Comment by echelon
7 hours ago
Buf's offering of protobuf registries and codegen SDKs for microservices seems less necessary in the LLM era.
I'm starting to question many of protobuf's advantages (perhaps not the wire format). Add to that monorepos and other fads of the 2010s given the rise of LLMs.
I used to be a big believer in this stuff, but I'm quickly having my core assumptions change out from under me.
A big differentiator is whether one imagines an LLM in-band with most/all future software. If there is, and we’re deferring until very late parts of a program that world have been load bearing, and we’re able to programmatically ands reliably squint and say “eh i know what you meant” … then yeah formalizations seem superfluous-to-counterproductive.
OTOH if LLMs are to write, but not supplant, much of software, then boundaries, delegation to deterministic layers, good compilers to bonk miscreant models on the head with error message seem essential.
At one point it would have been shocking to assert that the compiler would live in-band with the program too. and yet JS eats the world. It seems shocking today that we could have a universal prior over the world operating in the ms/us nJ/pJ range required. And yet … ?
The real argument shouldn't be about protocols becoming obsolete, but programming languages that are "less efficient" could eventually become obsolete in favor of highly scalable and performant languages due to LLMs when the main gap becomes knowing how the tech works at a high level, and not the syntax, why code in one language over another if you don't need to worry about messing up on syntax, only about reviewing logic for sanity and correctness against business rules as well as validating that it is stable code.
Anybody who thinks that you can just chuck unstructured data into LLM and YOLO the app is an idiot.
This works up to a point, and then it doesn't. And you're left with tons of inconsistently formatted data.
My company is built on protobufs from ground up :) We use it in the database, for remote calls, on the frontend, etc. The protobuf language is not great, but it's about the right balance between too expressive and too restricting.
And the best thing is that it's compact, compared to OpenAPI.
I feel exactly that way about REST. The assumption that LLMs make schemas obsolete misses how structured outputs actually work in production. When you have probabilistic models generating code, strict contracts become more critical, not less. It is no coincidence that several major LLM platforms rely on ConnectRPC and Protobuf for their own APIs.
You're right, but the adeptness of models to spin up clients and behaviors on the fly is remarkable. They're capturing the semantics of behavior at a deeper level.
If we do strict schemas, I'd like to see less ceremony around them. Tool calls instead of brittle build steps and protocol registries.
Perhaps we need new tools for this going forward.
why? are humans going to stop using services?
Hm... Maybe. In my view, an IDL is part of the input that you absolutely want humans to author or carefully review at least. In my experience, the ceremony around generating code is also performed very well by LLMs. But I do agree, there's definitely some changes that are needed to integrate Protobufs better. Some languages have built-in tooling to make it seamless, but it's definitely not universal.