Why LLMs Won’t Lead to Artificial General Intelligence

Imagine a world where machines think, understand, and learn like humans, acting autonomously. This is the vision behind Artificial General Intelligence (AGI). But can Large Language Models (LLMs) really take us there? Let’s explore this question.

My perspective on LLMs

Large Language Models, such as GPT-4, have changed how we interact with technology. They're powering everything from chatbots to automated customer service, and even writing essays. Influential figures like Ilya Sutskever and Sam Altman believe that scalability is the key to achieving AGI. But I have a different perspective.

LLMs are sophisticated algorithms trained on vast amounts of text to understand and generate language. They function as mathematical representations of language itself. During thousands of years, humans have developed language, meaning it inherits a part of human intelligence and is kind of a snapshot of that part of intelligence. Then the model is trained on the whole language and by doing that the intelligence that is inherited in language is transferred into the model. So the model is a snapshot of the intelligence in the language. Therefore the model is kind of a second derivative of real time human intelligence.

Now there are a few reasons why this is a problem.?

1. First of all because it is a derivative of the intelligence that is stored in language, there is no actual reasoning and thinking involved. There is an echo of reasoning that again is stored in language itself but that is not enough to actually reason in real time.
2. Secondly, because it is a snapshot, it can never ‘think’ or ‘be’ (sorry for being so philosophical). You see, we humans continuously get input of information that flows in and gives us a continuous synchronisation with the world.

Limitations of LLMs

LLMs are inherently limited by their architecture and the hardware they run on. They don’t understand the world, but merely recognize patterns. Unlike humans, they lack beliefs, intentions, instincts, or desires. To be truly aware of and understand their environment, systems need to be part of nature’s flow or be on the same frequency which humans subconsciously are. Digital computers are not. This is why I believe we will not achieve true AGI or ASI (Artificial Super Intelligence) with current silicon-based hardware. Real innovation needs to occur at the hardware level.

The AI we will have on digital computers will serve as a new way of processing and interacting with information. It will be an advanced and interactive search engine that will be able to perform computing operations for us and will contain most of human knowledge. It will help us solve many problems, but it will not solve them for us. It won’t be a god like Ray Kurzweil or Ben Goertzel ought to be, it will be just a very handy tool. Think about it this way: AI models are great simulations but not the real thing.?

The Complexity of Human Intelligence

Understanding human intelligence is still beyond our grasp. Assuming we can replicate it using mathematics and software is naive. I think the only way to really create human intelligence in a lab is with synthetic biology. However, there are many many issues of technical, scientific and ethical nature with that. And above all, we must ask ourselves, do we actually need and want a synthetic human level intelligence??

While LLMs represent a significant advancement in AI, they alone cannot lead us to AGI. The journey to AGI requires breakthroughs beyond scaling existing models, involving hardware innovation and a deeper understanding of intelligence itself.

So, what is your take on the journey to AGI? Can we get there with LLMs alone and is scalability enough? Share your thoughts in the comments. And don’t forget to follow for more deep dives into the world of AI, entrepreneurship and business.