ChatGPT and Beyond

I am sure by now everyone is aware of ChatGPT, the new big thing that has got Google worried and Microsoft happy. I have been always a GCP or AWS person but after the recent developments, i took the first step towards studying Azure. why Azure ? With Microsoft investing heavily into OpenAI, its just matter of time that the ecosystem will revolve around Azure !

So lets talk about ChatGPT.. to start off .. I asked ChatGPT to explain and this is the response - "Its a large scale language model developed by OpenAI that is trained on a diverse range of internet text and is capable of generating human like text. Its based on GPT (Generative Pre-training transformer) architecture and can be fine-tuned for a variety of natural language processing tasks such as translation, question answering and text summarization". The next question was what is GPT and it answered "GPT is a deep learning model that uses a transformer architecture to generate text and is pretrained on a large corpus of text" . This may sound complicated so let me simplify this, if the chatbots we see today in every other app is a toddler who can understand basic questions and respond then ChatGPT is like a young kid who will understand , reason and will respond with context. Its not perfect and will continue to learn and grow but the potential is enormous and may be a major milestone for our future way of life. The ecosystem and the businesses that will be built around ChatGPT and similar framework APIs has potential to revolutionize not just how we search for information( hence Google is worried) but how we interpret information and thereby build opinions and impact our behaviors. To put it in simple terms, i google for something and it throws out information which i use to build an opinion (yes i may be biased by the ratings on a product or a comments of fellow humans or bots) but i still feel that i am in control of building that opinion based on the information that google provided me (albeit the pages that google promotes may be biased and promoted) but you still get a sense of free will and i have a choice to accept or reject that opinion but will that same sense of free will prevail when you are having a conversation with ChatGPT? For now its a lot of fun and there is ton of content online about how this will revolutionize our world! I will not venture into prophesizing how ChatGPT or anything else that Google or Meta will roll out will change the world but this may kick start a bunch of revolutionary set of products. This reminds me of a quote that i read some where about a famous painter who said "A good artist copies, A great one steals" , Creativity triggers are always inspired by something, it just does not manifest out of thin air and that's what we see today with the evolution of this AI tech. The amount of information that's available in digital format is enormous and its only growing exponentially so deep learning and its successors will generate art, music, poetry , essays which for me is similar to how a human will be inspired to create art by watching a sunset but will a machine appreciate the art that it has created and "feel" the emotion? Again going back to my previous blog.. Machines have become more efficient and i dare say creative but intelligent?

In 2020, me and my team Sanjay Narvekar, Manish Kumar and Dhananjay Saraf presented a working prototype of a product in a hackathon that was well received and we were even one of the finalists. ChatGPT sort of reminds me of that idea that we worked on and i am sure my team will echo the same sentiment.

Our problem statement was "Currently any human relationship manager builds relation with the customer by having a conversation and understanding the customers needs. The biggest challenge for any chatbot is to give personalized responses and experiences specific to each customer. The human behind the computer should "not" realize that he/she is taking to a bot and that's the measure of our success and the vision of our product "AI Relationship manager."

The chatbots are trained based on a predefined set of corpus which makes it a very vanilla experience to the customer. The solution that we designed and prototyped was aptly called "ANNA" - Artificial Neural Network Agent.

Design

The simplest explanation of the inner workings of a chatbot model is that it is trained based on a set of questions and answers called "Corpus" and a machine learning model will train to select the best fit response from a set of corpus (set of questions and answers). This is called Natural language processing and there are tons of algorithms and ways to achieve it but the layman description would be that you take a question (set of words) and scan through a list and find the probability of which one matches and use that as the response. The solution for ANNA revolved around building dynamic corpus that is tailored specific to each customer.

Architecture

A simple chatbot comprises of 4 components

1) User interface channel

2) Machine learning model to process the natural language

3) Corpus or the data required to "train" the natural language and this data is captured based on how humans converse

4) Database for processing the actions to be processed by the bot

In a traditional sense we would train the chat bot based on different questions and answers but ANNA needed to be customized and the questions and answers would vary based on each customer. For example if a customer is predicted to miss their due date on the credit card, the bot should be asking questions pertaining to the due date and may be offer a loan product or if a customer is spending on specific type of product then may be offer a cashback credit card. Now the intent was to build a relationship hence the same interaction between customer A and B would feel different though the customer was chatting to check the balance in the first place.

How ANNA achieved it

This is not a prescription on the most accurate way of achieving it but the intent of this blog is just to give insights into a means by which we achieved this goal with limited resources and we were more interested in exploring, learning and solving this problem in the most simplest way with basic resources . The first problem we wanted to solve was to build custom corpus for each customer. we had a base corpus which is applicable for each customer but now we wanted to add specific questions and answers to this corpus for each customer and this is where we applied the first set of traditional machine learning algorithms. we had two sets of sample data 1) Customer transactions 2)Customers financial data, like balance, past dues etc. So to start off we built two traditional models based on supervised and unsupervised learning. The supervised model would predict if the customer would become delinquent and the second model would group the customers into different categories based on their spend.

These two models then formed the base of building the corpus, the idea was why not build micromodels like microservices? The dynamic learning model approach meant that the corpus questions related to loan default will be added to the base corpus for the customer who was predicted to default on the loan and the model would be trained and this would be the model/database used when the customer logs into the app. Similarly the corpus related to upsell would be added to the base corpus for the same customer, this meant that the micromodels used for each customer will be different. A combination of different feeds from a cluster of models will be the trigger to add the corpus which would be the fed into the NLP. It was really a proud moment for me and my team when this concept actually worked albeit at a very rudimentary level for a hackathon and it was fun to see how the conversation would differ for two different customers. There are much better chat bots with advanced features in the market and it would only get better. The reason to explain this was to give a context of the evolution of this technology and the wonderful opportunities it presents.

Our lives in the future?

Just for fun.. Believe at your own peril and don't throw brickbats if you don't agree. As OpenAI and ChatGTP like products and the ecosystem evolves, we will find improved products which will be a level up to the ones we have today

1. This may happen sooner but Alexa, SIRI, Home etc with Hologram. Won't it be cool to have a holographic AI to speak to instead of a speaker with whom we can have more realistic conversations? We have seen this in Sci-fi movies but soon this will be a reality. Imagine you are getting ready to go to work, Alexa kicks in and tells you that there is traffic on your normal route to office and you should take an alternative route. This is available today in maps but integrated with conversation engine without a trigger and with context will truly make home assistant solutions even more powerful. I long for that day when my refrigerator using computer vision warns me when i sneak in to get my tub of that ice cream at midnight.
2. Psychologist / Therapists - Tough today but possible in the future where you can talk to an AI that will help you with your metal health issues
3. Our Hackathon Idea - AI based Relationship Manager for financial advice and portfolio management
4. Real time translation - Imagine a device that will look like a mask and you speak into it in your language and the speaker blurts out the translated language like if you are in Mexico, you speak English but the person in front of you hear Spanish and what they speak, you hear in English in a headphone in your ears.. pretty sure its just around the corner that google or amazon will create such a device, with the power of deep learning and cloud computing supported by the widescale internet penetration will soon make this a reality
5. Teachers - I hope that we don't lose our teachers to technology but we have already seen that online learning is possible to some extent, it just needs powerful technology and with the power of platforms like ChatGPT and ever evolving Virtual Reality and MetaVerse , holographic projections and in next few decades it may happen.
6. Doctors - We have already seen that computers can fly planes so i am not even considering pilots here because they are a done deal, its just a matter of time that we won't have pilots but doctors are bit sensitive but if we leave our inhibitions aside then with the digitization of medical records and the growth of computer vision in medical realm in conjunction with a conversational engine definitely has the possibility that you may prefer going to a doctor who does not need an appointment and is available 24*7. I dare say, robotic surgeon will do a better job than a human one.
7. Lawyers - This has already started but may take bit time for acceptance like doctors
8. The list continues and the possibilities are endless, enormous and exciting

Conclusion

In summary these all are surely a possibility as the technology evolves but the bottom line is acceptance for us as humans, will we trust a machine over a human?

That's where these may seem bit far fetched but the generations that are born now and the ones in the future are born into this world of artificial intelligence and for them this will be a way of life. Our reservations stems from the fact that we grew up in a different world but the citizens of tomorrow would probably grow up and live in a world where this trust deficit may not exist and the optimistic in me feels that it may be a wonderful world to live in !

We may or may not trust the machine but will we trust the human who controls these machines ?