On +2026-05-05 14:58:03 -0700, Vagrant Cascadian wrote: > On 2026-05-05, pinoaffe wrote: > > Vagrant Cascadian <[email protected]> writes: > >> I think even the GNU Hurd development has started using some of these > >> technologies, > > > > Do you have any source / more specific information on that? I find this > > quite saddening to be honest >
Do you (pinoaffe) mean sad like you might be sad that conputers can now defeat the best human chess master, or what specifically is your "this" that saddens you? > https://lists.gnu.org/archive/html/bug-hurd/2026-02/msg00133.html > I read the pdf linked in that thread post [1] <https://lists.gnu.org/archive/html/bug-hurd/2026-02/pdffAOvuZusjj.pdf> It seems like a fantastically powerful (and therefore DANGEROUS!!++) tool! I sure would like to have that capability, but I would have a hard time using it without fear of what it could do to me if mis-prompted.... ... Which would probably happen with my first prompt :) (see example in PS below :) But since prompts are a form of coding in Brent's view, what if they, system owners at [2] <https://en.wikipedia.org/wiki/Anthropic> (however benevolent) have, or had, bugs in the prompts that generated their current production system? The wikipedia article [2] sounds hopefully benevolent, but money dependencies are dangerous, especially at that scale. Claude Code's capabilities are truly important to contemplate possible consequences of, given the example the WikiPedia article mentions: Chinese success in misusing it for cyberwar, by a simple ruse apparently (pretending to be researching defense to get Claude to reveal its knowlege of attacks). I guess it's a trusting trust problem to use any code generated by it beyond 25x80 ascii that is per-Tony-Hoare [R.I.P.] obviously correct ;-) As the first exception to zero-tolerance, though, IMO it would be reasonable to allow conditionally its use it to generate non-executable parts of code, like documentation headers. Though docs could be misleading and biased ;-) I might be nice to have a guile interactive ",Claude" along with ",d" -- assuming it could be done safely and snappily (or snappy promise of later result). > There did not seem to be much of any discussion on the list about the > implications, at least at the time... although I do not really follow > Hurd development, just saw it referenced via the fediverse... > Much of (valid) concern seems about preventing inclusion of plagiarized code and thereby violating author(s) rights. Perhaps the FSF could come up with a prompt that must be included in any code produced by Claude Code which would force it to include legally accepted proof of its right to publish the code, and make that a part of the code boilerplate, much like GPL or CC-by etc.? I would say Claude, in Brent's example, is WAY WAY beyond what I had seen come from Chat-GPT, which as far as LLM seemed like industrial scale Eliza[3]. Chat-GP was pretty amazing in comparison to Eliza, but not A _I_ according to my idea of intelligence :) Well, I probably only saw results of the freebie version, so probably Chat-GPT and others are near each others top priced levels. I haven't (knowingly at least) used any of them personally. [3] <https://en.wikipedia.org/wiki/ELIZA> Interesting times :) > > live well, > vagrant -- PS: sudo is powerful enough for me to get into trouble, which leads me to suggest a mitigation: My recent case: I wanted to preset the PATH set at boot time before logging in, which I did by creating /etc/profile.d/dumbbugsrc.sh This made it so I could get past decrypting the disk, and all the way to the login prompt, but after carefully typing in my user password, it would just blink and ask for a new prompt. But it apparently accepted my password, since otherwise if I typed it wrong it would complain, and it didn't. I haven't figured out where it gave up, but fixing the paths in /etc/profile.d/dumbbugsrc.sh fixed it. Mitigation: (?) If the boot sequence breaks after login password is accepted, detect this and substitute a path that will give access to the normal gnome post-login situation, or something helpful if that is also broken.
