Is GPT4 turbo really smarter?

Has anyone done any comparison of GPT4 Vs Turbo?

I had some inconsistency in a prompt output, had some tests that don’t pass all of the time.

I switched the prompt over to GPT4 Turbo out of curiosity expecting it to be more consistent but it wasn’t.

I admit it’s not a great test to expect better consistency from a prompt specially if that prompt has some ambiguity to cause it to vary it’s output.

So has anyone done any extensive testing or comparison for logic and reasoning? I am assuming “smarter” means this though now have some doubts :thinking:

From what I understand, the “turbo” models are ablated (reduced-size) models, where they follow the coefficients of each neuron/layer/parameter backwards, and those that have very little impact on the output overall, are removed. That way, the model can be made to run using less memory, and thus run faster, while giving approximately the same output as the pre-ablation model.

I personally believe ‘yes.’ I also believe it better understands my prompts. While I haven’t conducted any official tests, based on my usage, GPT-4-Turbo performs as well as GPT-4. The responses from GPT-4 and GPT-4-Turbo are subtly different, but acceptable. For instance, when using GPT-4-Turbo, you might encounter ‘’'html or ‘’'json at the beginning, but you can remove that when streaming or after completion from the backend or frontend. So… No significant issue.

I have the same issue. I have the feeling that the gpt-4 turbo is less smart than the original gpt-4… I sometimes encounter oddities or a lack of ‘reflection’ from GPT-4 Turbo.