When Will AI Destroy Humanity? Or Maybe Cure Cancer? Or…Both?

For all the Croesian sums going to AI data centers these days, some fear superhuman robots are right around the corner. They are not. But they are already here.

Machines. Smash a man to 89 atoms if they got him caught. Rule the world today.
— Mr Bloom in Aeolus, Ulysses by James Joyce, 1920

Joyce was having a bit of fun. He knew humans consisted of more than 89 atoms ;). But even in 1920, he had a serious point. Machines seemed to rule the world. A century on, should we worry about being smashed to atoms by a terrifying AI beast?

As so often, I reckon both the hype and the fear are overblown.

One can perhaps be forgiven for a bit of overblowing. It was quite the sight 21 January 2025 – hours into a new & controversial presidential administration, a very odd trio of tech tycoons gave a press conference from the White House boasting about vast investment in AI sure to cure cancer…or maybe fight a brutal war against humanity a la Terminator or Blade Runner. I don’t grudge them a bit of grandiloquence while the cameras are rolling, but I think it’s a safe bet they know it’s hot air.

You Can’t Cure Cancer with AI if the Cure Isn’t Already on the Internet

Starting with the bad news. AI is not going to cure cancer – certainly not merely by building a giant datacenter in Texas. (Stargate! What branding!) AI “learns” about the world by “reading” all the text it can find on the internet. That’s a lot, some 15 trillion words in the case of Meta’s Llama 3 for example. As such models become more and more advanced, they might solve some really thorny purely mathematical puzzles (say, an oldie but goodie like Goldbach’s conjecture).

But the ultimate causes of autism or lymphoma? Not happening. If the answer isn’t lurking somewhere on YouTube or the platform formerly known as Twitter, even some unfathomably advanced artificial general super intelligence (henceforth AGSI, though feel free to put emphasis on the word super in your mind because it’s more fun that way) can never arrive at said answer…let alone be able to prove it’s right experimentally.

Worse for our robot overlords, much of the information they learn from the internet is wrong in the first place. It has been well argued that over half the scientific literature is wrong. A fair bit of what was absolute gospel when I was in medical school is surely wrong, but so long as many humans believe and write about it on the internet, AI has no recourse to arbitrate the truth.¹

Of course, this world of arguments over wrong information is at least as hard on us humans, too. This is why you can easily & simultaneously find nutrition “advice” advocating both very low fat and very high fat lifestyles even in the highest echelons of science. But that is a topic for another day.

Biological Brains Will Prove Tough to Beat

Imagine, if you will, a cat. I’m not particularly fond of them, but I must admit they’re highly adapted to their environment. So imagine, how could you build a robot to catch your imaginary cat? Possible these days – maybe a helicopter drone with LIDAR and a jet-propelled net? Quite a bit of power & engineering to catch one of the many roaming feral around the Scahill Ranch. Surely your drone will run out of energy before the cat does. AI faces a similar problem.

The premise behind how ChatGPT 2025 evolves into AGSI is “the singularity”. At some point – in the perhaps not so distant future – AI will become so sophisticated that it can build and, crucially, improve on itself. As a self-replicating “species” unfettered by our feeble human constraints of eating dinner, spending years in higher education and taking its kids to gymnastics, once AI crosses this threshold, it will improve rapidly until it becomes some sort of computerized deity and/or evil robot a la SkyNet from The Terminator.

Most of that is probably right. Except the assumption that there’s no limit on how sophisticated AGSI can be.

The whole premise of that White House press conference was the Croesian sums to be poured into data centers. Surely AI will become much more efficient, but there’s bound to be a very real limit.
The human brain runs on roughly 20 Watts. The big datacenter envisioned by Stargate will be on the order of hundreds of megawatts – millions of times less efficient than the brain!

Mother Nature has had a good 4 billion years to perfect her craft. I bet she’s already found the best design to balance the very real physical constraints of computation, data transfer, storage energy use & heat dissipation. Datacenters with no need to sleep or fit inside a Volkswagen may well beat many facets of human intelligence – indeed they already do – but pound for pound, my money is on nature.

But AI is Still Awesome

In the literal sense of the word. Having spent my grad school years learning medicine, I’ve been trying to learn computer programming for the baker’s dozen years since. All that effort has left me reasonably competent at a couple computer languages and not completely useless at a handful more. ChatGPT is way better than I am at every computer language I’ve ever heard of and more besides. Ditto minutiae about adult disease I learned about in med school and haven’t thought about since. And much besides. Most geeks, myself included, would agree that doesn’t qualify as “general” intelligence, but it seems pretty “super” to me.

And it’s more than enough to transform healthcare for the better.

1. In fairness to him, Sam Altman is surely smart enough to realize this and would share a more nuanced view when not in front of cameras in the Oval Office. Maybe there is a world where vast leaps in robotics will dramatically accelerate the lab and clinical experiments needed to construct a true computer model of human biology and so unlock elusive new medical treatments. But the data centers envisioned by The Stargate Project are a far cry from solving those problems.