What exactly is the difference between gpt-4 and gpt-4-turbo-preview? As far as I can tell from the documentation, the ‘turbo’ suffix basically seems to mean ‘better’; am I missing a tradeoff, some reason one might prefer to still use plain gpt-4 after all?
1 Like
This should help:
https://help.openai.com/en/articles/7127966-what-is-the-difference-between-the-gpt-4-model-versions
2 Likes
Well that’s the thing, as far as I can see, that page seems to be saying that the turbo version has a larger context window than the plain version, and is more up-to-date, and is also cheaper, so that’s why I’m confused; what is the plain version for, given all that?
1 Like
I believe it has to do with “attention” and other factors… They don’t give out all details, so best thing to do is test and see if Turbo works for you and if not, try sans Turbo as it’s likely a little bit “better” at some tasks…
1 Like
It’s unfortunately not trivial to quantify.
Some people (myself included) believe this:
GPT-4-0314 is the best of the GPT-4 series.
It performs the best at solving abstract problems. GPT-4-turbo (1106, 0125) struggles more with logic reasoning.
However, GPT-4-0314 is more prone to hallucinations, and also more expensive.
GPT-4-turbo, on the other hand, is much more stable, and less likely to make stuff up, on top of having a gigantic context window that actually seems to work.
There are a lot of benchmarks out there, and they’re unfortunately hard to interpret.
GPT-4-turbo will be better for generating code and doing basic text transfomations with decent reliability.
GPT-4-0314 can have a chance of solving certain problems that gpt-4-turbo will never be able to tackle.
But what about gpt-4-0613?
I personally think 0613 is the most useless of the series. It attempts to be 0314 with the kookiness toned down, with some new bells and whistles I’m personally not interested in, at the cost of reasoning. I haven’t found a use for it yet.
2 Likes
Totally agree with this, I had prompts perfectly working with GPT-4-0314 and the results were big failures after using GPT-4-Turbo. I had to rewrite most of my prompts to get similar results.
Nonetheless I think it’s smart to use GPT-4-turbo, it’s faster and cheaper.
My advice to tackle complex problems is to start with GPT-4 to check if the task you are trying to solve is doable by the easiest model to prompt. If that’s the case, tinker your prompt to make it work with the turbo version.
2 Likes
GPT-4-1106-preview is far more powerful than the rest of them