As with most of what Sabine writes, the book is very one-sided, in that it gives the reader the impression that a free-thinking person can only come to one reasonable conclusion, i.e. complete agreement with the book. It also gives the false impression that the field is much more single-minded than it actually is.

And honestly, even if you do agree with the book in every detail, "just make correct predictions, bro" is not at all helpful advice for actual physicists. There is a massive set of possible predictions, only a tiny fraction will be right, and everybody is already incentivized to look as hard as possible for that tiny fraction, in as many different ways as possible. You might as well show up at a floundering startup's office, lean back in a comfy chair, and say "just make more money, bro".

Sabine acts as if the correct approach is obvious, but on the very rare occasions where she actually does mention what ideas she thinks are promising, they're just as contrived or difficult to test as everybody else's, and often more so.

great comment, also something unrelated looking at your website I was hoping to find some notes about how to write in latex I have seen some submissions at HN but I think you could give some awesome tips

IANAP but my understanding is we've reached a point where it's very difficult to test anything in physics. Eg, we've picked all the low-hanging fruit and everything left is something that requires large expenditures to test (more powerful particle colliders, etc).

Is there some alternative direction available where testing and verification are easier and cheaper?

>Is there some alternative direction available where testing and verification are easier and cheaper?

A plurality of physicists now work in condensed matter. See e.g.

https://physicstoday.scitation.org/doi/10.1063/PT.3.4110

This is the right question, although I don't know if there is a happy answer. If we need ever more expensive experiments to learn anything, progress is going to become very slow.

Maybe that's why there is so much math; math is cheap. And there is always the hope of coming up with something testable.

And there are also at least two other somewhat older books in the same general vein: Not Even Wrong and The Trouble with Physics. To the comments in a couple of other posts, it's not clear to me why one would write yet another book on this topic unless they're also proposing useful alternative approaches.

This book is quite good. Suggesting we follow the historical trend where advances in physics almost always reduce the number of magic numbers in the model. https://www.amazon.com/Mathematical-Reality-Space-Time-Illus...

>when they can, fail the test

I'm not seeing why this is a bad thing. Every new idea we learn to test for, even if we test it and it comes back false, advances our knowledge and understanding.

Could it just be the case we are so far from our everyday experiences of the natural world that advancing further in our understanding requires taking small steps?

> wasted too much time and effort trying to come up with elegant math and grand theories that sound plausible but can't be tested

Einstein did not believe the gravitational waves he predicted could ever be detected. Or so I recall. But stuff like many-worlds interpretation appears fundamentally untestable (add same disclaimer here).

I wonder if it is the same thing that stalled AI research for decades? After the initial burst of hacks that seemed really impressive (ELIZA for example), the field focused on finding formal mathematical solutions to the problems and effectively stalled for 30+ years. It's only in the past few years with a shift to doing statistical things on large sets of data--something that feels a lot more like the hacky original efforts--that it feels like we're making any progress again.

This is from a outsider's point of view so it may be completely wrong. But I did try to take an AI course in the late 90s and it was so far divorced from computer hardware that I joked it should have been in the math department.

Neural nets had been in development since the 80s. The hardware became powerful enough to take advantage of the research quite recently.

AI today is still pretty divorced from hardware-level details, unless you’re working directly with GPU code. When I took a machine learning course, it was much more math than coding

CS has historically been associated with the math department at many schools. At many others, it's in the engineering school but that's by no means universal.

I don't really see any disconnect between CS and math.