We do not have a good model for the structure of water.
What the author meant was that nothing that will be discovered in this domain will shake fundamental physics. Things that will shake fundamental physics won't come from 'everyday life', but rather from extreme environments like those found in particle accelerators, and through study of gravitational waves, CMB etc.
And you don't know that a priori. It's possible that with a hightened level of simulation, and accurate measurement of an unusual phenomenon, some fine structure emerges. I'm not going to hold my breath, but sometimes you are surprised. The yellow color of gold is a consequence of relativisitic effects, so it could have been entirely possible (if seemingly rather unlikely) that humanity discovered relativity from that angle instead.
But nobody said it was 'a priori'. It comes from thousands of experiments which probed relevant energy levels, over a period of decades.
>It's possible that with a hightened level of simulation, and accurate measurement of an unusual phenomenon, some fine structure emerges.
Sure, in fact we know the Standard Model is incomplete because it doesn't take into account dark matter.
If everything was already answered, then why offer $1M to solve an equation? Fun?
Impact on what? On the standard model and QED?
It isn't a contradiction to point out that we seem to have a very, very good model of the universe's most fundamental particle processes, but still can not simply derive from that model to even very simple things of interest. In principle, we could "just" simulate the standard model and "just" simulate a lot of water molecules under the conditions of interest and "just" watch what happens. In reality, that turns out to be many, many orders of magnitude more computation than we can do.
See also protein folding, where even without trying to run things all the way down to the standard model, things that we in principle fully understand require gobs of computation to approximate for what are in the real world simple things happening gazzilions of times per second in every living organism.
In the other direction, witness the difficulty we have trying to improve our understanding of how the universe works at the Plank scale. One imagines that if the smallest things we had were not multiple orders of magnitude larger than the scale we are trying to study that perhaps we'd have made more progress by now, but it's like trying to determine the rules of biology when the only capability you have is the ability to sling planets at each other really hard and see what happens. The miracle is that we've learned anything at all.
On that note, I've often wondered if the standard model really "works". Suppose someone with a lot more simulating power than us actually programmed it in directly and set it running, with a reasonable initial state that resembles our world. (My understanding is that inflation doesn't necessarily come out of it so we can guess that you can't just start from the big bang.) Does it really work? Is the result indistinguishable from the real world without a particle accelerator? Or is there some flaw that only comes out at scale, or does the universe crash if you create a black hole? How close are we, really? If someone waved a magic wand and said "Shazam, the standard model is now what the universe really and truly runs on!", would we notice anything different? Or would we be instantly dead because it turns out chemistry doesn't quite work right? Or perhaps at all?
Thinking about your final paragraph, I guess it gets very much at Carroll's point. I expect that he'd claim we wouldn't notice any difference at all: if we would, that would directly imply that something beyond the standard model was necessary to understand everyday life. But as he explains, we have every reason to believe we've exhaustively explored the relevant range of energies and interaction strengths. If something about chemistry required physics beyond the standard model, we would almost certainly have noticed by now.
"Playing Go", which does not fit this rubric obviously, I think, is conquered by deep learning because there is a discrete matrix of 'moves' which looks awfully lot like an image recognition matrix, and so a move which optimizes score ('classification') can be selected.
Fundamentally one of the problems is that we don't know how to score proteins. We also don't necessarily know how to score go, but it's trivial to generate solutions by running hypothetical matches because the game outcome is deterministic, solvable, and definite. While there is a decent subset of proteins whose known structures could be used to seed a database for scoring, it is very small in comparison to the universe of possibilities (20^n, where n is the length of the amino acid chain).
In short: Given an arbitrary "end-board layout" I can with our existing human knowledgebase tell you which go player has won. Given an arbitrary amino acid sequence and a proposed structure, I cannot with our existing human knowledgebase tell you if the structure is correct. Because of this limitation in creating a scoring system, while I won't say that deep learning won't be used to improve protein folding (maybe it will, indirectly) it almost certainly won't be used directly, because it is not in the same class of problem.
Im a guy that took graduate thermodynamics from the guy who discovered the statistical mechanics of the hydrophobic effect. There may be only 100 people in the world who are qualified to even describe water's molecular structure at varying conditions.