Wordy models
When you just can't shut them up
After a few repeated failures from the house model Gemini Flash 3.5, I had enough. The failures were appending weird repeating garbage to the end of a JSON structure. This sort of thing, fairly persistent across Google models since I can remember, is the reason for my Janky nickname for Gemini. Anyway, good opportunity to switch it for the new hotness GLM 5.2 and save some money in the process.
The model is clearly excellent but switching the house model is a kind of vibe check all on its own. It turns out GLM is a wordy bastard, and I say that as… a wordy bastard. I don’t really mind it but, well, I think it’s fairly safe to say that these days people don’t like to read as much as my generation. You might have a different view, but either way, content ought to be taut.
So I revisited the prompting for the deep dive with a few strategies such as naming fewer sections, and adding a word count and a revision/edit phase under certain conditions. That last move was suggested by Claude, the clever bastard. I should have thought of that.
Also, I think Deep Dives could always do with an abstract illustration even if there’s no data to visualised so I loosened the language around when/how to use the image tools offered to the model (a fixed datastructure one for doing things with Matplotlib etc., and an free-form vision model, OpenAI’s excellent GPT Image 2.0).
Can we tame the beast? Let’s see how it goes when today’s runs at 10:00.