Interesting how the article mentions the previous low volume in October 2020... that was right before Bitcoin climbed from 12000 to 60000.

And it could have worked just as well with just the top part for AC and DC. Tesla proposed this and implemented it in Model S, it worked great and everybody could adopt it, but nooooo... not invented here, too convenient. So Tesla had to outfit Model 3 with those horrendous double plugs.

The european plug would not be so bad if they would just use the AC pins for DC as well (like the old european Model S plugs did). But for some reason they decided to go with this ridiculously large double plug with separate parts for AC and DC, it's a total abomination.

I bet that DMCA complaint did wonders for the SDAROT website visitor count, though.

I always enjoy negative comments about new Apple products. Especially when I read them again a few years later. They age so well...

I've looked at optimized code, and while some optimizations indeed occur like you pointed out, there's still an awful lot of low hanging fruit to be optimized. I once stepped through a fused-multiply-add library call, which should have just been one instruction, but instead it was a function calling a function calling a function looking up the address of a function and then calling that, each of them saving a few registers on the stack. It was like 50 instructions instead of one. (On MacOS)

And I've seen plenty of other instances where I could rewrite a loop to be 50% faster just by avoiding some simple recalculations and things like that.

They shouldn't call it "new algorithms" though. It's micro-optimization which is definitely a good thing but doesn't actually change how the algorithm works (if I understood correctly).

I do believe this has a lot of potential: high level languages are designed with human programmers in mind, preventing them from making mistakes but also adding way too much code in the process. Manually optimizing that bloated code for a whole app, one instruction at a time, is basically impossible for humans (without making mistakes) but would be easy for AI. It can turn nicely structured but overly complicated functions back into good old spaghetti code that runs faster while producing the same results.

The point is that they are optimizing the output from a C compiler. Contrary to the article, they are not developing new "algorithms" but optimizing low level code.

I wonder how exactly then can know that the crystal is in a superposition state. It has to be isolated from the environment, otherwise the wave function collapses. So you can't exactly watch it, because as soon as you try to determine the vibration state it picks one of the two. How can they tell that it's in superposition, then? Do they shine a laser on it that can reflect in two different ways depending on the vibration of the crystal and let those two paths interfere with each other like in the double slit experiment? Something like that?

They started with only 3 out, but then more failed during the flight. But why abort if the area is clear anyway? They just continued to gather data, went past max Q (a pretty important milestone), and even allowed the rocket to tumble multiple times before finally blowing it up. That taught them a lot more than blowing it up at the first anomaly.

I couldn't help imagining Elon frantically pressing A/S/D/W to try and point the rocket in the right direction again and make it to orbit somehow. It really looked like one of my Kerbal launches and I was amazed they didn't blow it up sooner like what they usually do when a rocket even starts to deviate a little bit. They just let that thing tumble and continued to gather data, which is great.

GPT4 is already being connected to the internet. It's not smart enough yet, but that kind of connectivity could one day be enough for it to exploit zero days.

Right now I'm not worried yet. And halting progress is pretty useless right now because those rules will be disregarded by precisely by the wrong countries and we certainly don't want them to pull ahead. We're basically stuck on a very dangerous path with no way of getting off. What we should do is think FAR ahead, not just the next few years, so we can be ready when we need to be.

Personally, what I'm most afraid of is not AI taking over the world by force, but AI becoming so good that it will take over de facto because they will become so much more capable than us that it won't make sense anymore to keep humans in charge of anything. Then what will we do? Sit back and enjoy life as AI pets, hoping that they will continue to care for us?

Yes, but they are still running old classic computers. And Opportunity lasted for 14 years on Mars until a monster dust storm covered its solar panels. That's quite amazing for such a rudimentary machine, and I'm sure we can make them last a whole lot longer with future technology.

Neural nets are already beating humans at all sorts of different tasks. First one task at a time (AlphaGo,...), then multiple tasks at once (GPT answering questions about all sorts of totally different subjects, not without errors but probably with better knowledge than the average American), soon they will be unbeatable at pretty much anything. Probes with that kind of intelligence on board won't need constant control from earth anymore, just an occasional "we'd like you to try this or that". When it's smarter than an astronaut, it will be just like having an actual astronaut there. Not tomorrow, but it will happen.

A pattern matcher capable of not only passing the bar exam but actually surprising human professors with its thorough analysis of problems their students normally struggle to understand. Problems specifically written to weed out those who studied the course by heart but didn't thoroughly understand it.

Yeah, no creative thoughts, no learning.

And Midjourney images often look a lot like dreams, with very similar inconsistencies.

You really are underestimating where this is going. Like watching an Atari 2600 and claiming computers will never be able to do 3D shooters.