LifeFuel AI is going to get massively better

[nvrbegan]

First reason: hardware is overpriced right now, it could be 7x cheaper and would still be profitable. Nvidia is milking their customers until their competitors catch up.

Second: There are a lot of software optimizations possible. In 10 years nobody will do AI like now. A lot of papers and effort engage with this topic.

Third: Current AI is undertrained in relation to their model size, so just by training more (which will become cheaper automatically with better hardware) the models will get better, without raising VRAM or computing requirement.

Fifth: Datasets will become more abundant and easier to obtain and higher quality, just because more people will work on this issue.
Companies used to use whatever they scraped from the web, now they actively create high-quality content specifically intended for AI training, too.

Sixth: specialized AI-hardware. Right now, AI-chips are essentially modified GPUs. GPUs are good at rendering graphic and due to their parallelization they can be used fir linear algebra/vector math too, but if the entire chip were designed from beginning to end with AI and not graphics in mind, a speed-up of maybe 4x or more in the same chip size were possible.

Seventh: other random improvements and fierce competition will drive the development rapidly forward.

 

[CruxGammata]

Are you Mental Outlaw?

 

[nvrbegan]

Are you Mental Outlaw?

No, I don't know who he is.
Why do people keep asking this?

 

[CruxGammata]

No, I don't know who he is.
Why do people keep asking this?

Maybe it has something to do with your avi idk

 

[nvrbegan]

That's also the reason why local AI will become better relative to SAAS-offerings.
Right now there are so many low-hanging fruits that there's rapid progress (that can be bought with tons of money that only big corpos have).

Foe example, with datasets:
If the entire dataset is 10 TB and you add 10 TB, you've doubled it.
If it's 100 TB and you add 10, you only 1,1x-ed it.
The basis of open dataset will grow and grow and it will be harder and harder for corpos to edge out an advantage by creating new proprietary datasets.

 

[nvrbegan]

Maybe it has something to do with your avi idk

Does he has the same?
I downloaded this pic from 4chan, checked the file name and it's not from here.

 

[hierophant]

First reason: hardware is overpriced right now, it could be 7x cheaper and would still be profitable. Nvidia is milking their customers until their competitors catch up.

Second: There are a lot of software optimizations possible. In 10 years nobody will do AI like now. A lot of papers and effort engage with this topic.

Third: Current AI is undertrained in relation to their model size, so just by training more (which will become cheaper automatically with better hardware) the models will get better, without raising VRAM or computing requirement.

Fifth: Datasets will become more abundant and easier to obtain and higher quality, just because more people will work on this issue.
Companies used to use whatever they scraped from the web, now they actively create high-quality content specifically intended for AI training, too.

Sixth: specialized AI-hardware. Right now, AI-chips are essentially modified GPUs. GPUs are good at rendering graphic and due to their parallelization they can be used fir linear algebra/vector math too, but if the entire chip were designed from beginning to end with AI and not graphics in mind, a speed-up of maybe 4x or more in the same chip size were possible.

Seventh: other random improvements and fierce competition will drive the development rapidly forward.

I don't understand the technical aspects of what you mean, but the idea is clear.

 

[turbocuckcel_7000]

the thing is AI still seems really inconsistent for something like creating a convincing waifu experience, i get a bigger kick out of playing with koikatsu instead of AI simply because it makes a lot less mistakes with its 3d character rendering and can handle any angle and perspective
i don't really see this changing soon because too much is being asked of an AI when you just feed it some drawings and text, and then you expect it to understand humans, even if it's just cartoon humans

 

[nvrbegan]

Does he has the same?
I downloaded this pic from 4chan, checked the file name and it's not from here.

You can download my avi and search on desuarchive or 4plebs with the file hash, someone posted this pic ~2 weeks ago and when it came to uploading a profile pic I looked into my download folder.

 

[Linguinicel]

IQ mogs me but sounds good nonetheless

 

[nvrbegan]

Currently, when you want to create a model with x parameter, you create a model with x parameters and then initialize those parameters with random numbers. During training, this randomness vanishes and a meaningful model emerges. However, there's still a lot of randomness left that isn't trained out because it would be too expensive.
A model with x parameters is always the same regarding how much ressources (compute+ram) you need.
A well trained model needs as much as a bad model if both have the same parameter size. So just training more is a low hanging fruit that can be grabbed with better hardware.

 

[nvrbegan]

Currently, when you want to create a model with x parameter, you create a model with x parameters and then initialize those parameters with random numbers. During training, this randomness vanishes and a meaningful model emerges. However, there's still a lot of randomness left that isn't trained out because it would be too expensive.
A model with x parameters is always the same regarding how much ressources (compute+ram) you need.
A well trained model needs as much as a bad model if both have the same parameter size. So just training more is a low hanging fruit that can be grabbed with better hardware.

Thats what I meant in third.
Sixth basically means:
You can eat cornflakes with a spoon or a fork. Both need the same amount of metal and are equally cheap/easy to manufacture. Currently, we use a fork but switching to a spoon is easy. It needs a few years of research until it's done, but another low-hanging fruit.

 

[CruxGammata]

Does he has the same?
I downloaded this pic from 4chan, checked the file name and it's not from here.

hahaha /g/tards

 

[ars]

maybe porn search engines will finally be fixed then

 

[BlackPill Scholar]