Why Chat GPT is not the next Google
Looking into social media and even newspapers recently it was impossible to avoid the buzzword “Chat GPT”. Being an AI practioner myself I was very curious to try it out myself. I have to admit that I am amazed by its capabilities, namely:
- understanding the context of my prompt
- being able to refer back to questions and responses of the past
- how natural the resonses are written that it feels like you are talking to a human
I quickly realized that the service is giving quite acurate results if you do ask it easy questions, like simple facts to well known persons, places, etc.. But it failed way to often for information about niche topics and places, like DIY projects or my little hometown. Asking it for activities in my hometwon it came up with great ideas and places that sounded like a lot of fun, but unfortunatly they do not exist …
The biggest problems I faced with Chat GPT
- it is way to confident in its response, even when it is incorrect.
- It cannot provide any sources to its responses for factual questions, because the whole concept of Chat GPT is not designed to do that
- it is easy to trick:
A colleque of mine asked it the follwing
That looked like a good response … but what if we ask the following as a follow up question
I am aware that this kind of information is accessible via a Google Search as well, but you need to dig a lo deeper into the results to find this information and from my point of view a new way of searching should also incooperate new safety mechanisms that would block information like this.
Chat GPT is not a search engine …
One of the most hyped field of application for Chat GPT is being the replacement of the Google search engine. I do not know who on earth came up with that, but I am pretty sure nobody at OpenAI, nor at Google thinking about that for the following reasons:
- Chat GPT is a trained language model that has no connection to the internet → it produces results purely based on the training data
- it would need to be retrained on a daily basis which is not possible due to its massive size
- due to its massive size it is computationally too heavy to run it simultaneously for millions of users → it was unavailable during the beta phase a couple of times although it just had ~1 million users NOT at the same time.
- the output of Chat GPT varies too much based on the prompt → if several people search the same question with a different sintax they should get the same results (similar to a search engine)
- a search engine should never output just a single answer, because every person should make up their own mind what of the presented results they feel are good or not → we do not want to have an AI deciding what the single correct answer to a question is (even if it is factual question, than we can be happy to see that the Eiffel tower is 300m of height multiple times)
- how should OpenAI serve this model in a similar fashion like Google does with its search engine … for free? Google earns the money to run all its services for free by placing advertisement and giving higher search volume to paying websites → Chat GPT cannot do this
How could (Chat) GPT be used similar to Google Search?
If people are so keen on replacing Google search with a chat bot, a lot of stuff need to change:
- chat gpt need to be linked to a database that is daily updated to deliver up to date information → a (chat) gpt model could be used to understand the search prompt better than a classic google search entry. The prompt representation can be used as a query for the database
- chat gpt need to be able to output multiple answers to the same questions by default
- chat gpt need to deliver sources (and the link) to its responses
- a much smaller version of the model need to exist that is capable of 99% of what Chat GPT does at the moment to serve it to millions of users at the same time in real time
Do we need a new way of information retrival?
But, do we really want this? Are we so lazy, that we feel the need for an even faster way of retrieving information? Personally, I prefer the way it is done right now:
- enter a search string
- read through multiple responses
- make up my mind what I personally think and feel is the correct answer
How about you?
