While optional, it is also the default behavior.

It’s a bit directly on Microsoft, unless you go out of your way, bitlocker will upload the keys to Microsoft. They assume you want them to help recover your data if your tpm becomes unavailable.

Interesting fun fact, when I tried to swype type bitlocker it really wanted to put bootlicker instead.

Blasphemy, TempleOS requires 512MB, so everyone has to at least have that.

Well the good news about 365 suite on the web is they made it even worse… wait…

Yeah, the croplands came up in a discussion here…

A farm was shutting down because a datacenter operator bought the land, a fully functioning farm. It was more profitable to sell the land than keep it viable for food production…

Now the chances of that land ever being appropriate for farming again…

Now you have me imagining the volume of investment currently thrown at LLM datacenters instead being thrown at solar and energy storage and I’m even more disappointed. Areas that seem to have some legs where we haven’t pushed the physics quite as hard as we have computing yet.

In hopes of making you feel better, the cache amount consumed hardly matters. It’s evictable. So if you read a gigabyte in once that you’ll never ever need again, it’ll probably just float in cache because, well, why not? It’s not like an application needs it right now.

If you really want to feel better about your reported memory usage, sync; echo 3 > /proc/sys/vm/drop_caches. You’ll slow things down a bit as it rereads the stuff it actually needs to reuse, but particularly if your system has a lot of I/O at bootup that never happens again, a single pass can make the accounting look better.

You could at least do it once to see how much cache can be dropped so you can feel good about the actual amount of memory if an application really needs it.

Though the memory usage of VMs gets tricky, especially double-caching, since inside the VM it is evictable, but the host has no idea that it is evictable, so memory pressure won’t reclaim stuff in a guest or peer VM.

My guess is that they have email hooks into LLM and call each entry point into LLM invocation an ‘agent’ and I have seen in a lot of companies the easiest way for them to have an email is to just add them to the directory.

It’s still dumb as hell, but I am no stranger to non-human’s in an ‘employee directory’, though usually it is supremely obvious that it is a non-person so if it’s at all confusing it means they are being ‘cute’ about their accounts.

because they cornered and manipulated the hardware market

Makes me think of the ‘First Time?’ meme but with the housing market wearing the noose.

Yeah, very good analogy actually…

I remember back in the day people putting stuff like ‘Microsoft Word’ under ‘skills’. Instead of thinking ‘oh good, they will be able to use Word competently’, the impression was ‘my god, they think Word is a skill worth bragging about, I’m inclined to believe they have no useful skills’.

‘Excel skills’ on a resume is just so vague, people put it down when they just figured out they can click and put things into a table, some people will be able to quickly roll some complicated formula, which is at least more of a skill (I’d rather program a normal way than try to wrangle some of the abominations I’ve seen in excel sheets).

Using an LLM is not a skill with a significant acquisition cost. To the extent that it does or does not work, it doesn’t really need learning. If anything people who overthink the ‘skill’ of writing a prompt just end up with stupid superstitions that don’t work, and when they first find out that it doesn’t work, they just grow new prompt superstitions to add to it to ‘fix’ the problem.

Unless Nvidia gets in on remanufacturing, those GPUs are never going to repurposed for residential usefulness. B300 was designed from the onset for datacenter ai use exclusively, with no concessions for theoretical video out integrated to a board that overall demands over 15 kw. You couldn’t even power it with a 60A 220V circuit.

Some of the storage could get more consumer support, SAS is unusual but if there were a glut then various solutions could emerge. Similarly EDSFF cages aren’t really a hot consumer item, especially not e1.l, but I could imagine a glut driving home friendly adaptions.

DRAM modules are somewhere in between, though practically speaking they won’t be workable outside of their initial application.

There was a time when home and datacenter got closer together, but there’s been quite the divergence the last few years.

Yes, because Microsoft’s revenue growth is in fact the most important thing the folks at Davos had to think about…

Getting a dns name is straightforward enough, and let’s encrypt to get a tla cert…

But for purely internal services that you didn’t otherwise want to publish extremely, the complexity goes way up (either maintain a bunch of domain names externally to renew certificates and use a private DNS to point them to the real place locally, or make your own CA and make all the client devices enroll it. Of course I’m less concerned about passkeys internally.

They needed to find a successor with a more appropriate surname, like Ganon.

You don’t have to target every distribution, target a vaguely credible glibc, and of course the kernel, and you are covered.

As a distribution platform themself, they don’t have to sweat packaging N different ways, they package the way they want. Bundle all the libraries (which is not different then the way they do it in Windows, the bundle so many libraries).

They don’t get the advantage of the platform libraries and packaging, but that is how they treat Windows already because the library situation in Windows is actually really messy, despite being ostensibly a more monolithic ecosystem.

That you can’t plug into a traditional computer and that has not even pads for a video connector to be soldered to.

Folks just don’t realize how exotically different they have ultimately made the GPU packaging for datacenters. B200/B300 come in very specific packaging that is nowhere near a PCIe card.

Nuke it from orbit, it’s the only way to be sure

Broadly speaking, the private keys can be protected.

For ssh, ssh-agent can retain the viable form for convenience while leaving the ssh key passphrase encrypted on disk. Beyond that your entire filesystem should be further encrypted for further offline protection.

Passkeys as used in webauthn are generally very specifically protected in accordance with the browser restrictions. For example, secured in a tpm protected storage, and authenticated by pin or biometric.

For ssh, ssh keys.

For https, webauthn is the way to do it, though services are relatively rare, particularly for self hosting, partly because browsers are very picky about using a domain name with valid cert, so browsers won’t allow them by ip or if you click through a self signed cert

I could see a refusal to use codegen as a potential liability, but that’s not “skills”. The biggest thing about codegen is you have to review it and just lower your expectations that the code comes from a technique dumber than the dumbest human intern you have ever seen and approach it with supremely thorough skepticism. It’s exhausting how dumb it can be and how you have to be paranoid for every single piece of output. But it’s not a “skill”