It's no secret that I'm a generative AI fan.
SmartSuite already provides a few generative AI features. There are certainly more on the way. I don't have insider knowledge, but it's easy to see where this is going. Soon (I predict), there will be an AI field or possibly a formula field that can include AI prompts and field placeholders. We will soon be able to use generative AI to interpret data fields, such as extracting the sentiment of customer comments.
I'm deeply biased about generative AI, but for good reason. Here's one of them. This is an actual question here in the SmartSuite community.
It involves a requirement to transform a field of data from a number into a formatted number representing a user-friendly phone number.
... take that 9 digit number and separate it into three parts similar to a phone number. 123456789 -> 123-456-789
The only approach [today] is a formula like this offered up by D Mitchell.
When SmartSuite adds integrated generative AI capabilities, an AI-first engineer might approach this challenge by using a prompt in an AI field instead of a formula:
You are an expert phone number parser. I will provide you with a number. You will transform the number into a usual and customary phone number format output. i.e., 999-999-9999 You will output only the formatted phone number.
Examples: '
Number: 2125551212
Phone number: 212-555-1212Number: 212 555 1212
Phone number: 212-555-1212Number: (212) 555.1212.
Phone number: 212-555-1212Number: {{Number}}
Phone number: [output]'
The beauty of this approach is the agility it provides to clean the data. The examples in the prompt perform a zero-shot fine-tuning of the LLM (model) to create an expected and deterministic behavior. It serves as the guard rails for the LLM. While only four variants are documented in the learner examples, do not be misled into thinking this is the limit of this prompt. Here's why...
Imagine the AI encounters a messed up phone number. It has the intelligence to work through the mess and arrive at the correct answer without ever having seen the pattern and without the prompt engineer ever considering such a corrupt field entry might exist in the data set.
Formulas are rigid and deeply constrained. They are literal in every sense. They cannot be agile unless you write a lot more "code". By definition, formulas are low-code; they require skills to codify relationships and processes of variables and constants. This is code inside your no-code platform.
AI can eliminate the effort and skill required to low-code your solution. Indeed, AI is the ultimate no-code platform.
If you can "code" for all the possible edge cases without requiring you to consider all the possible edge cases, AI trumps code in many situations.
This is the magic of generative AI and why it may eventually render formulas entirely obsolete.
If you think this is a ways off, consider this thread in the Coda Community. It explores a similar number formatting challenge, but unlike the speculation of doing this in SmartSuite, we already do this in Coda using Coda AI, which can already supplant Coda's formula language (CFL).
This is controversial as the trust level for generative AI is generally low. However, if you use learner shots to fine-tune the LLM to meet your expectations, the accuracy rises significantly, and in many cases, it is 100% predictable. There are certainly cases where generative AI would be an unwise choice. But in many cases, this will work well, and developers who train themselves to become AI-first engineers will run circles around their colleagues.
