Comment by gabriel666smith | Hacker News Reader

Comment by gabriel666smith | Hacker News Reader

Parent Full thread

gabriel666smith

Huh, interesting. The company has amassed 50k hours of training data - that’s a very large dataset.

A side-project of mine recently has been building out this stack:

- Recording audio in proximity to me, and transcribing it

- Grabbing EEG data from cheap sensors on my own head

- Mapping it against the transcription

With a view to eventually train transformer models which might reliably understand, from EEG signals, what words I am thinking. The existing studies on this method don’t seem to show any drop off in efficacy at mind-reading (lol) as the dataset gets larger.

(If anyone wants to invest some seed money so I can justify spending more time strapping things to my head, do hit me up.)

Crucially, the Muse headband the article talks about doesn’t have a microphone. This seems (slightly!) better for user privacy, but it seems like it would hamper training somewhat, depending on your aims.

A less privacy-respecting company in this space with a popular wellness device could do some pretty dark stuff - it doesn’t really bear thinking about. A sleep tracker app with companion EEG sensor device, for example, could reasonably request microphone access without most users batting an eye.

Cool project. Is there any way to follow along with it? I'm also working on neuro-AI / BCI side projects but for images. Where are you based?

Thanks. I'm London-based. There's nowhere to follow it yet - I'm only a couple of weeks in, and I've currently got it categorised mentally as "weird hobby project", so it hadn't yet occurred to me to publicly document the process at all.

I'm London based too, add me on instagram (@aphynia) or email me if you want to chat more about these kinds of projects.

That’s an interesting project and if you ever succeed it will be massive. What I say shouldn’t stop you but just in case, EEG signals through the scalp are incredibly noisy. You can’t get the level of granularity you need like that. You’re putting a stethoscope on the side of the building and trying to guess what everyone in a few rooms is working on.

Yeah! Super super noisy - and maybe harder than you've described. At least you might overhear some meetings with a stethoscope-on-building approach.

That's kind of the motivation for trying a no-MEG, (comparatively) large dataset, ML approach - it strikes me as exactly the kind of problem to which transformer models might be particularly well-suited, if it's even possible to find a hint of anything in that noise at all.

how, what? without any input from you, ever? did you feed some personal parameters, first? like "fuck the Yankees, love red, may the Wall Street Soviet live on forever" so it can "map assumptions" bouncing around in your head to all that which you catch on your audiovisual peripheral and/or center of focus while you catch it all at different distances and pitches? Are you a person who feels bloodflow/tension in brain parts more or less specifically temporo-spatially ( pre-, during, release, post-release phases, ... ) or not and you just feel the neuro-chemistry vibe with the brain waves ( tired, awake, confused, intensifying, uuuh me so horny )... ?

This paper explains it better than I could:

https://www.nature.com/articles/s41467-025-65499-0

If you know that at [timestamp] a certain word was audible in my vicinity, it might be reasonably safe to guess that I also thought this word around that point, because I heard that word.

The EEG signal my brain creates might behave in a unique way when I heard (and thus, thought) that word. Over time, it might be possible, using machine learning, to notice that unique pattern repeatedly in the data an EEG sensor captures.

So my hope is that decoding EEG might be perhaps simpler than people currently assume. Most studies create their datasets using "lots of participants, participating for relatively short periods" rather than "one participant, over an extremely long period".

If it's possible to have my EEG reasonably-well mapped and decoded (this signal previously repeatedly meant I was hearing the words "uuuh me so horny", to use your example) it might then become possible to read mapped words in the EEG signals even when the mapped words were not audible.

This, in turn, might mean it was possible to read (some of) the words I am thinking.

As a sibling commenter pointed out, it's insanely noisy data, so I have no idea whether any of the above will ever prove possible.

But mapping assumptions in the manner you laid out isn't something I would have any idea how to do, not least because I often hate the sports teams I love when they are underperforming.

  >Are you a person who feels bloodflow/tension in brain parts

Feeling the brain is biologically impossible, since the brain has no nerve endings.

What you're feeling is muscle tension tighten and release in your head/face/neck/scalp, which does effect blood flow.

> is biologically impossible

But there are nerve endings all around the scalp, everywhere in the face and neck, ears, that organ that picks up sound, eye nerves? So some hypersensitivity might create some cerebral geolocationish triangulation of where vessels dilate and constrict? Or are blocked by clots, which gives the flow and dilation a different shape and timing, which also changes clock sync of what usually fires together at much more specific rates than during the blockade.

I'm asking. I'm curious and have not done a preliminary consultation with AI, yet ( why do I feel the need to say this? Probably the sustained increased oxygen supply. Lot's of additional neurogenesis going on in my head atm--19 months on nose plasters )

I used to have annoying TLE attacks which changed thinking patterns and nudged inner dialogues in and out of conversations with people. Even when I was super smooth after weeks of continuous stretching and mobility exercises ( hold two minutes, don't be a noob ) ... but that's about the only somewhat relatable example which I mentioned elsewhere some time in the past. Oh, I felt the TLE like I felt pinching or stuffed sinuses or swallowing unchewed chunks of food for the duration of the TLE.

Some people report tensions on the inside of their forehead even though the particular and surrounding facial nerves are quantifyably ( there's probably a correct term for that ) chill, when taking certain medications.

All kidding aside, a good face and head massage is genuinely life-changing. :)

The brain's mental mapping of nerves in 3D space is highly fluid and adaptive, which is awesome, but it's also why we get things like referred pain and phantom limb syndrome.

Reply on news.ycombinator.com