Veedrac

53 ms·

601.
by Veedrac 6y ago
GPT-3 also works differently to a Markov chain.
602.
by Veedrac 6y ago
> It's a little bit like some sort of Chinese room, or asking a non-developer to answer you programming questions by looking like something that vaguely resembles your prompt and then picking the most upvoted answer on stackoverflow
603.
by Veedrac 6y ago
> regurgitating approximate conditional probabilities of words given all the text it has ever seen and the prompt This is meaningless; you have only described the task. It is equally applicable to a superintelligence as it is of a Markov
604.
by Veedrac 6y ago
The article is meaninglessly cherry-picked, showing six bad answers out of 157, except those 157 examples were themselves cherry-picked to be bad out of a larger set. As usual, Gary Marcus is absurdly biased. For example, out of the larger
605.
by Veedrac 6y ago
NVIDIA uses tiled rasterization, but not TBDR. I believe the same is true of AMD.
606.
by Veedrac 6y ago
Did you sign up for the Dragon model?
607.
by Veedrac 6y ago
The point of a sandbox is that the only things in it are the things you've put in it. A can only touch B through mechanisms interfaces it has been gifted. Computer architecture is too broken to provide this guarantee to anything in the
608.
by Veedrac 6y ago
It's a pretty big architectural difference, and the effect travels all the way through the API. Apple had a few talks on this in their recent WWDC. https://developer.apple.com/videos/play/wwdc2020/10602
609.
by Veedrac 6y ago
Apple's going to be entering the mid to high range soon, and they use Tile-Based Deferred Rendering. Very different indeed.
610.
by Veedrac 6y ago
Apple's AI accelerator is meaningfully better than the Snapdragon 855's, but way behind the 865's. > I would agree with that characterization of Google. No. It's fine to dislike Google's stuff, but words have mea
611.
by Veedrac 6y ago
Apple's AI work is decent for what it is, and their AI hardware is fine too, but so much of this article is just weird. It starts by talking about how until recently Apple wasn't doing AI work where it needed to be. Then there&#x2
612.
by Veedrac 6y ago
> It is a bit odd that there is no mention of the awful security problems like Spectre and Meltdown which mostly hit Intel and was very poorly handled by them. I recall him mentioning exactly this.
613.
by Veedrac 6y ago
You are massively underestimating the performance of Apple's GPUs. They can beat the 5600M. https://www.reddit.com/r/hardware/comments/hgahzx/apple_gpu_...
614.
by Veedrac 6y ago
To the skeptics, Apple Silicon is about thin laptops and fat margins. It makes no sense for Apple to reinforce this. Apple's silicon team is exceptional, and they should prove it.
615.
by Veedrac 6y ago
The Sabrent Rocket 1TB, a reasonable, fast $130 SSD, has an endurance rating of 1665 TBW. To have hit that rating, you'd have to have written 760 GB to it every day over those six years. At 5%, that's still 38 GB a day, every da
616.
by Veedrac 6y ago
I don't think it'll be close to that long. Rumours are that they'll release an 8+4 core this year, which is enough for everything bar the top-end iMac Pro and Mac Pros, and I'm sure they can get out a 16+N core by 2021 o
617.
by Veedrac 6y ago
The performance of their chips is good enough that it wouldn't make much sense not to update everything once silicon availability is assured. It would be ridiculous to keep their iMac slower than their Macbook Pro.
618.
by Veedrac 6y ago
Doing a more sophisticated simulation, I think it's trending towards 1.4 exactly (42 to 43 is 1.400003). Idk how to prove it though.
619.
by Veedrac 6y ago
I prefer ‘silicon’ to ‘CPU’ because it makes it clearer Apple intends to compete on the GPU front too.
620.
by Veedrac 6y ago
I think you meant > you can tell he's not like a lot of other crony type Trump appointees
621.
by Veedrac 6y ago
Someone should collate all the instances of people saying stuff written by GPT-3 couldn't possibly be written by GPT-3, or that they couldn't possibly be representative output, or that they must be some form of barely-cognizant pl
622.
by Veedrac 6y ago
> A: (It's a fish, and also a brown fish with a tail and no legs.) It's a fish. It actually got a valid answer!
623.
by Veedrac 6y ago
It's more than instruction decoding, though admittedly that's a big part. There's also memory ordering (tho it's debatable whether that's problematic), the number of logical registers (particularly important for JIT
624.
by Veedrac 6y ago
AMD's previous design was its own thing. Arm's and Apple's designs are their own things. There's still flexibility in how to build a core. Skylake is clearly not even the optimal target; Apple's approach easily outc
625.
by Veedrac 6y ago
> zen 2 has better IPC Ice Lake has better IPC. The only issue with Ice Lake is that it's on a broken 10nm process, which primarily means that clocks are terrible. > Source for this? https://en.wikichip.org/wiki&#
626.
by Veedrac 6y ago
AMD came back from Bulldozer.
627.
by Veedrac 6y ago
2nm then 1.4nm. Not sure what's after that.
628.
by Veedrac 6y ago
> But Intel has rarely been the best in microarchitecture. When is the last time Intel was behind in microarchitecture? Ice Lake beats Zen 2, then you follow all the way back through Bulldozer, where AMD was utterly outclassed, then what
629.
by Veedrac 6y ago
It's not just that you're not arguing it isn't useful; as far as I can tell, neither of your comments contain an argument against ML at all . I have nothing to meaningfully argue against. ML is undergoing a Cambrian explosio
630.
by Veedrac 6y ago
GPT-3 can generate well-formed programs, so yes, it does things well beyond this complexity. > After 500,000 iterations I'm still getting invalid sequences. How frequently? If it's only the occasional issue it might be down to

More ›