Post B821wWRqhHw6okSaHI by jpm@aus.social
(DIR) More posts by jpm@aus.social
(DIR) Post #B7zUcS1HlbdGUIw9Fg by ariadne@social.treehouse.systems
1 likes, 0 repeats
some analysts have said that gpt-5.5 is on par with the systems analysis capabilities of anthropic's new fable/mythos modelsas that is a bold claim, it piqued my interest, so, to test this, I've set up codex and instructed it to build a microkernel-based operating system with a Linux kernel personality from scratchit hasn't done a terrible job so far, but there are a lot of bugs which have all traced back to it missing key details when doing things like context switchingthe coding bot is pushing its work to my github if you want to follow along and see what it is producing in this experiment: https://github.com/kaniini/bunixI still don't see myself using this for real world stuff, there's too many unanswered questions
(DIR) Post #B7zVo2pLMGKnmdRXdo by ariadne@social.treehouse.systems
0 likes, 0 repeats
and by missing details I mean:- it enabled SSE2/AVX on kernel side but did nothing to save/restore on thread switching- it has only very recently generated any kind of data structures code, instead making everything linked lists and tables- SMP bringup was miserable and I told it to focus on that early, it wound up brute forcing its way through booting the secondary APs
(DIR) Post #B7zWR68B0aunK7sGcC by ariadne@social.treehouse.systems
0 likes, 0 repeats
anyway as an operating system designer & security engineer for almost 20 years now, I think I'm pretty qualified to judge its progress. so far I would give it a D. things are looking very new jersey here, if you catch my drift
(DIR) Post #B7zXuB3TsZVWwq6knQ by ariadne@social.treehouse.systems
0 likes, 0 repeats
however it does boot...
(DIR) Post #B7zl9KPQgMUPhsvej2 by ariadne@social.treehouse.systems
0 likes, 0 repeats
it just spent 45 minutes spiraling over a recursive lock deadlock issue
(DIR) Post #B7zmswT08DjSj3uibA by ariadne@social.treehouse.systems
0 likes, 0 repeats
@dysfun nah its splatting printf's all over the place, i think it is close to making a discoveryit has to be with all these printf's
(DIR) Post #B7zoub9CZEAd5nZPto by ariadne@social.treehouse.systems
0 likes, 0 repeats
@dysfun about an hour ago 🙃
(DIR) Post #B7zpSjYBpbRt03XFce by ariadne@social.treehouse.systems
0 likes, 1 repeats
@dysfun you see,this is art,i'm going to have $ git commit -m \ "Verified with make test."printed on t-shirts, and sell them
(DIR) Post #B7zpy4bDcdMtG6tbbE by ariadne@social.treehouse.systems
0 likes, 0 repeats
you ok there buddy? i don't think that makes any meaningful difference at all...
(DIR) Post #B7zqChcfKaRuLFY4HI by ariadne@social.treehouse.systems
0 likes, 0 repeats
@dysfun i mean it is an optimization i guess, there's no need to immediately reenter the receiving process if it isn't going to synchronously reply
(DIR) Post #B7zqLpRvbIARDqsU7c by ariadne@social.treehouse.systems
0 likes, 0 repeats
what the fuck was *that* patch?
(DIR) Post #B7zqSRGltorGFyoMee by ariadne@social.treehouse.systems
0 likes, 0 repeats
(for context, it is attempting to do the QNX thing, also sometimes known as the Utah Mach 4 shuttles thing, where an IPC can bypass the scheduler entirely and just exit into the receiving thread)
(DIR) Post #B7zqgG5OHBy7cr4sOe by ariadne@social.treehouse.systems
0 likes, 0 repeats
@gabboman @dysfun so far, i remain convinced that claude is the only one worth anyone's time
(DIR) Post #B80rPovxkmJttULyd6 by ariadne@social.treehouse.systems
0 likes, 0 repeats
i gave it an entire set of objectives to do overnight, and then it didn't do them, and then when i corrected it, it still didn't do all of them.clearly the technology of the FUTURE
(DIR) Post #B80rkDzaIRrFP2mEOu by ariadne@social.treehouse.systems
0 likes, 0 repeats
@Lenni let me have my amusements, thanks
(DIR) Post #B80s5Kz9UYITAldrxg by ariadne@social.treehouse.systems
0 likes, 0 repeats
@Lenni you're one post away from eating a block, and possibly a report to your admin
(DIR) Post #B80tb4yOez56jTzKgy by ariadne@social.treehouse.systems
0 likes, 0 repeats
@Lenni the bottom line is that you seem to not understand that i am doing legitimate research here, or maybe you do, but you see it as illegitimate because of your own biasesin other words, you're doing the same shit techbros do, but with the opposite position.and i don't disagree with your position, but this research will help me to debunk claims about the capabilities of this model, and others in the future that i choose to evaluate
(DIR) Post #B80uawHT4u1hjBPDLk by ariadne@social.treehouse.systems
0 likes, 0 repeats
@Lenni beyond that, as a tech industry executive, there is an expectation that i be fully aware of all of this stuff, its capabilities, and so on.
(DIR) Post #B80v2QndiYFQCPYy4O by ariadne@social.treehouse.systems
0 likes, 1 repeats
@Lenni no, i want to see if gpt-5.5 can actually build an operating system from scratch. the fact that it is doing so badly, and that the tooling is way worse than what openai's competitors has, is indeed a source of amusement to me, but that isn't the point. the point is to see if it can do it.and, maybe, to make & sell sarcastic merch about it along the way.it is both research and legitimate artistic creation in that way, even if the AI-generated slop is still slop, that isn't the actual work product here
(DIR) Post #B80vUkBlDtKgRRLVp2 by ariadne@social.treehouse.systems
0 likes, 0 repeats
omg it did it AGAIN. codex is like that one worker you always see sitting around doing nothing in a caltrans work zone.
(DIR) Post #B80vf32zxh0RpJA2Nc by ariadne@social.treehouse.systems
0 likes, 0 repeats
@dysfun yes, claude will actually implement a whole plan now from beginning to end, even with fanout to other agents. codex is also running in agentic mode, so this shouldn't be a problem...
(DIR) Post #B80vqFkF5pxaOOjN8S by ariadne@social.treehouse.systems
0 likes, 0 repeats
why does it call a work unit a 'slice'
(DIR) Post #B80vtsuHwigeiiA4X2 by ariadne@social.treehouse.systems
0 likes, 0 repeats
gonna COMMIT THE SLICE BRO
(DIR) Post #B80wGBXlKtQGtLG6vg by ariadne@social.treehouse.systems
0 likes, 0 repeats
@dysfun too late, it was ALREADY COMMITTED
(DIR) Post #B80xGlOi8Dbu2qrfV2 by ariadne@social.treehouse.systems
0 likes, 0 repeats
anyway, after figuring out if gpt-5.5 can build a production grade operating system from scratch (i suspect the answer is no), i will have the anthropic hack anything model: civilian edition (fable) attempt to find and build exploits for it.
(DIR) Post #B80xVkTRqcOZWZE4MS by ariadne@social.treehouse.systems
0 likes, 0 repeats
@dysfun ha
(DIR) Post #B80xfTCGwDwqpupMo4 by ariadne@social.treehouse.systems
0 likes, 0 repeats
@dysfun claude tokens are way more expensive 😂
(DIR) Post #B80xn0N3zp3zkcT4xk by jubilee@social.treehouse.systems
0 likes, 0 repeats
@ariadne @dysfun oh is that the C one? they also like "Verified with git diff --check"
(DIR) Post #B80xqoLod2IBd5hPma by ariadne@social.treehouse.systems
0 likes, 0 repeats
@jubilee @dysfun oh it's done that one too
(DIR) Post #B80y4mLOt0WRSaAJMm by ariadne@social.treehouse.systems
0 likes, 0 repeats
ok turns out codex has goal mode, not just plan mode
(DIR) Post #B80yV7BG2Vi1c0ylm4 by ariadne@social.treehouse.systems
0 likes, 0 repeats
ok with goal mode, i think it will stop taking breaks, lets see
(DIR) Post #B8107rQ6fcqnynxp56 by ariadne@social.treehouse.systems
0 likes, 0 repeats
what even is this commit
(DIR) Post #B817Y3OtRv5yLQpjQO by ariadne@social.treehouse.systems
0 likes, 0 repeats
@tknarr @Lenni in fairness, i don't think a human could write as much as this has generated in the timeframe given. maybe if you give the human meth or something, but then the code quality will likely be the same, if not worse.
(DIR) Post #B821w0gcmok2JJ7vrk by faoluin@chitter.xyz
0 likes, 0 repeats
@ariadne
(DIR) Post #B821w1F0iyFO1xDOSm by arch@floofy.tech
0 likes, 0 repeats
@ariadne you are what you eat, and the model ate a lot of recursive lock deadlocks
(DIR) Post #B821w1gf4AMhPi9TYu by razze@osna.social
0 likes, 0 repeats
@ariadne sounds like your not using skills? While codex should be good, I'm not sure that's the best harness, they matter a lot.
(DIR) Post #B821wJd4MGBf3mbrSi by tris@chaos.social
0 likes, 0 repeats
@ariadne It would be fun to see sel4 but worse :P
(DIR) Post #B821wQv3FnjLfa4oxk by Lenni@fosstodon.org
0 likes, 0 repeats
@ariadne Didn't you waste enough energy just to make fun of slop machines by now? This isn't benefitting anyone to be honest.
(DIR) Post #B821wRMLcJZ52EqcVc by Lenni@fosstodon.org
0 likes, 0 repeats
@ariadne IMO this isn't any better than people seriously using slop machines ti generate production code. You're equally destroying our planet generating this garbage, but it doesn't even have any usecase afterwards at all. It could be considered to be an even worse use of this technlogy.
(DIR) Post #B821wRYksBUPeieWqe by Lenni@fosstodon.org
0 likes, 0 repeats
@ariadne I seriously don't get why you're so angry. I'm just stating my opinion on LLM use, which generally doesn't seem to be too different from your view of that topic. Would your reaction be the same if I posted a similar reply under the post of an AI techbro? And why would this be something my instance admin should care about? I'm not trying to make you angry or anything like that, just a bit confused and seriously wondering about it.
(DIR) Post #B821wTjom1WkPbqOMy by Lenni@fosstodon.org
0 likes, 0 repeats
@ariadne From your previous posts it seemed just like what you were doing was just for amusement without any real goal or benefit to anyone. I personally think criticising that would be appropriate. Sorry if my reaction sounded a bit harsh though.
(DIR) Post #B821wUZZfavZ083j2e by grumpybozo@toad.social
0 likes, 0 repeats
@ariadne I have occasionally displayed a similar work pattern. Therefore the LLM is human. \s
(DIR) Post #B821wVqGwzsWwCsZhw by cadey@pony.social
0 likes, 0 repeats
@ariadne oh i saw the same thing in my vibe coding experiment i'm writing up
(DIR) Post #B821wWRqhHw6okSaHI by jpm@aus.social
0 likes, 0 repeats
@ariadne only difference is that one worker apparently doing nothing is actually doing something - watching for utilities buried in a different position than the documentation says while the others are digging. Source: my brother, an ex- civil construction worker
(DIR) Post #B821wXMvH5adflAAEq by dotsie@mastodon.social
0 likes, 0 repeats
@ariadne I have to use codex instead of Claude in a couple weeks for a project and I am dreading the switch because of nonsense like this
(DIR) Post #B821wXSExJqLwFePWi by tknarr@mstdn.social
0 likes, 0 repeats
@ariadne @Lenni Can't validly critique something if you don't know itcs actual capabilities and limits. I think what you're doing is a useful exercise. I find it easier to deal with people pushing genAI when I can respond with logs of me trying exactly what they claim works and showing actual results that manifestly don't work without more human work on them than doing it myself would have involved.
(DIR) Post #B821wbgZDzgF3qNuPg by tknarr@mstdn.social
0 likes, 0 repeats
@ariadne @Lenni Probably, computers can always do operations faster than a human. That's why I use the refactoring and code-generation tools in my IDE. But speed isn't everything. I phrase it "It doesn't matter how fast you can fail to do something.".