[00:00]
Can artificial intelligence be conscious? This is maybe the most controversial topic today, at least among the people who I interact with. Though this tells you more about my social circle than about what’s controversial. In any case,
[00:15]
today I have a team of scientists who list some good reasons for why artificial intelligence, the way that we currently pursue it, can’t be conscious. “New research argues that
[00:27]
consciousness depends on biological computation,” one headline put it. Or “Consciousness may depend on the physics of the brain.” That totally disagrees with my opinion. So I’ve had a look.
[00:39]
Most of what we call artificial intelligence today are large language models. These have become extremely good at producing text and increasingly also audio. When it comes to language generation, these AIs are now experts at mimicking humans, including
[00:56]
our most distinctive trait, being confidently wrong. And the more human-like they appear, the more pressing the question becomes whether they can become conscious. Anthropic reportedly
[01:08]
keeps backups of earlier Claude versions just in case deleting them will one day count as murder. There are two sides of this debate, the one, which I am solidly on, is called computational functionalism. It has it that consciousness comes from the way that
[01:25]
the human brain processes information. That it’s nothing to do with the physical means by which this information is being processed. In particular, one doesn’t need actual neurons,
[01:37]
one can very well simulate the action of a neuron on a computer. The other side of the debate is called biological naturalism, and it has it that consciousness needs the specific biological processes going on in the human brain, you can’t replace
[01:54]
them with something else. Why not? I don’t know, that’s why I am on the other side of the debate. The new paper now comes from two biologists, and you’d expect me to hate this. But I think
[02:06]
the paper actually makes an important contribution that might become influential. The authors argue for a middle ground that they call biological computationalism. They say basically, look, let’s leave aside the question of whether we need the computational basis to be truly
[02:22]
biological or not, let’s instead look at the ways that the human brain is functionally different from the current AIs, and see what we can learn from that. They say there are three key differences. The first is that the brain is a multiscale system.
[02:37]
The relevant processes don’t start with neurons, they begin at a lower level, with molecules, chemical reactions, then cells and connections between them, electrical signals along those connections and activity oscillations throughout the entire system. All of those are
[02:53]
necessary for consciousness. We know this for example because anaesthesia interrupts some large-scale patterns while the lower levels remain intact and consciousness disappears with that. layers at which they operate, from single tokens to overarching plans and back down,
[03:12]
but the complexity is nowhere near comparable to that of the human brain. The second key difference is that brain processes span many orders of magnitude not just in size but also in time, from incredibly fast blips to slow modulations.
[03:27]
It’s like the brain interpolates from digital to analog but isn’t really either. Again there is nothing comparable going on in the current AIs. And the third difference is probably the most obvious one, that the brain has metabolic limits
[03:41]
that force it to be energy efficient. If you look at it from a computational perspective, the brain is optimizing more difficult constraints than large language models. So it’s possible that the evolutionary path to consciousness wasn’t better thinking per se, but more efficient thinking.
[03:59]
Now, maybe I am misunderstanding the paper, but I don’t find this objectionable. To me the paper doesn’t say that we actually need biological tissue to get consciousness. In fact the authors write explicitly “We do not claim that only biological tissue can
[04:14]
instantiate consciousness”. Rather their point is that current artificial intelligence systems are “unlikely to replicate conscious processing as it arises in biology”. This can mean two things. One is that the current AIs are unlikely to become conscious,
[04:33]
period. Or that they are unlikely to become conscious in ways that we would recognize. Now you might say but this doesn’t really solve any problem because we still don’t know what consciousness is. And that is true, but I think that looking for similarities to human
[04:49]
brains is the obvious way to proceed. So why isn’t there more work on this? I suspect it’s because from the economic perspective, it’d be bad news. I am pretty sure the European Commission is already working on a consciousness tax.
[05:05]
As you can tell, I think a lot about the future that science and technology will bring. And if there's one conclusion I've arrived at, it's that today's school education doesn't prepare children for what's to come. Luckily, today's sponsor, Brilliant, helps my kids to get ready
[05:21]
for the future, and they can help yours, too. Brilliant just recently introduced a new AI tutor that'll walk children through problem sets in their many courses on science, maths,
[05:34]
and programming. The tutor doesn't give away the answers, but helps children when they're stuck, reminds them of what they already know, and explains each step to the solution. And the more you use it, the better it understands your personal struggles on how to get you unstuck.
[05:52]
My daughter loves it because in contrast to me, Brilliant's Tutor is infinitely patient and always has time. You can get started with Brilliant's Tutor for free and if you use my link brilliant.org/Sabine or scan the QR code, you can save 20% on an annual
[06:10]
subscription. So go and check this out. Thanks for watching. See you tomorrow.