|
|
|
|
|
|
by wruza
642 days ago
|
|
|
AI genuinely replaces software engineers. If that's true, there's no reason to believe it wouldn't soon disrupt other engineering domains as well. If that’s true, then… How? Most engineering is physical and blueprint work which cannot be converted into embeddings and filled in the middle. Or if you’re assuming that something LLM/VLM/SD-like will appear for this domain, that’s a pretty wild guess. I even doubt that transformers are more suitable for that than now-classic boring ML. If you hope that the language will somehow “emerge itself” into building, schematics, industrial, etc behaviors, that’s also quite a guess. Make sure you’re not coping, because transformers generalize over (fuzzy) painting and text, the only data we have in abundance. These are common attributes of programmers, writers and graphics artists and barely anyone else, and the latter two are already screwed. We don’t have readily accessible logs for abstract thoughts, professional conversations at work or inner monologues. Maybe in ten years China will realize it and start to collect it at scale. |
|
|
To be clear, I happen to be of the opinion that AI/ML/LLMs/whatever cannot truly replace software engineers without the true advent of AGI.
But to play devil's advocate. I'm an embedded SWE and work with circuit designers/have dabbled in some PCB design myself. Taking circuit design as an example:
From an abstract logic standpoint, PCB design is not that different from software development. It essentially boils down to taking some input and feeding it through a cascade of various discrete transformations until you have some desired output. The fact that PCB designs are captured through schematics is irrelevant, as a schematic is just a visual representation of a netlist. There are even DSLs that allow for the design of circuits through code alone. The mechanisms that make LLMs work show some level of adaptability to domains outside of traditional language. It is totally conceivable today to finetune an LLM on netlists (or even on images of schematics with the right encoder model) and have it be able to generate circuit designs. The training data exists, though it's not as plentiful as say code or english texts. I'm not an ML expert but I believe it's totally possible to have something where, for example, the text "555 timer-based LED-blinking circuit", could map in embedding space to a netlist that describes said circuit. There are in fact companies working on this exact thing today (Flux.ai)
I don't see why this wouldn't work for other engineering disciplines aswell. There's lots of public scad code for various 3d models. It's totally conceivable to build a model trained on scad or even raw 3d object file geometry and text embeddings to generate mechanical designs. Lot's of research is going into generative meshes which could be seen as an early form of this.
I think my point being that the hard part of engineering, discipline aside, is being able to take a fuzzy set of desires from a customer and a fuzzy set of constraints and being able to use the hard math and science to create and iteratively refine a product/system/whatever to make all stakeholders happy. LLMs can't really emulate this process today.