Nick Smith

I’m a software engineer & researcher, working towards the future of programming.

[Nick Smith @NickSmit_]

@nilscmr How are you planning to overcome the hurdle of needing a new, dedicated programming language to maximise utilisation of your hardware? Even if it’s Rust-like, you’re leaving its ecosystem behind

[Nick Smith @NickSmit_]

@jimkxa @intelligenz_b IIRC Blackhole also has tensor cores bolted onto RISC-V cores with SIMD. What's the difference?

[Nick Smith @NickSmit_]

@seanbax @LParreaux @tigregalis @elricmann Not quite sure what your question means, but the borrow checking works differently. You borrow-check function calls, not reference construction. Your program is correct if the arguments don't mutably alias, and if the environments of different closure types don't mutably alias.

[Nick Smith @NickSmit_]

@seanbax @LParreaux @tigregalis @elricmann Yes, if you don't know the concrete environment that the reference captures then you need to pass its type around as a type parameter. This is basically equivalent to lifetime parameters in Rust, except it's "just" a type parameter. My model eliminates the concept of "lifetimes".

[Nick Smith @NickSmit_]

@seanbax @LParreaux @tigregalis @elricmann Not easy to explain in a tweet, but I'll have it all written up soon. 🙂

[Nick Smith @NickSmit_]

@seanbax @LParreaux @tigregalis @elricmann Yeah, I've learned the hard way how difficult it is to do any better! I've ended up with a design that is like closures on steroids. A "reference" is a value that "captures" data from an external environment. The environment forms part of the reference's type.

[Nick Smith @NickSmit_]

@seanbax @LParreaux @tigregalis @elricmann It's a hard problem but I rarely ever see people even *suggesting* that it's possible to relax Rust's borrowing rules. I think I'm close to having a working design. If I remember I'll send you a draft when I've written it up.

[Nick Smith @NickSmit_]

@seanbax @LParreaux @tigregalis @elricmann Thanks for the links! I don't think it's as simple as `shared T&`, but I believe there might be a practical way to model shared mutability and pointer invalidation. I've been working (casually) on a borrowing system for Mojo over the last ~9 months that aims to achieve that.

[Nick Smith @NickSmit_]

@seanbax @LParreaux @tigregalis @elricmann Any thoughts @seanbax? I value your opinion. 🙂

[Nick Smith @NickSmit_]

@seanbax @LParreaux @tigregalis @elricmann For your definition of `func`, if your language has a way to (optionally) declare that `s` may point into the vector's buffer, and a way for the compiler to understand that `push_back` may reallocate the buffer, then your language can ensure memory safety.

[Nick Smith @NickSmit_]

@clattner_llvm @lemire By "manually", I mean exposed as a library.

[Nick Smith @NickSmit_]

@clattner_llvm @lemire Has autovectorization been replaced with autofusion (with the same "what if it doesn't optimize" concern), or is loop fusion/kernel fusion done manually?

[Nick Smith @NickSmit_]

@Avamander @pcwalton Donald Knuth's famous quote seems highly relevant to this discussion:

"Beware of bugs in the above code; I have only proved it correct, not tried it."

[Nick Smith @NickSmit_]

@Avamander @pcwalton But yes, it would be beneficial to have somebody attempt to formally prove that the design is sound at *some* point.

[Nick Smith @NickSmit_]

@pcwalton I suspect you can come up with a working system by doing iterative design, i.e. making something and then trying to break it. You don't necessarily need to prove a system's soundness prior to implementing it.

[Nick Smith @NickSmit_]

@pcwalton "Any AOT compiled language is going to be." What if Rust compiled 100x faster? The problem isn't AOT, it's slow compilers, right? Even Python is partially AOT-compiled.

[Nick Smith @NickSmit_]

@jonathoda @dubroy We don’t need everyone to believe in the mission. We just need *enough* people that we can make progress.

[Nick Smith @NickSmit_]

@jonathoda @dubroy Said another way: creating software is currently easy for 0 people on Earth. It would be wise to make it easy for >0 people before attempting to make it easy for everyone.

[Nick Smith @NickSmit_]

@jonathoda @dubroy My hypothesis has always been that we need to make it easy for “experts” to create software first. I wouldn’t expect anyone to be able to invent a *broadly-accessible* (and highly general) programming interface when we haven’t even invented an adequate professional interface yet.