June 14th: OpenAI's new GPT upgrade keeps it with the front runners

While commercial providers battle it out for Large Language Model (LLM) supremacy, and the Open Source community develops innovative approaches to train their models - without requiring a zillion Graphical Processing Units that need to be run for half an eternity....OpenAI has released a new feature named Function Calling. This is available for GPT 3.5 and 4, in release '0613.'

Is it a big deal? Yes. With this upgrade, we now have the ability to derive the structure and related content of GPT responses, at a program level..and when we know that it is far simpler to take appropriate actions

Here's an example. Today, in Cambridge (UK), the weather is warm. It is late afternoon. I would like to know some countries that have an average temperature that is, say, 5C cooler than Cambridge.

Let's ask ChatGPT-4. Here's the query and part of its response:

It's attempting to provide a useful response, but as it can't access real-time data, it won't be able to determine the current temperature in Cambridge and therefore the rest of the response is not particularly useful.

Enter Function Calling (FC)

FC is the new OpenAI GPT feature that enables, within a program, the GPT response to return structured data. By structured, I mean it can extract the content of, say, the location (ie Cambridge), month (ie June) and the type of temperture (ie average).

Having extracted the relevant data, we can then call a program function, to pass it to a real-time weather service, and return the current temperature in Cambridge. Following is the result of such a call:

It states that the temperature is currently 19C, which certainly feels about right.

Now, I can prepare a subsequent GPT request, using the various extracted items of data; remembering to adjust for the 5C lower temperature. Again, this will be called within the program, so is hidden from users; they only see what they want - the outcome. The GPT-4 FC

response is:

Remember that GPT is interpreting the request to structure the data

And this means that sophisticated responses can be returned. For example, providing 2 country lists that meet the temperature criteria; one in English, the other in French - as shown in the sample programmatic GT-4 FC output here:

'in_english': 'Germany, France, Italy, Spain, United Kingdom', '

in_french': 'Allemagne, France, Italie, Espagne, Royaume-Uni'

Summing up

Although the ability to access GPT, programmatically, has been around for 2 months, FC increases the ability to exploit the deductive power of GPT versions 3.5 and 4, and extend the range of powerful GPT-based solutions.