Mr. Bloody Lemonade !0JqexkOGBc (OP) double-posted this 1 month ago, 1 minute later[^][v]#1,451,026
There are some major drawbacks to this. Mainly attributed to the fact that hardwired the weights of the model into the chip so the model can never ever be changed or updated.
Mr. Bloody Lemonade !0JqexkOGBc joined in and replied with this 1 month ago, 11 minutes later, 13 minutes after the original post[^][v]#1,451,027
Of course this isn’t a new idea. Chips have done the same thing with video codes for ages. Just hardwire the program into the chip and now it’s extraordinary fast… as long as you never ever need to change it.
Anonymous C replied with this 1 month ago, 19 minutes later, 2 hours after the original post[^][v]#1,451,036
@previous (Mr. Bloody Lemonade !0JqexkOGBc)
If the model fits it's purpose, the increase in speed can be worth it. A lot of workplaces use equiptment that they bought 20 years ago and still works fine.
The tradeoff works. They don't need neverending updates, they need the speed.
Mr. Bloody Lemonade !0JqexkOGBc joined in and replied with this 1 month ago, 1 minute later, 2 hours after the original post[^][v]#1,451,038
@previous (C)
It’s definitely useful. The problem is that facts change and if the model doesn’t change, it will be wrong at some point in the future. Imagine asking a model hard coded into a chip who the president of the United States is when you bought it in 2026 but now it’s 2030 for example.
Anonymous C replied with this 1 month ago, 3 minutes later, 2 hours after the original post[^][v]#1,451,039
@previous (Mr. Bloody Lemonade !0JqexkOGBc)
Which is why it depends on the application. For some uses, it will never need to know about current events.
Mr. Bloody Lemonade !0JqexkOGBc replied with this 1 month ago, 45 seconds later, 2 hours after the original post[^][v]#1,451,041
Like there weren’t LLMs 20 years ago, but hypothetically if they were and you spun up a 20 year old LLM and you started asking it questions like what’s the world population, who’s the president, who’s the richest person, etc. that might be kind of a problem. Even for coding tasks, if you asked a 20 year old LLM to write a website using react… that would cause some problems.
Mr. Bloody Lemonade !0JqexkOGBc triple-posted this 1 month ago, 4 minutes later, 2 hours after the original post[^][v]#1,451,043
Also you know IRL what they’re going to do is big tech companies would buy this and throw them all out every few years which would be incredibly wasteful so from an environmental perspective I really hope they don’t do this because this is worse than what they’re already doing which is not great.
Anonymous C replied with this 1 month ago, 1 minute later, 2 hours after the original post[^][v]#1,451,044
@1,451,041 (Mr. Bloody Lemonade !0JqexkOGBc) @1,451,042 (Mr. Bloody Lemonade !0JqexkOGBc)
You're still missing the point. You wouldn't use this in a situation where current events matter.
> You're still missing the point. You wouldn't use this in a situation where current events matter.
Yeah, I’m aware of that. The thing is, situations where current events matter is like… most situations. I’m not saying it has no use I’m saying it has a more limited use case.
Anonymous C replied with this 1 month ago, 23 seconds later, 2 hours after the original post[^][v]#1,451,058
@1,451,055 (Mr. Bloody Lemonade !0JqexkOGBc)
When you never acknowledge the qctual point of the product, and start talking about how it doesn't fit into unrelated niches, it seem like you don't understand the point of what they are doing.
> It’s definitely useful. The problem is that facts change and if the model doesn’t change, it will be wrong at some point in the future. Imagine asking a model hard coded into a chip who the president of the United States is when you bought it in 2026 but now it’s 2030 for example.
The chip has been around since February. It’s not new. I understand that it has a specific use case, my point is that use case isn’t most people’s use case so it will have a relatively small market share within the entire AI industry.
Anonymous C replied with this 1 month ago, 3 minutes later, 3 hours after the original post[^][v]#1,451,061
@previous (Mr. Bloody Lemonade !0JqexkOGBc)
Imagine someone made a new topic talking about paperclips, and how these paperclips can't hold together heavy machinery and did a bad job of pinning clothes.
It's not that it's wrong, it's that you come off looking too stupid to understand the point of the product.
No one needs it pointed out a paperclip can't hold heavy machinery together, just like no one needs you to point out that this is updateable.
You know when I asked if you were being annoying for the sake of annoying, you could have just said yes and that would have been more efficient, right?
> No one needs it pointed out a paperclip can't hold heavy machinery together, just like no one needs you to point out that this is updateable.
Yeah you did, you didn’t know about it until I mentioned it, you said the website wasn’t loading for you so you obviously never used it before. How would you know why it was 17,000 tokens per second if I didn’t tell you?
Mr. Bloody Lemonade !0JqexkOGBc double-posted this 1 month ago, 2 minutes later, 3 hours after the original post[^][v]#1,451,065
The website isn’t even a product. The website only exists to prove that they could generate text at 17,000 tokens per second and they’re getting bought by AMD.
Anonymous C replied with this 1 month ago, 2 minutes later, 3 hours after the original post[^][v]#1,451,066
@1,451,064 (Mr. Bloody Lemonade !0JqexkOGBc)
I don't need the website to know that a hardwired model can't be updated with a software update. That's by definition.
Now, you didn't show any understanding that it was a niche product for specific uses until I said something and you replied "I understand that".
It's inane to talk about how it can't be updated, when that's part of the design. Talking about the things it can't do, without any mention of the intent, clearly shows you dodn't understand.
Mr. Bloody Lemonade !0JqexkOGBc replied with this 1 month ago, 42 seconds later, 3 hours after the original post[^][v]#1,451,067
So basically this may or may not ever become a product. It could be that the work they did ends up becoming a part of some other product produced by AMD that isn’t a chip with a hard coded AI model but utilizes some technology that this company developed for producing a hard coded AI model on a chip. Or it could become a real product. But not every acquisition necessarily turns into a successful product.
Mr. Bloody Lemonade !0JqexkOGBc double-posted this 1 month ago, 58 seconds later, 3 hours after the original post[^][v]#1,451,070
Why are you arguing that I don’t understand something when it’s so obvious that you literally just found out that it exists from me telling you about it? That doesn’t make any sense!
Mr. Bloody Lemonade !0JqexkOGBc replied with this 1 month ago, 0 seconds later, 3 hours after the original post[^][v]#1,451,072
And I’m not even saying it’s a bad technology, I’m just pointing out that it has pros and cons. All technology has pros and cons. I think it’s really cool that’s why I posted it in the first place.
Mr. Bloody Lemonade !0JqexkOGBc double-posted this 1 month ago, 31 seconds later, 3 hours after the original post[^][v]#1,451,073
@1,451,071 (C)
You must have some mental disorder or something. Something’s wrong with you. You just argue for the sake of arguing. Don’t even talk to me.
> Of course this isn’t a new idea. Chips have done the same thing with video codes for ages. Just hardwire the program into the chip and now it’s extraordinary fast… as long as you never ever need to change it.
If you thought that wasn’t positive, explain why you think that.
Mr. Bloody Lemonade !0JqexkOGBc double-posted this 1 month ago, 4 minutes later, 3 hours after the original post[^][v]#1,451,078
I essentially said that given particular constraints (you never need to change it) this has been proven (given consumer chips have been doing this with video codecs for ages) this can achieve incredible performance.
That’s not negative, it’s only negative if you think that old ideas are bad and having software that doesn’t need to change is bad. The tradeoffs they made aren’t "bad." Nothing is objectively good or bad by itself they’re only good or bad for achieving the objective of some desired outcome. If the outcome is performance and you don’t care about modifications, then this is perfect. I just made the comment that’s not the most common use case of AI which is true.
Mr. Bloody Lemonade !0JqexkOGBc triple-posted this 1 month ago, 3 minutes later, 3 hours after the original post[^][v]#1,451,079
And the fact that the technology has cons isn’t bad. All technology does. The cons are only bad if you’re in a situation where they matter. I pointed out the biggest selling factor in the title of this thread which is that it can output 17,000 tokens per second which is insane.
Mr. Bloody Lemonade !0JqexkOGBc quadruple-posted this 1 month ago, 1 minute later, 3 hours after the original post[^][v]#1,451,080
And the fact that underneath this post I was being more critical was to demonstrate that I don’t just think this is some magic technology that will take over the world and change everything and that I understand that constraints exist even though I’m pointing out that this is really cool in one particular way.