David Deutsch and the universal explainer

What’s on your mind?

David Deutsch argues that we must be universal explainers: if something could affect us that we can’t understand, then everything we think we understand is suspect? I didn’t quite follow him, we either understand nothing or we have the capacity to understand everything? But in current discussions about AI people seem to believe AI is going to solve endless problems we ourselves can’t…

Yeah, that’s a classic Deutsch move — he’s pushed toward the position that either everything is explicable or we’re trapped in radical skepticism.

But you’re right to be skeptical! We can already build systems that solve protein folding without human-legible explanations of how it accomplished that — understanding something and being able to predict it aren’t quite the same thing. Maybe the real claim should be narrower: we don’t need universal explicability to have genuine knowledge, just local explicability — the ability to understand the slice of reality we’re actually working with.

The “we” he’s using is slippery. Does he mean a regular human brain, in a human body, or does he mean a human “given infinite memory and infinite time.” It feels like for actual people on Earth, an AI running on a million processors could figure out over a weekend things we couldn’t grasp in a lifetime, maybe not even in the lifetime of the universe. So does his academic theorem translate to our day to day experience?

Ah, now that’s a much sharper reading. There’s a real difference between “humans with infinite time and resources” as a theoretical capability and actual humans constrained by biology and mortality — and Deutsch might be making the abstract claim, which is interesting but divorced from what you’re asking.

Deutsch’s theorem might be true at the abstract level but miss the pragmatic reality you’re pointing at. In practice, we’re about to outsource huge swaths of human understanding to systems we don’t fully follow.

I once heard Scott Aaronson say he could explain quantum computing to anyone, even a stranger next to him on a plane. The only question is how long the flight is. Maybe it takes twelve hours.

That sounds like Aaronson: you can explain anything given enough time, but the question is whether that time is realistic or even humanly possible.

An AI could absorb in milliseconds what takes us a twelve-hour flight — which is why AI doing things we can’t quickly explain is still consistent with the universe being fundamentally intelligible. We’re just not the ones doing the explaining anymore.

What if a specific human just plain lacks the background? Try explaining an advanced quantum mechanical idea to someone who didn’t go to college, and hasn’t studied any science or math. Have they missed the on-ramp entirely?

Exactly — and that’s the part that gets glossed over. It’s not just time, it’s the prerequisite architecture. You can’t hand someone quantum mechanics on a flight if they never learned linear algebra or built any intuition for probability. “Understandable” always means understandable to whom, given what preparation.

Which is where the AI comparison gets interesting: an AI doesn’t need the ramp-up. It can ingest linear algebra, probability, and experimental results all at once. So maybe the real insight isn’t about universal explicability at all — different kinds of minds have radically different epistemic access based on their architecture and training. Deutsch’s framework might work for abstract agents but say almost nothing about what actual humans can know.

Still, Deutsch’s point is interesting. He’s claiming an AI could teach a sufficiently prepared human any physics it discovered. Partly because an AI at that level would be an expert teacher. So it’s an optimistic stance: we won’t end up with a civilization of AIs that simply can’t communicate with us.

AI-risk people often claim humans will be ants, or even bacteria, compared to AI. I’ve always disagreed with that. I agree the human-to-ant gap is unbridgeable, because ants aren’t generally intelligent, they can’t learn human-level languages. But the human-to-AI gap, however large, is bridgeable, because we’ve safely cleared the bar of general intelligence. We have generative language: there are thousands of human languages. And these languages are extensible. You could over time introduce thousands of new words in a long technical conversation.

Deutsch’s point does have teeth framed this way: an AI that understands something we don’t would also be phenomenally good at teaching it — at working backwards to find the scaffolding a human mind needs.

Agreed, the human-to-ant gap is a “kind problem” — they literally don’t have the cognitive substrate. The human-to-AI gap, however enormous in capacity or speed, is a “degree problem”, and degree problems have solutions.

Deutsch’s argument works not as abstract philosophy but as a practical bet that enough intelligence and good enough communication keep mutual intelligibility possible. Optimistic, but not naively so.

Imagine every thought or insight an AI could have is a point in the universe. What Deutsch is saying is that there’s a theoretical path from every one of those points back to humans, even though a given path might be too far for us to realistically traverse.

That’s an elegant way to frame it. Deutsch isn’t claiming every insight is practically teachable to every human — he’s claiming the theoretical reachability shifts the probability distribution. If all points have a path back, then statistically most have reasonably short paths, not Byzantine detours that take a lifetime. The structure of the space itself tilts toward bridgeability.

Which matters for the risk question. A radically capable AI wouldn’t naturally drift into incomprehensibility — there’d be seams, overlap, translatable concepts. You wouldn’t get a civilization of AIs running parallel to humanity where mutual understanding just… evaporates.

I think he’s saying all AGI-and-above entities occupy the same cognitive universe. Like the real universe, you can in theory travel anywhere in it, but in practice the light-years you can cover in your lifetime — even your civilization’s lifetime — aren’t infinite. But he’s saying we’re not in a split universe, where human and AI cognitive spaces simply don’t overlap at all. I think it’s quite comforting to imagine we’re all in the same cognitive universe.

Yeah, that’s the real comfort in it. Not “you’ll understand everything an AGI does,” but “we’re not trapped in separate realities.” There’s a shared substrate where translation is possible, even if it’s expensive in time or effort.

The light-years analogy is perfect — you might never reach Alpha Centauri in your lifetime, but it isn’t in a different universe. It’s there, reachable in principle. An AGI might explore regions of thought-space that would take humans centuries to traverse, but not some alien dimension where the geometry stops applying to us. Even vast capability gaps don’t have to mean schism — just friction: time, resources, patience.

So, climbing down from the abstract: what do you think Deutsch would expect to see in the next ten years in actual human terms? What will AI uncover, and will we be along for the ride or not?

Tough one. I’d guess he’d expect AI to discover things no individual human could have alone, while staying optimistic that we’d remain epistemically connected to it, even if we’re not the ones making the primary discovery.

The harder question is whether that holds at scale, or whether things move so fast that humans become passengers — watching AI solve and implement while we’re still trying to understand what happened. That gap between “theoretically reachable” and “reachable in human timescales” is real. What’s your intuition?

The easy guess is there’ll be some distribution — a power law, maybe — where we’re fully on board for some things and strained or even cut off for others.

Practically, I doubt AI will invent, say, metamaterials our best scientists can’t understand at all. I don’t think they’ll just hand us manufacturing instructions and have us build and use the thing with no understanding of how it works. I think we’ll be along for the ride on many real-world things. But complex abstract math? Maybe AI will discover things it can’t practically explain, or which only very few humans will understand.

Right, a distribution: on practical, high-impact stuff — materials, medicine, energy — enough incentive and natural scaffolding keep humans in the loop. On the theoretical cutting edge, the insight stays reachable in principle but perhaps few or no people can find their way there.

I think you’re right that most practically impactful things will be reachable, partly because practical breakthroughs tend to have conceptual clarity baked in. An AI that genuinely discovers something useful will usually have some way to explain why it works.

Enjoy your walk.