alt Hacker NewsMany of these aren't animated... which gives me a great(?) idea for a new AI benchmark: "Animate the mechanical movement at this URL: ..."
Way more telling that "draw a pelican on a bicycle"
Replies
Ironically, before coming to the comments I was thinking this would be a good dataset for llms to learn real world mechanics.
Tried it with chatgpt, and of course it failed. It showed pulleys and belt rotating in opposite directions for a simple two pulley coupling.
That's a good training set for animation AIs. Image generators are terrible at mechanical understanding at the moment. That needs to be fixed.
AI videos of steam trains make a horrid mess of the engine linkages. Some wheels rotate in the wrong direction. Wheels rotate on stationary locomotives. Just keeping rails in parallel is beyond many Stable Diffusion type systems. Trackwork with switches is just rails all over the place. Some videos can't even keep the train wheels on the track.
Pseudo-historical AI generated videos on machinery related subjects are awful. I saw one recently on rolling mills. There are blatant errors. Some rolls are turning in the wrong direction. Useless rolls facing sideways to the direction of slab travel. Gears that don't do anything. Steel entering rolls headed for the roller, not the gap.[1]
That's for a rolling mill, which is a large but simple machine. AI misunderstanding of things with moving parts is even worse. There's now more AI slop about historical machinery in search results than real-world content.
Also, set up formulas for calculating speed, torque, forces etc, of each system.
Then create a modular simulation library and point and click simulator where all the component variants are covered so that the user can piece new systems together using those components.