> compiling your code for every combination of architecture and platform you plan to deploy to
I mean, on the one hand you are arguing for C FFI and on the other worrying about compiling for every combination of architecture. Those positions seem to be contradictory. Although I guess you're assuming that other people who write the C libraries for you did that work. I guess you better hope libraries exist for every possible performance issue you come across in your cross platform scripting library.
And why limit your runtime to AWS Lambda? That is a constraint you are placing on yourself. Nowadays with docker you can have pretty much any Linux you want as an image. But why not just implement on top of cgroups from scratch? I guess we live a world where that is unthinkable to many. Probably just better to pay AWS. But if you do use docker, all of a sudden worrying about compiling for all of those architectures seems like less of an issue. And you can use ECS, so you can still pay AWS!
As for tooling issues, and there are definitely tooling issues with every language, it is a pick your poison kind of thing. I remember really liking Pascal tooling way back in the day. Smalltalk images have some nifty features. Who doesn't like Lisp, the language that taught us all REPL. Not sure I'd choose them for a project today though.
As LLMs get better, I just assume what constitutes "developer experience" is going to change. Will I even care about how unergonomic writing test cases in Go can be if I can just say "LLM, write a test that covers X, Y, Z case". As long as I can read the resultant output and verify it meets my expectations, I don't care how many characters of code or boilerplate that will force the LLM to generate.
edit: I misread your point about Go test cases but I'll leave my mistake standing. My overall point was the stuff I find annoying to do myself I can just farm out to the LLM. If the cost of writing an experiment is "LLM, give this a try" and if it works great and if not `git checkout`, then I will be ok with something less optimal.
I mean, on the one hand you are arguing for C FFI and on the other worrying about compiling for every combination of architecture. Those positions seem to be contradictory. Although I guess you're assuming that other people who write the C libraries for you did that work. I guess you better hope libraries exist for every possible performance issue you come across in your cross platform scripting library.
And why limit your runtime to AWS Lambda? That is a constraint you are placing on yourself. Nowadays with docker you can have pretty much any Linux you want as an image. But why not just implement on top of cgroups from scratch? I guess we live a world where that is unthinkable to many. Probably just better to pay AWS. But if you do use docker, all of a sudden worrying about compiling for all of those architectures seems like less of an issue. And you can use ECS, so you can still pay AWS!
As for tooling issues, and there are definitely tooling issues with every language, it is a pick your poison kind of thing. I remember really liking Pascal tooling way back in the day. Smalltalk images have some nifty features. Who doesn't like Lisp, the language that taught us all REPL. Not sure I'd choose them for a project today though.
As LLMs get better, I just assume what constitutes "developer experience" is going to change. Will I even care about how unergonomic writing test cases in Go can be if I can just say "LLM, write a test that covers X, Y, Z case". As long as I can read the resultant output and verify it meets my expectations, I don't care how many characters of code or boilerplate that will force the LLM to generate.
edit: I misread your point about Go test cases but I'll leave my mistake standing. My overall point was the stuff I find annoying to do myself I can just farm out to the LLM. If the cost of writing an experiment is "LLM, give this a try" and if it works great and if not `git checkout`, then I will be ok with something less optimal.