Changing the metadata profile to at least raid1 (raid1, raid1c3, raid1c4) is a good idea, especially for anyone, against recommendations, using raid5 or raid6 for a btrfs array (raid1c3 is more appropriate for raid6). That would make it very difficult for metadata to get corrupted, which is the lion's share of the higher-impact problems with raid5/6 btrfs.
I still got corrupted metadata with metadata raid1c3 on btrfs on a power loss. I never had this happen with ext4 alone or atop Linux raid.
I want to be clear that losing (meta)data in flight during a power loss is expected. But a broken filesystem after that is definitely not acceptable.
Some postgresql db endedup soft corrupted. Postgresql could not replay its log because btrfs threw IO errors on fsync. That's just plain not acceptable.
This should be at the top, using metadata DUP on a 3 disk volume is already asking for it, and of course you loose data when you just use it as jbod with data stored only once. Unless this are enterprise disks with capacitors anything can happen when it suddenly looses power. Not the FSes fault.
With the same configuration this can happen with ZFS, bcachefs etc just as well.
Will it render the whole filesystem inaccessible and unrepairable on those filesystems as well? One of the issues with btrfs is that it's brittle: failure tends not to cause an inconsistency in the affected part of the filesystem but bring down the whole thing. In general people are a lot more understanding of a power failure resulting in data corruption around the files that are actively being written at the time (there are limits to how much consistency can be achieved here anyway), much less so when the blast radius expands a lot further.
A few decades ago, XFS was notorious because a power failure would wipe out various files, even if they had been opened only for reading. For instance, I had seen many systems that were bricked because XFS wiped out /etc/fstab after a power failure.
Nevertheless, many, many years ago, the XFS problems have been removed and today it is very robust.
During the last few years, I have seen a great number of power failures on some computers without a UPS, where XFS was used intensively at the moment of the power failure. Despite that, in none of those cases there was any filesystem corruption whatsoever, but the worst that has ever happened was the loss of the last writes performed immediately before the power failure.
This is the behavior that is expected from any file system that claims to be journaled, even in the past many journaled file systems failed to keep their promises, e.g. a few decades ago I had seen corrupted file systems on all existing Linux file systems and also on NTFS. At that time only the FreeBSD UFS with "soft updates" was completely unaffected by any kind of power failures.
However, nowadays I would expect all these file systems to be much more mature and to have fixed any bugs long ago.
BTRFS appears to be the exception, as the stories about corruption events do not seem to diminish in time.
> Unless this are enterprise disks with capacitors anything can happen when it suddenly looses power. Not the FSes fault.
Most filesystems just get a few files/directories damaged though. ZFS is famous for handling totally crazy things like broken hardware which damages data in-transit. ext4 has no checksum, but at least fsck will drop things into lost+found directory.
The "making all data inaccessible" part is pretty unique to btrfs, and lets not pretend nothing can be done about this.
Are they bemoaning that science is being done, or are they bemoaning that the experimental results have not yet reached high enough confidence to justify the conclusions being suggested?
As someone who has made several comments consisting entirely of “…in mice.”, let me assure you that the reflex only kicks in after reading the paper and noticing that the experimental subjects were exclusively mice.
The problem is not mice experiments on arxiv, the problem is posting those papers for broader dissemenation to the public, with titles suggesting to the public that cancer has been cured, without prominently pointing out that the experiments were not about cancer in humans.
> problem is posting those papers for broader dissemenation to the public, with titles suggesting to the public that cancer has been cured
Fair enough. I'm thinking of cases where a good study that isn't turned into PR slop is dismissed because it was done in mice. Which is fine for most people. But not great if we're treating real science that way.
Dismissing good science is entirely the correct decision when the good science isn't ready for broad dissemination to the audience which it is being presented to.
I disagree. I think people understand studies have to begin in mice. It’s what the GP said. You can’t release those studies because there’s not a high enough confidence rate in what most people are interested in ie how it effect humans.
> You can’t release those studies because there’s not a high enough confidence rate in what most people are interested in ie how it effect humans
This is science by ignoramus. It isn't how science works, at least not when it works at its best. Someone advocating for censoring science because it might be misread is not on the side of science.
I’m not advocating for censoring them. I’m advocating for less hype in science media reporting around mice studies because let’s be frank. The vast majority of the population are ignoramuses that cannot make the distinctions themselves, and that has real political consequences through lack of trust in scientific organizations.
Where are you seeing people being told that AI is infallible? AI is being hyped to the moon, but "infallible" is not one of the claims.
To the extent people trust AI to be infallible, it's just laziness and rapport (AI is rarely if ever rude without prompting, nor does it criticize extensive question-asking as many humans would, it's the quintessential enabler[1]) that causes people to assume that because it's useful and helpful for so many things, it'll be right about everything.
The models all have disclaimers that state the inverse. People just gradually lose sight of that.
[1] This might be the nature of LLMs, or it might be by design, similar to social media slop driving engagement. It's in AI companies' interest to have people buying subscriptions to talk with AIs more. If AI goes meta and critiques the user (except in more serious cases like harm to self or others, or specific kinds of cultural wrongthink), that's bad for business.
> To the extent people trust AI to be infallible, it's just laziness and rapport (…) that causes people to assume that because it's useful and helpful for so many things, it'll be right about everything.
Why it happens is secondary to the fact that it does.
> The models all have disclaimers that state the inverse. People just gradually lose sight of that.
Those disclaimers are barely effective (if at all), and everyone knows that. Including the ones putting them there.
- serif was claimed to lead to better horizontal tracking... so better for long prose readability
- sans serif was claimed to lead to better spot-recognition of characters... so better for spot-character/word recognition and legibility
Those effects were never very strong, and varied depending on the exact fonts in use (and for digital, font rendering characteristics).
There's also probably an effect based on what you're used to. If most of the books you read are serif (which they would be for older people, since almost all printed books were serif), and your exposure to sans serif was largely via the internet, and you don't like most of what's written on the internet, that might sway you toward serif. Conversely, if you mostly read modern internet text, you might have the opposite bias.
I can confirm the later effect; prior to the internet, my primary exposure to sans serif fonts were government documents and forms, and advertisements, neither of which inculcated an association with any virtues.
Perhaps you should compare government documents and forms from different governments. UK government forms are extraordinarily beautiful, and welcoming, and are easy to fill out. US government forms, on the other hand, seem almost calculated to be unfriendly, and are incredibly difficult to fill out even when you use supporting instructions. It almost seems like they have been deliberately designed so that they cannot be filled out without the assistance of a lawyer. Canadian forms seem pretty neutral, and practical, but are nowhere near as pleasant to fill out as UK forms are.
No approval for AP/IB/SAT/ACT, as far as I can see.
No RPN. Every modern graphing calculator needs a mode (doesn't have to be the default) with RPN and a visible (4+ entry) stack. Once people actually learn how to use that for rapid, efficient calculations, they won't go back, but they never learn because all the major calculators don't even offer it as an alternate mode. That's the killer app for "graphing" calculators, because they can show multiple stack entries.
RPN may not be useful for math classes, which tend not to have as many problems involving many sequential calculations, but it's extremely valuable for science and engineering.
Funny you mention that. I had to get special permission to use my HP-15C on a "calculators allowed" exam in a first-year "calculus for non majors" course, and had to show the instructor how to reset its memory.
Which I found amusing, because the list of approved calulators included several graphing calculator models with computer algebra systems in ROM sufficient to produce correct answers to every problem on the exam, and they're worried about what I might have stored in the memory of a (then) 15-year old keystroke programmable calculator with a single line, ten digit, seven segment numeric display.
> Starting from a preliminary list of 200 titles created by bookshops and journalists, 17,000 French participants responded to the question, "Which books have stuck in your mind?" (Quels livres sont restés dans votre mémoire?) As Le Monde journalist Josyane Savigneau aptly clarified in her article, the list is not meant to encompass the 100 most distinguished French literary works of the 20th century, but rather to reflect the emotional connections of the French populace.[1]
Limiting the poll to 200 books (if I'm understanding it right... the cited le monde page is paywalled), selected by the elite French-literati, who then polled sub-elite French-literati, is a questionable basis for a list of "books of the century", even in France. Numbers of votes for each book would've been nice, to see how unanimous the top selections were.
8/100, in any language, are from the last third of the century (after 1967). Of those, 4 are well-known in North America (Styron, Eco, Rushdie, Solzhenitsyn). Of those, only two, Styron and Rushdie, are originally in English.
The most recent -en- works prior to those two were Kerouac in 1957 and then Nabokov and Tolkien in 1955.
Given the chronological bias, and few postwar -en- works probably due to distribution and translation challenges, it's pointless to mention all of the American and even British classics they left out from later in the 20th century. English-language books from the 80s and 90s, particularly science fiction, might have barely reached mainstream French consciousness in 1999.
You can't do that without already having the contents of the book, in which case getting an LLM to regurgitate it with partial prompting shouldn't be legally relevant at all. What it regurgitates will have errors, and if you try to chain that as prompt cues without re-basing each cue to the actual text (which you have separately), the LLM's output will rapidly lose coherence with the original work.
If its responses were perfect so that you could chain them, or if you could ask "please give me words 10-15 of chapter 3 paragraph 4 of HPatSS, and it did so, then you'd have a better case to complain. Still, the counterargument is that repeated prompting like that, explicitly asking for copyright violation, is the real crime. Are you going to throw someone in prison if they memorize the entirety of HPatSS and recite arbitrary parts of it on demand?
Combining both issues: that LLMs are only regurgitating mostly accurate continuations, and they're only providing that to the person who explicitly asked... any meaningful copyright violation moves downstream. If you record someone reciting HPatSS from memory, and post it on youtube, you are (or should be considered) the real copyright violator, not them.
If you ask for an identifiable short segment of writing, or a piece of art, and get something close enough that violates copyright, that should really be your problem if you redistribute it (whether manually or because you've coded something to allow 3rd parties to submit LLM prompts and feed answers back to them, and they go on to redistribute it).
Blaming LLMs for "copyright violation" is like persuading a retarded person to do something illegal and then blaming them for it.
Changing the metadata profile to at least raid1 (raid1, raid1c3, raid1c4) is a good idea, especially for anyone, against recommendations, using raid5 or raid6 for a btrfs array (raid1c3 is more appropriate for raid6). That would make it very difficult for metadata to get corrupted, which is the lion's share of the higher-impact problems with raid5/6 btrfs.
check:
convert metadata: (make sure it's -mconvert — m is for metadata — not -dconvert which would switch profiles for data, messing up your array)reply