Read classic rigorous pedagogical authors. I like Serge Lang from the older generation, Terrance Tao does similar things today on his blog. Lang is good for all the simple things, his generation formulated mathematics within set-theory from the ground up, in the Bourbaki project, which essentially should have been called the "Lang/Grothendieck project". Tao is good for more modern things, because he usually blogs as a prelude to some original work.
Read ALL the books on a subject, there is always one with a different good presentation that everyone else forgot or standardized away. For example, the "bisection" proofs of the intermediate value theorems and mean value theorem are good, but modern proofs introduce point-set concepts like connectedness and compactness. I read these in 19th century calculus textbooks that were collecting dust on shelves, but they made the basic ideas clear. For more advanced stuff, Milnor is good for geometry of the mid 20th century.
It is always best to read literature about a discovery from the era when the discovery was made, not the later "simplifications", and "generalizations". The generalizations first obscure the basic idea, which is always clear originally, because someone had to come up with it, and then the simplifications don't simplify the idea, they simplify the generalizations, which are obscurantist. This is how math books become opaque, through wrong-headed too early over-generalization, and simplification of the generalization. Don't worry about how simple a proof is to read, worry about how straightforward it is for the brain to understand.
When both criteria are met, when you have a simple, general formulation, you have what Erdos called a "book proof", a proof from God's book of theorems. Never be satisfied until you know a book proof for the theorem in question.
So always, always, read the historical literature. It is much easier than the modern literature, because people back then were stupid and ignorant (not their fault, all the interesting stuff was discovered later), and you can become fluent in it more quickly, This is good practice for becoming fluent in the modern literature.
You need to learn to "unpack" proofs into the construction that is involved, to know what the proof is saying really. It is no good to memorize the proof, you need to understand the construction, and this will motivate the proof.
I will give an example of unpacking: consider the proof of the Jordan curve theorem. This is proved in modern books always in the same way: by noting a paradox regarding some homotopies (I forgot the details of standard presentations, but I remember the idea, I unpacked it). The proof is obscure, so much so that mathematicians consider it difficult! They tell students it's a hard theorem. It's not hard, it's trivial.
First, you should prove it yourself in the differentiable (or piecewise linear) case, by using the original demonstration from the end of the 19th century. If you pick a random line from a point, the number of intersections with the curve is even or odd, according to the insideness or outsidenes. If you cross the curve, obviously this changes by 1 unit, and if you count intersections by their "sense" (the orientation which they happen by), you can't have a number of intersections which is different from 0 or 1 without the curve having a self-intesection, as you can see by turning the line by 360 degrees, and seeing how the intersections meet and annihilate (they have to come back to their original position at the end). This is a sketch, but it's easy to fill out to a proof in the differentiable or piecewise linear case, and the singularity sliding method is a baby version of more sophisticated later constructions in higher dimension using Morse theory, due to smale, which proved the higher-dimensional Poincare conjecture.
Why doesn't this work as a modern proof? Because the theorem is also true for continuous Jordan curves, which can be very wild, they can have positive Lebesgue measure (another easy 19th century construction you should do for yourself, it helps to know how to construct a space-filling curve). So you want a proof that works for continuous Jordan curves, where the number of intersections with a line is generically infinite.
The modern theorem is proved using a general method involving homology groups which look complicated, but only because they are generalizing to arbitrary dimension and to a general formalism for obstructions. If you read Munkres' proof, the only computation involved is of the winding number of a map from a circle to a circle, which is an integer, which tells you how many times the map went around the circle.
So the actual proof is just a simple winding number construction. What is it? It turns out that the winding number of a Jordan curve around a point can be easily seen to be either 0 or 1, depending on whether the point is "inside" or "outside", and this can be easily related to the differentiable case proof, because this winding number changes in the proper way when you go along a line and pass an intersection (this is easy to prove). So this is the generalization for the continuous case.
You can then explicitly prove the Jordan curve theorem in an ugly way using winding number, and construct the winding number yourself, using your favorite homegrown method. This will make the proofs in the textbook obvious and intuitive, although they will be annoying, because you will think they are obscuring something simple for no good reason.
To generalize to higher dimension, you need to learn how to define homology, so that you know the abelian notion of sphere-winding. This is NOT equivalent to sphere-onto-sphere homotopy classification, but it's the same thing when the spheres are equal dimension, and it's the "right" thing to study anyway, in that it is more regular, and any computation of homotopy proceeds through homology in any case. When you are finished, you have an ugly personal proof, but this is not the main goal. The personal proof has made all the literature stuff clear, because you see it is just standardizing the personal proof so that it can be applied without thinking to a large number of cases, and in a way that is completely standard between different authors.
This is the thing that makes mathematics difficult. The ideas can only be discovered by a personal process of ugly construction and half-baked personal proofs, but the final result is an elegant machinary, that you can learn in a half-assed way by studying the formal proof, without understanding any of it. The goal of mathematics education is to force you to break down and reconstruct all the theorems for yourself.
The easiest way to do this is simply to explain theorems to others. You can do this through teaching, you can also explain it to yourself, close the book, and present the theorem on your own, with no notes.
The mathematicians hide their concepts through this mechanism, but they expect each other to unhide their theorems, by reconstructing it themselves. The mathematicians have some historical presentations too, which help with things that have acquired modern obfuscation.
Also, you shouldn't bother with some theorems which are just make-work. For example, in knot theory, there is a notion of "isotopy" which is difficult to make precise, but whose only purpose is to prove that the reidemeister moves are fine for computing knot-motions. You should learn how to do the differentiable case, because it is easy and is the motivating thing, but the generalizations to the continuous case are not so insightful, and generally serve as make-work for mathematicians who are at the moment unable to find a new idea.
You also need to get over the hump of the political bullshit. So you need to learn infinitesimals, constructive/Soviet mathematics, ordinal analysis, and all the secret things that are politicially hidden. But this doesn't take long. You also need to trust your own intuition, because it is easy to snow a person with a lot of complicated symbols that don't mean anything. If you are intellectually honest, it is easy to ask questions and figure out what the gibberish means, and then it stops being gibberish.
Arxiv rejects cold fusion papers automatically. I will write up a formal paper soon, I need to do a bunch of calculations, for now the informal description is sufficient for priority. The calculations require some imulation for the nuclear rates, a potential model, and some estimates for the full spectrum of isotopes produced. But I know it works qualitatively for sure, and with no fudging, so this is why I am confident. It's easy to be confident with a phenomenon this crazy, because once you find an idea that sort-of works, and you know the experiments are correct, then it certainly really works, because the phenomenon is so crazy, there aren't going to be two different ways it can happen.
I am working on a bio project now for money that is taking all of my time (and more). The rest is eaten up by my relentless egotistical self-promotion here.
But yes, it is possible to make an impact, please share your ideas, it will not hurt you, the world is fairer today.
With Weinberg and Polchinski you chose the right mentors! But Weinberg might be hard for an intro because he leaves the path integral too late. Try Zee for a more path-integral heavy intro, Weinberg for more advanced methods that are arcane (Weinberg is a classic, great book, but the presentation is a little old fasioned). It is not too bad, you can zip through it if you have the time to devote yourself full time. It also pays to read the literature, if you have a library with Nuclear Physics B, flip through all the articles from 1969-1993 (the rest are on arxiv) finding the ones that you like. The same with Physical Review, but it's huge, and Physics Letters, Reviews of Modern Physics, JETP and JETP letters, Nuovo Cimento, these are where all the good stuff is.
Oh that's good too. I don't know why you're complaining. Try to get an institutional subscription, sneak into a library. You need Nuclear Physics B from 1970-1993 without any question if you want to do anything. It's on the shelf at most libraries, but it's also online, you just can't read it. It's a bummer.
If you can do the exercises you are getting a better education than most people who take the class.But field theory is only part of the story. Try to pick up Gribov's "The Theory of Complex Angular Momentum" for Regge theory, and this will allow you to read the Regge theory stuff that is foundational for the strings.
But also, there's condensed matter. There are good books here, but I found reading Anderson's collected papers was the most educational. He is phenomenal, everything he wrote is gold. There is also Parisi, Kadanoff and Wilson, and a bunch of people, I liked the turbulence papers, my old advisor Siggia has classic papers here, but there's also Kraichnan and there is a wonderful toy model by Forster Nelson and Stevens based on the Martin-Siggia formalism. You need to know that stuff, it is real physics, but it becomes easy with a path integral. But black holes is completely different physics, it's all Susskind and Vafa, Strominger, you know, the string theory guys. It takes a little while, but it's great stuff. If you find out some stuff, tell me, please.