We may finally have good customer support chatbots, so I imagine a reduction in the number of customer support.

That said I struggle to see it replacing jobs that require a few months of training. I work for a multi-billion healthcare company and most of the cost savings/human labour reductions is done through traditional software practices. An administrative error can be horrendously expensive.

Because OP is ok with paying for a TPU he just wants a flat fee per month.

> Surely the tensor factorization space could equally well been explored by other techniques such as evolution or even just MCTS.

I think you're underestimating the difficulty of exploring an enormous state space. The state space of this problem is bigger than the one in go or chess.

Reinforcement Learning specializes in finding good solutions when only a small subset of state space can be explored. You're quite right that Monte Carlo Tree Search would work here because that's exactly what they used ^ ^

> Similarly to AlphaZero, AlphaTensor uses a deep neural network to guide a Monte Carlo tree search (MCTS) planning procedure.

That said, you do need a good way to guide this MCTS, and a neural network is a great solution to evaluate how good a given state is. But then you've got a new problem, how do you train this neural network ? And so on. It's not trivial, and frankly even the best tools have quite some weaknesses.

But no, evolution algorithms would not be easier, because you still need a fitness function, and once again you can use neural networks for approximating it, but you run into training issues once again. As far as I know, evolution algorithms are just worse than MCTS at the moment until someone figures a better way to approximate fitness functions.