Hacker News Reader

Comment by smusamashah | Hacker News Reader

This video https://x.com/Sentdex/status/1845146540555243615 looks way too much like my dreams. This is almost exactly that happens when I sometimes try to jump high, it transforms me to a different place just like that. Things keep changing just like that. It's amazing to see how close it is to a real dream experience.

View on HN

I noticed that all text looked garbled up when I had some lucid dreams. When diffusion models started to gain attention, I made the connection that text generated in generated images also looked garbled up.

Maybe all of those are clues that parts of the human subconscious mind operate pretty close to the principles behind diffusion models.

I also lucid dream occasionally. Very rarely things are very detailed, most often the colors and details are just as bleak and blurry and keep changing as these videos. I walk down a street, take a turn (or not), its almost guaranteed I can't go back to where I came from. I usually appreciate when I can track back the same path.

I've watched entire movies while lucid dreaming. I've listened to entire concerts play out, eaten entire meals with all 5 senses.

One of the more annoying parts of growing older was that I stopped lucid dreaming.

I don't think lucid dreaming is a requirement for this. Whenever I dream my environment morphs into another one, scene by scene, things I try to get details from, like the content of a text, refuse to show clearly enough to extract any meaningful information from it, no matter what I try.

"I try" sounds like lucid dreams though? In my dreams there is no try, things just happen (including my actions). I think I get meaningful information sometimes, but it's selective

Also the AI generated images that can't get the fingers right. Have you ever tried to look at your hands while lucid dreaming and try counting fingers? There are some really interesting parallels between the dreams and diffusion models.

Of course, due to the very nature of dreams, your awareness of diffusion models and their output flavors how you perceive even past dreams.

Our brains love retroactively altering fuzzy memories.

On the other hand, psychedelics give you perceptions similar to even early deepdream genai images.

On LSD, I was swimming in my friend’s pool (for six hours…) amazed at all the patterns on his pool tiles underwater. I couldn’t get enough. Every tile had a different sophisticated pattern.

The next day I went back to his place (sober) and commented on how cool his pool tiles were. He had nfi what I was talking about.

I walk out to the pool and sure enough it’s just a grid of small featureless white tiles. Upon closer inspection they have a slight grain to them. I guess my brain was connecting the dots on the grain and creating patterns.

It was quite a trip to be so wrong about reality.

Not really related to your claim I guess but I haven’t thought of this story in 10 years and don’t want to delete it.

This may be a joke, but counting your fingers to lucid dream has been a thing for a lot longer than diffusion models.

That being said, your reality will influence your dreams if you're exposed to some things enough. I used to play minecraft on a really bad PC back in the day, and in my lucid dreams I used to encounter the same slow chunk loading as I saw in the game.

Playing Population One in VR did this to me. Whenever I hopped into a new game, I'd ask the other participants if they'd had particularly vivid dreams since getting VR, and more than half of folks said they had.

What’s amazing is that if you really start paying attention it seems like the mind is often doing the same thing when you’re awake, less noticeable with your visual field but more noticeable with attention and thoughts themselves.

This is a very interesting thought. I never thought of mind doing anything like that in wake state. I know I will now be thinking about this idea every time I recall those dreams.

Yeah I also hadn’t but it makes sense. Just like all output from an LLM is a “hallucination” but we have a tendency to only call it a hallucination when something looks wrong about the result, we forget that our conscious lived experience is a hallucination based on a bunch of abstract sensory data that our brain fuses in to a world state experience. It’s obvious when we are asleep that dreams are hallucinations, but it is less obvious that the conscious experience is too.

Agree that the whole senate field is a type of hallucination. But when it comes to thoughts appearing in our minds their transitions are much more classically hallucination like. I became aware of this during meditation. It has a strange intoxicating quality where you have to figuratively pinch yourself to notice how bizarrely one thought transitions to a thought in the periphery of your attention.

Its interesting how much dreams differ from person to person. Mine tend to be completely coherant visually, to the point that I have used google maps in my dreams, and while the geography was inaccurate, it was consistent. However, I have never been lucid within a dream, maybe that makes a difference

This is why I'm excited in a limited way. Clearly something is disconnected in a dream state that has an analogous disconnect here.

I think these models lack a world model, something with strong spatial reasoning and continuity expectations that animals have.

Of course that's probably learned too.

That looks way too much to the one time I did DMT-5

Machine Elves

IYKYK

How are you so sure this is like your dreams? If it was easy to accurately remember dreams why would they be all so smooshy and such a jumbled mess like in this video?

We are unconsciously (pun intended) implementing how brains work both in dream and wake states. Can't wait until we add some kind of (lossless) memory to this models.

Any evidence to back this lofty claim?

vibes

We have lossless memory for models today. That's the training data. You could consider this the offline version of a replay buffer which is also typically lossless.

The online, continuous and lossy version of this problem is more like how our memory works and still largely unsolved.

Reply on news.ycombinator.com