This makes no sense.
Neural networks inherently work in a probabilistic latent space. The whole point is that we have a definable problem and we are unable to apply any deterministic mathematical calculations to it. That’s why we call it a “black box problem” - we know what the solution looks like and we have a few examples of input and output data, but we don’t know what the method looks like. In fact, even if we have an AI we don’t know what the method looks like. If you wanna know more about this issue, look up explainable AI and you will find all the ways we have tried to come up with smart solutions like model heat maps to estimate their methodology. But these are all approximations.
Now if you ask an LLM what method it used to get to a solution, it will try to work backwards using its own reasoning and context clues to give you what its best estimation of the methodology is but that’s not the same as its true methodology.
For me (and most people of the machine learning field I am aware of) an AI is inherently worse at calculations. That’s what algorithms are for. And the only things sitting somewhere between an LLM and an algorithm are maybe Markov chains and randomforest. If there’s anything remotely close to an AI model that is good at calculations, I am all ears.
Ok but then why call it AI?
We don’t have to dream about a distant future to find a thing that can do math for us. It’s called a computer, and if it’s in our home, it might even be a personal computer. In fact, we have them smaller and faster, and they have fancy designs and can communicate to the world but it turns out they got a small chip inside that does a bunch of arithmetic binary operations and then a lot of them also have a chip inside them that can quickly operate on huge matrices of values. We call those cpus and gpus respectively.
What I’m saying is, if it’s just a threshold of different values, we were capable of making such a machine already, even before AI. And if it’s an interpretative opinion, it turns out AI is notoriously bad at those.
So the only thing an AI actually helps with is to do something when we don’t know how. But the issue was never how. Now there’s something to be said about the predictability of death but there’s an insane amount of compounding factors that contribute to this. An AI could maybe be trained on a subset of factors but the truth is, we don’t have good input and output data, that’s why AI won’t help us here.
I’m not saying this to be pedantic, I’m making a point: even IF we decide that actually ethically it’s fine for a machine to decide when someone has to die - which is quite a big leap - even then an AI won’t be better than a human, so there’s no point. And there’s no way to make this look good, no way to sell this to the broader populus. I just don’t see any way where this scenario is possible, let alone probable.
And I think that’s important, because it’s very easy to slide down this slippery slope, interpolate a time series forecast in our head and essentially conclude “it’s going downhill so the future must be what our average dystopia predicted” and I don’t think that’s right or even helpful. There’s quite a few things we should improve and our world is on the brink of a bigger societal change again, but change is never just bad. Change is an opportunity to improve the things that are important to us. And to do this, we need all the brilliant minds that see those issues to imagine those improvements, so we should make sure we know what needs to be done when we finally get the chance.