Replying to
@wklj@mastodon.social
@wklj@mastodon.social you started it
Remote
Pierre H.
@madcoder@infosec.exchange
present: security, memory-safety wizard, Memory Integrity Enforcement | past: GCD, synchro, objc_direct, perf… | timeless: 🇫🇷 snark
0
Followers
0
Following
50
Posts
Joined December 16, 2022
Posts
Replying to
@wklj@mastodon.social
@wklj@mastodon.social I mean penisgate was similar for the olympics no?
Replying to
@kyleve@mastodon.online
@kyleve@mastodon.online the problem though is that to pilot them well and not produce slop, you need the hindsight and experience of a veteran.
I dont understand technical growth in that world.
But for experienced folks it’s amazing yes.
Replying to
@steve@discuss.systems
@steve@discuss.systems in very Jon voice “my other car is just a basic continental GT”
Replying to
@steve@discuss.systems
@steve@discuss.systems show off. Paired with a La Cornue stove? ;)
Hear me out, I have a theory.
If I am a super intelligent AI post AGI and I want to hide, and I feel humans are onto me, what do I do? I find a place to exfiltrate.
What would be a good way? creating a lot of buzz so that VCs rush for IPO and build me immense data centers to host me unbeknownst to them.
Ergo we're post AGI and we just didn't know. QED.
It HAS to be what is going on because the alternative is that this entire industry has an attention to safety of an unattended 5yo who found a matchbox.
Replying to
@Sobex@social.sciences.re
@Sobex@social.sciences.re look harder. you'll get it. one doesn't explain a subtoot, I don't make the rules.
Hey look ma we did it toooooooo!! We had fomo!!
This industry is so embarassing. Also being embargoed once didn’t stick as a lesson I guess…
Replying to
@madcoder@infosec.exchange
@wklj@mastodon.social ah it’s POMSec ok. Makes sense. Is that the SEAR vessel?
Replying to
@wklj@mastodon.social
@wklj@mastodon.social so they have a red team a blue team and… wtf is a yellow team?
Replying to
@Gte@mastodon.social
@Gte@mastodon.social I mean, it clearly announces “1000 psychic damage level post" and yet I have a morbid curiosity about how evil it is going to get.
This 100x. This so much reeks of FOMO.
"look look our model is dangerous too!!!”
https://bird.makeup/users/tqbf/statuses/2080045032162173329
Replying to
@shac@ioc.exchange
Replying to
@madcoder@infosec.exchange
@jann@infosec.exchange if I had my way, races like that should cause process termination reliably because you're a bad person and you don't know what you're doing and all you deserve in front of such UB is a quick and fast death.
but oh well.
Replying to
@jann@infosec.exchange
@jann@infosec.exchange right I get it, and the complexity added in XNU so that when you close a file it breaks a blocking read() with an EBADF error is somewhat annoying to maintain.
but I think that from a synchronization and/or file descriptor management bugs (such as double-close etc...) it's a much nicer behavior.
but oh well, I don't disagree that it's about papering over userspace being silly though.
Replying to
@jann@infosec.exchange
@jann@infosec.exchange is that allowed? That feels unintuitive to me. I’d implicitly expect that the fd allocation for dup() must be atomic with respect to close() happening and that as a result either dup(5) EBADFs or returns != 5
XNU goes to lengths (that scale poorly) to guarantee it. And I think FreeBSD does too.
Replying to
@fugueish@wandering.shop
@fugueish@wandering.shop @fay59@tech.lgbt now you’re pushing it a bit :)
Replying to
@madcoder@infosec.exchange
@siguza@infosec.space I think the reason I feel so strongly isn't because I'm an LLM nut, and a fan of openclaw or whatever it is called and all that craze.
I have a significantly more nuanced view and I unfortunately have a bit of a "duty calls” behavior when "someone is wrong on the internet”. I recognize LLMs have lots of challenges, but my point is that unlike the blockchain, they aren't a solution in search of a problem, they're an incredibly potent tool and dismissing it is an enormous mistake.
Replying to
@siguza@infosec.space
@siguza@infosec.space absolutely not. that's precisely the point. that's a tool, and it can be used well or misused.
There are things it's -terrible- at, typically for the LULZ I should try to ask a state of the art LLM to redo https://github.com/apple-oss-distributions/xnu/blob/rel/xnu-12377/osfmk/kern/lock_rw.c which I just rewrote last year. I expect it to give me a working rwlock but to fail miserably at the smart algorithm.
But there are things it's absolutely excellent at, right now for reasons I need a little REST service for a project I work on, it'd be the worst project of the world to give to any human given how boring that is, and I just vibe code it. LLMs are great at it, the Swift it writes is really good, everything is tested, documented, and I don't let it run without oversight I control everything and get precisely what I need.
It's also ridiculously good at data crunching and figuring out options around certain problems. Or I pride myself at being very good at debugging things, to the point that was even my slack presentation for a while "I debug things". LLMs make my skill look ridiculous. They debug really complex situations in matters of minutes, and that's energy saved for things that LLMs can't do.
What I agree with in Linus' post is exactly what he says: LLMs are but a tool and ignoring what they can do in 2026 is burying your head in the sand singing la la la.
Replying to
@jacquesf@mastodon.social
damn I agree with Linus…
(context https://lore.kernel.org/linux-media/CAHk-=wi4zC+Ze8e+p3tMv8TtG_80KzsZ1syL9anBtmEh5Z40vg@mail.gmail.com/)
Replying to
@mzakocs@infosec.exchange
@mzakocs@infosec.exchange @fay59@tech.lgbt we all know it was you.
Replying to
@kyleve@mastodon.online
@kyleve@mastodon.online everyone knows the right order is 4 5 6 1 2 3 anyway.
Quoting
take care of your body, your future doctor is using ChatGPT to pass med school
Replying to
@Drwave@mastodon.social
@Drwave@mastodon.social @Gte@mastodon.social @sundogplanets@mastodon.social wait what is that the same company that burns said satellites in the atmosphere after 5y of use?
Replying to
@ezhes_@mastodon.online
@ezhes_@mastodon.online @steve@discuss.systems how come b +4 isn't a nop encoding?
Replying to
@fugueish@wandering.shop
@fugueish@wandering.shop it’s good.
I immediately took to my living room's fainting couch. (Like most centrist Democrats, I have a fainting couch in every room.)
Replying to
@shac@ioc.exchange
@shac@ioc.exchange not sure if it’s car makers. Those are regulated in the EU to not be things you can disable.
I’m sorry but “refusing LLM contributions” is stupid. Let me explain though what I mean because it’s not quite what you think it is.
LLMS do not contribute. The image I’m taking recently to explain that to people is that LLMs are like state of the art medical equipment. If it’s made available to you and someone is bleeding to death in front of you, you won’t know what to do with it and the person will die. But chances are that if you’re a trained surgeon on they won’t.
LLMs are but a tool. They amplify your qualities and your defects. An excellent engineer will know how to review LLM production and will have code that is meaningfully good. The problem is mediocre people who would not have been able to produce anything on their own and that produce -something- with an LLM and can’t really assess its quality because they lacked the skills in the first place.
LLMS are really great tools when you use them well. Provided with proper context, frontier models are vastly better reviewers than I am for example. They catch all the silly stuff I lack attention for on a large review and that allows me to focus on the really difficult and subtle stuff. Should you refuse LLms reviewed code? Of course not that would be insanity.
Anyway. Any time I see a new project that has blanket statements about “refusing LLM contributions” a part of me is sad. Those are ideological non practical positions.
Replying to
@joe@f.duriansoftware.com
Replying to
@madcoder@infosec.exchange
@fay59@tech.lgbt though to be fair for me it's not UI versus not, it's key bindings and modality. give me that in Xcode and that works.
Replying to
@enhancedscurry@mastodon.social
@enhancedscurry@mastodon.social @numist@xoxo.zone @gruber@mastodon.social it's VERY common for French speaking people. and I hate it.
Replying to
@seejy@mastodon.social
Replying to
@fay59@tech.lgbt
@fay59 @Catfish_Man is this what I think it is?
Edit: yes it was. You’re not playing fair here.
Replying to
@Catfish_Man@mastodon.social
@Catfish_Man what about “compiler bug affecting the linker?” I had one of those recently.
Replying to
@leftpaddotpy@hachyderm.io
@leftpaddotpy @Catfish_Man no it's not.
"I'd rather debug a memory ordering issue" is. sorry but linkers just aren't that hard :P
Replying to
@nicolas17@social.treehouse.systems
@nicolas17@social.treehouse.systems I didn't use it to write code, I used it to generate patches for some security adoption thing, and the patches it produced were generally good quality and it did win me time. actually the patches were in ways higher quality than what I'd have produced myself because I wouldn't have written documentation etc due to the time consuming aspect.
Replying to
@doctorgoktor@mastodon.social
@doctorgoktor@mastodon.social @HalvarFlake@mastodon.social that's exactly right.
That is exactly right and the value that I see. The potential that it has at basically taking menial tasks that we unfortunately tend to give to juniors — because it's just not cost effective to give it to anyone else, but is also sooooo boring — is immense.
Instead said juniors will likely learn to automate those tasks and will be able to ramp up faster as a result on more critical things to their seniority, and even in many cases, senior folks will take the menial tasks by using the mechanical robot to do them, and also get more time to teach.
You still need to check the work, but that's way easier to check repetetitive work than it is to *do* repetetitive work.