NPR on GPTChat

Absolutely everybody’s doing it.  We’ve all got to try out GPTChat. 

By now it’s abundantly clear that GPTChat produces undergraduate-level plausible BS.  This is newsworthy partly because so much of out world seems to be operating at this level of competence, which is scary (especially if you make your living producing plausible BS.)

But when you ask for something more, it becomes clear that the bot has no clue. 

So no one is surprised by the NPR headling, “We asked the new AI to do some simple rocket science. It crashed and burned,” [1].  Shockingly enough, general-purpose learning from texts and drawings isn’t enough to actually build a moon rocket.

As usual, the output looks “correct” initially, but closer examination shows that it is, to use the technical term, “wrong”.  In fact, it is basically gibberish.

The reporters also asked bots to generate images of how to build and fly a rocket.  The results are kind of interesting.  Even to a non-specialist, they look wrong.  Close examination shows they are, again, nonsense.

“When asked to provide a blueprint of a rocket engine, they produced complex-looking schematics that vaguely resemble rocket motors but lack things like openings for the hot gasses to come out of.”

([1])

To me, these images look a lot like illustrations in old, pre WWII popular and pulp media.  Lot’s of sexy looking, rocket-y imagery, but no evidence that the artist has any knowledge of real rockets.  In the old literature, the artists were portraying something that had never been built; so applying plausible imagination was not only OK, it was required.  But these days, we have actual rockets so there is no excuse for such mucked up imagery.  It’s just wrong.

Which basically shows that science and engineering is not just plausible BS, dressed up in impenetrable jargon. There actually is some knowledge there.  And, I have to tell you, some answers are right and some answers are wrong. 

---

1. Geoff Brumfield, We asked the new AI to do some simple rocket science. It crashed and burned, in NPR News, February 2, 2023. https://www.npr.org/2023/02/02/1152481564/we-asked-the-new-ai-to-do-some-simple-rocket-science-it-crashed-and-burned