GenAI isn't intelligent, it's auto complete on steroids

I have long been skeptical of the hype surrounding LLMs and Generative AI, specifically the claim that it will lead to AGI (Artificial General Intelligence). A term that has become so overloaded as to have lost its meaning. I personally define AGI as “human equivalent intelligence”, I am expecting sentience, agency, consciousness and more.

The recent paper by Apple, explored by this Guardian article, lends credence to my skepticism and after reading the article I am now more strongly confident that GenAI is not the path towards AGI. GenAI was already missing agency, something that humans have from the moment they are born, the inherent ability to pursue actions independently of other agents. Like for example, being inspired to write this blog post and doing it. GenAI lacks this, it is almost like a virus, a non-living organism which can only replicate inside other living cells. In the same way, GenAI can only respond to us if it is prompted. The inability of GenAI to solve simple puzzles, never mind novel ones, is another sign that it is not the path to AGI.

GenAI is effectively just an auto complete on steroids. The absolutely insane amount of data it is trained on enables it to recall and generate information in an impressive manner. But that doesn’t change the fact that it is just auto completing. It is not thinking or making intelligent decisions. It is only capable of novel thought in so much that randomly generated lottery numbers are a novel thought.

This all is not to say that GenAI isn’t a useful tool. Certainly for generating code it is incredibly impressive and makes software development far more productive. But at the same time, it isn’t without major problems. For one, the data it was trained on has likely been stolen and plagiarised.

I am mostly writing this as reference for myself in the future and as a playful bet against myself. My bet is that GenAI will not lead us to AGI. I would also make a further bet and say that it has already hit a wall, and that the findings of this Apple paper will not be overcome by training models with ever more data.

Let’s see where we stand in 2030.