@feltanimalworld: Master, your post kept me pondering for two whole days! I originally wanted to write a long piece, but there were too many threads and I didn't know where to start. To sum it up, this is why I've been on Twitter lately—I felt something crucial was missing in my development understanding; it's also why, besides my hardware repair work, I've had to look into text-to-CAD recently. I...

Master, your post made me ponder for two full days! I wanted to write a long reply, but there are too many threads and I don’t know where to start. Basically, this is why I’ve been on X/Twitter recently — I feel like I’ve been missing a crucial piece of understanding in my own development. And it’s also why, aside from my hardware repair work, I’ve had to dive into text-to-CAD.

Why have I been on X? Because I’m in a bad mood. I talk a lot when I’m down. When I’m in a good mood, I don’t talk — I just keep working.

Many industries, even most serious scenarios where AI needs to be deployed, lack IR and Verifier. The subsequent TD and Platform Config must be built on the premise that the earlier steps have been completed. The reason AI writing code is so powerful is that the chain from IR to Platform Config is smooth.

Take a look at my video. Zoo (text-to-CAD) created the IR called KCL, which combines what you called Specs and Impl in one place. The Specs are the front part of the code. The Impl and Verifier — like some constraints — are mixed together there.

This version is very well done: following your tip, I wrote my own Specs IR. I directly imported it into Fusion, added threads, and got a fully printable part.

If we want to make any industry “as code,” we probably can’t skip this stage.

Law, medicine, CAD, circuits — all need IR.

Otherwise, AI can only provide plans, not plan verification. Without a unified IR, it’s just a language machine — it can only talk! For example, in text-to-CAD, the model can generate KCL; but without geometry checks and constraint checks, you can’t even draw the shape.

Transformer is essentially a probability predictor: P(next token | previous tokens).

Code is so successful because the coding industry has the world’s most mature IR and Verifier.

Right now, the absence of IR in various industries is the fundamental reason AI feels floating in the language universe. Take medicine, for example. Isn’t medicine one of the most standardized fields in the world?

Medicine already has FHIR, DICOM, SNOMED — but most of these standards address data exchange (interoperability), not clinical reasoning.

If we truly want AI to do medical reasoning, we’ll need over a decade to actually build the IR and unify the standards.

Master, I bow to you.

IndenScale (@david0520782123):
Many industries actually lack a programming language (specifically, they lack Spec IR and Impl IR), lack Verifiers (Linter, Compiler, Static Analysis, Dynamic Analysis), lack Target Description (TD) and Platform Config.
Without these, AI struggles to be effective, because if you give a solution, you don’t know if it’s feasible.