🪄 Demystifying Einstein GPT for everyone
Let's understand everything from a layman's POV.
Note: This blog is not primarily intended to discuss features of Einstein GPT or AI in depth but rather to create an awareness around the general concept and SHIFT+DEL all jargons & AI fear from everyone's mind. Grab a cup of coffee, enjoy the read!
What is Salesforce:
A Cloud based CRM service that helps your business navigate Sales, Service, Marketing and E-Commerce operations fluently.
What is Einstein:
In 2016, Salesforce announced its AI services for its customers and called it Einstein AI. Businesses can make use of Einstein AI to predict possibilities and provide recommendations to their users. Evaluating insights at this level just by analyzing the business data (reports) manually is a challenging and tedious task. In Einstein AI, you define the dataset and criteria that will help Einstein make recommendations for your users efficiently. Implementing Einstein is a continuous and iterative process. It requires a minimum set of existing records to build the predictions or recommendations as a starting point. The recommendations and predictions improve over time, with the volume of data and corrections that you make by regularly monitoring its performance. This type of AI uses smaller models compared to GPT models, also called Foundational Models or Large Language Models, which use parameters (datasets) on a scale of billions.
What is GPT:
Many of you may have already used OpenAI's ChatGPT, which provides human-like conversational responses with correct (almost) answers. At its core, GPT (Generative Pre-Trained Transformers) are deep learning AI models that predict the next word in a sequence of words. For example, it can complete the sentence: 'Jack and Jill went up the ______.'
OpenAI developed ChatGPT by leveraging large language models with datasets containing billions of parameters, such as the case of GPT-3. The logical advancements and improved data collection methods contribute to the chat window displaying responses in a sequence of words, resembling a friend typing a response. This experience is derived from the fundamental workings of GPT.
What is Einstein GPT:
It is Salesforce GPT offering for businesses to make efficient use of the CRM + Data + AI, to improve efficiency and experience: Administrators, Developers and End Users. What I love about it is that it respects the data security at its core which every enterprise should follow while integrating GPT like technologies.
Having worked in a large scale Banking project, I am always concerned about PII (Personal Identifiable Information) and IP (Intellectual Property) protection of customers and businesses respectively. I must say Salesforce have nailed it with the way they have designed the architecture of this solution. It’s based on their guidelines of Trusted generative AI which I see it to align with their Trusted AI principles. Salesforce Einstein GPT keep users in control i.e. wherever GPT is involved, the output doesn’t goes to production/customer without consent/approval from users. This is something that solves some of the problems with GPT. One thing every user of GPT should understand is that it’s been trained on huge datasets and is being continuously trained by its users when they accept/deny an answer to be correct or incorrect. This introduces problems of bias as well as giving false positive answers as a response. GPT has a potential ability to go crazy and pretty wrong, just like humans.
Salesforce enables the use of their own GPT model, your own GPT model (you can bring your own GPT models), and public models (like OpenAI's ChatGPT model) with your CRM and Data to better serve the GPT use cases they are planning to offer. Going further, Einstein works to reduce bias, toxicity, and irrelevance issues before presenting you with a response. Here are some of the use cases covered in public announcements, to the best of my knowledge:
Content creation for Sales, Service, Marketing and Commerce Cloud. For example, knowledge articles, product descriptions, marketing contents/pages, along with images, etc.
It will generate responses for Emails/Chat Queries on the fly for you to review, modify and use it.
Helping developers develop code and test classes on the fly using code comments.
Assisting admins in creating flows using text prompts.
Einstein GPT will also be available in Slack and Tableau for creating relevant replies and interactive analytical dashboards on the go.
Note: In the case of metadata like Apex and Flows, I highly recommend to use it as a starting point or help, and it should reach production only through human approvals/consent, even if the solution works for your use case as is.
One thing that I am still confused about and would love to see a demo around from Salesforce is how they ensure the PII and IP information like confidential APEX codes or Customer's bank balance are not being exposed to external GPT vendors or also via their own GPT models to other Orgs. I am pretty sure that they have already addressed it looking at their security pillar of Einstein GPT approach. However, since its in a closed BETA trial, right now I can just imagine that it would be something like what they did for Einstein AI. In Einstein AI they allowed users to exclude Fields which are confidential from Training Data Set configuration used for making Einstein predictions or recommendations.
Putting that aside, I am extremely excited about this paradigm shift in the Salesforce ecosystem. I was listening to Jayesh Govindrajan, the SVP of AI/ML at Salesforce, where he shared an interesting experience related to a delayed delivery of an Avenger merchandise that he bought. He received an apology for the late delivery. Jayesh gave an example of how this apology letter could be made even more engaging for fans by being written in the Asgardian language. This demonstrates a cool use case of using GPT to enhance experiences and align your copy with your brand. I would also love to see these capabilities being made available to ISV partners in the future.