TRACTOR - C to Rust AI Compiler By DARPA

Do not hallucinate.

You have to assume that whatever gets adopted by the government will never be upgraded or replaced, unless you're fine with a great deal of pain and expenses. Which is why ANYTHING to do with AI in this decade should NOT be touched by any government on Earth.

LowLevelAgen

The DoD has a history of attempts of translating code from language A to language B. All of these attempts failed, as a shocker to nobody. In the mix were C, Ada or Java in the early 2000s. At some point they realized they cant do without a low level systems programming language like C++, so they employed the man himself Bjarne to write some safe code guidelines for what is now known as the Joint Strike Fighter program. Every experienced devs will tell you, code is only as good as it is maintained. And that is the core issue of these humongous DoD/government code bases: they get shit out by LockMart and Co but years down the line its extremely challenging to maintain it. In a safety context, where even the compiler needs to be certified to create the correct output, you cannot just feed it with a statistical error creating machine that is a LLM and call it a day.

bare Metal to Rust? There's a joke in there somewhere.

imagine hating C so much you’d prefer a dream-powered transpiler to Rust

Ha! Twenty years ago I was hired for a 9 months contract to rewrite some thousands lines of bare metal x86 assembly language in C and adapt it to run under a real-time operating system. It worked very well and is still in service today. However still today people occasionally ask me questions about how it works or why it works as it does. I have no idea, never did have. There were no requirement docs or tests for the original assembler version. There were almost no comments in there either. As far as I know nobody ever dared working on that new shiny C code.

I agree with the general pessimism on the outcome but I don't think this is an instance of government misunderstanding the technology. DARPA's mission is to invest in moonshot projects that would be bolster national security if successful but would struggle to get funding elsewhere (usually because they're too risky). This is exactly the type of project they should be investing in.

The known universe is based near entirely on C ABI boundaries

“70% of exploits are memory exploits” Social engineers: 🙂