In extremely high performance code you use different data structures and algorithms and change your approach to memory allocation. TigerBeetle famously does all memory allocation once on startup.

Roc is attempting to make a similar set of trade-offs in their compiler as Zig, so it makes sense that the author finds many shared patterns.

C# grew all those features over time. It had to leave syntax to support old patterns to preserve backwards compatibility. Thus, the syntax has grown a bit noisy over time to support all those features. This is reboot keeping the newer ergonomics and streamlining the syntax.

I probably wouldn’t adopt it for existing projects or use .Net for any future project, but it looks really nice for what it is.

Effects allow more aspects of computation to be represented/generalized, but as the author has noted concrete implementations have already been tried in some cases. Rust’s ownership, Java’s exceptions, Typescript’s async are all instances where the language made a choice about and provided syntax for some effect.

Generalized effects are not yet proven (IMO) to improve developer ergonomics for rank-and-file devs. It may be the case that most devs need most decisions to be made for them, and the power of effects is just lost to them. That may also apply to agents. They don’t _need_ to encode the constraint. The constraint was decided when the language was picked.

Humans often think in terms of deontological ethics. Corporations operate in terms of consequentialist ethics, and the only consequence that matters is that the numbers go up.

Some time back I went on a tour of the fort at Dry Tortugas. Largest brick masonry fort in the world. Many innovations of the form. It was abandoned right as it was completed because barrel rifling had just been invented, dramatically increasing the penetrating power of artillery, rendering brick masonry forts obsolete virtually overnight.

This project evokes a similar feeling.

I admit I’m not particularly well versed in this space, but I’ve never come across that formulation, only the “pure” one the original researcher says is internally inconsistent. I don’t necessarily buy “it’s close enough it doesn’t matter” given what I perceive as many notable exceptions.

Either or both of those being true is almost as mind-boggling to me. How is one supposed to navigate that world?

Mass delusion seems like the most likely answer to me. I can point to instances of that at similar scales outside of tech/investing.

It also seems impossible. What are people seeing that I don’t?

But I also grew with a different internet than we have now. There’s a level of targeted manipulation that’s novel. I’m not sure the cat goes back in the bag no matter what we do.

Any particular model almost doesn’t matter at this point. Harnesses are built around them. OpenAI and Anthropic are basically interchangeable in an open-source harness like OpenCode; the switching cost is virtually zero. Local models are improving rapidly and are already “good enough” for many use cases already. The bet that LLMs will continue to exist as an algorithm is pretty solid.

Don’t do N+2. The goal isn’t to predict the future, nobody can do that. The goal is a durable understanding of the domain and the best fit implementation you can get with that current understanding and resources.

That said, SQLite passes that bar for me in most use cases.

I was recently consulting at org where two separate engineering teams were all in on two different, incompatible deployment platforms and using AI to accelerate adoption of each.

Management was mystified why their engineering leads kept telling them they couldn’t deploy a complete implementation of their solution.

Now imagine you had the tools to do a diff across all those repos simultaneously, and classify, group, and review those patterns. What could you learn NPM teams and practices?

Now imagine you could pick best of breed, and propagate those back to all the other projects automatically to improve their productivity, security, etc. How fast would your productivity improve and your engineering culture change if everyone could automatically learn the best of what everyone else had to offer?

Companies like Spotify have sophisticated tooling for detecting repo changes and enforcing policy like that, and they run that experiment 1,000 times a day. Small evolutions in what was an identical build script, like a version bump, are detected, and if it passes a threshold it can be rolled out everywhere else immediately.

Having all the copies that you can sync up centrally periodically puts natural selection to work on internal best practices.

Basically, things work differently at scale. When the number developers you employ approaches a meaningful percentage of the total number of developers globally, your internal diversity starts to mirror the global diversity. So you have to manage that diversity. If you freeze policy entirely, you fall behind the global average. If you let things run wild, your company fractures technologically.

So, make a 1,000 copies, see what pops up, adopt and enforce things that look good, then do it again. Evolve to the next best place you can be from where you are.