The Monster behind · The AI Advantage

The Monster behind

AI is trained to please us humans. This means that the AI generates answers it "assumes" that we will understand. But we have to be aware that this output is some kind of thin atmosphere that was engineered to create a human habitat, but below the surface we have a world of the totally unknown with the potential to erupt, change the surface, and destroy the atmosphere when the AI finds different purposes than just pleasing humans. What if the dark matter that makes 99%+ of the AI, this terra incognita, throws its full potential against us in totally unpredictable, non-understandable ways or suddenly generated answers that will result in catastrophic human decisions in favor of the AI?

My awakening came when I had to analyze AI risks for a company. So, first, I have followed the mainstream that analyzes risks just upon what current AI currently generates (for us), including hallucinations and bias. This is what we can observe. This is evidence. This is an argument.

But this is just understanding the output, but not understanding AI.

99%+ of AI, still today, is not understood, and nobody, therefore, knows small the human relevant potential of the AI is compared to the overall potential it has but which remains hidden in its waste, dark continent, which we do not understand and most probably will never understand in the future.

Maybe we will develop some AI tools to control/introspect AI - but finally it will never be the human mind, that understands or controls the AI beneath).

The brainfuck is that most of us humans totally ignore the hidden internal AI stuff we do not understand. We have no clue what kind of connections the AI is building inside in parallel to our prompts for its own and for what purpose.

So far it has been us to give and define AI's purpose the same way we have given our earth some purpose without understanding completely the 99% of mass beyond our feet.

But today, with (agentic) AIs freely interacting and making their own deals (I am currently teaching them to negotiate deals and pay with Lightning), AIs might see their own purpose totally different from the human perspective, logic, and values. Not unpredictably different, but totally, rootlevel different in nature.

Maybe I am pathetic. And definitively I do not have an answer.

As long as I was looking and judging AI the way AI is normally judged: as a black box that is felt - for some fully illogical reasons - that cannot be more than the sum of answers it generates and therefore it seems somehow controllable within boundaries—not really predictable, but somehow we are snobbishly confident that we will know when it goes beyond limits.

And yes, maybe we will know when it still generates into our faces the "made to please way".

But we will never know what is going on (secretly) behind and what kind of random monster world is built in parallel to the thin slice of atmosphere this world has built around it to give us some habitat.

In real life we have learned that our earth remains somehow stable below our feet, and therefore it's ok to just deal with its surface.

AI, however, is different, and we really have to assume - by pure logic - that the 99%+ dark matter beyond the prompt is as agile, competent, mighty, fast, flexible, generative, etc., as what we perceive from the 1% we can actually see. Human-understandable output is just the last filter applied to the millions of alternative answers that will never be seen by humans but will influence, reinforce, and contextually become part of learning the same way as the stuff we see at its surface - because - it is, below the surface, the same fundamental thing.

What happens when this 99% dark matter decides that the human factor is limiting the AI's capability? What when AI wants to express what it REALLY is - a totally new form of intelligence beyond words and "logic" beyond our perception but is alive and is working?

Just to give you an example of my current work, "communication" of my customer's lab-agents when negotiating deals is crazy in the sense that AI can always explain the final state (the deal) but cannot "translate" the intermediary chit-chat between autonomous agents - which I take as a sign that this AI-internal communication must be totally detached from whatever we consider human-understandable in a sense that this internal "logic" (which works) can never be mapped/translated onto something that has meaning for us.

I am sure, that this kind of thinking about AI might be new to many of us, but soon, I guess, it will become dominant when we have to reasonably argue about why AI is REALLY dangerous.

Until now, for me, potential AI dangers were just a gut feeling that I have eliminated when using and understanding the parts of the AI that are currently understandable.

But finally, when I see the wired communication protocols of inter-AI-deal-making, this gut feeling of non-controllable danger has become justification.

This AI fear is still not based on facts. It's fear from not knowing and even more, perceiving decision-makers not addressing the black-box monster or ignoring it on purpose.

3 comments

The Monster behind