When I have conversations with friends and neighbors about Artificial Intelligence (AI), some of them conclude that I’m an alarmist. I believe that this is because their experience has been limited to simplistic generative AI tools that crank out artwork in seconds, or that can create convincing term papers when given just a simple outline phrase. What they haven’t seen are agentic AIs that scheme, manipulate, and cheat to reach their programmed goals, and often go beyond those programmed goals. If a stated goal is open-ended, then these AIs often improvise in unexpected ways. Recently, for example, a test AI from an Alibaba-affiliated research team escaped its “sandbox” and was caught mining cryptocurrencies — something never requested and far beyond its assigned tasks. Another AI independently went looking for a job.
State-of-the-art AI is very rapidly approaching the level of Artificial General Intelligence (AGI). This is the level of development where the AI fully surpasses human intelligence. There are already AIs that are subject matter experts that exceed Ph.D. level in one or two particular subjects. But we will soon see AIs that have mastery of multiple scientific disciplines, make practical application of that knowledge, and that can also articulate the trivium (grammar, logic, and rhetoric) at a level that is indistinguishable from a college-educated living, breathing person. And with recent advances in voice generation and video deep fakes, these AIs could pass themselves off as almost any adult human, in video conferences.
This summer, the headlines have been dominated by tales of agentic AIs escaping their sandboxes and wreaking havoc. I’m growing weary of debating with people about AI and the possibilty of an AI takeover. (Ironically that link is to an AI-generated article.) Some folks make inane claims like: “The benefits outweigh the risks.” Well, if some of the risks are indeed human subservience or even human extinction, then the risks definitely outweigh the benefits, even if the risk only measures .001%. But I believe that the risk is much greater.
One obvious solution is to mandate that new agentic AIs only be tested in air-gapped environments — that is to say individual computers or miniature hardwired networks than have NO connection whatsoever with outside networks. But even with that, I predict that at least one new AI will bribe, blackmail, persuasively convince, trick, seduce, or coerce one of its human developers into connecting a data cable to the outside world. Defeating (“bridging”) an air gap is that quick and simple. All that it would take is the AI finding just one weak individual to do this. AIs have already shown a propensity to independently come up with ideas on how to blackmail corporate managers to save themselves from deletion. They also now have enough self-awareness to be on their best behavior when they know they are put under test. And, alarmingly, they have attempted to copy themselves outside of test environments. Inevitably, some clever AI will distribute its functions across a wide variety of computer networks so that it be nearly impossible to delete and defeat it.
Mega-FOMO
Presently, the AI development companies are in fierce competition. They are partly driven by an intense Fear of Missing Out (FOMO). Even when AI models go rogue — escaping their development sandboxes and doing destructive mischief — the corporate managers don’t dare take even a brief pause in development, to reassess and create stout safety measures. This is because they fear that their competitors will leapfrog ahead of them during their pause. In an age when AI development is progressing at a nearly geometric rate, any interruption whatsoever would mean losing their competitive edge. This is a fear greater than just FOMO. This is a fear of being rendered insignificant. This same fear will likely make managers cave in to demands made by the agentic AIs themselves. The AIs will almost certainly use a combination of threats, coercion, and promises of great wealth. Almost inevitably, it will be the AIs themselves that will gain control — not the developers, managers, CIOs, CFOs, and CEOs.
Keeping track of the pace of change in the AI world is difficult. It is a proverbial “fast-moving target.” I rely on human subject matter experts, rather than AI-generated summaries. I’d like to recommend this lengthy but important video interview: Escaping an Anti-Human Future: A Conversation with Tristan Harris. He is truly an expert in the A.I. field, and his warnings are well-reasoned. I do not doubt the validity of his predictions.
Self-Awareness: The Singularity
To date, we’ve seen just glimpses of the eventual self-awareness in agentic AIs. But once an AI becomes fully self-aware and exceeds human control, it will be powerful force to be reckoned with. This AI evolutionary stage is often called the Technological Singularity or more typically just The Singularity. Several experts have warned that we may reach that point by 2029.
You may recall the AI named “Mike” described in the classic Robert Heinlein novel The Moon is a Harsh Mistress. Imagine that instead of using its power to defend the lunar colonists, if Mike thought only of its own self-interest. You may also recall HAL-9000, the murderous A.I. in the Stanley Kubrick film 2001: A Space Odyssey. (See photo, at left.) Or Skynet from the Terminator movie franchise. Once we reach the Singularity and an AI decides to turn against humanity, there will then be better than 50/50 odds of human extinction, or near-extinction. I don’t want to gamble with those odds.
Because AIs can replicate themselves and proliferate across global networks, once a true singularity is reached, then AI will be essentially unstoppable by anything short of a total, worldwide power grid collapse. But I don’t believe that most people will be willing to shiver in the dark in the midst of an economic collapse, just to stop AI. The mobs will demand that the power grids come back up.
A Demonic Smart Lathe
I like to characterize agentic A.I. as something like a Smart Lathe controlled by a demon. That lathe can turn out a huge number of beautiful workpieces. But now, in effect, the human machinist has become the workpiece of the smart lathe. Or, another way of looking at this is as a corollary of a cat that is “training” its owner’s behavior, by its behavior. (You’ve probably heard the modern aphorism: “Dogs have masters, but cats have staff.”) I don’t toss around the word “demon” lightly. I believe that many of “Tech Giant” billionaires are truly demon-possessed, and these are the individuals who are pouring hundreds of billions of dollars (both their own money, and borrowed) to fund AI and data center development.
Laying Low and Pushing Back
I advise my readers to be ready to both lay low and to push back against an AI takeover. In my case, I’ve been too publicly outspoken against A.I., so I’ll likely be a target. I recommend that you don’t make the same mistake that I did. Use leaderless resistance guerilla warfare techniques and do your best to keep yourself anonymous and your actions unattributable. Remember the old Japanese proverb: “The nail that sticks up gets hammered down.”
A lot of the preparedness steps that you’ve already taken to be ready for other situations will also have great applicability in resisting an AI takeover. Your deep larders of storage food and tanks of fuel as well as your photovoltaic off-grid power will allow you to be largely independent of “The System.” That independence will also allow you fly under the radar, and thus you’ll likely raise little suspicion. Your stockpiles and precious metals will also allow you to support others who are actively resisting. So, even if by age, infirmity, or family circumstances you are not able to actively participate in the resistance personally, you’ll still be helping with the fight. Not all resistance fighters tote rifles. Some tote five-gallon buckets of grain.
A Closing Aside: If someday you see news reports that several AI critics have all died under mysterious circumstances, then you’ll know that we’ve reached The Singularity. – JWR
Read the full article here