A "software factory" is ultimately scaffolding. If Models generalize everything, then over time, they will generalize the factory too. Do you need a factory when your prompt materializes into a working app?
@jtregunna Yeah, coding sessions typically have high task variance where you build a bunch through deep interactions with the agent. Different from data mining a large corpus..
@raffi What this tells me is “run your own representative evals, on your own workloads, with your own configuration”.
Otherwise we will learn the same lesson we learnt from micro benchmarks again.
This is not software that runs on microcontrollers or ffmpeg. There are a ton of software that can benefit from wholesale re-writes.
There certainly is a use case for same code, 6x slower, but safe.
Perhaps what’s missed in the Fil-C vs Rust debate is that tonnes of legacy software are on C/C++. Being able to ensure memory safety for them through Fil-C is very valuable.
Legacy software is software with features for which rationale is lost to time, and the difference between intent and bug is unclear.
In such situations, a 6x slowdown for running the same exact thing safely matters a lot. Focused bug fixes to address safety is easier.
Today, we launched our second-generation driverless fleet.
Without a person behind the wheel, these trucks are ready to help us serve growing customer demand.
Built for commercial deployment with next-generation hardware. Learn more: https://t.co/jZtlLOA6de
@penberg@kulkalkul the Rust argument has always been that it's enforced at compile time. And that other "swiss cheese" approaches are more error prone. Fil-C's claim is that it can do even more; and eliminate the unsafe {} blocks as well. So in some ways it out rusts rust.
The roadmap to support this looks really good too https://t.co/NiB8LYJlV9 makes MCP stateless and adds cleaner discovery.https://t.co/MlBCKKyzoM extends it beyond tools into skills.
Doing that gives you an operational layer that you set policies and secure once. You retain real diversity in the cognitive layer (claude, codex, your bespoke agent.. whatever). This scales up to agent identity and other concerns really well.