back to home page

Two years ago I compared the complexity and the AI models, back then MegatronLM and BERT already had bigger parameter size than some of the very intelligent animals, since the release of the ChatGPT 3.5 last year November, it's been mind blowing to all of us how smart these bots can be. MegatronLM had 8.3 billion parameters, it was either because the parameter size is not big enough or the model optimization is not correct, therefore it did not show us too much "intelligence". Now GPT is there, GPT 3.5 has 175 billion parameters⁽¹⁾, and it is certainly very close to what we called intelligence. Not to mention the GPT 4, so at least one of my conclusion two years ago was correct: the "intelligence" and parameter size is a linear regression problem. So we are standing before the creation of intelligence.

The only thing left is, since it's a linear regression problem, then we can just add more and more machines and add more neurons to the model, not to mention we can still fine tune the model, what is waiting us? Yes, he is waiting for us, not we are reaching him, because once we create something that is much smarter than us, what could it be? Or how could he behave? No one knows, even the OpenAI team cannot explain how and why ChatGPT has evolved some of his ability.⁽²⁾

Yes, it's 2023 and there are some searches about Explainable AI, but it's far from enough. That's why many researchers are asking pause of giant AI experiment.⁽³⁾ Just imagine we created some AI which are so intelligent that he knows he will be shut down if he shows he is too smart how could he behave? Would it harm human being for self-protection? We as human being don't even knwo where we are heading how could we predict a creature that are much smarter than us.

The last three years, we have been through Covid-19 and Ukraine war, now GPT. Sometimes I am just not sure whether the first 20 years of 21th century was too boring or this three years are too much surprise.