211–220 of 415 posts
GN did a documentary on the situation from the perspective of consumer-facing companies. Seems pretty dire for them, and it's hard to see the long-range consequences, but the idea of consumers being priced out isn't too far out, which to me is a little alarming. https://m.youtube.com/watch?v=zyQwAhppWj8
Earlier quoted context omitted.
I was picking up DDR3 16GB sticks for $5/piece on eBay last year. The world has gone mad.
To be fair, DDR3 is still kinda cheap. It's really only DDR4 and DDR5 that are massively in demand. DDR3 is a bit too old for it to be in high demand by consumers.
AMD just brought the popular 5800X3D back out of retirement to give people maintaining the DDR4 based platforms something to buy. Last I checked used DDR4 was half the price of used DDR5 after the prices of both shot up.
Everyone re-releasing their old crap. Nvidia is remaking the 3060 too I believe because they can use cheaper older RAM on it.
Same thing with GPUs, is kind of insane having so much processing power and yet requiring more and more. What purpose for? What's the limit? Does it really really pay off such investments?
For me as a non-AI developer (I don't use any models of any kind, nor I train LLMs at all), a system with 16Gb seems to be more than enough for a vast number of applications.
Time to migrate all those electron "native" apps to actual native code. I bet with some decent optimizations 4GB will be more than enough for casual user, and still with some free memory to spare.
Earlier quoted context omitted.
You're wrong. There is demand. More and more people are exploring AI and getting real work done with it.
It's like how some people like listening to Ed Sheeran. So yes, there is demand. But nobody is "getting real work done" with these toy AI models. Real AI is a geo-political threat, and will not be allowed to exist for the average person. So, enjoy your toy AI models, because that's all you're getting.
Earlier quoted context omitted.
> that local models are good enough for whatever they're peddling they are not. Unless you are satisfied with plausible, but mostly garbage output.
They are actually quite a bit better than you might think. Qwen3.6 27B is pretty capable at coding. For non-coding work, they are more than good enough. A lot of the ways my non-technical family members have interacted with AI would be perfectly served by using a local model. After all, people were more than satisfied with the results from GPT 3. That has long since been surpassed by open weight models.
Earlier quoted context omitted.
Yup. Once production ramps up (and then subsequently overshoots) in the next couple years we're going to have GPUs with 96GB VRAM for the price of a 4090 today.
They're not ramping up capacity though.
Earlier quoted context omitted.
Yesterday I did a price check on the PC I built two years ago. It went from $2300 to $3650. The bulk of that increase was that the ram went from $210 to $940. Its now more expensive than when DDR5 was new.
Those are rookie numbers https://stocks.sjer.red/
But then again I remember spending hundreds of dollars as a high school student to upgrade my family's 8Mb desktop to 40 Mb.