1 reply
0 recast
0 reaction
@xh3b4sd.eth
đź’Żall things are relative.
Which is why I question currrnt techniques for storing it.
The way I see it, we have two options for storing metadata. One is seen in a relational database via Buckets.. or “categories”.
ex. lets say there’s a set of t-shirts
“t-shirts: [{
id. 0
“color”: “green”
“size”:“small”
},
{
id: 1
“color”:”blue”
“size”:“large”
},
{
id:2
“color”:”lapis”,
“size”:“medium”
},
{
id: 3,
“color”:”green”,
“size”:”large”
}
]
this is fine… if we are comfortable leaving us set and stone for all eternity banking on nothing to change. But.. if you’re moving on with this, you will find:
- Placing that column ties data to a specific format (english) or arbitrary label
- wasted space, subjective placement of data
- difficult to add new dimensions
Whereas, we could’ve just looked at each dimension itself AND organized the tshirts on a SPECTRUM for each dimension… i.e. an undirected graph DB. 1 reply
0 recast
0 reaction
id’s: 0,1,2,3
size:0,2,1,3
color:0,3,1,0
This tells the same story as the labels, as well as allowing us to say that “azul is similar to blue”… without restricting us to the labels.
If we want to add labels… sure blue dimension is 1,0
It’s data wrapping around itself! it kinda looks like a spiderweb in my head.
Move to LLMs… Why take data and restrict the neural network to arbitrary values.. when
-Data is surprisingly easy to organize and place
- you could help the human understand by showing us how it is ordering / connecting the data on spectrums.
With this model, we ourselves can add to the “LLM graph”, instead of us being a passer by as the AI algorithms get bigger and bigger by throwing a bunch of shit at it… without us knowing what the hell is going on.
My dream is LLMs without stupid fucking numbers. On a blockchain 🙏 0 reply
0 recast
0 reaction