And worse, the top one is the best at very low signal levels. Good down to -65dBm into 50Ω at 1MHz. The bottom one is the worst at this signal level, reading 3.1dB too high.

@azonenberg You're going to become the man who shouts at satellites!

It was the bottom one (by correct I mean closest to what I have my calibrator set to)

@azonenberg indeed, my next job is to look at an old OCXO that is substantially off-frequency, coincidentally.

@azonenberg Well, yeah…

@u0421793 All that training and no actual results. Every time I think I'm seeing real progress it turns out to be something that's basically easier and often more reliable to do without involving LLMs. But LLMs would do well in any situation where quality and resilience doesn't matter or, more interestingly for me, where plagiarism is kind of accepted as part of the process. That's why they got traction in businesses where the C-suites are aggressively removed from any of the real world processes, which is most big businesses and even some small ones. The human input on the big LLMs is substantial, I reckon, but not transparent and would probably hurt the investment attractiveness of that business. As for energy cost falling, nah, it can only rise as they create often useless complexity at superhuman rates. Think about the data surge that digital cameras, then camera phones, then apply it to every document that already exists. I'm already finding that most electronics data, even old stuff, is available secondhand and thirdhand thorugh regurgitated LLM output - the defining characteristic being that it's usually subtly wrong. This is starting to impact the only thing LLMs did reasonably well, summaries. They're now regularly summarising garbage documents that only exist to pull in clicks for onlne advertising. Quality is getting worse fast. Local use is a bit like local use of food for final preparation and consumption, there's a lot upstream that could affect it. I'm more sceptical than ever as it gets worse, but more prevalent.