> We are completely overwhelmed, of course.
-
thank you Meta, I proclaimed, filling the Linux kernel with a quote unquote overwhelming amount of fucking slop
@zzt
Identity scans for everything on the internet. Thank you Meta! -
> We are completely overwhelmed, of course. The glimmer of hope is that we secured sufficient LLM budget and access (thank you Meta!) to run reviews with multiple frontier models on each patch. This eliminates some hallucinations.
jesus christ https://mastodon.catgirl.cloud/@lina/117130089920503564
@zzt feeling like gandalf when he read that journal about the fall of moria, only it's about LLMs coming to the linux kernel
-
@hipsterelectron@circumstances.run @zzt@mas.to @lykso@tiny.tilde.website @c0dec0dec0de@hachyderm.io
Your list of operating systems ain't exactly large if LLM extruded code is a complete dealbreaker. Especially if you're exposing something on the internet.
-
@hipsterelectron@circumstances.run @zzt@mas.to @lykso@tiny.tilde.website @c0dec0dec0de@hachyderm.io
Your list of operating systems ain't exactly large if LLM extruded code is a complete dealbreaker. Especially if you're exposing something on the internet.
@daemonspudguy @hipsterelectron @zzt @c0dec0dec0de Yeah. Longer-term plan is developing something on top of seL4 with (probably) Idris 2, but I need something for hosting in the meantime.
I'm also planning to do this in a private-ish environment (hence the tilde), releasing everything under a license that precludes use by any entity that is not worker owned and controlled, and restricting access to the source code along the same lines.
-
@zzt right but previous maintainers had to type a git command. don't you see, now clod can do it. after reviewing it. we trust clod's review, don't we?
well that's fine because we crosscheck it with chatgpt. we trust chatgpt's review, don't we?
well that's fine because we crosscheck it whatever the fuck meta is cooking. we trust meta's meth lab's review, don't we?
-
@hipsterelectron@circumstances.run @zzt@mas.to @lykso@tiny.tilde.website @c0dec0dec0de@hachyderm.io
Your list of operating systems ain't exactly large if LLM extruded code is a complete dealbreaker. Especially if you're exposing something on the internet.
@daemonspudguy @zzt @lykso @c0dec0dec0de that's why i'm making my own kernel
-
@daemonspudguy @zzt @lykso @c0dec0dec0de that's why i'm making my own kernel
@daemonspudguy @zzt @lykso @c0dec0dec0de operating systems like linux are extraordinarily inefficient
-
@lykso @hipsterelectron @daemonspudguy @zzt FUCK, tmux, is nothing sacred?!
-
@daemonspudguy @zzt @lykso @c0dec0dec0de that's why i'm making my own kernel
@hipsterelectron@circumstances.run @zzt@mas.to @lykso@tiny.tilde.website @c0dec0dec0de@hachyderm.io
I'm curious, is it a microkernel?
-
@lykso @hipsterelectron @daemonspudguy @zzt FUCK, tmux, is nothing sacred?!
@hipsterelectron@circumstances.run @zzt@mas.to @lykso@tiny.tilde.website @c0dec0dec0de@hachyderm.io
No. Sacredness was killed in the year 2020 along with all of humanity. 2020 was actually the apocalypse and this is actually a purgatory for all of humanity. /j
-
@hipsterelectron@circumstances.run @zzt@mas.to @lykso@tiny.tilde.website @c0dec0dec0de@hachyderm.io
I'm curious, is it a microkernel?
@daemonspudguy @zzt @lykso @c0dec0dec0de yes but not in the L4 style. a "monolithic" kernel actively abdicates security boundaries or protocols of any sort and is impossible to optimize to application-specific access patterns. ring 0 is for memory allocations and not for hardware. explicit message buffering (unlike seL4) enables strong guarantees around data persistence and data copies can be used to safely transfer across security contexts as well as memory hierarchy regimes. the bullshit linux does to pass around pages to avoid copies is incredibly archaic and it's very cache-unfriendly
-
@daemonspudguy @zzt @lykso @c0dec0dec0de yes but not in the L4 style. a "monolithic" kernel actively abdicates security boundaries or protocols of any sort and is impossible to optimize to application-specific access patterns. ring 0 is for memory allocations and not for hardware. explicit message buffering (unlike seL4) enables strong guarantees around data persistence and data copies can be used to safely transfer across security contexts as well as memory hierarchy regimes. the bullshit linux does to pass around pages to avoid copies is incredibly archaic and it's very cache-unfriendly
@daemonspudguy @zzt @lykso @c0dec0dec0de most of POSIX can be supported but not fucking pthreads. the entire concept of a mutex is completely broken and that's why deadlocks are even allowed to happen. haven't figured all that out yet but i think any constructs built on top of c11 atomics need to have clear semantics for progress. the ring buffer i've built supports stuff like "block until N bytes are available to write" and it makes parallelism so much easier when data structures provide guarantees around data flow rather than control flow
-
@daemonspudguy @zzt @lykso @c0dec0dec0de most of POSIX can be supported but not fucking pthreads. the entire concept of a mutex is completely broken and that's why deadlocks are even allowed to happen. haven't figured all that out yet but i think any constructs built on top of c11 atomics need to have clear semantics for progress. the ring buffer i've built supports stuff like "block until N bytes are available to write" and it makes parallelism so much easier when data structures provide guarantees around data flow rather than control flow
@daemonspudguy @zzt @lykso @c0dec0dec0de alexia massalin did the absolute coolest fucking shit ever with only an atomic compare-and-swap instruction in her synthesis kernel https://pages.cs.wisc.edu/~bart/736/papers/synthesis.pdf i'm doing a lot more with filesystem i/o and scheduler control because the scheduler is a security boundary and no process should have a global view of all files
-
@daemonspudguy @zzt @lykso @c0dec0dec0de most of POSIX can be supported but not fucking pthreads. the entire concept of a mutex is completely broken and that's why deadlocks are even allowed to happen. haven't figured all that out yet but i think any constructs built on top of c11 atomics need to have clear semantics for progress. the ring buffer i've built supports stuff like "block until N bytes are available to write" and it makes parallelism so much easier when data structures provide guarantees around data flow rather than control flow
@hipsterelectron That sounds fantastic, TBH. Really looking forward to seeing what you end up writing!
@daemonspudguy @zzt @c0dec0dec0de -
> We are completely overwhelmed, of course. The glimmer of hope is that we secured sufficient LLM budget and access (thank you Meta!) to run reviews with multiple frontier models on each patch. This eliminates some hallucinations.
jesus christ https://mastodon.catgirl.cloud/@lina/117130089920503564
@zzt “we are drowning in slop, the solution is obviously more slop”
-
if only we had lots and lots of examples of easier systems projects using similar techniques that entirely failed and now don’t exist as functioning projects
what are bun and rsync, why do you keep saying those names
@zzt it will be vindicating seeing linux go on that list
-
@zzt@mas.to @c0dec0dec0de@hachyderm.io
Even when translating to what the phrase is poorly trying to pretend to be different from, I have absolutely no fucking clue what "social warrior project" is meant to convey because "social justice warrior project" is just not a phrase I can make any sense out of. It doesn't scan at all.
@daemonspudguy @c0dec0dec0de @zzt it is the right-wing nutjob framing of “ethical project”
-
@zzt This is just the natural outcome.
The real red flags where when IBM (that is working into AI from more than a decade) buyed Red Hat and when they pushed madly a broken systemD as the usual "it's the FuTuRe" in the same guise they are pushing Wayland... that isn't really ready yet either for a lot of cases.
With these precedents why the LLMs push is so strange? Is just a normal development for these people.@raster I can not read the conversation here, thanks to @zzt for the block /s
I see that he enjoys the phrase "fucking slop", so you can point him and his fucking awful.systems to the open (non-blocked) anti-AI discussion linked below.
Alternatively, with less profanity:
― sloppy human systems.
https://www.reddit.com/r/freebsd/comments/1vtgx7t/antiai_open_source_has_an_enemy_in_common_but/
-
good thing the linux kernel has multiple frontier models, none of which are capable of generating a C compiler that fucking works
thank you Meta!
@zzt i am wondering if any of the benchmarks is "generate compiler for arbitrary language and architecture" with a testset that is just arbitrary code to compile amd execute (which is not oart of the context).
I guess not, that could lead to embaressment and admitting they set a trillion dollars aflame for spicy autocomplete
