This Stanford seminar video can provide some references:
That video shows examples of new models with one-shot performance better than old models with fine-tuning. It doesn't compare new models with fine-tuning to new models without fine-tuning.
Right, the page 28 of this article mentions "On some benchmarks, task-general models (not explicitly trained to perform a task) surpass prior state-of-the-art performance held by a task-specific model" but they don't provide data for fine-tuning of their very large models:
What’s your take on the argument that fine tuning might be destroying model calibration and causing it to overfit to aspects of the finetuning dataset eg: causing it to appear more confident, rather than becoming more correct?
See https://twitter.com/animaanandkumar/status/16816906540601876...
That tweet is about a specific type of fine tuning: instruction tuning and RLHF. It causes to get better at the thing that it's being fine tuned for, which is making users feel better about the answer. Whether that's actually desirable in all cases is the question -- but that's not really related to whether fine tuning is a useful tool in general.
Here's an example of a paper that shows good results from fine-tuning for instance: https://arxiv.org/abs/2305.02301