Artificial Intelligence (AI)

[ALC Fox]

Regarding the energy demands of AI, how feasible is it really that data centres could be put in space?

 

Is the energy needed to put them there less than that required if they were situated on the  Earth instead, bearing in mind that they may need to be repaired and/or upgraded periodically?

 

Also what about ever-increasing space pollution and the increasing likelihood of debris in space becoming a danger for other satellites?

 

I'd genuinely like to read about the pros and cons, understanding that once the data centres are up there, they're powered by limitless solar energy.

[Chelmofox]

11 hours ago, ALC Fox said:

Regarding the energy demands of AI, how feasible is it really that data centres could be put in space?

 

Is the energy needed to put them there less than that required if they were situated on the  Earth instead, bearing in mind that they may need to be repaired and/or upgraded periodically?

 

Also what about ever-increasing space pollution and the increasing likelihood of debris in space becoming a danger for other satellites?

 

I'd genuinely like to read about the pros and cons, understanding that once the data centres are up there, they're powered by limitless solar energy.

Not my area of expertise, but i thought the heat dynamics dont really work in space. Sure its cold, but with no atmosphere you cant get the heat away from the hardware. I also image that real world latency would be a massive issue, as well general maintenence. 

Edited 9 hours ago by Chelmofox

[danny.]

17 hours ago, Chelmofox said:

Not my area of expertise, but i thought the heat dynamics dont really work in space. Sure its cold, but with no atmosphere you cant get the heat away from the hardware. I also image that real world latency would be a massive issue, as well general maintenence. 

Latency isn't an issue at all. APIs are nowhere near instant now, even the quickest models have perceivable delays of seconds not ms. Starlink V2 latency is < 5ms, assuming AI satellites are in that ballpark or even double that - it wouldn't be noticeable for gaming let along LLM API calls.

[leicsmac]

18 hours ago, Sampson said:

It's the 21st century equivalent of the nuclear Arms race between the US and China in the 1950s and 60s.

 

Just hope we in Europe decouple from it all.

And it's potentially much more dangerous for the whole world in terms of destructive power than nuclear weapons are. 

[Zear0]

20 hours ago, leicsmac said:

Just listened to some American talking head going on about how the US is totally right to go all in on AI and data centres for reasons that boil down to basically "we good, China bad, we do it or they do" and that "the wars of the future will be fought with AI".

 

That would be the most insanely hubristic and self destructive sentiment I had heard today if I hadn't heard about the possibility of Leicester appointing Martin as manager. 

Was a good chat with Jack Clark I saw and it’s not even US (good) China (bad).

 

All the leaders of these companies are massive distrusting of one another and are racing each other. Jack mentioned there’s a perception they all think Elon, Sam, Mark Demis and Dario are bad and only they can be trusted with the pinnacle model.

 

 

[leicsmac]

6 minutes ago, Zear0 said:

Was a good chat with Jack Clark I saw and it’s not even US (good) China (bad).

 

All the leaders of these companies are massive distrusting of one another and are racing each other. Jack mentioned there’s a perception they all think Elon, Sam, Mark Demis and Dario are bad and only they can be trusted with the pinnacle model.

 

 

It was some talking head on Reuters that I saw. I'll have to look at this interview here. 

 

You're right about the competition element, but I think it's a distinct possibility that these people are messing with forces they don't really understand in the name of personal legacy and prestige. And that doesn't end well.

[Chelmofox]

1 hour ago, danny. said:

Latency isn't an issue at all. APIs are nowhere near instant now, even the quickest models have perceivable delays of seconds not ms. Starlink V2 latency is < 5ms, assuming AI satellites are in that ballpark or even double that - it wouldn't be noticeable for gaming let along LLM API calls.

It is. We are talking data centers here that may need to communicate with each other. The latency between you contacting an LLM and getting the result isn't the only thing going on.  To get results from an LLM requires communication across a huge amount of nodes, and when these are hit at scale there needs to be fast, transactional compute ensuring the huge amount of requests can happen in a timely manner. Then there is all the transactional logging going on, and also some more traditional steps happening across lots of different aps.  Also, there is black magic i dont understand with saving thought requests from LLM calls, hence why you can resume conversations over long context periods.  

 

Not seen anything to suggest Starlink latency is <5ms. Their own docs suggest 25->60 in 'good' areas. God knows what node to node would be like. https://starlink.com/legal/documents/DOC-1470-99699-90

 

Data centers do more than just output LLM requests. I spent several years working in the digital ads space. Real time bidding requires a huge number of request across supply/demand nodes to complete all steps in 100ms. Sometimes this arms race meant Demand Platforms paying big bucks to get the physical infrastructure as physically close to the supply sources as possible to gain an advantage. Can't see anything like that happening when the latency of one request is 60ms in the best case scenario from satellite to earth. 

Edited 55 minutes ago by Chelmofox