Re: ChatGPT

[Ken Burtch <koburtch@gmail.com>]

On Saturday, April 1, 2023 at 3:39:51 AM UTC-4, Dmitry A. Kazakov wrote:
> On 2023-03-31 23:44, Anatoly Chernyshev wrote: 
> > Data science people swear it's just a matter of the size of training set used...
> They lie. In machine learning overtraining is as much a problem as 
> undertraining. The simplest example from mathematics is polynomial 
> interpolation becoming unstable with higher orders. 
> 
> And this does not even touch contradictory samples requiring retraining 
> or time constrained samples etc.
> > I did also a few tests on some simple chemistry problems. ChatGPT looks like a bad but diligent student, who memorized the formulas, but has no clue how to use them. Specifically, units conversions (e.g. between mL, L, m3) is completely off-limits as of now.
> One must remember that ChatGPT is nothing but ELIZA on steroids. 
> 
> https://en.wikipedia.org/wiki/ELIZA
> -- 
> Regards, 
> Dmitry A. Kazakov 
> http://www.dmitry-kazakov.de

For what it's worth on the subject of the chatbot, " "Produce Ada code for solving a quadratic equation."  is a terrible choice for a test of ChatGPT as one is asking if it can do a Google search.  To test its abilities, you have to pick a challenge that cannot be solved with a Google search.

My short assessment of ChatGPT, with the history of chatbots, are available on my February blog post.  I gave it a simple programming problem and it failed 3 times out of 4.  It's not surprising as I've learned since February that the chatbot doesn't actually understand programming: it uses examples off the Internet and tries to predict what you might have typed based on keyword patterns.  It is an imitation of an imitation, smoke and mirrors.  This is why Vint Cerf denounced it.  You can read my thoughts on my blog:

https://www.pegasoft.ca/coder/coder_february_2023.html

Ken Burtch