On this thread and similar, I'm noticing that some strong opinions about $LLM_PROVIDER are coming from accounts without much post history. With so much on the line, and the way that HN can influence developer behavior, I wonder what ways we can responsibly consume opinions in a thread like this.
Not to cast too much criticism. HN is extremely well-moderated (thanks team!). But think we-developers need to be very wary.
I asked it what the cheapest train fare would be for my partner to get somewhere and it hallucinated the two together railcard rules to the point it would have got us a fine. That said, British train fares are arguably more convoluted than even the most complex software application.
I think the community on this site these days, much like other comment sections on the web, just read the headline and make a low effort comment. Regression to the mean I guess.
As an update to myself, the comments did eventually sort themselves out. I guess the initial "reaction" commenters and voters are just more interested in participating than in SNR. Good opportunity for me to finally start blocklisting users, and I'll probably block some of these large, reactive thread authors.
In the past I wrote a version of HN that uses modern CSS rather than tables that populates stuff from the API. There I built a little blocklist of my own that prunes a comment tree the moment it encounters a blocked user (inspired by posts from another HN user, arjie.)
I've been thinking of making a purely algorithmic filter for myself but at that point I might just ditch the fake HN interface and make something. I've been thinking of building atop Mastodon/ActivityPub clients.
Data point: Albertsons/Safeway/etc. is rolling out new card readers that have a camera in them. Software support likely isn't in place yet, but that's definitely something they are thinking about long term.
Counterpoint - the benefits of wireless are there but the worry of your unit dying in the middle of a ride now replaces the concern of whether your derailleur is tuned and ready to go. It's easy for a shop to assemble, but now I'm worried about shorting the control unit of my di2 which would be a pricey fix. I have two bikes: one with and one without di2. Both work just as well and one costs much more.
Honestly, routing the cables through the headset introduces so many tight bends, I suspect you'll get better long-term reliability with wireless. Plus, there are no connections to make waterproof aside from the battery.
You will need to take it apart any time you need to change the tube... "no connections" is a fantasy.
I've never seen anything wireless being more reliable than a mechanical analog. Anecdotally, I moved into a house that had a bunch of this wireless garbage: wireless thermostat, wireless doorbell, wireless light over the backyard gate. All this garbage is dead and dysfunctional in just a little over two years.
I agree with you that all this electronic stuff is doomed for the dumpster, but I suspect that in this highly specific situation, it prevents wires from getting chafed through and failing. And let's be serious: for the time being the people buying wireless electronic groupsets are replacing them every couple years when something sexier comes out. If it makes it three-five years, it'll be long enough.
We agree that there will be no present day equivalent of all the beautiful old Campy Record in the used bike shops of 30 years from now. All this shit will be long since busted.
Tube is the rubber "doughnut" that goes inside the tire. The part that you usually need to replace if you run over a sharp object and puncture a hole. It's probably called different things around the world.
Flipped through the language overview on the website and noticed that matrices are limited in how large they can be because they're stack allocated [1]. Ergonomics of the language otherwise look solid but for my use case that would be fairly constraining
Actually we implement it manually ourselves the exact same way that the extensions work. This is because we have to support multiple different versions of LLVM which don't have those extensions.
Matrices, like any other primitive data type, are stack-allocated. Do you want matrices to be allocated on the heap? If you're looking to do Pandas-style data science, you'd have to write your own implementation, where you could adjust for exactly how you want to do the multiplication. The builtin matrices are typically pretty small (~4x4 or 3x3), a very common use case in graphics or games programming.
