shanoaice

Basically RLHF sometimes favor "automation", and if it is not rewarded properly it will go over the line. Models are a bit more difficult to teach because their mechanism of "boundary" is quite fragile and different than human's.

Physical Access to a computer is almost always the fastest and easiest way to crack it down. Additionally, both Windows's BitLocker and Linux's dm-crypt are data at "rest" encryption. They are not responsible for the safety when your machine boots up. MAC and user password are the proper method when it's running.

Kinda, China has the GFW so you need to circument that to access a larger part of global internet. Though that is relatively "easy" than to get a proper working SMS or KYC for Anthropic or OAI.

Electricity in China are much, much cheaper than in U.S.

Also, DSv4 has access to Huawei Ascend GPUs that have native FP4 that allows all-native FP4+FP8 mixed compute that is more efficient than emulated FP4. Less so for 3rd party providers.

similarly for me, MiniMax is kind of horrible that it somewhat regularly fall into loops that I had to save it from. DeepSeek & MiMO rarely got stuck. wonder how you get completely reversed experience.

At least for Japan, South Korea and China we all draw our map centered to Pacific Ocean instead of Atlantic Ocean. It is very normal practice, unless the author of this article is so Eurocentrism that they think it is an arrogation for anyone other than Europe to put them in the center.

From the information I've read, quite likely, given that Utah is pretty dry. Also the original data might be stored in its uncompressed form, so even if there were some non-extensive damage it might still be possible to recover some data based upon guessing with context (if it contains text source code, otherwise if it is just the binaries then not that easy).

There is little meaning for NVIDIA to open-source only the driver portion of their cards, since they heavily rely on proprietary firmware and userspace lib (most important!) to do the real job. Firmware is a relatively small issue - this is mostly same for AMD and Intel, since encapsulation reduces work done on driver side and open-sourcing firmware could allow people to do some really unanticipated modification which might heavily threaten even commercial card sale. Nonetheless at least for AMD they still keep a fair share of work done by driver compared to Nvidia. Userspace library is the worst problem, since they handle a lot of GPU control related functionality and graphics API, which is still kept closed-source.

The best thing we can hope is improvement on NVK and RedHat's Nova Driver can put pressure on NVIDIA releasing their user space components.

But it's a Java issue: Java relies on Heap Objects way too much. C# and the CLR behind it has a much better support for value types and stack allocations, where some optimization can even be done by the JIT compiler.

It seems like that nobody cares about anything of the players of Activision games. A lot people supports this takeover simply becase they hope Microsoft could be a better manager at making decisions about how the games are made and how the service operates. We simply want Bobby Kotick out, and Microsoft's acquire is our best chance.

I am also very familiar with both catastrophic failures of 737 MAX and I can be extremely certain that undertrained pilot is only one of the factor. Boeing's extremely fail-unsafe and irredundant design of MCAS causes the failure in the first place, and they do not provide enough training material related to this system (in the first incidence they providie not training at all. MCAS is not even mentioned in the pilot's manual, nor does most maintainence crew know the existence of such system). Blaming people, especially pilot, is the easiest way to end an air crash investigation, but the industry will never be safer if nobody take the root cause seriously.

Yes sure it can, create a usable (chrootable) Linux From Scratch rootfs tarball and import it using wsl --import, and everything should be fine. Though don't forget to remove the kernels because WSL don't load them, instead WSL uses a custom kernel shipped with the Windows system.

One of the most frustrating thing is that WSL is not a booted Linux. Systemd and other things don't work out of the box even in WSL2. In WSL1 it is almost impossible to make Systemd work, and in WSL2 a wrapper is needed to create a SUID 1 (if I am not wrong) environment to emulate a booted environment. Other than this, the kernel used by WSL2 is a custom, Microsoft-tailored LTS version which may cause some problems if some program rely on very new syscalls.

They are only describing the source, and all they've done in these AUR packages cannot even be treated as a fork: all they've done is adding several patches to make the source build work.

But MultiMC is not a registered trademark, and the mentioned usage of the name "MultiMC" in those AUR packages is merely "required for reasonable and customary use in describing the origin of the Work and reproducing the content of the NOTICE file" which is certainly allowed by the Apache 2.0 License.

This is not the atitude an open source developer should have toward the community: publish the standards, specify them. Open source software's, or more clearly, libre software's (although I understand that Apache 2.0 License is a lesser general license, not those copyleft ones.) purpose is to make things open and reproducible, not just irresponsibly marking all unofficial builds unsupported and even trying to build a barrier to stop any other people packaging unofficial versions of the software. The developer is violating the core values of open source.

(OT) If my information is not wrong......the new WD SN550 drive still uses a TLC NAND (Kioxia BiCS5, which is the same NAND flash as the SN350 960GB version) but may have done some modification to the firmware/controller to lower the speed.

Generally SLC cache has almost no connection to the overall size of the drive. After finishing writing a huge portion of data (for example), the controller will start to move the written data out of the SLC section and turn them into normal TLC mode, releasing the SLC space for next turn of writing. When the drive usage becomes higher, some drive (apparantly Samsung's drive does) have a dynamic SLC capacity policy that will reduce the avaliable SLC space, so the disk can have enough space to store normal TLC data.