AI Mode with Gemini 3 gets past the hedging to a welcome specificity

Gemini 3 shows that LLM progress is continuing in the ways that matter; education might want to prepare for that.

As I think people know, I do a series of video walkthroughs. I take something I want to look at more deeply in the morning and walk through it using Google’s AI Mode. It’s a way for me to think about how to structure pedagogy around these tools, and as a side effect I learn new things.

I spent the last two episodes of that series looking at the supposed “first gay kiss” in film history in the film Wings (1927). And I demonstrated how AI could help students break out of their modern assumptions about what something in the past represents. In particular, using AI Mode reminded us that the norms of male physical intimacy were actually more permissive before WWII, and that scene could be read differently in that context.

AI mode did wonderfully at breaking us out of our mindset with that general reminder once we added a follow-up. But it’s first response was far from perfect. Let’s take a look:

This is good in some ways, but very hedgey, leaning on the “matter of debate” line so many AI Mode first responses gravitate towards. Subsequent follow-ups got it to lock in better, and highlighted the social norms of the time.

In a twist of fate, when I posted this to Bluesky, the daughter-in-law of the director and writer of this film turned out to be a follower of mine. She highlighted something that AI Mode had missed completely, that this was seen in the context a very special type on bond between friends in combat.

Again, AI Mode, when using follow-ups, did correctly identify the idea of changing social norms for us, and you can watch that initial video here. But like a lot of responses, it did not drill down to this level of specificity — not just the context of the 1920s, but the context of male friendships in WWI specifically, and the way that incredibly specific lens added to this perspective.

So I was interested in seeing what would happen when I put this into AI Mode and used the (currently paid) option to use the new Gemini 3 model underneath it.

It was impressive. Here’s a full response you can look at.

The difference is the specificity. Here’s the beginning of the response from AI Mode with Gemini 3, zeroing in immediately on the combat intimacy dimension:

Compare that to plain AI Mode:

Do you see the difference? It’s hedging vs. detail.

When we add follow-ups to the Gemini 3 version, it shines even more. Here’s a bit of what it tells us when we ask it to think like a historian. It immediately zeros in on combat intimacy again:

But then goes further into detail — look at the lenses and concepts we get to look at this with here. There is the “deathbed exemption” in film, where characters are allowed to express otherwise inappropriate emotions, there is “trench intimacy” which was a fascination of post WWI audiences, and seen as a protective defense against the horrors of war. And there is also just the necessary physicality of silent era acting, which had to make everything very dramatic and intense.

This is an incredible toolbox of concepts to apply to this!

The full response goes into depth and I encourage you to read it. But the thing it most suggests to me is that the definition of small and large leaps is somewhat badly formed. I know we’re not comparing here to the previous thinking mode you could select in AI Mode, but even so I think many people might look at the difference between these two responses in this post and think well, that’s small. But it’s not small at all. Specificity of contextualization is actually huge, and if your take has been that progress with these tools is slowing, I would suggest reconsidering that.