The first time I heard of Lojban was about 10 years ago, and since then attempted to learn it a few times, but I kept switching to something else after a day or two. Recently I've finally went through la karda, a few crash course lessons, wave lessons, and started reading la snime blabi. Decided to write the first impressions here, along with relevant links.

Together with the lessons, it's useful to check the sentences in a parser such as camxes (there's a few versions, the linked one seems to be more up-to-date than others), and there is vlasisku to translate lujvo (compound terms, a basic stardict dictionary won't handle those). I'm also using emacs lojban-mode, and stardict.el with a lojban dictionary, which is a handy combination:

A Lojban text in emacs-w3m

A reference (by the same John Cowan who work{s,ed} on XML, Unicode, R7RS, and other things) is quite helpful, as is the #lojban channel on Freenode/Libera.chat.

Pleasant bits and expectations

At first it seemed like the language mostly consists of selbri, or verbs, or relations – which can be combined into tanru (used as adjectives), turned into sumti (nouns), or modified in a few ways. It looks uniform and simple.

There's a PEG, which is nice too: one can parse the language unambiguously, and following a set of rules. It also allows to implement structured editing, and observing the parse tree helps to learn the language.

Time and space constructs, as well as logical connectives, are designed, hence are not as chaotic and messy as in natural languages.

Morphology is nice and helpful. Plenty of good things, actually.

Unpleasant bits and disappointments

It's not quite minimalistic or elegant: the linked PEG file is 84+ KB, for instance. There's a lot of cmavo, and no [semi-]formal rules by which they are defined (hence a lot of special cases): neither syntax nor semantics follow any basic rules, it seems, making it to look and feel much more like a natural language than like a constructed one (as I would imagine such a language before, at least). Likewise with logic (connectives, quantifiers) in it: though it escapes ambiguity, it appears to be designed to resemble common constructs of natural languages.

Maybe it all makes more sense to linguists, and maybe it would be awkward to use selbri instead of most of the cmavo. Or it's just a bunch of arbitrary decisions, as in most of the programming languages, where users are not expected to learn and follow the rules strictly, but rather to read and write things intuitively.

Besides, there are informal dialects with relaxed rules, apparently, which also seem to work like a natural language evolution, eliminating the potentially nice properties.

Overall impressions

Probably I'll continue learning Lojban someday, though being less enthusiastic about it now.