Jonathan Birch

Yeah, these are very hard cases. I suppose when I started writing the book around 2020, not sure the large language models were even on my radar at all. And then they've jumped onto everybody's radar through things like ChatGPT. And I suppose I've been on a journey like everyone else during that time.

Yeah, these are very hard cases. I suppose when I started writing the book around 2020, not sure the large language models were even on my radar at all. And then they've jumped onto everybody's radar through things like ChatGPT. And I suppose I've been on a journey like everyone else during that time.

I initially thought, well, these are next token predictors and the sector has been moving away from brain-like forms of organization. So it's been taking out things like recurrent processing that on many theories of consciousness are absolutely essential, but transformers take that out. So I thought, well, here is something that is conspicuously unlikely to be sentient.

I initially thought, well, these are next token predictors and the sector has been moving away from brain-like forms of organization. So it's been taking out things like recurrent processing that on many theories of consciousness are absolutely essential, but transformers take that out. So I thought, well, here is something that is conspicuously unlikely to be sentient.

But then I suppose I'm not sure that's the correct view anymore, I suppose, because I've been quite astonished by the feats of reasoning they seem to perform today. where it's, well, it's reasonably evident that we do not understand how they work. They're incredibly opaque to us. We don't know how they do what they do.

But then I suppose I'm not sure that's the correct view anymore, I suppose, because I've been quite astonished by the feats of reasoning they seem to perform today. where it's, well, it's reasonably evident that we do not understand how they work. They're incredibly opaque to us. We don't know how they do what they do.

And there seems to be some element of acquiring algorithms during training that were never explicitly programmed into them. So in a way that architecture that was programmed into them, the transformer architecture, no reason at all to think that would be capable of sentience.

And there seems to be some element of acquiring algorithms during training that were never explicitly programmed into them. So in a way that architecture that was programmed into them, the transformer architecture, no reason at all to think that would be capable of sentience.

But when you have these very, very large models where they've acquired algorithms during training, we don't know how and we don't know what they are. We don't know the upper limit on what algorithms they might acquire. And we don't know what algorithms are sufficient or not for sentience. And so we're not really in a position to be so sure anymore. that they couldn't acquire those algorithms.

But when you have these very, very large models where they've acquired algorithms during training, we don't know how and we don't know what they are. We don't know the upper limit on what algorithms they might acquire. And we don't know what algorithms are sufficient or not for sentience. And so we're not really in a position to be so sure anymore. that they couldn't acquire those algorithms.

So, for example, if you think a global workspace is what it takes to have sentience, as many have suggested, we don't know that they couldn't acquire a global workspace.

So, for example, if you think a global workspace is what it takes to have sentience, as many have suggested, we don't know that they couldn't acquire a global workspace.

Well, this is Stan De Haan's theory. His book Consciousness and the Brain is a nice exposition of it. But it's this quite popular idea that consciousness has to do with a network that puts the whole brain on the same page, as it were, by taking inputs from many, many different sensory sources and integrating them into something coherent and then broadcasting that content back

Well, this is Stan De Haan's theory. His book Consciousness and the Brain is a nice exposition of it. But it's this quite popular idea that consciousness has to do with a network that puts the whole brain on the same page, as it were, by taking inputs from many, many different sensory sources and integrating them into something coherent and then broadcasting that content back

to the input systems and onwards to other systems of motor planning, reasoning, etc. So it's the the bit where you know, the central coming together of everything in the brain. And well, of course, that is designed as a theory of consciousness in the human brain. But the basic architecture

to the input systems and onwards to other systems of motor planning, reasoning, etc. So it's the the bit where you know, the central coming together of everything in the brain. And well, of course, that is designed as a theory of consciousness in the human brain. But the basic architecture

where you have lots and lots of input processes competing for access to this workspace, where once a representation gets in, the integrated content will then be broadcast back and onwards. There's nothing about that architecture that is inherently difficult to achieve computationally. And so we did a big report on this last year, 19 of us.

where you have lots and lots of input processes competing for access to this workspace, where once a representation gets in, the integrated content will then be broadcast back and onwards. There's nothing about that architecture that is inherently difficult to achieve computationally. And so we did a big report on this last year, 19 of us.

It was led by Rob Long and Patrick Butlin and had some top AI experts in there, including Yoshua Bengio. And our conclusion was there's no obvious technical barriers for why AI might not achieve something like a global workspace in the near future.

It was led by Rob Long and Patrick Butlin and had some top AI experts in there, including Yoshua Bengio. And our conclusion was there's no obvious technical barriers for why AI might not achieve something like a global workspace in the near future.