ChatGPT - the myth, the feast, the opportunity
If you’ve missed the hubbub over the last couple of weeks on ChatGPT where have you been hiding? Now that the hype is starting to die down, I thought it would be good to look at where there could be opportunities for ChatGPT usage. In order to do this, we first need to explore what this technology is not.
The myth: ChatGPT has been lauded, with comments such as, "this is THE chatbot!", "I now don't need Google!" and "I can't believe other tech providers haven't released this before!". I will just refer to the man managing the team who created ChatGPT, Open AI CEO (Sam Altman).
ChatGPT is incredibly limited, but good enough at some things to create a misleading impression of greatness.
— Sam Altman (@sama) December 11, 2022
it's a mistake to be relying on it for anything important right now. it’s a preview of progress; we have lots of work to do on robustness and truthfulness.
It’s good enough to mislead you into an impression of greatness for sure. We shouldn't however, rely on it for critical aspects of business. ChatGPT won’t replace Google. Even if it could, there is a key difference between being presented with a set of results based on algorithms assessing the truthfulness of the source(s) and making open decisions on my personal critical thinking from just being provided ONE answer and taking this as the absolute truth. We also need to remember that most of the information fed into ChatGPT, is only a few months old, so it won't give you the latest result of the World Cup. Finally, it can be tricked to say very inaccurate things, as shown in a very simple example by Andrew NG.
The feast: All that said, it is an impressive technology. The feeling of intelligence is impressive (even so, it is nowhere near human intelligence). Its knowledge however, when correct, is the best we have seen to date. Its understanding of context is the most remarkable I have come across. So can I trick it, ‘yes!’. Will it get it wrong from time to time, ‘yes!’ but so are my friends during human to human conversation. I believe we will see a lot of evolution and improvement from where ChatGPT is right now, but it definitely helps us to imagine the art of the possible. So, I do congratulate the team behind it.
The opportunity: I have been asked a lot over the last few days, about what we can and can't do with it. We have to bear in mind its limitations; The data is not up to date and can't be updated easily. Output cannot be controlled, and domains can't be tailored to a business need. Because of this, it makes it impossible for businesses to even consider this as their chatbot. We discussed the case of general information retrieval (aka Google) but we have seen the pitfalls of this use case too. Can this be used in a human augmentation scenario, especially when creativity is required? Technically, ‘yes!’, as long as the output is used as a draft for a human, to use, correct and improve upon.
In my world of AI NLP training, it could be used for experimenting with augmented training data (QBox.ai does this already with GPT3/Microsoft Turing and our own technology). All this can give you good results, but here is the trick - GPT3/ChatGPT and all LLM (Large Language Model) technology have created an impression of greatness as Sam mentioned earlier. The reality is it is only possible with quality input from the users of these technologies. Over the last few months, you might have heard of a new skillset called ‘Prompt engineering’. This is how you formulate your questions to LLM systems to get a good answer. This is today, the key ingredient is to use part of these technologies in a human-augmented capacity for the enterprise. So, my advice is, go explore, find your use case, think about the human in the loop, and work on your prompt engineering skills.
ChatGPT is sometimes amazing, and sometimes hilariously wrong. Here, it explains to me why an abacus is faster than a GPU. 😀
— Andrew Ng on LinkedIn
