Reply to why wouldnt AI replace engineering?

Forum Index » College and University Discussion

Reply to "why wouldnt AI replace engineering?"

Subject:

Emoticons





More smilies

Text Color: Font: Close Marks

[quote=Anonymous][quote=Anonymous][quote=Anonymous][quote=Anonymous]AI doesn't even understand physics. See this 
https://www.thealgorithmicbridge.com/p/harvard-and-mit-study-ai-models-are

It gets the physics completely wrong but gets the orbits right -- but people did that too in the distant past before they figured out Newtonian mechanics (epicycles). AI (or at least the current version of deep networks) optimize for prediction and matching, not concept abstraction. I keep on top of this research for my work, and AI is too far away from this. The situation is far worse in biology.

And before someone jumps in to say that it's only a matter of time, hardly anyone (in academic research or at companies) is optimizing for this. AI firms have bet on AGI, but most of those designs are just beefed up transformers -- which are powerful but have serious limitations.[/quote]

It's only a matter of time before someone decides you're too expensive and decides to focus the AI on your job. It doesn't need to understand the full physics—you have all types of computer programs that don't "understand" the physics but are fantastic tools for design work. AI will piece them together and make them more useful. And it learns. So what it doesn't get today, will be "understood" in 6-12 months.[/quote]

Did you guys read the article? This is elementary mechanics -- if it fails miserably on this, I don't want it designing bridges (or ICs for that matter). The current computer programs that aid engineering, design etc. are all built with the underlying physics that embody decades of theory and modeling. AI can't even run a proper web query -- I am not that hopeful that it will abstract design specs into appropriate physics (or control theory or systems modeling) and "ask" the right questions of the "dumb" programs.

All the tall claims about AI weather modeling, physics inspired neural networks work well up to a point but fail miserably on edge and not-so-edge cases because of lack of abstraction. Most of the stuff at NeurIPS and other conferences is not quite there and it's not for lack of trying. The frameworks just aren't appropriate. AI/ML performance in biology is even more miserable. The designer proteins don't fold properly, don't express in cells etc etc. We are pretty far from the promised AI utopia.

[/quote]

If you can insert physics into CAD design, you can teach AI. And AI's capacity for learning is increasing extremely quickly. You're talking about current limitations—it's improved past those limitations already. You're cooked. [/quote]

Options

	Disable HTML in this message
	Disable BB Code in this message
	Disable smilies in this message

Review message

Search Recent Topics Hottest Topics