AI: awareness is all you need
Last week I spent 25 hours behind the wheel and listened to the latest podcasts about AI [1, 2, 3, 4, 5], and a while ago I read the entire modern scientific base on AI. I suppose the scientific background of h-index>0 and understanding how MLP/GPT/LLM work on a deep level allow me to consider myself a person who understands something about it. So, we live in the most magical🪄 and somewhat pre-apocalyptic time, and we need to prepare for a high level of turbulence in everything.
Why is AI, GPT, LLM, AGI, and AI alignment possibly the most important thing to happen to humanity (sic!) and how are they already defining the future of humanity?
In short, what happened in recent years/months: in 2017, the “Attention is all you need” article was published, which proposed a new architecture of neural networks — General Purpose Transformer (GPT). The highlight was that this architecture is based on the Multilayer Perceptron (MLP), which is a simple-as-a-board approximation of how a neuron in the brain works at the biological level. To put it even more simply, MLP is a mathematical expression which you can take a derivative from, and in mathematics, everything that has a derivative can be optimized for some utility function. MLP has been used for half a century and the majority of the scientific community was skeptical that it was a promising way to develop research in the field of AI. However, MLP’s simplicity and a few improvements made a breakthrough⚡.
GPT as an architecture added the ability to have connections between different parts of the neural network and assign different weights to input data through the attention mechanism. Thanks to the simplicity of implementation (matrix operations), it all started working and scaling very well. The final touch was that this GPT architecture works remarkably for the next token prediction task for the language. Humanity has generated a lot of text information throughout its existence, and the dataset of all this text, if it’s broken down correctly, contains both X (input data) and Y (output data) and is self-sufficient for learning without a human. This is how the so-called Large Language Models (LLM) appeared, which optimize the function of what can be the next symbol of the sequence.
Imagine that this is a kind of Michael Scott from The Office who can always finish any sentence “that’s what she…” with a good joke. Moreover, in recent months it turned out that most problems in the world can be reduced to the next token prediction, which LLMs have started to cope with well. To describe it very inaccurately, the LLM/GPT architecture became possible because we have powerful GPU processors that were able to train very large, but architecturally simple neural networks based on all the textual information in the world.
That is, there was a breakthrough in the field of artificial intelligence systems, which became possible due to:
- the scientific apparatus that allowed it (GPT)
- a large dataset you can learn from (“the Internet” and all the information it contains)
- the capabilities of modern processors to perform fast and scalable calculations that were previously unavailable (GPU)
Additionally, it turned out that very large neural networks can solve problems that were not in the training data and that it becomes very similar to Artificial General Intelligence (AGI)🧠.
AGI is a concept of artificial intelligence that can be as versatile and powerful as the human brain. Most likely, GPT is not AGI yet, but this is not certain.
Not for certain, because it passes all the intelligence tests possible, but subjectively it doesn’t look like it’s there yet. And all would be fine, but the problem with AGI is that theoretically, if we can physically make something as powerful as the human brain (we’re not there yet), then nothing will prevent us from scaling it horizontally, for example, by 10 times and getting something 10 times smarter than the human brain. Or 1000 times, or 1000000 times. Even if we imagine that AGI can only be several times smarter than a human, then AGI can improve itself to become even “smarter”. Therefore it is quite possible that as soon as we get beyond the human level of intelligence in the computer, further development will be exponential and very fast. This is the so-called transformation of AGI into Artificial Super Intelligence (ASI), the so-called AI takeoff. Nick Bostrom writes well about it in his book Superintelligence.
The problem is that we cannot even comprehend how humanity can coexist with something exponentially smarter than us. How do we program the right motivation into artificial intelligence so that we don’t go down the Terminator movie plot when the machines want to destroy humanity? Who will program this motivation and won’t the first AI researchers optimize to gain power and control over the rest of society? If you want more complex thought experiments, google the “Paperclip maximizer” or the “Simulation argument”…
On the bright side, superintelligence can find cures for all diseases, for example, cancer. The downside is humanity doesn’t yet understand what motivations and levels of consciousness superintelligence can have. This is the so-called AI alignment problem. This is the reason many smart and not-so-smart people today like Yuval Harari, Elon Musk and Steve Wozniak wrote a letter to the governments with a request to stop AI research because we, humanity, are not ready for it.
The issue is that AI research is like a nuclear race (💥Manhattan Project). The first government to succeed in this will be the superpower of the new world. Therefore, even if the US stops its research, China, Iran or North Korea can continue their programs. A game where everyone loses yet is forced to move quickly towards AGI.
The good thing is if we assume that the AGI -> ASI transformation will not happen in a few hours or days, then humanity will have a good chance to improve labor efficiency and the GDP of all countries will grow. Then we can possibly even get to a universal basic income. Many people will lose their jobs, but the economic impact of AI will create hundreds of millions of new, more creative jobs 💸.
As a Ukrainian I also ask myself “What does it mean for the war in Ukraine?”. The world is likely to face turbulence in the near future. We can’t predict what will be affected the most: possible big economic shocks with “white-collar workers” losing their jobs, or disinformation domination and a wave of synthetic reality created by artificial intelligence that will distort the space of the news and truth. It must be taken into account in all scenarios of the war as well as in our lives.
Sadly, we are unable to predict whether AGI will be good or bad for humanity. The majority of people don’t notice this revolution yet. So awareness is all we need 😊.
