ChatGPT: All that glitters is not gold
Pyrite

ChatGPT: All that glitters is not gold

Perhaps, just like everyone else, you too have been impressed by how smart and amazing ChatGPT is, right?

In fact, ChatGPT seems to give really good answers, but this is not the first time that we see an "intelligent" chatbot doing it and impressing many people.

In the beginning of 2000, a chatbot used to roam the cyberspace. It was called A.L.I.C.E. (Artificial Linguistic Internet Computer Entity), a creation of Richard S. Wallace.

Compared to AI development at the time, A.L.I.C.E. seemed very smart too, but of course if you push her a little bit she will reveal the truth. She will eventually say that this information was not available in her directory.

Do you think ChatGPT is very different from A.L.I.C.E.? Better think twice! Of course, in these more than 20 years, AI has evolved a little and, most importantly, processing power has increased a lot (thanks to the number of parallel units inside each GPU). But make no mistake, ChatGPT is just spitting out text based on a huge database of text used to train it (the "entire" Internet).

Reading "The Elements of AIML Style" by Richard Wallace is a good way to understand fundamental things behind the "intelligence" of a chatboot. It all starts with the way we use words to communicate. Human texts are not random, they will fit in the "Zipf curve".

I don't want to accuse ChatGPT of not being smart without proving it, so I'll do that in a few minutes. But before that, I want to suggest a thing to "AI researchers": start focusing on how the human brain and even the primitive nervous systems of animals (i.e. Caenorhabditis elegans) works. Recently, scientists discovered that memory is stored in RNA inside neurons: https://www.scientificamerican.com/article/memory-transferred-between-snails-challenging-standard-theory-of-how-the-brain-remembers/

It means that not only the weight of each synapse of a neuron is important, but how the data is stored inside each neuron could modify the neurons response dynamically.

So, as I promised, this is a little test that an average 5-year-old kid who knows the letters V and W could answer correctly, but ChatGPT failed because it never saw this question on the Internet before:

No alt text provided for this image
Marcelo Varanda

Firmware Architect

2 年

More concerning are the models they are feeding AI: https://www.bitchute.com/video/DWzprRWPI68/

Marek Matej

Embedded Software Developer

2 年

Except the logical errors or - so to say - lack of imagination, there is also intrinsic bias (training data) which is less funny then logical stumble. Here are few examples https://u.pcloud.link/publink/show?code=kZF9QTVZSTq8XkUFY6X1XcD3ELAS75VqyEv7

Marek Matej

Embedded Software Developer

2 年

Maybe people will understand the difference between ML and AI.

Alan C.

Embedded System Engineer

2 年

Thank you Rodrigo! Yes, I like it too, this is a great tool. It needs improvements and it is biased: https://twitter.com/leighwolf/status/1620744921241251842?s=48&t=g25lQJICjdcG5idfeCwJqw

要查看或添加评论,请登录

Alan C.的更多文章

社区洞察

其他会员也浏览了