18+ years' globalization experience working in the ERP/Supply Chain/SaaS industry.Special private deployment solutions for small and medium-sized enterprises i
SpaceX has almost finished writing V1.0 of an in-house AI training stack in C that exact-maps to 220k GB300s with 800G NICs, making heavy use of pipeline parallelism and getting as close to bare metal as possible.
The potential speed improvement vs JAX for large training runs is over an order of magnitude.
Everyone talks about token costs and data as AI bottlenecks. The real constraint is human’s mental models. Software engineering industry is a great example.
Most software engineers operate from deeply ingrained paradigms about how code gets written, tested, and deployed. These paradigms worked for decades. Now they're becoming liabilities.
My observation is that only the minority of engineers have truly restructured their thinking for the AI era. They understand that their role is shifting from code implementer to engineering manager or engineering director. The majority is still trying to fit AI into old workflows. They use Claude to complement them. They review AI-generated code as if it is written by a human programmer.
Everyone is evolving of course. I evolved my view on how to think about code review in the last month alone. I am not sure I am evolving fast enough myself.
I do believe the productivity difference between paradigm-shifted engineers and traditional engineers is becoming impossible to ignore.
The engineers who figure this out will become incredibly productive. Those who don't will become increasingly irrelevant, regardless of their previous expertise.
This isn't about learning new tools - it's about reconceptualizing what programming means when machines handle implementation and humans handle intention.
#ParadigmShift #SoftwareEngineering