No it’s not, at least I don’t think it can improve it’s learning during use (besides considering earlier prompts and responses) but I am actually not sure.
And just to be certain, I asked
Q: Are you able to learn from other discussions to improve continuously?
GPT: As a large language model trained by OpenAI, I am not capable of learning or improving in the same way that a human can. I am designed to generate text based on the input I receive, but I do not have the ability to learn from my interactions or incorporate new information into my responses. My knowledge and abilities are fixed based on the data I was trained on, and I am not able to browse the internet or incorporate new information into my responses.
The reason I wonder is that the two last days GPT has aced the problems. Or maybe the problems are getting easier to reason with.
Whatever the reason, GPT aced the first part of Day 4. First, we primed it with the normal prompt1. Producing a working solution on the first try after the normal prompt.
Similarly, it aced part 2. Working solutionon the first try.
Nothing else. If you are a software developer who spend most of your days writing code, maybe you should start thinking about those other career choices you ignored to become a (hopefully) well-paid programmer.
Jokes aside, it’s fascinating to see how GPT can be impressive one minute and then screw up the next thing completely.
I have personally not played very much with the earlier generations, so I don’t know whether the latest ChatGPT is better at not screwing up or not. But I wonder if large language models are coming close to what they will ever be capable of, or if this is just the beginning.
At least this is still the beginning of Advent Of Code, I hope we the next tasks will be more of a challenge for GPT as I feel I “learn” much more when it fails than when it succeeds.
Let’s end with an example of the other recent “wow it’s great, except when it isn’t” ML development, Stable Diffusion.
Hi again, we’re doing advent of code together. I need you to produce code or use any common command line tools which will solve the problems I give you. The problems will always include reading input from a file and producing an answer.