Social classes are conventionally mapped to easily observed factors like wealth and social status (e.g., certificates of educational achievement).
But I think we could also examine the productive work that people perform, and derive classes from that. This produces a different distribution than the more material or social class categorizations. Such a classification by labor rate can group billionaires together with ordinary and even low-means individuals. For example, does Elon Musk tap in Peter Thiel's phone number, that he memorized, into his cell to call him? Or does he look Pete up in his Contacts lists, or is Pete in his Favorites? Or does Elon use voice commands like "Call Thiel"?
As technological automation progressively distances users from the time+energy necessary to accomplish their goals, classes that seek automation, or that have automation forced on them, do less work by themselves, and this eventually interferes with the accuracy of their conjectures about how long and how much energy/effort is involved in a given task. For example, generative LLMs leverage the manual labor applied to add definitions to individual words in their training dictionaries. This labor is then eventually reused very many times within the generator's perceptron layers and attention matrices, and those products resold.
The chat LLM is essentially performing calculations whose output is (at least nominally) the equivalent of finding an expert to fulfill the input prompt; here the expert does some work everything they're asked a question. Since the same question can be asked many times, it makes sense to caches the expert's responses, reusing their work. Caching and cache lookup-and-retrieval are themselves energy-consuming processes, but they (hopefully) use less energy than a novel search for an expert and the expert's effort. But current chat-type generative LLM services *don't do much caching*, because they're already spending on hardware for their processing. Without observability, which is to say, tracking original source meta-data thought each of the thousands of transformations performed along the way, caching would be wildly inefficient. But maintaining the meta-data back-references also massively augments generative processing.
Older raw-search technologies performed these functions much more efficiently, because they distributed the total work more evenly. A user would have to do some thinking to compose an effect search query, and, on success, their results would be an expert's opinion, one which the expert had already produced.
So, which net-labor class are you in? Maybe you're in a "higher" class than you thought. The problem with Elon is that he leverages his wealth to stay stupid; the ultimate luxury, but his judgement is visibly impaired (his trivialization of unassisted vehicle piloting a case in point).
My concern is the people in positions to develop governmental and industrial (and other) policies are more or less of this same class. Their minds are mush, their best intentions thereby malformed, unrealistic and unrealizable. But to people in a similar tech-automation class, this may not be recognizable.
Grimm 
Shoshana 🏳️⚧️
Phil Giammattei
Alice Averlong🏳️⚧️
Molly White
Anne Deschaine