I'm going to go out on a limb and say that this entire conversation is... dumb. The concept of "sentience" is so poorly defined, with barely even a consistent definition let alone a rigorous test, that for some reasonable definition I think we've all met even
humans who are so flat or irrational that its debatable if they'd qualify as possessing sentience. It's basically become a word that we use to mean "oh that thing is special", whether we're talking about a human, cat, or robot in a movie. It's not useful and I think we should stop using it.
Here's a couple thinks I have off the cuff:
1) We phrase cognition as "input, process input, output", which is obviously a simplistic take but I think even at a simplistic level it is severely underestimating biology. The fundamental structure of our brain changes in response to the input, and the output is formulated based not only on the input, but also the structural biology changes created from that input, in real time. It seems like this is a gap in AI, where the training is generally front-loaded, independent of the input and output. Its not just: "here's a picture, insert that into memory, now what do you think about this input". But its ALSO not just: "here's a picture, insert that into memory, here's an input, integrate that input with your memory, ok what do you think about the output". It's even further than that: more similar to: here's a new picture, now subtly change what "think" even means in response to it, now what do you "think" about that? Maybe there's a model of AI which does this, but it seems like a gap; and a potentially critical one.
2) Most "training" in AI is an extremely expensive, large process; requiring hundreds of pieces of coordinated silicon across a data center, or even multiple data centers, before you get this convenient little "package" at the end that can actually reasonably run on hardware that more resembles the physical size of a human brain. Consider, for a moment: the speed of light; or more accurately, electricity, which I know I know doesn't have a "speed" but let's keep it simple and useful. Throughout the history of animals, all of the brains biology & evolution have produced are, roughly, the same size (within an order of magnitude); and all substantially smaller than a data center. Why? Who knows, but an idea I find tantalizing is that the speed of light, cause-and-effect of the electricity which powers The Brain and Computers: matters. It could be that coordination between neurons is so difficult that it can only happen if messages/electrical activity can move between the coordinated components extremely fast; so fast that evolution eventually settled on, eh, about six inches, maybe a foot or two but there will be sacrifices. If we eventually figure out that the input-processing and training components of an AI model have to be one continuous process, like described in (1), we may hit a second problem: our silicon is such a poor analogue for brain biology that we overpower the problem with "more distributed silicon", but that distribution is a critical blocker for generalized cognition to evolve. Its like virtualizing x86 Linux on an M1 Mac; there's going to be loss, and what if that loss is actually the Spark necessary for higher cognition?
3) Two other humbling observations I think AI engineers need to recognize. First; per the second point, we may need Better Silicon; something which can be a better analogue for brain biology. But we REALLY do not understand the brain; how the brain's structure and processes coalesce to form higher thought. Near-totally beyond medicine right now; and it seems unlikely to me that breakthroughs in this domain will happen in AI/CS, but rather in Medicine itself. But, let's assume they do; I think many of these researchers don't respect Conway's Law enough. Assert that our plan, Google, OpenAI, whoever, is to build a Human In A Box. Conway's Law asserts, nearly without fail, that software systems mirror the organizations who build them. We're at an impasse; I actually don't believe its possible for Corporate Engineering Teams to build a Human-In-A-Box, because anything they build will inevitably mirror themselves; a corporation. Maybe more wishfully; if a major innovation in this arena happens, it really feels like the news article announcing it will have the line "genius in her garage" somewhere in there; not "multi billion dollar well funded AI research lab at an advertising company". More realistically, the problem domain is so complex, requiring broad, deep knowledge on insanely disconnected topics, that it may never happen.