vcmj

Not an answer to the question, but in case performance is the goal, Torchaudio has it here

Yes, forgot the exact details apologies

You can change those to /dev/disk/by-uuid/XYZ ("ls -an" that directory to see the symlinks to your current drives)

Basically just look for things like root=/dev/sda2 in the kernel command line. You can get it at runtime by running "cat /proc/cmdline" having /dev/sda etc in your fstab might also be a problem

Yes if you have multiple drives some buggy BIOS may not enumerate them in the same order every time. Most modern distros do UUIDs by default but when manually setting up a bootloader it is easy to succumb to such temptations to use the much simpler device paths as the UUIDs are a pain. If you're not sure how to change the kernel parameters most likely you're good on that front actually, its in your grub config as others have mentioned. I'll leave this comment around in case some poor soul who did it manually comes across the thread.

Depending on if you wrote the kernel cmdline yourself I imagine this might happen using /dev/sdN style device paths? BIOS might change things up every now and then for fun, so using partition UUIDs would be a better way if so.

Yes. Samsung allows 2 managed profiles and the owner profile. There is the hidden folder, and additionally the work profile which you can activate with something like Shelter. So you can in fact install 3 instances of Twitter

Maybe use a flexible struct? You can have an indeterminate array at the end (https://www.geeksforgeeks.org/flexible-array-members-structure-c/) I'd rather just use the variable in the struct though. The packing isn't guaranteed to be right next to each other

Ah, even then it could just be a consequence of training samples usually being chronological(most often the expected resolution for conflicting instructions is "whatever you heard last", with some exceptions when explicitly stated) so it learns to think that way. I did find the pattern also applies to GPT trained on long articles where you'd expect it not to, so wanted to just explain why that might be.

Or I should explain better: most training samples will be cut off at the top, so the network sort of learns to ignore it a bit.

Yes, that's by design, the networks work on transcripts per input, it does genuinely get cut off eventually, usually it purges an entire older line when the tokens exceed a limit.

I was a curious child, and things spiralled out of control from there...