• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

AMD WEEKEND ONLY DEAL 7900 XT FROM £729 !!!!

Lightnix said:
I think it's here to stay in the long term, but in the medium term we're still probably going to see another couple of AI winters/hype cycles before it really beds in.
Click to expand...
As AI improves, the demand for faster processing power is going to increase, there won't be any winters because of how useful AI is going to be for assisting people with workflow, it's not like crypto where it will have waves of investment and periods of decline, it's going to be being used constantly and improving constantly
 
Minusorange said:
As AI improves, the demand for faster processing power is going to increase, there won't be any winters because of how useful AI is going to be for assisting people with workflow, it's not like crypto where it will have waves of investment and periods of decline, it's going to be being used constantly and improving constantly
Click to expand...

Exactly, unlike crypto AI has the potential to put thousands (+) out of work, so of course it’s going to be invested in heavily and all the big companies are going to jump on it (they already are).

‘I can get a machine to do this persons job and they never need a sick day or holiday and they do it perfectly always…?’

AI is going to be huge.
 
Minusorange said:
As AI improves, the demand for faster processing power is going to increase, there won't be any winters because of how useful AI is going to be for assisting people with workflow, it's not like crypto where it will have waves of investment and periods of decline, it's going to be being used constantly and improving constantly
Click to expand...

I think investors are not excited by "workflow improvements" and if they decided that's the limit of what e.g. GPT can do they would pull out. The money is all in it for its perceived potential to displace entire workforces but it's not quite there. Don't get me wrong, workflow improvements are great and useful and valuable in their own right but I think they're unlikely to sustain valuations and investment rates at the current levels.

And sure, theoretically there are improvements to be made but more and more I'm seeing evidence that significant improvements could become uneconomical. It will be interesting to see what happens if exponentially increasing (or worse) training demands conflict with Huang's law, for example.
 
Last edited:
nomadd said:
So is crypto. But the bubble still burst.
Click to expand...

Crypto makes no one money.

AI however, results in driverless cars, better search engines, better games, advanced scientific research, better financial, economic and technical modelling, better processing of large amounts of data. Successful companies using AI sell better and more products and services.

AI is like the invention of computers, or more aptly when computers were fast enough for people to use what is now called machine learning.

Blockchain is a different way to create a database (in many ways, worse).
 
Last edited:
Take GPT for example. Everyone agrees that it's fairly impressive, it's improving, it has some uses but is limited in some important fields. But it'll improve to beyond our wildest imaginations in a few iterations right? But here's the thing:

We had GPT-2 in 2019, this allegedly cost around $50k to train. The price of a nice car, lovely.

GPT-3 released in 2020 and cost around $4 million. Ok we're in "quite a nice house" territory.

GPT-4, the most recent model, released in 2023 and is thought to have cost more than $100 million to train. With the advantage of that extra few years of hardware improvement.

This is today. At this rate of cost increase we might have. The thing is the costs go up almost 2 orders of magnitude each time.

Imagine that they carried on down this path

GPT-5, costs £1b-£10b. Somewhere in the ballpark of an entire state of the art microfabrication plant.

GPT-6 would then cost between £100b-£1tn. Here we're at "nation state big ticket item/catastrophe" spending. At this point they've probably decided it's better value to design and manufacture their own custom chips. Maybe they literally have built their own fab. Not sure where this leaves Jensen.

GPT-7: £10tn-£100tb - essentially unobtainable.

And this is factoring in that between each of these steps hardware should have improved. If the rate of change is worse than between when GPT-3 and GPT-4 were trained things get even worse.

So this is the thing, there's a hard ceiling on how good contemporary approaches can get.

To do better needs even more breakthroughs of the sorts that don't necessarily come very easily.
 
Last edited:
GhostDog1981 said:
Lol. Jensen said something very similar recently….
Click to expand...
I know:

1*k7oCTgIzPTHuV4MRG9JChg.png
 
Lightnix said:
Take GPT for example. Everyone agrees that it's fairly impressive, it's improving, it has some uses but is limited in some important fields. But it'll improve to beyond our wildest imaginations in a few iterations right? But here's the thing:

We had GPT-2 in 2019, this allegedly cost around $50k to train. The price of a nice car, lovely.

GPT-3 released in 2020 and cost around $4 million. Ok we're in "quite a nice house" territory.

GPT-4, the most recent model, released in 2023 and is thought to have cost more than $100 million to train. With the advantage of that extra few years of hardware improvement.

This is today. At this rate of cost increase we might have. The thing is the costs go up almost 2 orders of magnitude each time.

Imagine that they carried on down this path

GPT-5, costs £1b-£10b. Somewhere in the ballpark of an entire state of the art microfabrication plant.

GPT-6 would then cost between £100b-£1tn. Here we're at "nation state big ticket item/catastrophe" spending. At this point they've probably decided it's better value to design and manufacture their own custom chips. Maybe they literally have built their own fab. Not sure where this leaves Jensen.

GPT-7: £10tn-£100tb - essentially unobtainable.

And this is factoring in that between each of these steps hardware should have improved. If the rate of change is worse than between when GPT-3 and GPT-4 were trained things get even worse.

So this is the thing, there's a hard ceiling on how good contemporary approaches can get.

To do better needs even more breakthroughs of the sorts that don't necessarily come very easily.
Click to expand...
It's the tech industry, they don't care how much things cost, that's what angel investors are for

You're also not factoring in at some point AI will reach a point where it can help created more advanced tech that will lower costs substantially, whoever owns the AI that does this first is going to be stupidly rich
 
Minusorange said:
It's the tech industry, they don't care how much things cost, that's what angel investors are for

You're also not factoring in at some point AI will reach a point where it can help created more advanced tech that will lower costs substantially, whoever owns the AI that does this first is going to be stupidly rich
Click to expand...
Costs are coming down rapidly as capable AI focused processors are coming to market. But you are right that the number of parameters has been a source of great functional leaps recently, at a large compute cost.

Further, you don't need to retrain your model. You can use it as a foundation model for other models with 0 cost.You can also give it to other people knowing they can't figure out how it was trained, so it's not terrible that it costs to initially train a model.
 
Last edited:
Gibbo said:
4090 sells in great volumes, end of the day the customer buying a halo card is less concerned about cost, its all about performance or having the best, so 4090 sells incredibly well for us and has done since launch, were also the ONLY UK store with them in stock from £1499. :)

Since the start of the month AMD have had some very strong deals, 7900 XT and XTX are both consistent strong sellers for us, but since the deals cards like the 6950 XT, 6800 XT, 6700 XT and 6600 are absolutely flying now in large volumes, the prices we have hit have certainly got the consumer spending again at OcUK on AMD and thus AMD sales numbers are very strong here at present.
Click to expand...

Yes, I had heard that things are picking up. It's interesting because it does mean that there are buyers there, but they are perhaps more price influenced than NVIDIA and Mini-Me-AMD realised.

I certainly do hope that AMD gain some market share, although they don't really deserve it. They have been terribly slow to take advantage of NVIDIA's blunder.
 
SpellowHouse said:
Yes, I had heard that things are picking up. It's interesting because it does mean that there are buyers there, but they are perhaps more price influenced than NVIDIA and Mini-Me-AMD realised.

I certainly do hope that AMD gain some market share, although they don't really deserve it. They have been terribly slow to take advantage of NVIDIA's blunder.
Click to expand...
I just hope NVIDIA just give up and leave the market tbh.
 
You must log in or register to reply here.
Back
Top Bottom