ossopite

I'm curious how google think about the damage this does to the perception of the quality of their AI tools. Every second google search I do reinforces the association "google AI is crap" because you can't help but see it there at the top of the page. Surely this perception hurts Gemini adoption across the board?

As I recall, C# supports this in a completely sensible way by distinguishing a[i,j] and a[i][j]. If I understand right, in C, a[i][j] means what C# would spell a[i,j], which does seem rather surprising and inconsistent

The docs/summaries part I can get behind if reviewed and improved by a human, but at least when working in pre-existing codebases, I tend to steer models away from writing comments, because I find that almost all comments they write are "not even wrong".

That is: they either reiterate what the code does, or would if the code were slightly clearer, or they tell half truths that are more confusing than helpful. Mostly they fail to emphasise the salient things, like the why over the what, that are not obvious from the code.

The idea that we could create a world where 'a big part of the future of hobbies and entertainment' is people listening to meaningless words made up by machines that help them feel good about themselves sounds horrifying. How could anybody feel ok about that? What would it say about the society we've built?

Right. Besides getting this incantation right, as gp did only after editing their comment, you also have to cast to create values of NewType. But generally you want to avoid casting in typescript if you care about type safety, so now everybody has to remember the rule that in this particular circumstance it's the right thing to do.

There are helper libraries to ease this (zod supports branded types, I think?), but I guess my general point is that while typescript might give you the ingredients you need to implement type safety in cases like this if you try really hard and remember all your rules everywhere, it doesn't come naturally so it's hard to maintain at scale.

In a way I agree with you, and I'm not sure that what popular languages embrace or make it easy to follow this philosophy. My sense is that Erlang is still the leader.

But I did want to add something the article also touches on: types can be not only about ensuring safety or correctness at runtime, but also about representing knowledge by encoding the theory of how the code is supposed to work as far as is practical, in a way that is durable as contributors come and go from a codebase.

Admittedly this can come at the cost of making it slower to experiment on or evolve the code, so you have to think about how strongly you want to enforce something to avoid the rigidity being more painful than valuable. But it's generally a win for helping someone new to a codebase understand it before they change it.

Edit: another thought I had is that type mistakes do not always causes crashes. Silent corruption can be much more insidious, e.g. from confusing types which mean something different but are the same at the primitive level (e.g. a string, number or uuid)

I'm not convinced it really works well in typescript. the lack of nominal types requires you to remember some pretty hacky incantations if you want something like a newtype wrapping a primitive type

my experience is that ocaml is more powerful than rust for enforcing this sort of type safety, because you have gadts that give you more expressive power, and polymorphic variants and object types (record row types) that give you more convenience. and the module system and functors of course.

you also avoid some abstraction limitations/difficulties that come from the rust borrow checker for places where garbage collection is just fine

Did you write it or did an LLM?

I find some irony in seeing the telltale tropes of conventional LLM writing there

I think the send/recv with a timeout example is very interesting, because in a language where futures start running immediately without being polled, I think the situation is likely to be the opposite way around. send with a timeout is probably safe (you may still send if the timeout happened, which you might be sad about, but the message isn't lost), while recv with a timeout is probably unsafe, because you might read the message out of the channel but then discard it because you selected the timeout completion instead. And the fix is similar, you want to select either the timeout or 'something is available' from the channel, and if you select the latter you can peek to get the available data.

Single core performance isn't just clock frequency. It must be multiplied by average IPC, but really it's more difficult since you have to account for factors like new SIMD instructions. Effective IPC improvements are where a significant fraction of single core speedup came from in this period

Ok, that's true, but my claim is that recursive descent parsing does not have to use the visitor pattern and indeed using recursive descent parsing is not the same as using the visitor pattern (you can do the former without the latter and I claim that you usually do)

I think I'm missing something here. if you have a grammar rule R with children A and B, and a function in your recursive descent parser that corresponds to R, why can R not call the parser functions for A and B, which return AST nodes themselves, and then construct another AST node using the result of those? Where was the visitor pattern required here?

I see that you've found an example of how recursive descent parsing actually can be implemented with the visitor pattern, which I've never come across before, and I didn't read it carefully enough to understand the motivation - but that doesn't mean they are the same thing - the recursive descent parsers I've seen before just inspect which tokens are seen and directly construct AST nodes

as an adendum, the reason I don't understand the motivation is that the visitor pattern in the way I described it is useful when you have many different operations to perform on your AST. If you have only one operation on tokens - parsing into an AST - I'm not sure why you need dynamic dispatch on a second thing, the first thing being the token type. Maybe the construction is that different operations correspond to different 'grammar rules'?