[seqfan] Re: Sequence from ChatGPT

[Alex Meiburg]

The "sycophantism" aspects of ChatGPT significantly limit its ability in
areas where it should fail, or point out an error in the user's work. This
unfortunately greatly limits its ability as a tutor: every student is
great, gets an A+, and GPT is sorry for having a different answer -- its
mistake!

GPT-4 is impressively better at long form reasoning, and a better critic
when things are wrong. Also, I'm looking forward to the plugin architecture.

Is anyone looking to build a superseeker plugin? Save it the pain of
needing to hallucinate an answer.

-Alex Meiburg


On Sun, Apr 2, 2023, 13:41 Marc LeBrun <mlb at well.com> wrote:

> >=israel at math.ubc.ca
> > Asking for an explanation or a proof is likely to return
> plausible-sounding nonsense.
>
> This has been my experience too.  This morning I tried to coax it into
> developing an elementary identity without giving it obvious search keywords
> (like "Legendre") to latch onto.  While it saw through my attempted
> camouflage and figured out that the topic of discussion was partitions,
> even with very targeted hinting (like reminding it that sets don't have
> repeated elements, that zero was not a positive integer, etc) it only
> seemed able to produce "proof-shaped" snippets of gibberish.
>
> As a friend put it, "it's like having an eager over-confident intern who
> will lie to make you happy".
>
> I don't think it's quite ready to be an OEIS contributor.  At least this
> iteration...
>
>
> --
> Seqfan Mailing list - http://list.seqfan.eu/
>