One of the biggest changes in Blender over the last two years has been major changes and reorganization of its user interface. The goal was to address Blender’s glaring usability issues, and indeed many of them have been addressed. Yet Blender remains a dauntingly difficult program to use, and — worse — a difficult program to remember how to use. I’d say that to me (and I am hardly an expert user of any) Blender is now still somewhat harder to use and more disorganized than, say, Modo (I have Modo 601) or Cinema 4D (I have v11 XL) but that’s not saying much.
All serious 3D applications, and Blender unquestionably qualifies as a serious 3D application, are plagued by the intrinsic complexity of what they’re trying to do, combined with a lack of agreement on the “right way to do things”, and the fact that real technical improvements are constantly being made in 3D graphics, leading to proliferation of overlapping tools and terminology (consider that there are multiple implementations of subdivision surfaces — such as Catmull-Clarke and Stam-Loop — and it’s important for a 3D program to support as many as possible if it’s to be compatible with other 3D programs. Similarly, there are many different tools for modeling organic objects and curved surfaces, including NURBS, subdivision surfaces, deformation lattices, and multiple approaches to sculpting (including voxels and displacement).
The “Right Way” isn’t obvious
If you go back to the era of DOS word-processors there was a similar lack of agreement — Word Perfect, for example, had a completely different paradigm for editing and formatting text than did Microsoft Word, and on the Mac side MacWrite and FullWrite used a “ruler” formatting paradigm while WriteNow and Microsoft Word used “styles” (but each wrapped them in a unique user interface and terminology). Pages and Microsoft Word, by comparison, are virtually identical in how they solve most problems, the differences coming down to nuances. The Mac toolbox (along with excellent free software such as MacPaint and MacWrite, and — eventually — written Human Interface Guidelines) standardized basic text editing operations (selection, inserting the caret, clipboard operations, and undo) but it took close to ten years for the world to settle on the next level of convention — word and paragraph styles with “spot” overrides. (And indeed, Windows still hasn’t quite settled on clipboard or text edit field behavior standards.)
And then let’s return for a moment to the point that — unlike with word-processing — almost every aspect of 3D technology is constantly churning. With word-processors and even page layout programs there have been some technical improvements to be sure (e.g. typefaces have become more sophisticated with things like multiple master fonts and automatic ligatures, there’s Unicode of course, and compositing has improved markedly) but these have been relatively simple changes conceptually, and in general they have just made things easier (e.g. instead of having to render a drop shadow in Photoshop, export it as EPS, and then embed a reference to it in Quark XPress, you can now perform the entire operation in a click or two in your page layout program or word-processor). Contrast this with the world of 3D which has seen  major new technologies emerge in pretty much every piece of core functionality from modeling to rigging to texture-mapping to lighting to rendering — and all these different technologies have to live side-by-side, inter-operate, and somehow fit into a single user interface.
As you’ll discover if you visit Blender’s website right now, Blender is now 20 years old, and that means it has 20 years worth of functionality clogging up its user interface. It supports NURBS and subdivision surfaces and multi-resolution displacement sculpting. Similarly it has support for rigging using bone envelopes, vertex-painting, and heat-mapping. It has three built-in renderers. It has two different node-based shader architectures (or is it three?). I’m sure you’re getting the picture.
Left-Click to Select
Blender remains unusually bad at a fundamental level because it refuses to adhere to such cross-application conventions as do exist. The most hilarious example is left-click to select. It’s almost impossible to find another application anywhere that uses left-click for something other than selection (unless there is no concept of selection at all). It’s simply bizarre. What’s worse, left click is used to position the 3d “cursor” object — which is actually a very clever concept — except that it’s a far less common operation than simple selection, and doing it by accident is really annoying.
(If you consider the behavior of a text cursor, clicking somewhere in a body of text positions the cursor rather than selecting anything. The cursor position determines where stuff you type will appear. This is the exact behavior of the Blender cursor. But it’s not the behavior of any 2D or 3D graphics program I’ve ever seen. But to select text you left-click and drag or double-click — this is definitely not how Blender works, so justifying Blender’s weirdness by analogy to text doesn’t get you very far.)
Now I need to note that Blender does allow you to make left-click into select as a preference, but the program really doesn’t work properly in this mode (and you need to reset the preference when you upgrade Blender) so that’s not a satisfying solution. Just to give you one example: there’s a shortcut in Blender to allow you to “lasso select” using ctrl+left-mouse (why left-mouse when selection is normally with the right mouse?) — this becomes ctrl+right-mouse if you’ve set left-click to select. What I want is to select stuff with the left mouse button. Not flip a ridiculous set of defaults and create a different ridiculous set of defaults. On the whole, I end up sticking with right-mouse to select simply because it means not having to reverse engineer every tutorial I find.
If it were up to me, selecting something (with the left mouse button) would also, by default, position the cursor at the selected object’s local origin (sometimes referred to as its “pivot”). You could explicitly position the cursor with the right mouse button (or perhaps setting the cursor could be a mode or tool), and there would also be an option to lock the cursor (many 3d programs effectively have a “cursor” locked at the origin). This would actually maximize the conceptual mapping between the 3d cursor and the text cursor — left clicking sets your insertion point – and not compromise users’ familiarity with programs other than Blender.
More Annoyances
Blender’s improved user interface — and it is definitely improved! — has two more major annoyances from my point of view. First, contextual tool panes have been added to the 3D view (which is great) but the process of showing them and hiding them is asymmetric. To disclose them you click on a little + widget — which then disappears. To hide them you either need to learn the keyboard shortcut, or use the mouse to scale them down to nothing (which causes the little + widget to reappear).
Second, with the advent of the new Cycles renderer, to get a live-rendered viewport you need to switch to the Cycles renderer (one step) and then in a completely different place you need to set the viewport mode to Rendered. For your trouble you are rewarded with a wonderful live-rendered view that gives you no UI affordances (e.g. you can’t tell what’s selected) but remains live for editing purposes. Ugh. So to make this workable, you then need to create a new viewport and tweak its settings.
Finally, to modify the quality of Cycles renders — and spent quite a bit of time googling to figure this out — you need to delve deep into the rendering panel and — slightly hilarious — the place where you determine how many samples to obtain for a render is the only place I know of in Blender’s interface where the preview settings are at the bottom and the final render settings are at the top. If you’re used to Blender’s “conventions” everywhere else (e.g. subdivision settings) it’s always the other way around.
These are most of the first order problems in Blender today (as of 2.69). Clutter and ambiguity are also first order problems — I’d suggest that a lot of that could be improved simply be making a decision as to which slabs of functionality to hide by default. I hope we can some day get to second order issues, but on the positive side most of the new stuff getting added to Blender seems to have a much higher standard of usability than older features. E.g. to do a physics simulation you simply click on the Dynamics tab, select your objects, assign them some properties (e.g. passive rigidbody) and then click the play button in the timeline panel and it just works. More of that please! Similarly, the node-based material system associated with Cycles is very easy to work with (indeed, you don’t really need to use the nodes at all) although right now it’s missing a lot of functionality.