News for hackers. Stuff that matters.
Re:OpenAI's Codex sure knows a lot about HN [video] by YeGoblynQueenne on Sunday August 15, 2021 @06:09PM (#28193002)
You asked it to do something with "the video on youtube" but what does "the video" refer to? It seems the most likely url associated with the phrase "the video on youtube" is, well, that.

So basically it failed at anaphora resolution.

Seen another way, you asked it for "the video" and so it gave you the video.

Re:OpenAI's Codex sure knows a lot about HN [video] by riveducha on Sunday August 15, 2021 @07:14PM (#28193447)
I made a similar little GPT-3 toy (for Linux/bash) that also ended up generating a Rickroll URL.[0]

In vague cases, GPT-3 functions approximately as well as a markov chain - it will give you the most “probable” sequence of tokens. There’s some implementation details regarding how it deals with tokenization, but unless you really increase the randomness (“temperature”) you’re likely to get very popular YT video IDs.

In other cases, it will “hallucinate” things like URLs, UUIDs, hashes, and other things that are basically a random string of characters. In my experience it will make UUIDs that have the right number of characters but seem suspiciously non-random and don’t fit any of the defined UUID formats.

Fun stuff.

[0] https://youtu.be/j0UnS3jHhAA

Re:OpenAI's Codex sure knows a lot about HN [video] by codetrotter on Sunday August 15, 2021 @07:21PM (#28193490)
> it will make UUIDs that have the right number of characters but seem suspiciously non-random and don’t fit any of the defined UUID formats

As predicted by Scott Adams in October 2001 ;)

https://dilbert.com/strip/2001-10-25