As AI systems get more powerful, the idea of existential threats from sentient algorithms seems to become more real. This emotion reverberates strongly in the recent open letter calling for a six-month pause on AI research for systems more powerful than GPT-4. If the researchers are freaking out about rogue AI and AGI risk, it means the layman must be shaking in his boots, right? 

Well, it seems that the risk associated with releasing AI systems is far less than it is perceived to be. Even if prominent research houses publish stories that over 300 million jobs will be replaced by AI systems, the truth, as always, is much more nuanced. 

However, the key takeaway is clear — AI won’t go rogue. 

It’s just a software

Recently, an article proposed the idea that the game ‘Elden Ring’ would soon become sentient and an existential risk to humanity. While this proposal seems outlandish at first, it is just a metaphor to illustrate an equally impossible scenario — that of modern-day AI taking on the role of Skynet. 

Currently, what we call AI is just another piece of software, with the field associating the name ‘artificial intelligence’ because it is the closest description of what is being created. Modern AI is nothing but a precursor to an actually intelligent program that can pose an existential threat to humanity.

Take the show Westworld for instance. Set in a near-future scenario, it explores the possibility of a theme park operated by human-like robots. When these robots gain sentience, the operator of the park raises an interesting question — what is consciousness? In a conversation between a programmer who knows intimately how his machines work and a machine that knows its own true nature, he states, “There is no threshold that makes us greater than the sum of our parts, no inflection point at which we become fully alive. We can’t define consciousness because consciousness does not exist.”

This brings up an important point to be answered by AI researchers today — how can we be scared of a generally intelligent program, also known as AGI, when we haven’t even defined what consciousness is yet? 

According to this paper published on the nature of consciousness in AGI, there are still four definitional types of consciousness that need to be solved before we reach AGI. In this matter, our current algorithms are just simple instruction-following programs, whereas AGI is a completely new technology going beyond algorithms. 

While GPT-4 and Midjourney are leaps ahead of ELIZA or GANs respectively, they are still non-intelligent programs. The problem with the field of AI isn’t the technology they’re researching, it’s the branding. 

AI has a branding problem

The real problem in AI is not AGI risk, but the fear associated with artificial intelligence. For decades, depictions of thinking machines in popular media have served to strike fear in the hearts of the audience. 

However, the path from the current state of AI technology to such capable machines is not only unclear, but necessitates a paradigm shift in computing itself. Computers today cannot even produce random numbers due to the binary nature of their foundations. 

Steve Ward, professor of computer science and engineering at MIT, stated, “On a completely deterministic machine you can’t generate anything you could really call a random sequence of numbers, because the machine is following the same algorithm to generate them.”

If our deterministic computers cannot even produce random numbers, how can they have sentience, consciousness, or even intelligence, which are purely non-deterministic concepts?

Let’s take a look at GPT-4, the most-advanced LLM and the subject of the aforementioned open letter. LLMs are just word generators, programs trained to predict the next likely word and fine-tuned to speak like humans. 

There’s nothing truly intelligent about GPT-4, except if you count its prediction capabilities — an echo of what is possible with real, human intelligence. While the disruptive potential of AI cannot be dismissed outrightly, it seems that social systems have already adapted to technology replacing humans. To create real thinking machines, AI research has to go beyond what is possible today. 

The real path to AGI

Current algorithms have gotten to this level mainly by aping human biology. Neural networks, the basis for GPT-4, are a faint echo of the structure of the human neuron. Reinforcement learning tries to apply a small facet of humans’ knowledge acquisition methods to computer programs. However, creating an artificial brain, or rather the recreation of the ‘thinking’ parts of the human brain, is a task best left to future researchers and scientists. 

This undertaking would not only necessitate advances in the understanding of the human brain, neuroscience, and psychology, but also breakthroughs in philosophy. To create a truly artificial intelligence, humanity needs to reach a consensus on what intelligence and sentience actually mean. Until then, researchers can figure out how to solve information hallucinations or how to make generative AI draw hands effectively.