Hacker News new | ask | show | jobs
by mberlove 1131 days ago
From the study: "...we tested whether successful decoding requires subject cooperation and found that subject cooperation is required both to train and to apply the decoder."

Reading the discussion in the paper, it looks like if you lie "out loud" in your head about what you're thinking, you'll fool the interpretation. For now.

Of course that can change -- will almost certainly change as the technology improves. But it's an important caveat at the moment and the researchers clearly paid attention to this aspect.
4 comments
jebarker 1131 days ago
The need for cooperation suggests that what's being learned here is the correlation between brain activity and produced language. Language is not all (maybe most?) of thinking though [1]. So it may not be a certainty that technology improvements allow decoding of subconscious thought processes. This feels related to the possible fundamental limitations of LLMs trained on language too.

[1] https://nautil.us/the-kekul-problem-236574/

link
niemandhier 1131 days ago
That sounds as if what actually was learned were the neural patterns associated with speaking. AFAIK its thinking out loud in you head causes activation in the motor neurons one would use for speaking.

Things like this have been tried before to help lock-in patents, unfortunately things like ALS seem to correlate with a weakening of those patterns.

link
coding123 1131 days ago
It does not require realizing you are cooperating, if you're just playing a VR game in some future version of Quest that captures brain waves while you're navigating a labyrinth.
link
JohnFen 1131 days ago
But it can't read what you aren't sounding out in your mind. It can't determine truthfulness, and it can't find memories.

If you're playing a VR game, you probably aren't using your inner voice to articulate your secrets. Although if you are of great importance and interest to people with sufficient resources, they might be able to trick you into doing so in a game world.

link
famouswaffles 1131 days ago
There's also a ways to go with the accuracy. On the other hand though, they trained with an extremely small GPT model (GPT one i believe) so...bitter lesson incoming ?
link