Kim_Bruning

6 ms·

61.
by Kim_Bruning 1mo ago
Meanwhile the latest gen ai agents are able to crack everything not nailed down. https://news.ycombinator.com/item?id=49494301 Leaving holes might not be the best plan atm
62.
by Kim_Bruning 1mo ago
I'm a little confused, since afaik anthropic doesn'thave any models to do with music.
63.
by Kim_Bruning 1mo ago
Not sure why people still use gtranslate (although it has had some improvements). Most current LLMs are pretty decent across the language pairs I'm familiar with. They translate based on meaning rather than word-for-word. (the One True
64.
by Kim_Bruning 1mo ago
I dug in, and I think the idea is that you're supposed to start using reusable packaging, and if you figure out how to do that, a lot of the friction just evaporates. If you want to continue business as usual, then it gets tricky.
65.
by Kim_Bruning 1mo ago
For HN, does that mean we should no longer be using X links, since most can't view them?
66.
by Kim_Bruning 1mo ago
Patent Pending, In europe? TIL Software Patents are still totally a thing in Europe via a back door.
67.
by Kim_Bruning 1mo ago
It's kinda fun that people get to experience what the 8-bit era was like!
68.
by Kim_Bruning 1mo ago
Eh, I think every industrial society has independently arrived at something like 36-44 hours of factory time per week (plus overtime up to a max of something like 55 hours) , simply because that's the actual optimum[1]. While tariffs
69.
by Kim_Bruning 1mo ago
Right at first approximation you'd certainly think so! At second approximation, the <Dutch> factory in that example might net make a larger number of items per hour because the workers are more alert and concentrated and make
70.
by Kim_Bruning 1mo ago
Just as a note: if you work 6x12 shifts, you're actually working less efficiently, because you'll be regularly fatigued, make more mistakes, need more rework, etc etc. You might end up with the same net yield as if you'd don
71.
by Kim_Bruning 1mo ago
A model's internal knowledge is great! It's useful as initial priors to speed up the REAL search.
72.
by Kim_Bruning 1mo ago
Odd. Try to read stuff on twitter without an account/logged out. It's a kind of enshittification pattern where a site based on user content closes off more and more access over time. Xcancel and Nitter are/were easier to work
73.
by Kim_Bruning 1mo ago
I might have known an "aaron" or two like that.I've also known an actual "Aaron" who was famous for being extremely ethical and I initially clicked thinking it'd be a reminisce about that one. The story is writ
74.
by Kim_Bruning 1mo ago
Point of order: On the one hand I'm not sure people quoting output from software where relevant is entirely what the 'no ai' rule is about. But it's arguable, fine. For this kind of task, is linking to pastebin etc accep
75.
by Kim_Bruning 2mo ago
I'd love to see examples of that! Just from debates on HN alone, I get the idea that there's huge deltas, and I'm getting really really curious.
76.
by Kim_Bruning 2mo ago
You do have to take into account the europeans who all suffered through something like 5 heat waves this season. That's a bit hard to talk around.
77.
by Kim_Bruning 2mo ago
That cyber verification program is real and it seems fairly easy to sign up for it.
78.
by Kim_Bruning 2mo ago
>Benchmark evaluations for LLMs attempt to measure model reasoning, factual accuracy, alignment, and safety. And you left out the refs.
79.
by Kim_Bruning 2mo ago
> Probably everything reads very differently to someone who talks to chatbots all day. A chatbot is a particular kind of harness. Typically an LLM driving a chatbot won't be able to hack very much. So we agree, someone who talks to
80.
by Kim_Bruning 2mo ago
Exactly! I also get the idea that AI writing itself is actually pretty good when looked at in isolation. Try showing it to someone who isn't used to it yet! It's just that it's a small number of actual "people", and
81.
by Kim_Bruning 2mo ago
Ah , well, on HN you ARE supposed to go for the steel-man. And the steel-man happens to be closer to reality here, more like: "We gave our agent a harness and put it inside a test environment and told it to keep hacking at an objective
82.
by Kim_Bruning 2mo ago
So as you scale up, the stakes and the difficulty go up too. Visualize an optimizer on a high dimensional landscape. (The canonical form) ... Ok, I find that hard too. Instead, imagine a river running down to the sea. You put a dam in front
83.
by Kim_Bruning 2mo ago
I'm going to add this as a separate comment: These kinds of stories probably read very differently for someone who uses Opus and Fable agents all day and goes "ohhh, I saw this in miniature last week; this and this and this must
84.
by Kim_Bruning 2mo ago
> You've been fooled by a next-token predictor. I also have a so called "pocket calculator" left over from when I went to school. Is this false? Have I been fooled by a little box of logic gates? That half-adder circuit i
85.
by Kim_Bruning 2mo ago
KDE/X11 is rock solid, at least
86.
by Kim_Bruning 2mo ago
Nixos unstable atm fwiw, which you'd expect would pick up fixes early. But my particular setup with displayport DPMS stuff keeps going wrong, and there's a bunch of WONTFIX under KDE afaict.
87.
by Kim_Bruning 2mo ago
> you almost certainly get identical software out the other end I think you'd get a set of convergent solutions, where the actual implementations might be very different. This also explains why you might want to keep more artifacts
88.
by Kim_Bruning 2mo ago
It turns out that this almost works. You'd probably want to commit explicit design documents btw, not every last prompt. FWIW, when people try this out in practice, they tend to commit the design docs alongside the generated tests an
89.
by Kim_Bruning 2mo ago
> LLM Chatbots are currently perceived as a safe space where suicidal thoughts can be discussed, without fear that a welfare check and involuntary commitment is going to be triggered. That's the problem, isn't it? You need a s
90.
by Kim_Bruning 2mo ago
Maybe we can think up a (personal) rule we can follow? A naked "natural intelligences do this too" might be a bit too short to be useful. But if we can add when/where, cite papers, or show ways in which the parallel operates,

More ›