Evidence (p=.05) that ≥15% of Super-intelligences are Safe?

With recent advances in AI, many of us wonder if a safe (i.e. won’t destroy humans) super-intelligence can be created. If it's possible, what's the likelihood that any given one would be safe? If it's impossible, would a technological society inevitably make a bad one anyway?

It’s an unprecedented question and seems like we only have thought experiments to guide us.

However, there may be real empirical evidence on this existentially-important question, though I have to warn you, it’s a bit out there. However, the stakes are so high, and existing evidence on this topic is nonexistent so we have to take what we can get.

I propose that however inadequate, some evidence does exist—in the form of the US Navy UFO reports. Yes, it's weird but here’s why you should at least temporarily come with me to crazy town:

  1. Many smart people have thought long and hard about AI, so most of the low-hanging fruit has been picked. If you want some really novel fruit, you might need to visit the orchard that everyone avoids because it’s where the crazies hang out.
  2. Even if you don’t think that the UFO reports should be taken seriously, Sam Altman, CEO of OpenAI and arguably the single most important person in AI development, says there are interesting lessons to learn. So regardless of what you think of the evidence, if you believe that what Sam Altman thinks is per se important, then the reports are important.

The question should be “Technology?” not “Aliens?”

Whether the UFOs are aliens is downstream of the unresolved debate of whether they are technology, as opposed to a hoax, psy-op, natural phenomenon, etc. The technological implications of this question are so huge that assigning them specific actor like aliens is a premature distraction.

If the UFOs are technology, then they represent a technology far beyond anything we currently have in 1) propulsion, 2) power generation 3) materials and 4) control systems. However, of these, the thing we’re currently the closest to is the control technology, with the use of machine learning to land spacecraft.

We’re farthest along the intelligence track: we think we might make an AGI just by continuing to scale LLMs (prediction markets are saying within 10 years), while we don’t even know if other technologies are even physically possible, let alone how we’d accomplish them.

This implies 2 possibilities:

  1. The UFOs were made by or alongside a super-intelligence.
  2. The UFOs were made by a society that had the technical capability to make one but didn’t.

In some sense, an entity with such a high level of technology would definitionally be super-intelligence, but at the very least we can conclude that UFOs constitute evidence of a super-intelligence level society, and therefore achieving that level of technology is indeed survivable. I’ll continue to talk about super-intelligences, but if you think it’s actually option #2, then just substitute that in because the math is the same.

Conclusions from a single observation

We can consider super-intelligences as either safe vs. unsafe, a binary value that would have some binomial distribution. That is to say, there’s some proportion between 0 and 1 of super-intelligences that are safe.

If we conservatively consider all the UFO reports as a single entity, then we have a single observation to go off of.

Confidence intervals for binomial distributions can be calculated in different ways, but for this scenario, the Jeffreys interval gives a 95% confidence a lower bound on the proportion of unsafe super-intelligences as ~15%,

(The most conservative Clopper–Pearson interval gives it as 2.5%; you can run the numbers yourself here.)

USS Nimitz: The best evidence we have

Most evidence for UFOs is very poor and I would encourage dismissing all of it, with the single glaring exception I’ve discussed before:

the 2004 USS Nimitz incident described in the NYT seems to defy normal explanations. It was separately detected on radar by a Navy plane and ship, filmed on a fighter’s infrared, and witnessed by 4 Airmen.

You may not find this convincing, but I encourage you to at least read the internal Navy report. Even if you’re unmoved, it still may be challenging to explain away such professional and separately corroborated accounts. One of the more popular explanations is that it's a clandestine American/Chinese/Russian drone, but that still implies a level of technological sophistication generations ahead of where we thought we were. If secret reaction-less, exhaust-less, super-ships existed in 2004, then surely secret big LLMs exist in 2023 and we’re already coexisting with AGI.

If nice aliens existed, they would have stopped the Holocaust

If the nice aliens were already here, they would have stopped the Holocaust. That's a valid argument against the existence of God. It's also a valid argument against the existence of nice aliens. And un-nice aliens would have just eaten the planet.

-Eliezer Yudkowsky

There are several issues with this line of reasoning:

  • If they did stop the Holocaust, how would you know? Depending on how visible their behavior is, there could have been many Holocausts averted but we don’t know to count any of them because they didn’t happen. Thus, the requirement for aliens is actually the much higher “if aliens were real, they would have stopped every Holocaust”. However, even if they did stop every holocaust, we could then move the goalposts and say “if Aliens were real they would have stopped the [next level down of bad thing]” ad infinitum until we are saying “me stubbing my toe proves aliens aren’t real.”
  • Perhaps they did stop some holocausts, there are reports as far back as the 60s of UFOs interfering with nuclear weapons.
  • We have evidence that sometimes when “enlightened” societies find a comparatively technologically primitive one it simply leaves them alone, holocausts and all. Various low-contact societies have committed their share of mass murders (such as the Waorani, where 42% of all mortality is from intra-group homicide), modern civilization believes that all in all the best course is to simply leave them alone.

As kooky as this all sounds, what I'm hoping for is a shift from dismissal to analysis, even if critical. It's literally vitally important and do remember that biases against extra-terrestrial origins have been broken down before—scientists only widely came around to believing in meteorites 1803.

Subscribe to Atoms vs Bits

Receive our weekly posts by email