2 ms·
It's one of the things LLMs do really well. Summarization, translation, and guided blithering for generating ad copy work very well. Sometimes a bit of understa
by Animats 3y ago
It's one of the things LLMs do really well. Summarization, translation, and guided blithering for generating ad copy work very well. Sometimes a bit of understanding seems to appear. Sometimes that's illusory.
Here's a video where someone tries to use GPT-4 to read a drawing and generate G-code for a milling machine.[1] He puts in a simple textbook problem drawing, and reasonable G-code comes out. The trouble is, GPT-4 plagiarized the answer from a Chegg homework assignment which has the same design of 4 holes in a square. This is clear because the problem didn't specify the material thickness, and the GPT-4 result used the material thickness from the Chegg homework assignment.
For harder, but still easy drawings, GPT-4 does worse. It's clear there's no underlying geometric model.
[1] https://scanalyst.fourmilab.ch/t/programming-a-computer-numerical-control-machine-with-gpt-4/3860 https://scanalyst.fourmilab.ch/t/programming-a-computer-nume...