ChatGPT and the rise of the machines
Maybe they're not quite ready to take over the world just yet...
I don’t know if you have heard about ChatGPT, but it is a language model trained on a lot of input data that does quite a good job at answering things like a human would. You can have lots of fun with some of its answers. For example, you can often times get good results from asking it to write a short play… though all these plays seem to have a very neat ending where everyone ends up friends… or perhaps at least not enemies. Hoping for something like this, I asked ChatGPT for a short play that included Constantin Gurdgiev as a guest on the Vincent Browne show…
I think Constantin’s reaction sums it up…
In any case, while writing this post I noticed that I had spelled Vincent Browne’s surname incorrectly… so I thought I would ask again…
It seems like 6 weeks on they have tempered the model’s willingness to write plays… though perhaps the outline itself is more accurate.
A more interesting use of ChatGPT is asking it to write computer programs… it’s actually quite good… though perhaps we still need a human to check its work!
What most of the general population perhaps do not know is that most of the so-called Machine Learning technology is just linear regression dressed up in fancy language. Linear regression is really quite good at fooling people into thinking that you have created an artificial intelligence when all you’re really doing is finding the best straight line to go through all the data points and extending out a little. Just how good is ChatGPT at writing a program to do some linear regression…
Now this looks to be a good answer… in fact it’s almost correct… but there’s a mistake.1 If I didn’t know what the answer should look like, I could perhaps have just used this sample code directly... what is impressive is that if you know there is a mistake, ChatGPT will fix it!
I especially like its explanation of the mistake… however the mistake is still a mistake! If you are interested, I have a twitter thread where I get it to write unit tests for its code too… and it makes more mistakes there also!
To round out this post I should really mention another of the openai.com projects which aims to generate images based on text… I asked Dall-E to generate an image to use for this post about ChatGPT and the rise of the machines:
I guess we can still laugh at these bots and what they produce… as, at least for now anyway, they’re probably not ready to take over the world!
There’s actually a second mistake. The two pass algorithm used is prone to numerical overflow errors. See https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance for some details of this issue in the case of just computing the variance of a single observation.