Wouldn't all binaries be in the training data, rather than the context? And output context could be in pieces, with something concatenating the pieces into a working binary?
Is it feasible that if we have a tokeniser that works on ELF (or PE/COFF) binaries, then we could have LLMs trained on existing binaries and have them generate binary code directly, skipping the need for programming languages?
Hashicorp has go-plugin: https://github.com/hashicorp/go-plugin. It does similar with support for net/rpc and grpc. With grpc, you could have the external process in Python. Unix domain sockets and TCP are supported. The framework handles spawning the process and managing it.
It is used extensively within hashicorp's products - nomad, packer etc.
The device maker controlling an app store made no sense always.
Its like saying the browser maker controls what websites you can visit.
We have so many efforts at keeping the web open, shouldn't we apply that to all platforms?
Good point on having developers in loop in a software business. Tesla might be a counter example, where it is often touted to be run like a "tech" company. Are developers in loop in the design of non-software stuff? (is chassis design an example? or maybe not?)
The other point on triangular communication with middle managers I think is known problem that Musk has publicly spoken about..
ChatGPT claims its possible, but not allowed due to OpenAI safety rules: https://chatgpt.com/share/68fb0a76-6bf8-800c-82f7-605ff9ca22...