I completely agree with albaby1's comments about Gemini. I admit to laziness in popping a question to the easiest AI at hand (my free time at Claude hadn't cycled yet). Gemini has the bad habit, in addition to albaby1's observations of losing its track of thought in a short period of time and expressing the same line of loquacious BS as a substitute for thought that I found so annoying in the past with ChatGPT (which either by design or by my ever more aggressive prompts), but which is now thankfully minimized. (I am looking for information, not a new friend).

What I should have done is, rather than immediately posting the (reasonable sounding) results from Gemini is to ask both Claud and ChatGPT what they thought as well and then integrated the answers.

Jeff, this certainly wasn't a criticism of you personally. I just thought it interesting how constrained Gemini is, and I had just posted on TMF about it.

However, I think it's worth keeping in mind that this general architecture appears to be common to all the free models. They all have a core set of training data about which they have a lot of knowledge, but anything recent just involves some web searches - and very limited web searches, far more than I think we'd all expect.

Claude's better than Gemini, but it's still pretty nerfed. The training data's current through June 2026. But if you ask it a question that needs more current information, it will still only do 3 to 8 web searches, look at snippets from ten results from each search, and only occasionally actually pull the full article/reference from any of those hits.

I hadn't used Claude yet today, so I ran your initial prompt through it, asked for an answer and the search metrics. It did only seven web searches, only 63 hits, and only looked at the snippets for all of them. It did not pull even a single complete article to look at. Better than Gemini, of course, but still only a very casual amount of actual research.

Obviously it's better to compare the returns from multiple models than to just look at one - but all of them are going to be working of a very small knowledge base for things involving current events or recent developments.