Can LLMs ie generative AI be discriminative ?

There are broadly two classes of AI - generative and discriminative.

Models like classification are discriminative . ex when you classify between spam or no spam you are performing a discriminative function, These models are able to determine a decision boundary and assign a new data point to the appropriate data pint based ion the decision boundary. On the other hand, generative models are able to discern an underlying probability distribution and then generate new datapoints based on that distribution. see generative v..s discriminative .

all this is good .. but the question is:

Can LLM based large language models be disriminitive?

The surprising answer is yes - LLM based generative models can also be discriminitve based on the ability to generate code

The best example is vipergpt

The problem is .. I want to divide 8 muffins between two poys

I take a picture and prompt the image about this question

Using computer vision, we can identify the two boys and the eight muffins in the image

We parse this question

Next, we generate code to work on this problem of dividing the muffins and present the correct answer

some images below and also the comparison of the two stacks