ChatGPT simple math calculation mistake

Hi @daisyyedda

Further to my earlier post about ChatGPT being a large language model, I ran an experiment for you last night:

  • Fine tuned a base OpenAI davinci model using 16 n_epochs with the following JSON training data:
{"prompt":"241 - (-241) + 1 ++++", "completion":" 483 ####"}

This fine-tuning baked and was ready to test when I was having morning coffee (just now):

So, ran a completion using this model, as follows, and got the correct “answer” you are looking for.

This is because ChatGPT and the GPT APIs in general are simply “fancy” auto-completion engines based on a large language prediction model. So, I fine-tuned the model to increase the probability that when the davinci model sees the exact text you were interesting in:

241 - (-241) + 1

The model will predict that 483 is the next text in the sequence of natural language.

Note, my new fine-tuned model is not doing any math! The model is also not “calculating” at all.

It’s simply predicting text.

To prove this, I will slightly alter the prompt to:

242 - (-242) + 1

Now we re-run the completions, and it’s not surprising, here is what we get:

This is because we cannot train these GPT models to perform math, we can only train them to predict text, given prior test.

In the experimental case above, the model simply predicts the next sequence of text as 483 because it was trained to predict this and this is the closest match, statistically speaking.

There is no “calculation” happening at all, @daisyyedda

ChatGPT cannot make a “math calculation mistake” because it’s not performing math calculations, it is just predicting the next text in a sequence based on a language model.

Hope this helps you understand what GPT is and does.


See Also: