@tjrobertson52

There is no best large language model. There's only the best model for a given task. For some tasks, you want the smartest model, for others you want an opinionated model and for some tasks, you just want a model that'll follow instructions.

So I wanna talk about which model is best for when you know exactly what you need the model to do and exactly what the output should look like. So for this comparison, we're gonna be talking about Gemini, ChatGPT and Claude. So let's start with Claude.

Claude is actually my favourite model and by far the one I use most for work. We use Claude Sonic for any kind of writing and we use Claude Opus frequently for analysis. While Claude is generally good at following instructions, it tends to be very opinionated.

It's also a bit of an overachiever. Claude is especially good at understanding your goals and then extrapolating from there. It's gonna do what you ask it to do and then some.

This is especially true of Claude sonnet, which tends to be quite verbose. Opus tends to be more analytical and to the point. It also loves presenting things in tables.

I think Claude is the best model at thinking through tough problems and drafting presentable documents. It's a great partner when you don't know exactly what you want, but it's not the best at following instructions. Let's talk about Gemini.

According to the benchmarks, Gemini is the smartest model. I also Think it might be the most obedient. In other words, from my point of view, Gemini is the least likely model to give you push back.

Now, sometimes that's a good thing, but in most cases, at least for me, I kinda want the models to push back. The problem with Gemini is that it's kind of lazy. It's gonna answer before putting a lot of thought and deliberation into it.

And if you explicitly ask it to think and deliberate, it's gonna take shortcuts. It seems to do whatever it can to limit the amount of tokens it spends. It might be the model best at following instructions for very simple tasks, but overall, the model I find to be the best at following instructions is ChatGPT.

When you're able to articulate exactly what you want the model to do and exactly what you want the output to look like, I think this is where ChatGPT really shines.