Incredible research. I wonder how close we are to outputting the universal language into it's own reasoning context (which skips encoding layers). Then using the later decoding layers to lazily inspect the reasoning context.

You can avoid the serialization in WASM by pushing structured bytes to the SharedArrayBuffer, then do serialization in JS which should be relatively cheap compared to pushing JSON strings across the boundary.

It's not an unsuitable application for WASM. They could've drastically reduced the WASM boundary impact if instead of mapping to JSON in Rust they streamed out structured bytes to JS then mapped to JSON there. And the streaming fix was language independent.

So it's more so a story about architectural mistakes.

> But I need to have everything in a mono repo for agents to properly work on in.

Why was this a problem with Deno? Up until recently you had to use package.json and npm/pnpm for it to work, but even then it was better than Bun or Node since you could use import map to avoid compiling packages for testing etc (Node and Bun's type stripping doesn't work across local monorepo dependencies, and tsx produces mangled source maps making debugging a hassle). Now Deno has built-in workspace/monorepo support in deno.json.

Focusing on needs of providers isn't a very good long term strategy if you believe compute will eventually move to self hosted and on premises solutions where large batch sizes aren't needed.

That's a foolish take.

That's like gamers thinking most of Nvidia's revenue coming from gaming GPUs, so Nvidia should prioritize gamers.

Inference is ruled by inference providers, not local. Local inference is a rounding error, and will remain as such unless there is economic incentive otherwise.

Unless they are someone that values your opinion there's nothing you can do other than move on.

Some comments here equating it to people who blindly believe things on the internet, but it's worse than that. Many previously rational people essentially getting hypnotized by LLM use and loosing touch of their rational thinking.

It's concerning to watch.

I could see something like this being useful if it could be passed to workers to replace any fs access inside the worker.

There's also a module expression proposal, that would remove the need to use blob imports.

https://github.com/tc39/proposal-module-expressions

Recursive Self Improvement

> yes I know what I'm doing and this over-cautious compiler case genuinely doesn't apply to me and there's no way to configure that so it actually lets you do your job

I'm curious why you didn't use `unsafe`?

In general people are really bad at knowing when the strict safety rules are actually being too strict, but if you're confident they are then using `unsafe` seems like a valid path to explore.