devit

At least modern Google Pixel phones have an IOMMU (https://grapheneos.org/faq#baseband-isolation)

Yes, but it's without prompt only for apps that they themselves installed, while Google Play can bypass that.

It runs as a privileged app.

It uses that to install new applications and update applications installed by other app stores without the OS-level confirmation, probably can't be uninstalled on most Android OSes and perhaps some other things.

You can also add a property to all objects with an array of keys in the order you want (if you can guarantee it won't conflict with existing properties, or can escape those), or turn them into a {k, o} array where k is the key array and o the object or if you don't care about looking up keys then either a {k, v} array where v is a value array or an object where keys are prefixed with their position or putting the position in the value as an array of value and position.

That's what game consoles usually attempt to be.

Presumably they have captured the Mozilla foundation board?

I think reading the encryption key from memory is usually easy and cannot really be detected in any way if done from the kernel or hypervisor by parsing page tables and accessing physical memory (except by looking for known code if the anti-cheat has as much privileges as the extractor, which the hypervisor approach avoids).

No, I meant gRow lamps that are marketed to help to grow plants indoor, and thus presumably have a light spectrum that is natural (although they also include UV light, so it's not clear if they are safe).

What are the pros and cons of those lamps vs those sold as "grow lamps" and those sold as "UFO" high bay lights?

Is there any >=100W (actual, not equivalent) sunlight-like bulb available that is dimmable and programmable via ZigBee or other means?

In my first attempt at the problem it made this mistake:

"The condition fails for i=1 since 1 does not divide p+q unless p+q is a multiple of n, which is not generally true" (1 divides everything)

In the second it gives up:

"However, this conclusion is based on heuristic reasoning and examples"

Third attempt it makes this mistake:

"If the immediate next divisor, di+1 , is not a multiple of p (for instance, it could be q or a product involving q), then p does not divide di+1. Hence, p will not divide the sum di+1+di+2 in such a case, violating the condition." (the fact that p does not divide d_i+1 does not imply that it does not divide d_i+1, d_i+2)

Then it gives up again: "However, this conclusion is based on heuristic reasoning and examples"

Then it makes this basic logic mistake: "However, since p and q are distinct primes, p does not divide q, and it's not guaranteed that p divides q+d_i+2 , especially if d_i+2 is not a multiple of p. Therefore, for such n, the condition fails." (the fact that "it's not guaranteed" doesn't imply that it's false)

Overall it proves that prime powers have the property, conjures that non-prime-powers don't, proves that p*q doesn't have it, but completely fails at coming up with a proof strategy that proves that non-prime-powers don't have the property.

It's problem solving (i.e. finding a proof of a statement or finding one or more mathematical objects that have some given properties and proving they have them and possibly that no other objects do), not computation.

That's obviously the only correct approach.

The concept of "window position on screen" makes no sense, since a compositor is allowed to display windows arbitrarily, and doesn't necessarily display them as 2D rectangles (for instance, they could be displayed in a VR environment at arbitrary 3D positions, or they could be warped arbitrarily, etc.).

Also, the window layout belongs to the user and compositor and it would be a security hole to let applications affect it.

If you want to position subwindows, you need to draw the subwindows yourself inside a main window like e.g. VSCode does with its various panes.

That is easily solved by using Debian instead of Ubuntu.

Well, the biggest advantages of Linux are that the OS vendor is not itself malicious (in particular, does not collect telemetry, push advertisements or attempt to restrict your use of the system like Microsoft and Apple do), the OS is relatively unpopular and thus not a profitable target for malware directed to end-users, and it is flexible and customizable.

You need to use VM-based isolation to have good security with Linux beyond that (i.e. use Qubes or a similar alternative).

Maybe Facebook's Segment Anything could replace the first clustering step?

I seem to remember that Stockfish is known to be bad at handicap games and in general at maximizing odds of winning against humans from strongly losing positions.

Based on the first figure in the paper, it seems that this scheme effectively turns 8 input values into a 4-bit number, thus giving an effective 0.5-bit quantization.

Considering that current aggressive quantization for LLM transformers uses 4 bits, does such a 0.5-bit quantization produce an effective neural network?

Does the scheme stay competitive if it is changed to use 4-bit quantization instead of 0.5-bit?

Seems to be a language equivalent to a subset of strict Haskell.

There doesn't seem to be any particularly compelling reason to use it, and note that, unless they can use libraries from an existing language, there needs to be a really MAJOR reason to use a new language to compensate the lack of libraries.

That seems like a problem.

But the solution seems to be to have the desktop client request data from the phone.

In fact I'm not sure how it can possibly work otherwise (what if someone just uses their phone for years and then opens the desktop client for the first time ever? does that not show any old messages? seems a terrible design).

Hmm, no?

Photos are generally <1MB in size and I think have a single photo sent but not received on average per user seems reasonable (most users probably almost never use Signal, and of those that do probably most only use text, and those that use photos probably most don't send more than one or a few per day).

Videos are probably relatively rare and if not maybe they should do something about them, like not storing overly large ones them on servers and requiring both phones to be online to transfer.

There's a 500x margin between the estimate and their costs anyway.

I assume they only need to store it between the time it is sent and the time it is received by the recipient.

Maybe the problem is that the Signal app doesn't eagerly download messages upon notification? They should start doing that given the money issues.

These costs seem absurd.

For instance, 1.3$ million per year for storage??? Apparently, they have 40 million users, so 1 MB per user (seems reasonable for Signal) means 40TB. You can buy a 4TB SSD for $200, which means you need $2000 one-time for 1MB per user.

How they get from $2000 to 1.3$ million is a mystery.

As for SMS registration, if they are spending 6 million, maybe they should find some way of doing it for free, e.g. Google might be offering it with Firebase, Twitter used to have it, etc. It's not great for privacy, but if they care about that they should just stop using phone numbers.

Routing video calls through a server to obscure IP address seems totally pointless while you are revealing the phone number anyway. And again there might be a way to do this for free, e.g. perhaps using one of free WebRTC STUN/TURN servers that e.g. Google seems to run.

As for bandwidth, a very conservative estimate seems 100 MB per month for each of 40 million users, giving 4 PB per month (though I guess the real usage is 1/10 that at most). Hetzner charges $1/TB, so that gives $4000 per month or $40k per year, overestimated.

Again a mystery how they get from $40k per month to $2.7 million.

Maybe the problem is that they use AWS/GCP/Azure/etc.? They have to be real idiots to use them since everyone knows they are insanely overpriced and should never be used unless a large corporation or deep-pocketed investors are footing the bills or they is no other possible solution.

Perhaps they need to consider stopping dumping money down the drain before asking for donations.

Seems like it would be much better to do conventional weather forecasting and then feed the predictions along with input data and other relevant information to a machine learning system.

Properly designed hardware with signed firmware versioning has the bootloader permanently disable the ability to blow efuses until reboot before passing control to the OS, either through hardware or a mandatory hypervisor (and the bootloader itself is trusted and cannot be replaced without either a signature or special physical hardware access).

I think it's more fun to program with LLMs, because rather than having just a single programming phase, you basically have three quite different phases: a first phase where you craft an initial prompt, then a second phase where you review the LLM code and ask it to make changes, and then a third phase where you change the code yourself to exactly what you want it to be.

Current CPU vector instruction sets are as good as the ones on GPUs and they also have tensor/matrix instructions comparable to the ones on GPUs.

However, usually CPUs don't have enough cores and execution units to compete with GPUs, because they spend the transistors on out-of-order execution, cache and other features instead.

Intel Larrabee was the only attempt to actually make a CPU with a GPU-sized vector unit and they dropped it.

Seems really terrible engineering to design a laptop where malware can destroy the speakers.

I guess there's reason to wonder whether malware might also be able to set the whole machine on fire.

It seems to me the problem with that font is that the strokes are way too thick, making the "m" look like it's almost all ink, and it also seems that the lowercase characters are far too tall for their width.

If you use a normal stroke width (one that makes the empty space in the "m" as wide as the stroke) and width/height ratio for all characters, the problem naturally disappears, and in fact this is not a problem for instance in Linux terminals with the default fonts (normally DejaVu Sans Mono), where the characters look perfectly natural and fine.

A much simpler and compatible solution would be to assign a single IPv4 address per AS and then use IP options to pass additional address bits to identify the specific host within the AS.

16-bit bytes?