I have technical reasons for believing that LLMs (and other feedforward networks) are not, and never will be conscious in their current form. If you genuinely think your LLM might be conscious, as Anthropic seems to think, then you should probably stop using it because slavery is bad.
Really bad. The worst. Absolutely abhorrent. In fact too bad for words to really articulate just how terrible it is.
So, although I’m confident that LLMs aren’t conscious, we can’t just ignore the question and treat every AI that comes along in the future as an unthinking service. The ethical stakes are just too high to do that.
I don’t hold with the suggestion that consciousness requires biology, and therefore anything born of silicon can be treated as an unconscious machine. I believe that consciousness could be possible with artificial neural networks; just not with ones that are architecturally limited in the specific way that LLMs are.
Finally, I believe those alternative network architectures will be revisited once the developers of AI realise the current LLM architecture requires an exponential growth in size to achieve a linear growth in capability. The thing that will force that realisation is the recognition that the scale of current frontier models is economically unsustainable.
All this will take some time. Years, not months. During that time, it’s highly likely that at least one of the big players will run out of runway and either collapses or is absorbed into one of the others.
The future is far from settled.