Dear readers, for the past few years AI/ML was making noise around us and were solving specific problems using various data-science technologies. Now suddenly the whole world is amazed with the entry of ChatGPT by OpenAI. Within just 2 years it has crossed more than 100M users which is the fastest user registration on any platform. Upon zoom-in on ChatGPT, we learn the role of LLM and Generative AI in their capabilities.
For a long time I wanted to write something a very simple write-up for a community who do not know anything about AI & GAI from the non-IT world. So let me try to make them understand, and I will keep this article very simple and with minimum jargon.
Whole concept of AI is based on four terms Model (प्रारूप), Data (आँकड़े), Training (प्रशिक्षण देना), Prediction (अनुमान या आकलन करना). Once we understand these 4 terms, we shall be able to correlate with AI & GAI.
Note : each term’s meaning is also given in Hindi which is contextually most closed. There is one more term “Machine Learning (यंत्र को प्रशिक्षित करना)”, which is self explanatory i.e. transform the machine into “self-learned-object” which can perform certain tasks with great efficiency.
Now let’s understand what is Model
To understand “Model”, consider a baby’s upbringing, how a baby is grown up from birth till higher education. Let’s understand the same from below table, each row of which explains learning at different stage of her/his life and what level of maturity the baby obtains considering a scale of 1 to 20 and 1 maturity point at each stage.
What baby learns at different stages | Maturity level (assuming 1 for each stage) | |
Baby takes birth on this beautiful planet and looks into her mother’s eye and learns her motherly face. | 1 out of 20 | |
Baby crawls with her mother’s support and learns to touch & feel things in the surrounding. | 2 out of 20 | |
Baby learns to walk small-small steps with her mother. | 3 out of 20 | |
Kids learning activities in the play school etc. Kids learn their parent’s name, teachers name, location, color etc. | 4 out of 20 | |
Kids learning the subjects of 1st standard e.g. alphabets, counting, tables etc. | 5 out of 20 | |
Kids learning the subjects of 2nd standard e.g. words, 2 digits calculation etc. | 6 out of 20 | |
Kids learning the subjects of 3rd standard e.g. complex words, sentences, N digits calculation etc. | 7 out of 20 | |
Kids learning the subjects of 4th standard e.g. science, math & english etc. | 8 out of 20 | |
Kids learning the subjects of 5th standard e.g. science, math & english, environments etc. | 9 out of 20 | |
Kids learning the subjects of 6th standard e.g. science, math & english, social science | 10 out of 20 | |
Kids learning the subjects of 7th standard e.g. science, math & english, arts etc. | 11 out of 20 | |
Kids learning the subjects of 8th standard e.g. science, math & english, history, geography etc. | 12 out of 20 | |
Kids learning the subjects of 10th standard e.g. physics, chemistry, math, zoology, botany, psychology, home science, human science, & english, history, geography, arts etc. | 13 out of 20 | |
Kids learning the subjects of 12th standard e.g. physics, chemistry, math, zoology, botany, psychology, home science, human science, & english, history, geography, arts etc. | 14 out of 20 | |
Students learning the streams in graduation e.g. Science, Arts, Commerce, Engineering, Medical etc. | 15 out of 20 | |
Students learning the specialization in streams in graduation e.g. Science, Arts, Commerce, Engineering, Medical etc. | 16 out of 20 | |
Students doing research in the specialized streams in PhD e.g. Quantum, Astronomy, Aerospace, AI/ML, Politics in one region etc. | 17 out of 20 |
Note : during the learning process of a baby, I considered only happy situations, no consideration of poverty issues, social issues & family issues or other factors.
So take-away from the above table, is that at every stage the baby is getting knowledge starting from when there was no knowledge at the time of birth. Now the readers are requested to correlate above baby grooming journey with themselves and let’s try to answer below
- If a kid with maturity level 13/20 (10th std) is asked the question, [describe your mother]. What do you think, is kid able to answer? Off-course the kid would be able to describe her mother so beautifully to make her mother proud.
- If a kid with maturity level 5/20 (1st std) is asked, [how does the Einstein equation [E=MC(2)] work?] What do you think, is kid able to answer? Off-course the kid would not be able to explain or you may get a very different answer to make you laugh or no-reaction.
So this way now we can understand that [Maturity level ∝ knowledge], the knowledge given to the baby /kid /student.
During each stage the baby (our AI-Model) learns the knowledge from surroundings, mother, teachers, friends or so, this process of learning is actually equivalent to “Training the model” and the “Knowledge” is like “Data” being used to groom the model. After these many stages the baby has turned into a mature student & research scholar with lots of knowledge, this is equivalent to that AI-Model (baby) is trained (groomed) with lots of data (Knowledge). He would be able to answer any sort of questions for which he has been groomed throughout his life. In a similar way getting responses for given set input parameters (questions) is equivalent to “Prediction”.
This is all about AI/ML’s to a non-technical person !!!
Now coming into the present, the floor is heavily hit by “Generative AI”. To understand “Generative AI” in a better way, let’s see what was before “Generative AI”. So we shall segregate AI into two
- Conventional AI, which was before “Generative AI”. In conventional AI, we classify the data and tag them with some label, so when the “Conventional AI” model is trained with a huge set of data, the model would be able to predict the outcome within its known data and maturity-level
e.g. if we train a model with data (knowledge) like [country, age, gender, city, weight, health-vitals, health-status (very-good, good, average, bad, worst)]. The model would be able to predict the health-status of any citizen from any country. This way “Conventional AI” is also known as “Predictive AI” or “Discriminative AI”.
- On the other hand “Generative AI” originated from the word “generate” i.e. it generates TIV (Text, Image & Video). “Generative AI” models use “neural networks” to identify the patterns and structures within existing data to generate new and original content. “Generative AI” models are called LLM (Large Language Model) which uses a “Vector database”. (Note : To keep the things simple, I will explain vector database some other time)
“Generative AI” models have the ability to leverage different learning approaches e.g. unsupervised (like teacher provided knowledge) or semi-supervised (mix teacher & self) learning for training. On the contrary of “Discriminative AI”, “Generative AI” trains the model (LLM) with huge data which need not to be labeled. So in a nutshell “Generative AI”, enables the thinking capability like humans.
Understand “Generative AI” by real-life example
Suppose we have invited a French person to India. The French guy does not know anything about India and its culture & history. So if I ask him anything from the epic Ramayana like
- Who was the King Rama,
- Who was the Queen Sita,
- Reason behind war between Rama & Ravana,
- Reason of Sita’s exile,
- Reason behind war between Rama & Lav-kush
- and many more,
the French guy would definitely not be able to answer any of them, even-if he tries that would be something nonsense.
Now let’s try one approach, we give him the book Ramayana translated in French language. Assuming 30 days are sufficient to read the whole book, so let’s meet him after 30 days. Considering the happy situation only, to your surprise the French guy is able to answer most of the questions related to the book. But asking anything which is not from inside the book or any irrelevant context, it is not certain what answer you will get e.g.
- Why did Lav-Kush kill Ravana?
- Why is King Rama not speaking on today’s global-warming issue?
So in this case the “French guy” was our “Generative AI” model (LLM), where the “French guy” was trained for all context through reading process sequentially, keeping those context into human memory (neural network i.e. vector database).
Popular LLMs in the market
These popular LLMs are also called “GPTs i.e. Generative Pre-trained Transformers (GPTs)”. GPT models have revolutionized natural language processing including by content processing (images & videos). OpenAI has done fabulous work in NLP processing and released 4 versions of GPT models (GPT-4 is still awaiting). GPTs can create content from text, write stories, songs, poems, hell lots of things.
Mock-1 : Let’s train a LLM (your own text GPT model)
Provide textual context (NLP inputs) to your LLM
- Jawahar Lal Nehru was 1st prime-minister of Independent India, his tenure was between 1947-1964.
- Gulzarilal Nanda was 2nd prime-minister of Independent India, he was made as an acting prime-minister in 1964 after the immediate death of prime-minister Jawahar Lal Nehru.
- Lal Bahadur Shastri was 3rd prime-minister of Independent India, his tenure was between 1964-1966.
- Gulzarilal Nanda was 4th prime-minister of Independent India, again he was made as an acting prime-minister in 1966 after the immediate death of prime-minister Lal Bahadur Shastri.
- Indira Gandhi was 5th prime-minister of Independent India, her tenure was between 1966-1977.
- Morarji Desai was 6th prime-minister of Independent India, his tenure was between 1977-1979.
- Charan Singh was 7th prime-minister of Independent India, his tenure was between 1979-1980.
- Indira Gandhi was 8th prime-minister of Independent India, her tenure was between 1980-1984.
- Rajiv Gandhi was 9th prime-minister of Independent India, his tenure was between 1984-1989.
- Vishwanath Pratap Singh was 10th prime-minister of Independent India, his tenure was between 1989-1990.
- Chandra Shekhar was 11th prime-minister of Independent India, his tenure was between 1990-1991.
- P. V. Narasimha Rao was 12th prime-minister of Independent India, his tenure was between 1991-1996.
- Atal Bihari Vajpayee was 13th prime-minister of Independent India, his tenure was for one year in 1996.
- H. D. Deve Gowda was 14th prime-minister of Independent India, his tenure was between 1996-1997.
- Inder Kumar Gujral was 15th prime-minister of Independent India, his tenure was between 1997-1998.
- Atal Bihari Vajpayee was 16th prime-minister of Independent India, his tenure was between 1998-2004.
- Manmohan Singh was 17th prime-minister of Independent India, his tenure was between 2004-2009.
- Manmohan Singh was 18th prime-minister of Independent India, his tenure was between 2009-2014.
- Narendra Modi is 19th prime-minister of Independent India, he is prime-minister since 2014.
Ask question from your LLM
E.g. your question is [Please provide list of all prime-ministers of india]
<source : ChatGPT> As of my last knowledge update in January 2022, here is a list of the Prime Ministers of India up to that date:
- Jawaharlal Nehru (1947-1964)
- Gulzarilal Nanda (acting) (1964)
- Lal Bahadur Shastri (1964-1966)
- Gulzarilal Nanda (acting) (1966)
- Indira Gandhi (1966-1977)
- Morarji Desai (1977-1979)
- Charan Singh (1979-1980)
- Indira Gandhi (1980-1984)
- Rajiv Gandhi (1984-1989)
- Vishwanath Pratap Singh (1989-1990)
- Chandra Shekhar (1990-1991)
- P. V. Narasimha Rao (1991-1996)
- Atal Bihari Vajpayee (1996, 1998-2004)
- H. D. Deve Gowda (1996-1997)
- Inder Kumar Gujral (1997-1998)
- Manmohan Singh (2004-2014)
- Narendra Modi (2014-present, as of January 2022)
Please note that this list only includes Prime Ministers up to January 2022. There may have been changes or new appointments after that date. You may want to check for the most current information to see if there have been any updates since then.
Mock-2 : Let’s train a LLM (your own content GPT model)
Provide content context (with Images & NLP inputs) to your LLM
India | This is India’s national flag. The flag has 3 horizontal stripes with colors orange on top, white in middle and green at bottom. In the middle on white stripe circle with 24 lines represents that the country provides equal rights to all the citizens for all 24 hours everyday. |
Ireland | This is Ireland’s national flag. The flag has 3 vertical stripes with colors orange on the right, white in middle and green on the left. |
Italy | This is Italy’s national flag. The flag has 3 vertical stripes with colors red on the right, white in middle and green on the left. |
Hungary | This is Hungary’s national flag. The flag has 3 horizontal stripes with colors red on top, white in middle and green at bottom. |
Iran | This is Iran’s national flag. The flag has 3 horizontal stripes with colors green on top, white in middle and red at bottom. In the middle on white stripe circle there is a symbol of “Allahu Akbar”, or “God is great” |
Switzerland | This is Switzerland’s national flag. The flag is of red color and has the sign of + (plus) in white color. |
Finland | This is Finland’s national flag. The flag is of white color and has the sign of + (plus) in blue color. |
Note : your GPT will respond only based on the above training or knowledge provided.
Ask question from your LLM
How many flags in the world have horizontal stripes | As of my last knowledge update with above table here is the list of national flags having horizontal stripes:IndiaHungaryIran |
How many flags in the world have + sign for different colors | As of my last knowledge update with above table here is the list of national flags having + sign:SwitzerlandFinland |
How many flags in the world have vertical stripes | As of my last knowledge update with above table here is the list of national flags having vertical stripes:IrelandItaly |
Generate new flag with horizontal stripes and + sign | Other sample4Other sample5 |
So now the readers should be able to understand how LLM are trained with context, so that it can prompt your various types of queries within the given context.
Impact of Gen AI on Industry
The IT industry is seriously exploring the benefits of “Generative AI”. For sure this is not going to replace the IT world but the way we communicate with the system and the way the system responds to us is going to change.
Examples – 1
Suppose there is a dispute in my mobile bill due to wrong roaming components
Current approach
- Customer will land on approach channel e.g. IVR, Chatbot, Mobile-App
- Navigate to the appropriate option and provide your maximum input..
- At last you will talk to a call-center-agent if virtual channels are not helpful.
- Then call-center-agent will navigate and assess all information manually to assist with better information.
“Generative AI” approach
- In this customer simply will say [there is a dispute in my mobile bill due to wrong roaming components]
- Using this information first system will identify the intent and supporting info
- In this case, intent is [Billing dispute] and supporting info is [wrong roaming component].
- Just to make you understand, these two input are sufficient to respond on behalf of Level-1 & Level-2 support staff
- System will verify the customer’s unpaid bill for the roaming component, will check roaming CDRs and rating as per country agreement.
- If the system finds no-mistake in their calculation, GPT model will write a beautiful informative response with all details and request to connect customer-agent if not satisfied.
Examples – 2
Suppose a customer is looking best health insurance plan for his family
Current approach
- Customer will land on approach channel e.g. IVR, Chatbot, Mobile-App
- Navigate to the appropriate option and provide your maximum input..
- At last you will talk to a call-center-agent if virtual channels are not helpful.
- Then call-center-agent will navigate and assess all information manually to assist with better information.
“Generative AI” approach
- In this customer simply will say [customer is looking best health insurance plan for his family]
- Using this information first system will identify the intent and supporting info
- In this case, intent is [Interest in health insurance plan] and supporting info is [Family].
- While processing the intent & supporting information, system will get to know more supporting information required i.e. [Family members]
- GPT model will generate beautiful query for required supporting information required i.e. [Family members].
- Assuming intent and required supporting information is provided to GPT model, GPT model will trigger product-catalog API [Get-Best-Insurance-Plan <input parameters>]
- The GPT model will get the best plan from API [Get-Best-Insurance-Plan <>] and prepare a beautiful personalized response for the customer.
Conclusion
So “Generative AI” will introduce a new virtual-level prior to Level-1 or Level-2 support staff, so that telecom operators can improve their “customer satisfaction”, “average handling time”, “operating efficiency” and TCO.
Popular use-cases, food for thought for enthusiastic readers
In the end leaving you with possible use-cases in different domains, keep exploring and share with the community if you find something important to be known.
About author
Profile : Rajesh Verma – Brief profile
Source : link for this article here
Series : S1 (Architecture & design)
Episode : S1E4 (How will I explain AI/ML & Gen AI to a non-technical person)
Author’s approach : Rajesh wants to share his learning & experience gained throughout his career from various sources. Author started the series on architectural topics and this article is the first episode in that attempt. Author feels that lots of information is available on various forums, but scattered here & there. Episodes in this series will be designed for most of the relevant topics in architecture-&-design, published gradually and organized in logical sequence. Principally episodes will have linkage with other episodes, so that readers can have proper connection among the topics and would be able to correlate with ongoing activities in their software life. Topics for example will be related to functional architecture, integration architecture, deployment architecture, microscopic view of mostly architecture-building-blocks (ABBs), security guidelines & approach to comply, performance KPIs & engineering, git branch & DevOps enabled automation strategy, NFR aspects (e.g. scalability, high-availability, stability, resiliency, etc.), commonly used architecture styles & design patterns, cloudification approaches, multi-tenancy approach, data migration, channel-cutover & rollout strategy, process standardization & simplification, greenfield rollout & brownfield transformation journeys, etc.
Thank you for reading the post, please stay connected.
No responses yet