You’ve probably noticed the pattern. Every week, another AI lab unveils a model that can write poetry, ace a math test, or generate a video of a cat riding a unicycle. And every week, you quietly wonder: Does this actually help anyone build anything?
There’s a reason that question never gets answered. The people running the world’s most celebrated AI labs have never actually made anything in the physical world. They are world-class researchers—brilliant, driven, and utterly disconnected from the messy reality of how things are manufactured, assembled, or delivered.
They treat software as if it’s just code. But software is the result of humans compromising on a shared worldview. That’s not a line of code. That’s a political negotiation, a supply chain trade-off, a decades-old legacy system that nobody wants to touch.
Let me say it plainly: these labs are building solutions in search of a problem. They release models, then scramble to find a use case. The benchmarks they hype are nonsense—internally consistent, externally meaningless. They never stopped to ask the most basic question: What are we actually working towards?
One of the rare exceptions is Yann LeCun, who seems to have a grip on the real world. The rest? They’re trapped in a self-referential bubble where November’s benchmark becomes December’s forgotten footnote.
This disconnect isn’t just academic—it’s dangerous. When you build AI that writes software, you assume software is a technical artifact. Wrong. Software is the output of human compromise—people talking, arguing, and finally agreeing on a fragile worldview. Ignore that, and your AI will generate code that works in a vacuum but fails in the factory, the warehouse, or the hospital floor.
But here’s the twist: this leaves the field wide open. The real opportunity isn’t for the labs that can generate the most impressive demos. It’s for the builders who understand manufacturing, logistics, and the gritty physics of how things actually get made. The people who know that ‘if you don’t understand a process, you don’t know shit.’
So stop asking which AI model is best. Start asking: Does this team understand my world? If the answer is no, you’re not being served—you’re being used as a testing ground for someone else’s resume.
FAQ
Q: But aren't AI models already useful for many tasks?
A: Yes, for narrow, well-defined tasks. But the big claims about rewriting manufacturing or logistics are pure hype. The models are solutions looking for problems—they work great in benchmarks, poorly in the real world.
Q: What should builders and engineers do differently?
A: Start with the problem, not the model. Go to the factory floor, talk to the people who actually make things, and understand the constraints. Then—and only then—ask if AI can help. The labs are doing it backwards.
Q: Isn't software just code? Why does the human side matter?
A: Because every line of code represents a decision, a compromise, a trade-off between competing priorities. AI that ignores that history will generate brittle, unusable outputs. The best systems are built by people who understand the context, not just the syntax.