What’s all the noise about?
To fully understand what the GPT-3 model is all about, let’s first check what we can expect from the spring update of the Dynamics 365 family of products.
“Use machine learning technology to get content ideas and create compelling emails in no time.”
So, it’s about a completely new function that no longer deals only with the analysis of the recipient’s interaction with content, but it will create content based on what it has learned. Seemingly nothing extraordinary, but as Microsoft assures, the generated text should sound and look as if it was written by a human. Therefore, there is no room for errors in the used language, punctuation, or a rigid textbook style. All you have to do is use keywords or phrases that you want to be included in the email. After content is generated, we can freely edit it to better suit our preferences, which will be taken into account in subsequent e-mails. Microsoft does not spare promises and quite boldly informs that the new function will not only provide interesting ideas and facilitate the creation of e-mails, but will also help in discovering completely new ways of expressing thoughts.
What’s different in GPT-3?
The first question that arises after hearing such revelations is usually “how is this different from what we have now.” First of all, numbers. The previous model, GPT-2, consisted of almost 2 billion parameters. Its successor already has 175 billion parameters[1] , making it the largest language model to date. As the creators declare, GPT-3 has been trained on 45 TB of text data. For comparison, the Millennium Bible in pdf format, containing 6637 pages, takes up 42 MB.
It is also worth mentioning that in 2020 the model was made available to scientists who could freely use it in exchange for providing their comments. As a result, more than 30 researchers monitored the accuracy of the model for over a year.
However, to avoid misunderstandings, it should be noted that GPT-3 has no knowledge and has its drawbacks. Its main task is to create content that is deceptively similar to text written by a human. The creators themselves talk about the weak points of the model, which are the synthesis and analysis of the text, repetitions, or loss of consistency in long forms. For example, ChatGPT does not cope with the interpretation of Polish law and often even gives incorrect paragraphs, confusing the meaning of words in the code with their meaning in everyday language[2]. This underscores, even more, the purpose of the model, which creates content for ordinary people as if they wrote it themselves.
GPT+CRM=?
Last week, Microsoft unveiled a brand new AI-powered Viva Sales. It will be used to generate e-mail content for sellers who have contact with current and potential customers. As we can see in the presentation video, just 3 clicks are enough to generate the content of the email. The text will contain all the necessary data downloaded live from the connected CRM system.
In short, it looks like this:
– I receive an email from the client
– In the inbox, I click on the Viva Sales banner
– I select one of the ready-made topics from the side panel or enter keywords (e.g. “delivery 7 days” is enough)
– Viva Sales automatically downloads data from Dynamics 365 and Azure OpenAI Service and GPT and enters them in the body of the email
– I make corrections, if needed
After sending an e-mail, Viva Sales will update data in Dynamics 365, thanks to which we will always be up to date, and additionally we will save time. The module will also work when we get an e-mail with an inquiry. Based on the customer’s questions, Viva Sales will create a personalized answer based on the company’s offer, and can even propose a meeting. The new environment will also ensure that we do not miss any opportunity to complete the transaction. When we do not initiate actions ourselves, we will receive a notification and a proposal for next steps along with the ready content of the email. Viva Sales will take care of such details as product descriptions, current prices and conditions, and even suggestions and dates for the next steps.
While we can enumerate flaws in the GPT-3 model itself, it’s hard to find them in the way it is implemented in Viva Sales. Microsoft focused on the strengths of GPT, combining it with the possibilities offered by Dynamics 365, Microsoft Cloud, Office 365 and Microsoft Enterprise Mobility + Security. The risk of errors has been almost completely eliminated. Emails are short and concise, and all data is taken from a separate database. In addition, the final version is supervised by a man who makes any stylistic corrections or refines the text in a way that is characteristic of the brand.
[1] Language Models are Few-Shot Learners, https://arxiv.org/pdf/2005.14165.pdf
[2] https://www.dogmatykarnisty.pl/2023/02/rozmowa-ze-sztuczna-inteligencja/