Given that pre-paid plans are the most popular way to subscribe to Claude, it quite plainly is a "the less tokens you use, the more money Anthropic makes" kind of situation.
In an environment where providers are almost entirely interchangeable and tiniest of perceived edges (because there's still no benchmark unambiguously judging which model is "better") make or break user retention, I just don't see how it's not ludicrous on its face that any LLM provider would be incentivized to give unreliable answers at some high-enough probability.
I feel like this has to be a toolchain issue, there's no reason the pin number -> register table couldn't be resolved at compile time, similar with conditionally compiling certain things based on the CPU features.
I'm not saying it's not a real or an easy problem, just that I wonder if it truly is the reason Arduino is "bad"
I actually have no idea what you mean with the example, all the toolbars on the page fit 4 or more buttons, I tried viewing it in various window widths, can you be a bit more specific?
As for the article, I'm also a bit confused because I'm really not sure whether people write that sort of code at the beginning "very commonly" - match and `ok_or` to handle None by turning them into proper Errors is one of the first things you learn in Rust.
Somewhat tangential, but there are much better options if you're looking for opportunities for optimization. It's literally trying to improve efficiency by skimping on safety features, like trying to save on vehicle weight by removing unnecessary seatbelts or crumple zones. Eliminating side channels concincingly is very difficult, you're just better off taking the tiny performance hit and virtually* eliminating that vector instead of trying to come up with a novel low-density seatbelt.
(I say virtually, because even constant time crypto isn't bulletproof - GoFetch, a recent Apple M-series CPU vulnerability inadvertently broke the "constant" part because of a quirk of the prefetcher. Side channels are hard, no need to make it harder.)
In an environment where providers are almost entirely interchangeable and tiniest of perceived edges (because there's still no benchmark unambiguously judging which model is "better") make or break user retention, I just don't see how it's not ludicrous on its face that any LLM provider would be incentivized to give unreliable answers at some high-enough probability.