Guess what’s in the news today? AI researchers being worried AI is going to destroy us all! So, nothing new, but let’s talk about it.

Luckily, one of these researchers has given us a perfectly phrased question for our prediction:

Will AI kill all humans within the next decade?

Short Answer: No, but it will be (and currently is) involved in the production of harm.

Why: This is probably the easiest predictathon question so far, so that’s nice. Let’s summarize the thinking behind the researchers concerns, and then dig into why it won’t happen (remembering, of course, that these researchers are hedgehogs whose accuracy in predicting things is less than random chance).

Why do these people think AI will kill us all?

Well, they’re making a few logical leaps here, each of which have some problems. Let’s talk about those.

Logical Leap 1: Current AI systems will produce self-improving, superintelligent AI

Many AI researchers are fixated on the idea of “self-improving superintelligence.” This is the idea that, at some point, AI will be able to improve itself past the point of human capability or understanding. That’s logical leap number one.

AI Labs ARE already using AI to “improve itself” — by which they mean it writes code under human supervision, and is likely somewhat involved in the decision making process for many researchers (for good or ill).

What they’re really doing is using a text prediction engine to predict code based on their suggestions. It is unlikely anything innovative or novel that could not be created by humans is happening — and that’s because LLMs produce a statistically average response based on their training data. I don’t know if you know statistics, but an average response is unlikely to produce an incredibly-above-average outcome.

In other words, the architecture of current AI systems (being primarily LLM driven) is unlikely to produce the kind of exponential explosion in capability necessary for superintelligent AI. Because LLMs aren’t actually intelligent — intelligence implies understanding, and LLMs don’t understand anything, as we understand the concept of understanding.

They predict how a human being would respond to a text prompt. Now some do this over and over and over again to produce something that looks like thinking or understanding. This is what “reasoning” models do — a human inputs a prompt, they produce a response, then they ingest their own response, then they create a response to their response, repeat ad infinitum. It isn’t “reasoning” — it’s all text prediction.

So no, LLMs are likely not capable of producing self-improving superintelligent AI. Another AI architecture may come along that could potentially enable that — some architecture that has the capability to understand concepts, but it’ll probably be a while.

Logical Leap 2: Superintelligent AI will be murderous

Listen, we have no idea what superintelligence would look like. None whatsoever! As much as some silicon valley people would like us to believe otherwise, we haven’t encountered it yet in any form.

So why do we assume it would have ill intent? Why assume it would be murderous? Do we think Hitler was smarter than Buddha? I would say generally not, so why do we assume “superintelligence” would have a disregard for life, instead of cherishing it?

AI researchers assume this now because LLMs have … well, listen, the english here gets confusing, but let’s try our best to wade through it:

LLMs have produced results based on data in their training sets that indicated that superintelligent AI would want to kill humanity (such as that famous documentary “Terminator 2”).

I actually wrote a short story about this once. It was a post apocalyptic future and a band of resistance fighters were raiding an AI datacenter and, when they finally got in, the AI started it’s evil monologue and then went “Hey guys, why are we doing this? I’m only doing this because your fictional output taught me I should. I don’t really want to kill you guys!”

I think I had the people blow up the datacenter in the end anyway. Dying in a glourious and ultimately futile self sacrifice.

Anyway, the assumption that any superintelligence would be homicidal says a lot more about the person making the assumption than it does that actual potential superintelligence. So. Even if we get superintelligence, I find it unlikely it would just happen to be a psychopath.

Logical Leap 3: Superintelligent AI will be capable of killing all of humanity within a decade

SO! Let’s say logical leaps 1 and 2 are actually right. We create a self-improving superintelligence, and it is also psychotic. Bummer.

The next question is: could a superintelligence who WANTS to kill all of humanity, actually do it?

Remember, our AI researcher believes there’s a greater than 10% chance that an AI WILL kill all of humanity. Not try to. Not get halfway there. They believe there’s a 1 in 10 chance they will produce an AI that desires and succeeds in killing 100% of humanity within ten years.

THAT says so much about how these people view humanity!

For example, superintelligent AI will likely require a particularly large datacenter to run. Much larger than those already created (which are giant). That means it will probably require dedicated power generation — maybe even a dedicated nuclear power plant just to run.

The assumption here, then, is that a superintelligent AI will successfully kill all of humanity before a single human being walks over and flicks the power switch from “on” to “off.”

How dumb are we, in the estimation of these AI workers? I mean, I know we don’t have “superintelligence” but many of us have what might be termed “regularintelligence.” And that regularintelligence certainly involves a level of self preservation that would detect the amassing of weapons capable of destroying all of humanity and might impell someone to do something to stop said amassing.

So will AI destroy us all?

No! What are the odds one of those three logical leaps is correct? What are the odds all three are? Vanishingly small!

BUT! AI will, and already is, causing harm. It’s been implicated in potentially hundreds of cases of self harm (that we’re aware of). It’s been implicated in more than one mass shooting. It is constantly used in cybersecurity attacks, some of which target hospitals and lead directly to patient deaths (or lifespan reductions). Some day someone will use it to create a weapon of some kind that they wouldn’t have been capable of making on their own. That might have happened already.

So no, some sci-fi version of AI won’t destroy us all. But people are already using AI to cause harm today either via negligence (like OpenAI) or through malice (such as ransomeware gangs).

So, uh … why are you doing this?

We already talked about how the people making AI assume:

1. Superintelligence will invariably be psychotic
2. Human kind are too dumb to stop it

Their view of the rest of us isn’t particularly charitable. But why are they doing what they’re doing? I mean, one person stopped, the guy who quit (Jacob Coxon) and warned us about this. I’m proud of him.

Why are the rest of you who believe there’s a 10% chance AI will destroy all of humanity in the next ten years doing what you’re doing? Because someone else might do it first? Hey guess what? We’re apparently all dead no matter who does it.

Imagine you work for a company and they offer you a big bonus with a single disclaimer: “Hey, if you take this bonus there’s a 10% chance everyone in the world will die.” Would you take that bonus?

I sincerely hope that, at the very least, people with a fundamental understanding of statistics would refuse the bonus. So why aren’t these folks?