Chat GPT- The Good, The Bad, and The Ugly!
Since everyone is playing with the shiny new toy; it was my turn to try #chatgpt . And here is my The Good, The Bad, and The Ugly findings about it!
What is ChatGPT?
It is the latest & coolest AI tool from OpenAI built on the top of GPT3. Technically speaking is a fine-tuned version of the GPT3, so it is also called GPT3.5. What made it the best thing ever in AI is the versatility of this chatbot. It can answer any question and provide contextual explanations, and more than regurgitating what is available on the internet. It can generate new text; almost writing new essays. Lots of cool applications have been built using ChatGPT in just the last couple of weeks which shows how it can be used to solve coding questions, provide better search results than Google and write high school papers on topics without being detected by the plagiarism tool.
So I also took this new toy for a spin and here is what I found.
1. The Good
First of all, as a data scientist, when you start playing with ChatGPT you will be simply blown away. What was once shown in sci-fi movies as all knowledgeable AI is now almost a reality!
But then AI answering questions is not the newest thing. Many such services like Galactica or Microsoft had to be shut down due to concerns around ethics, inappropriate answers, and the potential for harm.
However, #chatgpt is doing an amazingly good job on various ethical questions. For conversations related to violence or racism, its responses are more responsible than just regurgitating what’s available on the internet.
2. The Bad
It is still not free from implicit bias.
For example when asked the question “who won the World Cup?” — it simply assumed it was the soccer World Cup — and that too of men. And the answer is without any context. The true response would be to ask — which World Cup? Soccer or rugby or cricket? And if men or women?
This is a harder problem to solve. Firstly such an unconscious bias exists in humans in any typical conversation; so asking a machine to be better than humans is going to be a tough ask. Also, this area falls under “Deductive Reasoning” where instead of giving an answer to a question, one should ask a follow-up question. Few humans are as good as Sherlock in such deductive reasoning capabilities and so for machine learning, this is the biggest unsolved problem.
3. The Ugly
Still lacks common sense when it comes to real-world metrics. Google is somehow better at this.
I was not able to figure out why ChatGPT was so bad at this. Is this a bug?
Overall ChatGPT is the biggest advance in conversational AI since Watson won the Jeopardy challenge back in 2012. We came really far in just 10 years!
