to words, in the extreme case tens/hundreds of buttons/words, and they can even construct rudimentary sentences. It doesn't seem insane to me that we could perhaps do a very rudimentary version of this for dogs, given a large enough training set.
Might be of use.
https://news.ycombinator.com/item?id=48685558
Also, in the Meta result here, "while actively typing" is actually quite different than passive mind reading because the motor cortex sits nicely near the top surface of the skull, and the muscle memory from past typing makes for a nice well formed signal to measure and classify. It's the same trick over and over since the Brown BrainGate days where you can have people perform or imagine movements and get a decent but not good classification result, and it never gets much better after that trick is exploited. Project dies, VCs and grant writers forget or never appreciated the effect, time goes by, a grad student or corporate research lackey rediscovers it, media puts out an article claiming mind reading is here, and the cycle repeats...
How realistic would it be to make a smaller device?
Source: spouse works in a sleep lab studying dreams with MRI
p.s. come on you guys - this is not what HN is for. You may not owe $megacorp better but you owe this community better if you're participating here.
When the bad guys try they just get the lyrics to Yoko Ono music.
Meta has shown remarkable disregard to users' and employees' privacy.
Why should that not come up when discussing an entirely new and dystopian technology that allows them to do this at scale?
https://news.ycombinator.com/newsguidelines.html
Better than text-stripping the internet - this thing will soon be pulling the logits as well.
"Don't be snarky."
https://news.ycombinator.com/newsguidelines.html
1. The post was obviously bullish / optimistic on the technical capabilities. Not in the least dismissive.
2. The economics extrapolation is obvious. See current precedent for paid access for purchased screen-casts of dev work: https://pdoom.org/open_calls/04_crowd_cast.html
they'll just put it buried on page 450 of the meta glasses 3 or something
There are no layoffs in Ba Sing Se.
As an aside, disappointed by the very low quality of comments on this article here.