3 ms·
Listing and explaining is essentially repeating what someone else has said somewhere. Labeling a schematic requires understanding what you're saying (or copy-p
by mcswell 1y ago
Listing and explaining is essentially repeating what someone else has said somewhere. Labeling a schematic requires understanding what you're saying (or copy-pasting a schematic, so I guess we can be happy that GPT-5 doesn't do that). No one who actually understood the function of the major components would mislabel the schematic like that, unless they were blind.
- jvanderbot 1y agoIts true if you expect general intelligence. Its forgiveable for expecting general intelligence given the hype. But there's no real reason we should have expected a language model to create an image that is for some reason a perfect bicycle schematic (other than hype). And I'm not sure that imagery is actually a required format to demonstrate intelligence. I bet it could generate assembly instructions and list each part and help diagnose or tune. And that's remarkable and demonstrates enough fake understanding to be useful.
- mcswell 1y ago"fake understanding" is exactly the right term. And the image is just fine, it's the labeling that's bonkers. What it illustrates is that the LLM can repeat words, but it has no idea what it's saying. Whereas any pre-1817 engineer, reading descriptions of bicycles (which GPT-5 obviously has access to), could easily have labeled a picture of one. (1817 is the date the first real bicycle is believed to have been invented, but driven by the rider's feet on the ground. Bicycles with chain drives weren't invented until decades later. But an engineer would have understood the principle.)