Post B8MgVe7NkJFbL07Uhs by ariadne@social.treehouse.systems
(DIR) More posts by ariadne@social.treehouse.systems
(DIR) Post #B8McYkueXYazSmORRA by ariadne@social.treehouse.systems
0 likes, 0 repeats
the agent session which was processing the stack of bunix explorations has completed the explorationsin the final notes, it stated that one of the success criteria was satisfied because the test (download 100mb file) was satisfied in 47 secondsfriends, I have a multi-gigabit connection and can download that same 100mb file in less than 1 second on any other TCP/IP stack...
(DIR) Post #B8MchktBzsytzq09dA by ariadne@social.treehouse.systems
0 likes, 0 repeats
I kinda want to reject that solution and say, no, the download needs to be performant, because I don't really consider taking 47 seconds to download 100mb from a CDN on a capable connection to be a success
(DIR) Post #B8MdLgqIShrR1rgns8 by syn_rst@norden.social
0 likes, 0 repeats
@ariadneI don't have multi-Gb/s here but even here it shouldn't take longer than 2 to 5 seconds.Which is a tenth of the runtime.Gee, that is seriously bad performance.Definitely Enterprise-grade software.
(DIR) Post #B8MdTDnyJ4U5HtC2aG by ariadne@social.treehouse.systems
0 likes, 0 repeats
so far my conclusion is that these tools are useful for code review (which, notably, is not a generative task per se), if you are willing to tolerate false positives, and building proofs of concept that are intended to be thrown away
(DIR) Post #B8MdmLMKRd4aDbNi9g by icing@chaos.social
0 likes, 0 repeats
@ariadne They are very good at bean counting PRs. Checking the stuff that makes human eyes glaze over.They, of course, have no idea if a PR is a good idea, sound design or propose alternate approaches.useful bean counters😌
(DIR) Post #B8MdqzTNFFphmqtP7Y by whitequark@social.treehouse.systems
0 likes, 0 repeats
@ariadne yeah, in terms of pure utility isolated from any externalities, using them for code review personally (as opposed to making someone else read it) is probably fine. i've seen people integrate that into their workflow in a disciplined way where it did not appear to lead them into a worse placei don't think the isolation is reasonable in general, but it is still a useful thought experimenti don't entirely agree on prototypes. i think building a skillset that makes prototyping faster (which is something i've done explicitly because i noticed that genAI is used for that task) is a lot more useful in the long term. there were many things i haven't been doing that it forced me to reconsider
(DIR) Post #B8MduXrPqfmhJ7pi8e by ariadne@social.treehouse.systems
0 likes, 0 repeats
@dysfun I think it's also ok for "splat a function that does XYZ here, and hook it up in the appropriate places" as long as XYZ isn't terribly complexbut why pay for that when GPT-OSS can do those autocomplete type tasks locally.
(DIR) Post #B8MeNZJHXS0cCVhrpw by ariadne@social.treehouse.systems
0 likes, 0 repeats
to me, a very senior engineer with a background in operating system design, the ability to request a proof of concept and come back in 20-30 minutes to a complete proof of concept is a legitimately useful thingfor real work, deterministic planning is essential for driving quality
(DIR) Post #B8Mf24GbrQrbnMlgmm by chrisvest@mastodon.social
0 likes, 0 repeats
@ariadne to reduce false positive rate, I organize sub-agents into reviewers and verifiers, where the verifiers responsibility is to refute findings produced by reviewers. It helps a fair bit, leaving mostly clear findings and judgement calls for me to look at.
(DIR) Post #B8MfcJaUNOtRjiLnEW by ariadne@social.treehouse.systems
0 likes, 1 repeats
@duarte yuck. as someone whose job is literally to lead a research group, having an agent "auto-research" simply will not yield quality research.the operator needs to develop the thesis themselves, for a number of reasons, but the largest one being that dialectic analysis is key to refining theory.
(DIR) Post #B8MgBUDArv9iozDTBw by nyc@discuss.systems
0 likes, 0 repeats
@ariadne Hmm. I have a relatively difficult kernel project that I tried to carry out using LLM GenAI. It runs userspace, passes LTP and mm-stress in qemu for 20+ architectures, but eventually has trouble with some races on real laptop x86 hardware. Could I get away with rebranding the attempt as a Proof of Concept? The thing is a large, difficult project, so it would help to get some kind of mileage out of it.(To wit: In 2023 I used the subject of the project as an example of something that would be difficult for LLM GenAI to do because it involves a sort of new conceptual distinction needing to be imposed as a retrofit that can't piggyback readily on the type systems for common kernel implementation languages because it involves something akin to frequent dimensional analysis. It's unclear how to interpret the result of attempting to do it anyway. It feels close enough that blaming the user is well within reach as well as the usual conclusion that LLM GenAI isn't actually capable of it. So, in principle, I could also rebrand it as doing my 2023 example as an experimental test of LLM GenAI, with the major proviso of not really having had much methodology to the time spent on it.)
(DIR) Post #B8MgFWFdE9fQJ0HxjM by ariadne@social.treehouse.systems
0 likes, 0 repeats
interestingly when you ask it to debug the code it generated, it has a tendency to use printfs everywhere, even if it is a problem that printf debugging cannot solve such as performance analysis(i wound up rejecting the TCP stack it generated, because 47 seconds to download 100MB of data on a multi-gigabit connection is absurd)
(DIR) Post #B8MgVe7NkJFbL07Uhs by ariadne@social.treehouse.systems
0 likes, 0 repeats
@nyc i built something quite similar with gpt-5.6 (i was granted early access), and it is basically in the same state. i wound up shelving my plans to build a harness for it to test porting to a real board because what it generated is too fragile, even if it illustrates the concepts i asked it to prove.
(DIR) Post #B8MggOKAo5CxXXE76m by ariadne@social.treehouse.systems
0 likes, 0 repeats
and for the most part, conferenceware has always been slop my entire career so this isn't really much of a change 😂
(DIR) Post #B8Mh3S5OzF4brpJVcO by codonell@fosstodon.org
0 likes, 0 repeats
@ariadne What do you think about things like oss-fuzz-gen which effectively task the LLM with exploring the solution space and locking in incremental progress with commits that regardless of their validity increase deterministic coverage numbers?
(DIR) Post #B8MhQbqUroOenRNski by ariadne@social.treehouse.systems
0 likes, 0 repeats
@whitequark for software engineers, i agree that it is better to spend time on developing fundamental skills, which in turn leads to faster prototypingwhen i am talking about prototyping for researchers, i am talking about people who do not typically write code for a living, because it isn't a skill directly relevant to their research.sometimes, those researchers have a luxury of having a staff software engineer in their group who can build their prototypes for them, but i suspect with generative AI, that type of software engineering job is likely to become extinct :(
(DIR) Post #B8MhUXR5Cl34HHN212 by ariadne@social.treehouse.systems
0 likes, 0 repeats
@icing agreed
(DIR) Post #B8MhXwEHTnBG2g0Hr6 by whitequark@social.treehouse.systems
0 likes, 0 repeats
@ariadne I acknowledge this line of reasoning. I don't know how I feel about it, but I agree that this is essentially what's happening
(DIR) Post #B8MheJS63BvxnuVOjI by ariadne@social.treehouse.systems
0 likes, 0 repeats
@whitequark it's a damn shame because it has been kind of an unwritten secret that a lot of OSS maintainers fall into those kinds of 'staff research software engineer' jobs
(DIR) Post #B8Mhw0NPiB0Rr3keES by dngrs@chaos.social
0 likes, 0 repeats
@ariadne @whitequark the fundamental problem of LLMs having no idea about facts or truth (while at the same time extruding confidently wrong text) remains unsolved, as a psychologist I continue to view LLMs as deeply unethical (in the sense that they're accidental master manipulators and nobody is immune to gaslighting). How do you (target audience: non-programmers!) distinguish "tool took 47 seconds for the job" vs "tool just extruded text claiming the job was done"?
(DIR) Post #B8Mi9qY33BJzQVKFvc by ariadne@social.treehouse.systems
0 likes, 0 repeats
@dngrs @whitequark great question. in my case, i am a domain expert in the task i asked it to do, so i can see "tool took 47 seconds for the job" and say "LOL".but that is not most people using this tool. they do not know.
(DIR) Post #B8MigkVQXJMCF7my48 by nyc@discuss.systems
0 likes, 0 repeats
@ariadne Two projects of similar dimensions. Could the similar experiences be worth writing about together?The LLM GenAI attempted to write up the subject of the project at:https://git.sr.ht/~nadiayvette/pgcl-testscripts/tree/main/item/PGCL-TECHNICAL-WRITEUP.texwith BibTeX at:https://git.sr.ht/~nadiayvette/pgcl-testscripts/tree/main/item/pgcl-writeup.bibThe source for the LLM GenAI implementation attempt is at:https://git.sr.ht/~nadiayvette/linux-pgclThe impoverished models for the correctness of the VM system in the Tessera sub-project to use formal verification tools to bughunt also give some big hints of where it might be falling down. It suggests that basically bolting so much GOFAI atop it all that the LLM GenAI is a minor player might eventually get working code out of it all, albeit only with substantial human effort to avoid directly writing the code.
(DIR) Post #B8MiooLo7okmU58LUO by ariadne@social.treehouse.systems
0 likes, 0 repeats
@dalias @whitequark the reason why i instructed it to build bunix as a C project was explicitly because there should be a lot of training data in its corpus relating to building C kernels :p
(DIR) Post #B8MjBd2aLF964E1TTk by ariadne@social.treehouse.systems
0 likes, 0 repeats
@codonell i haven't tried oss-fuzz-gen yet, but generally i don't think increasing deterministic coverage numbers simply for the sake of increasing deterministic coverage numbers is useful...
(DIR) Post #B8MjWJSoK7rjQo9WLI by codonell@fosstodon.org
0 likes, 0 repeats
@ariadne I'm thinking specifically of things like libm where I know I have untested cut points or conditionals and would like to find them and add coverage at lower human cost. My question is more along the lines of "Can LLMs search for things that would be expensive for me to do?" In the context of threading I want something to explore the total global ordering of operations.
(DIR) Post #B8Mjb0Gp6yBTLtmYYi by mcc@mastodon.social
0 likes, 1 repeats
@ariadne Realizing with horror the only thing I can do to push back against AI adoption is continue to use printf debugging in my open source code, so the AI plagiarizing me does it too
(DIR) Post #B8MmPL8O0uxCRpxwZc by Kroppeb@tech.lgbt
0 likes, 0 repeats
@ariadne very much this. In the beginning I also used to use it to make my emails sound more human (????) but haven't done that in a long time.I only use generated code in stuff I don't really care about. (That's why I only use it at work and not for personal projects :p)
(DIR) Post #B8MmPLNd6F9BD767Ki by ariadne@social.treehouse.systems
0 likes, 0 repeats
@Kroppeb something something, boss gets a dollar, i get a dime, that's why i use claude on company time
(DIR) Post #B8MnEUFnro2VV6p8sa by ariadne@social.treehouse.systems
0 likes, 0 repeats
@dalias @poliorcetics @mcc yeah, i hate this because it makes life shitty for humans
(DIR) Post #B8Moiaxra646xHLxIm by ariadne@social.treehouse.systems
0 likes, 1 repeats
@BoydStephenSmithJr @dngrs @whitequark everyone says that, and then 2 hours later they are buying horse paste from tractor supply because they have convinced themselves that ivermectin can treat COVID because while the LLM system prompt had safeguards, a person with confirmation bias will inevitably find ways to bypass them when sufficiently motivated.
(DIR) Post #B8Mp3LdpkNkxVqsdGK by ariadne@social.treehouse.systems
0 likes, 1 repeats
@BoydStephenSmithJr @dngrs @whitequark like the problem with these LLM chatbots isn't just that the LLMs themselves will gaslight you, it's that the customer *wants* to be gaslit. so when they *want* to be "right", they will *willingly* go be gaslit by the LLM.
(DIR) Post #B8MpG6TZRYXsF5A6aG by danlyke@researchbuzz.masto.host
0 likes, 0 repeats
@ariadne 100% this. I have, several times in my life, thought "oh, wow, this salesperson is really good", and then bought the product anyway.Even being consciously aware that the deception is happening isn't always enough. @BoydStephenSmithJr @dngrs @whitequark
(DIR) Post #B8MstwGx6CCA2gwjse by poliorcetics@social.treehouse.systems
0 likes, 0 repeats
@ariadne @dalias @mcc oh I fully agree, I’m not seriously proposing doing this, but the idea was funny if not applicable to the real world
(DIR) Post #B8MuLktR6HO2oRm5tg by wizbean@tech.lgbt
0 likes, 1 repeats
@ariadne @BoydStephenSmithJr @dngrs @whitequark and all the combined advertising infrastructure that's been built up over the last century are being singularly put into telling the sceptical people "give gaslighting a try, you're gonna love it!"
(DIR) Post #B8Mx6dM78r60xEfaPg by datenwolf@chaos.social
0 likes, 0 repeats
@ariadne @whitequark That job is extinct already. Technically I am a researcher. However, instead of doing research I spend most of my working time doing staff software engineer tasks. It's absurd. My colleagues, who are not software developers, vibe code a crapton of one-off tools for simple tasks. I see the benefit for them, and am infuriated that the proper tooling is so underdeveloped.
(DIR) Post #B8N7aCcw5940HnHBQW by ariadne@social.treehouse.systems
0 likes, 0 repeats
@FritzAdalis @wizbean @BoydStephenSmithJr @dngrs @whitequark if you say so
(DIR) Post #B8N7lGyGGtjOwaWtuK by FritzAdalis@infosec.exchange
0 likes, 0 repeats
@ariadne (It was intended as a joke, I certainly think gaslighting is real.)
(DIR) Post #B8NCHCAO4MT57ug8mm by ldcd@social.treehouse.systems
0 likes, 0 repeats
@ariadne @whitequark I'm a graduate student but i've been functionally a staff SWE for a couple years at this point; my boss has made a lot of stuff with LLMs at this point and absolutely none of it has been successfully integrated into production (and the two things that were close were basically GUI wrappers around a library i wrote).Scientists who need an LLM to prototype just aren't going to be able to effectively review that code
(DIR) Post #B8NCSc9TnSQIu7hPfM by ariadne@social.treehouse.systems
0 likes, 0 repeats
@ldcd @whitequark i think it is likely because academia is commonly under-resourced, which puts pressure on PIs, department heads, etc, to be more efficient.
(DIR) Post #B8NCcxM3osMFYlW0lk by ldcd@social.treehouse.systems
0 likes, 0 repeats
@ariadne @whitequark yes it does, but i think we're seeing an illusion of efficiency rather than an actual increaseWe get a lot of slop repos that nominally "solve" the problem but can't be pushed the last 10% of the way there to get them to be actually usefuland then me, as a graduate student, i suddenly feel like i should not share ideas i have openly because someone might throw together an LLM prototype while i'm doing the actual hard work of fabricating a prototype
(DIR) Post #B8NCjmXiPu1I3PDnA8 by ariadne@social.treehouse.systems
0 likes, 0 repeats
@ldcd @whitequark yeah, i agree with you on the illusion of efficiency, and it's sad that it is changing the culture in research labs :(
(DIR) Post #B8NHccg5CD5FCZXay0 by nmeum@chaos.social
0 likes, 0 repeats
@ariadne @whitequark with research prototypes the thing is that the outcome is not just the prototype, it's the things you learn along the way while building it. For example, you often become aware of edge cases where your approach does not work as intended, during the implementation, and iteratively refine it accordingly. The prototype implementation is just part of a larger research process and not implementing it yourself can be detrimental to that process.See also: https://www.williamjbowman.com/blog/2026/03/05/against-vibes-when-is-a-generative-model-useful/#3-artifact-vs-process
(DIR) Post #B8NHzqk74JvjU4V9Ae by ariadne@social.treehouse.systems
0 likes, 0 repeats
@nmeum @whitequark it depends. many research prototypes are created just to validate a hypothesis. i know this, because much of my programming career has been working in labs writing code for researchers to validate their hypotheses.