As the AI revolution marches forward, it seems that most people are starting to form camps around these things. In one camp are the people who swear to never use AI, and in the other camp are people who want to use AI in everything. To me, both camps have the wrong mindset. The future of AI when it comes to software, and especially business applications and use, will always be a blend of AI and deterministic software approaches.
AI Is Non Deterministic
There is a reason why one side of the equation is called Deterministic Software, and the other side is AI. If you are expecting AI to be 100% accurate or to produce the same exact output 100% of the time, then you might as well also have the expectation that a unicorn is going to fly to your corporate headquarters and present itself as your new brand ambassador.
The power of AI comes specifically from its prediction and statistical analysis capabilities. If you want to start putting boxes and rules around that, these are all additional requirements for the AI model. That translates into worse and worse results. Sure, you could attempt to sit there and hammer a square peg into a round hole for hours, or you could simply accept that AI is not deterministic.
Traditional Software And Coding Is Not Getting Replaced
Let’s look at a common example that a lot of people want to use AI for nowadays, text summarization. This particular task has already been solved using deterministic software. There is a lot I can summarize, and I have almost complete flexibility over the process. I can summarize by keyword. I can determine exactly how long the summary will be. If you can think of a rule for it, I can include that in my deterministic software approach.
If you asked an AI model to perform these same things, under the exact same conditions, it would not be able to complete the task in a million years. Every single time I tried to have an AI model alone summarize a short piece of text, I simply ended up with a rewritten version of the text. If you adjust the max length of the output, it does not actually make the summarization smaller, it just cuts off the text at whatever your max length is. No matter how far AI comes, there are simply some parts of software it will not replace.
The Best of Both Worlds
While deterministic software is great, it also has its limitations as well. For my text summarization task, there is only so far that I can go with rules. I can use a rules based system to tell me what the keywords are within a text piece of text for example. It will do an adequate but nowhere near human level job at this. If I were to attach an AI model to the equation though and ask it to critique the keywords chosen by the deterministic software, it would produce and return inhuman results.
I could also use the deterministic software and its rules to force some limits onto the outputs of the AI model. In this scenario, I would in fact be able to control the maximum length of the summary with much greater precision for example. What exactly would this look like in practice?
Let’s take this piece of text as an example: "A deterministic software application can provide additional functionality and control over the summarization process, such as setting the length, format, or structure of the summary, filtering out irrelevant or redundant information, or ensuring factual accuracy and consistency. A deterministic software application can also integrate the LLM with other components, such as a user interface, a database, or a feedback mechanism."
Using a blend of AI and deterministic software in conjunction with each other, allows me to summarize the text as follows: “A deterministic software application can provide additional functionality and control over the summarization process.”
Want to experiment with these things for yourself? Check out my free Google Colab notebook that I made that specifically blends AI (a pretrained BERT model) and rules based software (you can control various aspects of the output) for text summarization: https://colab.research.google.com/drive/1o_2O31dcVWRXjHznSWkIDRW3vrovdNYD?usp=sharing
