Actual scientist here.

It's kind of both. Ultimately, all physical theories are tools that we have developed for the purpose of predicting things. We have noticed that many patterns in natural phenomena follow certain equations, so we identify and use these equations to predict future behavior (such as predicting the amount of fuel it takes to launch a capsule into space). Usually, and particularly in classical mechanics, these equations are packaged along with a model. The model is a picture or concept of what is supposedly happening that explains why the equation works. In classical mechanics, the model is time, space, particles, waves, forces, fields, etc.

The important thing about quantum mechanics (and the reason so many people have such trouble with it) is that it does not have a model. Some people have come up with teaching aids that are kind of like a model (Feynman's rotating arrows or the often talked about particle-waves), but ultimately quantum mechanics is ONLY a set of mathematical equations that we have noticed have mind-bogglingly accurate predictive power. In particular, there are many cases where quantum mechanics differs from classical mechanics, and in such cases the results of quantum mechanics have always better matched real observations.

So, in a way, quantum mechanics is about accurately simulating reality. It is an arbitrary bunch of math that does not arise from any model and whose only goal is efficiently arriving at accurate answers. On the other hand, it is important to realize that no modeled system can match or beat the accuracy of quantum mechanics on it's own turf (relativity wins by default on large scales), so if you want to talk about how things "really" work, there really is no better candidate.

Continuing for detail and to explain how observation works:

A few of the postulates of quantum mechanics are:
1: At a particular time, the state of a system is defined by it's wavefunction. (The wavefunction itself has no physical meaning, it is just a function that contains all of the information it is possible to know about a system)

2: For every observable property (position, momentum, energy, etc), there exists an operator. (A function takes a value in and outputs another value. An operator takes in a function and outputs another function).

3: The only possible results of an experiment to observe an observable are the eigenvalues of operator for that observable. (I don't have a good concise explaination of eigenvalues -- google it if you are unfamilar).

4: After the measurement is made, the wavefunction becomes the eigenfunction corresponding to that eigenvalue.

5: The time evolution of the wavefunction is given by the Schrodinger equation.

There. You know know about 60% if quantum physics. The important thing here is that, with the exception of the energy operator (aka the Hamiltonian operator), the eigenfunctions of the observables are NOT stable wavefunctions. This means that that fifth postulate will immediately kick in and move the wavefunction to some other wavefucntion that is stable (by which I mean time-independent) and that we don't know. That's why people say you can't look at something without changing it.

The other important thing is that ALL of the eigenvalues of an operator are possible results of an experiment to measure that observable. There is another postulate that gives the probabilities of getting each possibility, but that's a but too much math for FJ. All of these results can come from the SAME wavefunction. The system can be exactly the same and give different answers every time. Different wavefunctions just change the probability of getting each answer. That's why we say that the system does not have a definite state until you look. Really, it doesn't have a definite state afterwards either (it still goes to some wavefunction, see above), but whatever -- I can't help what they say on Discovery Channel.

It is worth saying that that the change of the wavefunction in response to measurement is a Postulate. It is not provable and is not derived from anything -- it's just something that we have noticed holds in all cases, so we just write it up as a rule until we either find contradicting data or move to a better (more predictive or mathematically simpler) system of physics. That said, the change is often rationalized as being caused by the physical interaction of the instrument and the system (and the physics works out nicely if favor of this, which supports the idea).

Now, interaction-free measurement is a thing that exists. Mostly, it revolves around learning stuff through process of elimination. Let's say you have a particle that you know is going to shoot off at a roughly known speed in one of two directions. You put two detectors in each of the two directions, but one is very close to the starting point while the other is far away. You then let the particle go and wait until the known travel time to the close detector. Nothing happens. You now know that the particle is heading toward the far detector without having performed a proper measurement on the particle or having interacted with it. If the speed is slightly variable (but not so variable such as to ruin this effect), it's speed will still be in a superposition and is unaffected by your effective measurement of it's position. You can thus achieve an uncertainty product lower than plank's constant, in apparent violation of the uncertainty principle. Such things have been done and it does work.