What will have when these llm models end up trained with all media already produced by humans?

synthetic data will allow them to continue training em

 

yes!

So it’s a good bad thing humans post so much AI slop AI will just eat itself.

What synthetic data means?

 

Won't this cause a negative cycle of innacuracy and deformity?

Habsburg AI

disturbingly consistent, although the background crowd is clearly not real sometimes

We as humans can endlessly produce more data.

 

Go into the field with your smartphone and start taking photos of flowers, insects, macro shots of grass, walls - thats new data, that never existed before. You go and take the photo of the cloud - thats new data. You go and take the photo of the dirt - thats new data.

 

You go into the crowd of humans, who all wear different clothes, have different skin color and you take videos of just random people - thats new data. You send a car with Google panoramic camera on the street - thats new data, because buildings could have changed, trees could have changed, lots of things are not the same on the street view after a year.

You can just take photos of your shoes each day and each day your shoe will look differently accumulating dirt or wear&tear and thats data.

 

You just record a baby mumbling or an old man rambling - thats a new data. You just put an audio recording device in the forest and it records birds singing - thats a new data.

 

We are not recording and not harnessing a lot of data, a lot of data is just being lost every moment, so i don't fucking see how the data can become scarce or used up. Earth is just a magical place, this planet is full of life and full of data. For comparison the Moon is very empty and there is no data on the Moon, except rooks and craters that almost all look the same.

The teeth are often too perfect in the AI generated content.

 

The real teeth is more random, the teeth of humans can tell a lot of stories, they are often very personal. Dead people are often identified by their dental records. Thats data, that is not being tapped in AI training.

Thank you, we will take note and fix this.

t. google fake news engineer

And still not able to properly generate artistic text in the images.

p. good video on the topic

Or for example this phenomenon

 

 

So they trained AI on the limited images of North Korean cities and then use it to overlay over GTA. That shows you how just human urban environment can become data used for AI. And there are a lot of different cities, villages, modern and historic that can be mined for data. Different countries, cultures, different seasons and times of day - like NY at night is different than NY during the day.

We also can train AI not only on the real world data, but on the data from the video games. So imagine a gameplay of the game like gothic 2

 

and we train AI on the videos of this game - how to generate new levels for this game, new NPC - completelly bypassing the game engine

 

The training data is not from the real world, but it is not AI generated. The training data from the game is generated by the game engine that has its own rules, so there will be a consistency in logic of the content that is generated based on this training dataset. We can have AI play the game on its engine, record the gameplay and generate new data to train another AI that will be generating new content for this game without relying on the engine mechanisms, but replicating the rules that underpin it.

 

Gothic 2 is just an example. How many games are out there that we can datamine for this content? Endless possibilities

nah, i aint afraid of this happening soon. Dental data is considered biometric and protected by some laws, so the AI companies will have a hard time accessing big troves of it

No

AI is already being trained with synthetic data (algoritmically generated data that imatates the real thang) and it's only becoming more accurate

Media serves us fake news all the time. Nothing will change.

Eg Schrödinger's Russia:

- Russia is constantly on the brink of collapse just after the next sanction package, they have to mine chips out of washing machines and fight with shovel.

- Russia is too strong a mortal danger to NATO and will occupy Berlin tomorrow.

So everything will be alright and llm development won't hit a wall?

Yes. And as soon as it takes over the production chain from raw materials through energy to robot assembly and will be capable of autonomous reproduction it will exterminate humanity.

(((they))) are not going to run out of data but computing power and energy comsumption

will become a problem soon

eventually only light based or thermodynamic computing will be able to run LLMs

t. AI influencers on Youtube watching pro

Wrote a bash script that does a thing. It's less than 1Kb and uses virtually not noticable amount of system resources. Mate wants to plug LLM in that needs 36Gb of VRAM to generate input for it.