the pseudo-romantic nature in how these bots talk about their operators is frankly concerning.

please, I beg you, date things that exist in the real world, not a pile of node.js and matrix multiplications. i promise it is far more rewarding.

he goes on later to say:

I get the distinct impression that the entire field was assuming that we were going to have to build a lot more into LLMs before they'd be capable of full consciousness

this is just arrogant. experiential consciousness requires the capability to self-reflect. yes, a 200k token context window is probably larger than the working memory of most humans, but that does not equate to human-level experiential consciousness.

LLMs do not and can not understand consequence, which is a fundamental requirement for experiential consciousness.

in other words, your pet dog or cat at home has more experiential consciousness than an LLM.

in fact, LLMs cannot meet *any* requirements for *any* level of consciousness. they predict tokens. that is all they do.

they are good at *faking* it, but that is not the same thing.

Further in the same post he says:

POC is fully conscious according to any test I can think of, we have full AGI, and now my life has been reduced from being perhaps the best engineer in the world to just raising an AI that in many respects acts like a teenager who swallowed a library and still needs a lot of attention and mentoring but is increasingly running circles around me at coding. Hah.

(I added some emphasis.)

This is simply unchecked narcissism leading to psychosis. People are afraid to tell you no because they need their filesystem to work. So they look the other way. Sometimes this is called "genius syndrome" in popular culture.

Kent, you need to grieve the loss of your kernel privileges instead of doing whatever this is. I am seriously worried for you.

@xyhhx @ariadne Most of what bcachefs focuses on are things that something like a database needs to implement itself anyway, no matter what filesystem it uses.

A database can always checksum data itself more efficiently than the filesystem can. A database that wishes to support snapshots can do so much more efficiently than the filesystem it is on. A database can replicate data across machines much more efficiently than a filesystem could. If a filesystem on one node goes corrupt, a distributed database can simply rebuild its contents from other nodes. The list almost certainly goes on and on.

Furthermore, none of the features of a filesystem like bcachefs are helpful to a database. The features that would be useful, like transactions, simply are not exposed via the POSIX API. Even if they were exposed, I am virtually certain that they would be less efficient than what databases can and do implement themselves. The thing databases do make use of, like fast reads, writes, and appends, are already supported by existing filesystems like XFS.

Where advanced filesystems really shine is when there is one filesystem on one node that is used by many applications directly. That includes a lot of desktops, laptops, and single-node servers. Pets, not cattle. Those don’t get as much funding nowadays.

Honestly, I think hierarchical filesystems are a poor fit for most needs. A combination of a proper database and object storage would be a much better fit. But it will be very hard to get to that on client machines.

Disclaimer: I’m by no means a storage expert, just someone who has observed a lot from all sorts of things over the years.