The pitch is "just change one line and it works". It is not "just change one line and you will get peak performance on the TPU".
From the text of the blog post:
"Portability doesn't eliminate hardware realities, so TorchTPU facilitates a tiered workflow: establish correct execution first, then use our upcoming deep-dive guidelines to identify and refactor suboptimal architectures, or to inject custom kernels, for optimal hardware utilization."
Once you download it does anything actually work beyond viewing? Can you search? Can you ask for directions? I do download maps, and pay for roaming data, but I still would never completely rely on it in a new place because I'm bound to be out of network coverage at some point.
Yes, which means downsizing. Need to downsize will put people off selling. People being put off selling reduces supply. Small supply keeps prices high.
In the SF bay area people are throwing $4, 5, 6 million at houses in cash all over the place. No contingencies, 7 day closing. The high end of the market has slowed a lot, though, and houses are having to repeatedly lower prices to get bids. The market up to about $4M has not slowed noticeably, possibly because supply has slowed more quickly than demand; people are still bidding hundreds of thousands over asking.
The problem with that argument is that in places that are setup for it, it clearly works because people do it. So we have to ask whether there is something unusual about the people in those places, or the infrastructure. It apparently isn't something wrong with being young or old, within obvious limits.
The taxes are skewed towards the rich, but also towards people who want to move for work. It's a wealth transfer towards the asset rich and immobile from the workers and the mobile. It's the economic distortion that's the problem as much as the lower taxation, but certainly part of it is shifting the taxes even more heavily onto the high income earners.
It is bizarre. Here in California you didn't used you buy a car with plates at all. Then they started to require paper temporary ones.
Why they don't just have plates fitted on the lot I don't know. I can't see how it saves any time, the dealer would just have to have the plate ready for each car like in most of the world. It seems that the simplest solution was too hard to implement.
Does Singapore make more sense? Papua New Guinea? Alaska, for that matter? All borders seem pretty arbitrary when you get down to it, but it's hardly alone in being a weird bit at the end of a bigger bit where one or other is part of a detached bit.
That's still a hidden charge that makes the menu pricing mildly dishonest. Adding X% to all the prices is clearly the right solution. The problem is that all restaurants need to do it to make it work.
To be fair, here in Menlo Park, CA, children of this age bike or walk half a mile to school all the time as well. It's not very different from when I lived in the London suburbs and walked myself 30 years ago. This isn't really a US/Europe split.
Data has to be stored somewhere. You can end-to-end encrypt it, but that has constraints and adds clunkiness for the user so it's not even obviously a feature. The easy access to the web interface for messenger over whatsapp or signal if a feature of a sort, depending on your point of view. If you don't want a company to have access to the data you store with them you have to keep the encryption key yourself. If you do that, you risk losing access to the data you store with the company. That's a big downside.
So yes, it's a product choice, but it's not a product choice out of some strange effort for the company to have access to your data, but because it provides the right product mix.
You might as well argue that your bank having access to your account records to hand over is a product choice. Yes, there are ways to work around that. Maybe some blockchain variant will give that level of privacy. It will not do it without significant tradeoffs.
We have a Roomba, though I think it's back in its box in a cupboard now. In the end it needed so much babysitting that I gave up and bought a cordless stick vacuum. That is the best home appliance purchase I've ever made. The lower hassle of not needing to extend and roll up the cord compared with my older vaccuum massively outweighs any benefit from the Roomba. Grabbing the stick, quickly buzzing it around for a couple of minutes and putting it away again is far less hassle than babysitting the Roomba.
Clearly one can write web services in C++. Meta does not really, though. It is Hack code that talks to most web clients (and python for Instagram). The C++ code sits behind that. All of the databases, the caches, the indexing services, the ads ranking and selection are written in C++ and talk Thrift to each other and to the Hack code.
Yes, because getting to the right answer is not the point of the interview. Apart from anything else, getting to the right answer may mean you memorised it and are incapable of doing anything else. Always show your working and thought process. Asking questions and showing that you understand tradeoffs and that users have different requirements is a good way to do that.
I'm in favour of cyclists occupying a lane, especially on narrow roads to take full control and I do that myself often. What I dislike is when they have their own lane and then ride 2 or 3 abreast so they don't fit and stray a foot or two into the car lane as well. It's the worst of both worlds. Even worse when they drift into the car lane without looking back to overtake other cyclists, which seems to happen regularly around here.
I did just that to a cyclist who role through a stop sign, causing a car that I had stopped for to have to stop mid turn to avoid the cyclist. Sure, stopping completely randomly isn't necessary, but a lot of cyclists take the cycle-through-stop thing far too far and assume that they also have priority, irrespective of the state of other rode users.
From the text of the blog post: "Portability doesn't eliminate hardware realities, so TorchTPU facilitates a tiered workflow: establish correct execution first, then use our upcoming deep-dive guidelines to identify and refactor suboptimal architectures, or to inject custom kernels, for optimal hardware utilization."