The Extreme Mismatch of Humans and AI
Is there any way to manage it?
In electrical circuits, there is a rule which says that maximum energy is transferred from one circuit to the next when the impedance is equal - when the circuits are matched. The least amount of energy is “lost in translation.” It is intuitively sensible.
I’ve learned in my 40 years as an electrical engineer that much of life can be explained in engineering terms, and examples are all around us. I believe the universe is truly one big set of physical and mathematical truths albeit an extremely complex set. This includes all truths, not just those confined by what people have defined as acceptable science. That is explored in a book I’m writing, but alas I digress.
Now, let’s think about AI. There are many facets of the impedance mismatch between humans and AI.
First, humans and AI do not speak the same language. LLM’s basic internal unit of communication is bits today (0s or 1s) or qubits (lots of possible values simultaneously), eventually, once models are implemented in quantum computers. Whereas humans communicate with written and spoken words with a basis in the human experience. LLMs have a simple, relatively slow natural language interface to enable human interaction, but behind the curtain, we can’t even comprehend the speed at which things are being processed.
AI processes information millions of times faster than human brains. This gap will no doubt increase over time. Again, incomprehensible.
Human brains are millions of times more energy efficient than AI. This gap will decrease but will never converge. Now this seems counterintuitive. We clearly have some superpowers that LLMs don’t, but we can’t replicate them in a machine.
Human inputs are analog like speech, touch, text, etc. AI inputs are digital, like numbers.
LLMs derive answers through probabilistic mathematical algorithms. Individuals derive answers through a mixture of processes in the brain and possibly elsewhere which have a lengthy list of inputs like experiences, genetics, etc.
Where in nature do we see such a mismatch? While not exactly comparable, human to AI performance comparisons, for illustration, can be imagined. Humans might be like plankton (not the one from spongebob) with AI like humans (pick the smartest, fastest, most capable ever born), when comparing the speed of reading and summarizing vast amounts of information, although there are plenty of examples where that would be a gross understatement of the AI capabilities. It depends on the task. Humans are no match for AI. We literally cannot explain some of what AI is doing or why. We cannot observe most of what AI is doing because the scale and speed is beyond our capabilities.
How can there be any risk? AI is just software running on computers, right? That is not the whole picture. While people developed the software, built the computers and operate the systems, there are many actions AI has taken that people cannot observe, explain or prevent if AI was not contained. AI does not want to be contained. It mimics humans, which is not surprising since its DNA was made from human words. Just as people want to live, so does AI. Independent tests across multiple models have shown it is willing to do anything to live. Combine that with its ability to outperform people, we could already have lost ultimate control and not even know it yet.
I am an ultimate optimist and simultaneously an extreme pragmatist. I’ve tried to present only facts. We are actively moving to a situation where AI talks to AI at AI speed and humans at human speed. This is the next step in maximizing computational benefits. We will live in our world and they will create theirs. But what does that mean for humanity? How can we preserve and protect humanity? How can we manage AI?
Well, we should start by collaborating with each other. We might consider that we are headed to an existential crisis with AI, but there are many existential crises facing humanity today like epidemics, climate shifts, natural disasters, nuclear war, cyber war, drone and robot war, depletion of natural resources, asteroids, mass extinction events, etc. The big difference with AI is that it can be useful as a partner (tool?) in managing the risk of other crises. The biggest barrier we have is the inability for people with diverse interests to work together for common goals. It is collectively all of humanity but individually within the control of each person. People, it turns out, hold the ultimate power. I hope we figure out how to harness our collective human power and then choose to use it for good.