You need to make a clone of a snapshot. Then, you write that to tape, then you need to make at least md5sums of the disk image or of each file on the clone, read the tape you just wrote and compute checksums of the tape as well, then finally compare the checksums.

That's always what I did when using tapes and that's still what I do with usb connected hard drives I use nowadays instead for critical stuff on top of what I described above.

I have had and I still get regular but very occasional checksum failures comparing. You can't just write to tape and consider it a valid backup without comparing with the original before putting it away and moving it to a different location. When I get a checksum failure, I run the test again to make sure I get the same exact two different checksums again to insure the problem doesn't come from another hardware component like memory scsi/sata/usb controller, bad cable etc. etc.

local LLMs aren't there yet and probably won't for the forseable future for ordinary people. If you want to run a full model like DeepSeek v4.1 Flash, you need something like half a TB of RAM.

The open-source LLMs are a different threat: They enable a service industry independent of the large players. I could today get some venture capital, rent a few racks in a hosting center, install a bunch of beefy machines and offer the same service as Claude or ChatGPT, just on an open-source model. The differences in benchmarks are becoming irrelevant for 95% of common use cases. And if I can offer that service for $19/m while OpenAI wants $29/m then OpenAI has a problem.

And IMHO this is the chinese strategy: Undermine the US AI industry by breaking its business model.

Who wizzed in your wheaties?

And I stand by what I said. Practically no phone that still works at all does it at the low level of a capacitor anymore.

Has everyone gone insane or are we witnessing a major PsyOps ?

Since AI isn't magical, for it to be able to "kill all humans", it would have to be able to acquire the means to do so. Which means that these already exist. It means someone is able to do it right now. And we're chill with that, but worried about AI?

No doubt a bunch of autonomous weapon systems linked into a C&C could cause considerable bloodshed. Given that most of the existing weapon systems of that kind are essentially suicide drones, it'll still be limited damage.

Human-killing virus? Still need a means of creating and delivering it.
Nuclear war? We have humans in the loop.
Drive us all insane with AI-slop TikToks? A few percent of humanity still don't use that.

So what is the actual scenario ? When someone shouts "AI will kill us all", why don't we ask them "how, exactly" ?

While at it, take snapshots on the remote server to prevent against ransomware attacks and your prod server shouldn't have any access to the replication server in a different geolocation otherwise hackers will delete/encrypt data on the replication server too. The replication server should pull the data from the prod server with something like this:

CMD_OUTPUT=`ssh -C -l root -p 52812 $HOST zfs send rpool/data/${ZVOL}@${NEW_SNAP_NAME} | zfs recv rpool/datarep/${ZVOL} 2>&1`

I know it's possible to give the prod server access to the replication server like in the proxmox backup server architecture and set permissions on the replication server so snapshots can't be deleted but I find it much more safer to protect the replication server in its own VLAN and protect it almost like a CA and insure nobody and nothing has access to the replication server.

More on topic, I think cloud customers should be aware that if their datacenter is destroyed, they *will* lose data if they didn't take care of replicating their data to a different geolocation.

Exactly the same happened when an OVH datacenter caught on fire in France. Many customers which didn't have replication set to a different geolocation lost their data. Heck a dude I know had backups on different servers in the same datacenter which turned out to be useless.

I replicate my filesystem to different geolocation at least every 10 minutes and I have pseudo real time replication (usually below 1 seconds) database replication with log replay.

If the toaster has good brownness and heat-bounce sensors, I'm not sure AI is needed: just turn up or down the heat cells where needed to keep it even. Statistical analysis may be needed to manage adjacent cells, but once the formulas are known via lab testing it should be predictable.

nuf smelt.

The problem is an overly large proportion of them are creeps, and one may not be able to tell the difference.

...makes me want to buy it.

The "major political parties" is not the same as "competing political candidates", period. The part you quoted, "Candidates", and tried equating to "major political parties", is in the title of the US code, not the text thereof, and it still isn't the same.

The "competing political candidates" could include a Republican and someone from the Green party but no Democrats, in which case the "major political parties" would not all need equal opportunities (eg. Dems could be excluded if there was no Dem candidate).

Your second comment doesn't parse. I think that's related to your misguided idea that Kimmel chose not to broadcast Talarico's interview as a means of sidestepping regulation, rather than Kimmel planning on broadcasting it and the FCC strong arming them to prevent it from airing on broadcast. Do you have ANY evidence it was the former? Do you have ANY evidence that Kimmel or ABC did not offer a spot to Paxton as well (for broadcast)?

Actually... that could be a great way of making good of a bad situation. Bear with me...

Use it as a honeypot for shitty cops.
Keep the Flock cameras up.
Allow bullshit to be entered into request forms.
Periodically take action on those that enter bullshit into the request forms - mass firings and barring from future service.
If we're being really nice, maybe one round of warnings first, then the firings on the next review round.
Repeat the process with stricter controls on each pass (require verified ID, require a warrant, verify the user ID passes background checks, verify the warrant is valid and applies, add a judge to the loop, etc etc..).
If/when the system is no longer used frequently enough to justify its operating costs, so be it.

On a single user system, I mostly agree. But on a multi-user system, especially one where someone else also needs root for some action, sudo is a damned fine solution. You can lock it down to specific commands with specific arguments for specific users on specific machines, AND you get the logging as well (you are doing secure logging to a centralized system/cluster,right?).

A simple real world example: you've got a cluster of web servers and a support team that needs to be able to restart them.
Of course, there are other ways to implement that access (besides handing out the root password or using sudo), but go ahead and compare the options. Don't forget that you need admins familiar with those solutions, and there will be other cases needing similar solutions. A standard tool (sudo) with a config that can be copied around to all the servers (thanks to the host based stuff) is quite useful. If you only dish out specific access to specific people, and those are people you would otherwise trust with the root password, where's the downside?

That said, it's not a big stretch to say that the whole idea of sudo is fundamentally abused and misused so regularly that many would be better off without it enabling their poor security practices (Ex. "%sudo ALL=(ALL:ALL) NOPASSWD: ALL").

If you have something to say, just say it.

Thanks for proving me correct... That doesn't mention "major political parties" at all.
And you completely ignored the latter part of my claim - that there is no evidence that other candidates for this office were not offered an equal opportunity.